text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Unique Perfect Phylogeny Characterizations via
Uniquely Representable Chordal Graphs
arXiv:1305.1375v2 [cs.DM] 8 May 2013
Rob Gysel
Department of Computer Science
University of California, Davis
1 Shields Avenue, Davis CA 95616, USA
[email protected]
Abstract. The perfect phylogeny problem is a classic problem in computational biology, where we seek an unrooted phylogeny that is compatible with a set of qualitative characters. Such a tree exists precisely
when an intersection graph associated with the character set, called the
partition intersection graph, can be triangulated using a restricted set
of fill edges. Semple and Steel used the partition intersection graph to
characterize when a character set has a unique perfect phylogeny. Bordewich, Huber, and Semple showed how to use the partition intersection
graph to find a maximum compatible set of characters. In this paper, we
build on these results, characterizing when a unique perfect phylogeny
exists for a subset of partial characters. Our characterization is stated in
terms of minimal triangulations of the partition intersection graph that
are uniquely representable, also known as ur-chordal graphs. Our characterization is motivated by the structure of ur-chordal graphs, and the
fact that the block structure of minimal triangulations is mirrored in the
graph that has been triangulated.
1
Introduction
An X−tree is a pair Tx = (T , φ) where T is a tree and φ is a map from X to
the nodes of T , such that every node of T with degree two or one is mapped to
by φ. We will call the range of φ the labeled nodes of Tx , and these nodes are
labeled by φ. The underlying tree of Tx is T . An X−tree is free if φ is a bijection
to the leaves of T , and it is ternary if every internal node of T has degree three.
Given A ⊆ X, we will use Tx (A) to denote the minimal subtree of T containing
the nodes φ(A). Two subtrees Tx (A) and Tx (A0 ) of T intersect if they have one
or more nodes in common, and if v is a common node of Tx (A) and Tx (A0 ) we
say that Tx (A) and Tx (A0 ) intersect at v.
A partial character for X is a partition χ = A1 |A2 | . . . |Ar of a subset X 0 ⊆ X.
Each Ai is called a cell of χ. If Tx (A) and Tx (A0 ) do not intersect for every pair
of distinct cells A and A0 of χ, then Tx displays χ. A perfect phylogeny for a set
of partial characters C is an X−tree Tx that displays each character in C. When
C has a perfect phylogeny, we also say that C is compatible. The perfect phylogeny
problem (also called the character compatibility problem) is to determine if a set
of partial characters is compatible.
(ab, χ1 )
d
a
u
(ac, χ2 )
(de, χ3 )
(ab, χ3 ) (bde, χ2 )
(cd, χ1 )
v
e
b
c
(a) Tx
(b) int(C)
Fig. 1. An X−tree for X = {a, b, c, d, e} and a non-chordal partition intersection graph
int(C). The characters C = {χ1 , χ2 , χ3 } are χ1 = ab|cd, χ2 = ac|bde, and χ3 = ab|de.
Edges of int(C) are given by solid edges, and the dashed edge represents the fill edge
required to obtain the triangulation int(C, Tx ) of int(C). Both χ1 and χ3 are displayed
by Tx , but χ2 is not because Tx (ac) and Tx (bde) intersect at u and v. This intersection
induces the dashed fill edge of int(C), which breaks χ2 . The edge uv is distinguished
by χ1 .
The perfect phylogeny problem reduces to a graph theoretical problem that
we detail now. Given a set of characters C, one can construct the partition intersection graph int(C) as follows. The vertex set of int(C) is
{(A, χ) | χ ∈ C and A is a cell of χ} ,
and there is an edge between two vertices (A, χ) and (A0 , χ0 ) if and only if A
and A0 have non-empty intersection. For a vertex (A, χ) of int(C), A is the cell
of (A, χ) and χ is the character of (A, χ). Observe that if χ = A1 |A2 | . . . |Ar
is a partial character, then every pair of distinct vertices (A, χ) and (A0 , χ) are
non-adjacent in int(C).
A graph is chordal if every cycle of length four or more has a chord, that is, an
edge between vertices of the cycle that do not appear consecutively in the cycle.
In general int(C) is not a chordal graph, and we are interested in adding edges to
int(C) to obtain a chordal supergraph H of int(C) that is called a triangulation
of int(C). The added edges are called fill edges. If no subset of the fill edges yields
a triangulation of int(C), it is a minimal triangulation of int(C). When each fill
edge is of the form (A, χ)(A0 , χ0 ) and χ 6= χ0 , the resulting triangulation is a
proper triangulation of int(C). The following classic result reduces the question
of determining compatibility to finding proper triangulations of the partition
intersection graph. It was originally phrased in terms of proper triangulations,
but from the definitions it follows that int(C) has a proper triangulation if and
only if it has a proper minimal triangulation.
Theorem 1. [6, 21, 26] Let C be a set of qualitative characters. Then C is
compatible if and only if int(C) has a proper minimal triangulation.
∼ T 0 , if there is a bijective
Two X−trees Tx and Tx0 are isomorphic, writing Tx =
x
0
map ψ : V (T ) → V (T ) that has the following properties:
2
1. it preserves labels, meaning that φ0 = ψ ◦ φ; and
2. it is a graph isomorphism, that is, uv ∈ E(T ) if and only if ψ(u)ψ(v) ∈ E(T 0 ).
A set of characters C defines a perfect phylogeny if it is the unique perfect
phylogeny, up to isomorphism, for C. The unique perfect phylogeny problem is
to determine if a set C of partial characters defines a perfect phylogeny. If Tx =
(T , φ) displays χ and uv is an edge of T such that u is a node of Tx (A) and v is
a node of Tx (A0 ) where A and A0 are distinct cells of χ, then uv is distinguished
by χ. If every edge of Tx is distinguished by at least one character of C, then Tx
is distinguished by C. The following characterization is due to Semple and Steel.
Theorem 2. [24] Let C be a set of partial characters on X. Then C defines a
perfect phylogeny if and only if the following conditions are satisfied:
(a) int(C) has a unique proper minimal triangulation H; and
(b) there is a free ternary perfect phylogeny for C and it is distinguished by C.
Further, if Tx is the unique perfect phylogeny for C, then Tx is a free ternary
X−tree distinguished by C, and int(C, Tx ) = H.
This result is the impetus of our current work, and one of our main interests
is to re-formulate condition (b) in terms of combinatorial structures that play a
significant role in the study of chordal graphs and minimal triangulations.
Chordal graphs are characterized by the existence of trees that represent the
adjacency structure of the graph. Suppose G is a graph with vertices V (G) =
{x1 , x2 , . . . , xn }. A tree representation of G consists of a tree T and subtrees
T1 , T2 , . . . , Tn of T such that two trees Ti and Tj intersect if and only if xi and
xj are adjacent. Here, the subtrees are in one-to-one correspondence with the
vertex set of G, and this correspondence is made explicit by mapping each subtree
Ti to the vertex xi of G. Observe that a node v of T defines a clique K(v) =
{xi | v is a node of Ti } of G. Notationally, we will write a tree representation as
an ordered pair Tr = (T , K) where K maps nodes of T to cliques of G satisfying
the following properties:
(Edge Coverage) a pair of vertices x and y of G are adjacent if and only if
there is a node v of T such that x, y ∈ K(v); and
(Convexity) for each vertex x of G, the set of nodes {v ∈ V (T ) | x ∈ K(v)}
induces a subtree of T (i.e. a connected subgraph of T ).
We will frequently refer to the convexity property throughout the paper. As
with X−trees, we will call T the underlying tree of Tr . Often we will define a
tree representation by only specifying the underlying tree T and a collection of
subtrees of T , which together implicitly define K.
Let T1 , T2 , . . . , Tk be a collection of subtrees of T . If each pair Ti , Tj of subtrees intersect at a node vij , then by the Helly property for subtrees of a tree [9],
all of T1 , T2 , . . . , Tk intersect at a common node v. This property manifests itself
as a statement about cliques of G and nodes of Tr in the following way: for
any clique K of G, there is at least one node u of T such that K ⊆ K(u). In
3
particular, this is true when K is a maximal clique of G (i.e. no proper superset
is also a clique), and therefore K(V (T )) contains the set of maximal cliques of
G. If the maximal cliques of G are in one-to-one correspondence with the nodes
of T via K, then Tr is a clique tree of G. See Figure 2 for an example.
Theorem 3. [6, 8, 27] The following statements are equivalent.
(a) G is a chordal graph.
(b) G has a tree representation.
(c) G has a clique tree.
Observe that if Tr = (T , K) is a clique tree of G and uv is an edge of T ,
then because K(u) and K(v) are both maximal cliques of G, there is a vertex
x of G in K(u) − K(v). In general, a chordal graph has an exponential number
of clique trees. An algorithm to enumerate clique trees, along with a formula to
count them, appears in [16].
We will often be analyzing a tree representation Tr = (T , K) of a triangulation
H of int(C). Given a vertex (A, χ) of int(C), we will denote the subtree of T that
it corresponds to by Tr (A, χ). Observe that v is a node of Tr (A, χ) if and only if
(A, χ) ∈ K(v). Given a set of characters C on X and an X−tree Tx = (T , φ), a
chordal graph int(C, Tx ) is given by adding an edge between two vertices (A1 , χ1 )
and (A2 , χ2 ) if and only if Tx (A1 ) and Tx (A2 ) intersect. This construction, along
with the fact that int(C, Tx ) is a triangulation of int(C), is well-known in the
phylogenetics literature, and will be discussed in detail in Section 2.
A chordal graph G is uniquely representable if it has a single clique tree, or
ur-chordal for short. A ur-chordal graph is ternary if each internal node of its
clique tree has degree three, and its leafage 1 is the number of leaves its clique
tree has. Let H be a proper triangulation of int(C) and Tr = (T , K) be a clique
tree of H. An edge uv of Tr is incontractable with respect to χ if there are distinct
cells A and A0 of χ such that u ∈ Tr (A, χ) and v ∈ Tr (A0 , χ). We say that Tr is
incontractable with respect to C if each edge is incontractable with respect to at
least one χ ∈ C. Now we present our first main result.
Theorem 4. Suppose C is a set of partial characters on X. Then C defines a
perfect phylogeny if and only if the following conditions hold:
(a) int(C) has a unique proper minimal triangulation H;
(b) H is a ternary ur-chordal graph with leafage |X|; and
(c) each edge of H’s unique clique tree is incontractable with respect to C.
Further, if Tx is the perfect phylogeny defined by C, then Tx is a free ternary
X−tree distinguished by C, and int(C, Tx ) = H.
Let C be a set of partial characters on X and χ ∈ C. Suppose H is a triangulation of int(C) with fill edge (A, χ)(A0 , χ) where A and A0 are distinct cells of χ.
1
In general, the leafage of a chordal graph is the minimum number of leaves that a
clique tree of the graph can have [19].
4
Then we say that (A, χ)(A0 , χ) breaks χ, and χ is a broken character of H. For
a triangulation H of int(C), its displayed characters are the characters of C that
are not broken characters of H. Bordewich, Huber, and Semple [3] proved that
it is possible to find a maximum-sized compatible subset of C using the partition
intersection graph.
Theorem 5. Let C be a set of partial characters on X. Then C 0 is a maximumsized compatible subset of C if and only if there is a triangulation H of int(C)
that has C 0 as its displayed characters, and any other triangulation of int(C) has
at most |C 0 | displayed characters.
A subset of partial characters C 0 ⊆ C is a maximal defining subset of C when
C 0 defines a perfect phylogeny, and there is no compatible set C 00 such that
C 0 ⊂ C 00 ⊂ C. Our second main result is the following.
Theorem 6. Suppose C is a set of partial characters on X and C 0 ⊆ C. Then
C 0 is a maximal defining subset of C if and only if the following conditions hold:
(a) int(C) has a unique minimal triangulation H that has C 0 as its displayed
characters, and no other minimal triangulation of int(C) has at least C 0 as
its displayed character set;
(b) H is a ternary ur-chordal graph with leafage |X|; and
(c) each edge of H’s unique clique tree is incontractable with respect to C 0 .
Further, if Tx is the perfect phylogeny defined by C 0 , then Tx is a free ternary
X−tree distinguished by C 0 , and int(C, Tx ) = H.
2
Chordal Graph Preliminaries
In this section, we detail known results on chordal graphs that are necessary for
the remainder of the paper. Suppose G = (V, E) is a graph and S ⊆ V . Let G−S
denote the graph obtained from G by removing S and all edges incident to at
least one vertex in S. If there are vertices x and y of G that are connected in G
but not in G − S, then S is an xy−separator, and if no proper subset of S has
this property, it is a minimal xy−separator. When S is a minimal xy−separator
for at least one pair of vertices x and y, it is a minimal separator. Minimality
in this definition is relative; it is possible to have containment relationships
between two minimal separators2 . The maximal connected subsets of G − S are
the connected components of G − S. Let C be a connected component of G − S.
The neighborhoood of C in G, denoted N (C), is the set of vertices of S that are
adjacent to at least one vertex in C. If N (C) = S, then it is a full component of
G−S. The following useful characterization of minimal separators is well-known,
and left as an exercise in [9].
2
Dirac [7] called these sets relatively minimal cut-sets, which is perhaps more descriptive, but this term has stuck of the modern literature on chordal graphs and minimal
triangulations.
5
K(v)
v
(b) Tr
(a) G
Fig. 2. A chordal graph G and clique tree Tr = (T , K) of G. There are three clique trees
for G, one of which is obtained by removing the bottom-leftmost node and its incident
edge, and then attaching it to v. The maximal clique map K is defined by the triangle
drawn inside each node. Arrows indicate vertices that consist of the intersection of a
neighboring node’s maximal clique, and each such intersection is a minimal separator
by Theorem 7.
Lemma 1. Let G = (V, E) be a graph, and S ⊆ V . Then S is a minimal
separator if and only if G − S has two or more full components.
A minimal separator has multiplicity k if G − S has k − 1 full components.
Interestingly, clique trees contain detailed information about the minimal separators of the graph it represents, which will be useful for our proofs in later
sections.
Theorem 7. [16, 20] Suppose G = (V, E) is a chordal graph, Tr = (T , K) is a
clique tree of G, and S ⊆ V . Then S is a minimal separator of G if and only if
there is an edge uv of T such that S = K(u) ∩ K(v)3 . Further, the multiplicity
of S is the number of edges of T with this property.
We will not need all of the following characterizations of ur-chordal graphs,
but we list them for completeness.
Theorem 8. [14, 18] Let G be a chordal graph. Then the following statements
are equivalent.
a) G is uniquely representable.
b) If S is a minimal separator of G, then there are exactly two maximal cliques
K and K 0 of G such that S ⊆ K, K 0 .
c) Each minimal separator of G has multiplicity one.
3
This theorem also implies that a minimal separator of a chordal graph is a clique.
In fact, chordal graphs are characterized by having only clique minimal separators,
which is one of the earliest results on chordal graphs [7].
6
d) There is no minimal separator of G that properly contains another minimal
separator of G.
e) The number of minimal separators of G is the number of maximal cliques of
G minus one.
Chordal graphs can be recognized in linear time [2], and the maximal cliques
and minimal separators of a chordal graph may be computed in linear time [1].
Using property e) of Theorem 8, this allows the recognition of ur-chordal graphs
in linear time.
3
X−trees, Clique Trees, and Tree Representations
In this section, we define two operations to facilitate the discussion between
X−trees and tree representations, and prove or state results that will be useful in
later sections. Both operations are commonly used in the literature (see [24,26]),
but do not seem to be named as we will do here. The results in this section will
be useful for proving our characterization for maximal defining subsets.
Given a set of characters C on X and an X−tree Tx = (T , φ), construct the
chordal graph int(C, Tx ) having:
• vertex set identical to that of int(C); and
• an edge between (A, χ) and (A0 , χ0 ) if and only if Tx (A) and Tx (A0 ) intersect.
This graph has a tree representation Tr with underlying tree T and subtrees
obtained by defining Tr (A, χ) = Tx (A) for each vertex (A, χ) of int(C). Therefore
int(C) is chordal by Theorem 3. Each edge (A1 , χ1 )(A2 , χ2 ) of int(C) has cells
A1 and A2 that share at least one member of X, say a, so Tx (A1 ) and Tx (A2 )
intersect at φ(a). Therefore Tr (A1 , χ1 ) and Tr (A2 , χ2 ) intersect at φ(a) as well, so
(A1 , χ1 )(A2 , χ2 ) is an edge of int(C, Tx ) by subtree intersection. Hence each edge
of int(C) is also an edge of int(C, Tx ), so int(C, Tx ) is a triangulation of int(C).
We will say that Tx derives the tree representation Tr = (T , K) of int(C, Tx ) and
Tr is derived from Tx . In general, Tr is not a clique tree.
Observation 1 Let C be a set of partial characters on X, Tx be an X−tree, and
Tr be the tree representation induced by Tx . Then the underlying tree of Tr is
the underlying tree of Tx , and for all χ ∈ C and cells A of χ, Tx (A) = Tr (A, χ).
Further, int(C, Tx ) is a triangulation of int(C).
Lemma 2. Let C be a set of partial characters on X, and Tx be an X−tree that
displays C 0 ⊆ C. Suppose that each edge of Tx is distinguished by C 0 . Then the
tree representation of int(C, Tx ) derived from Tx is a clique tree of int(C, Tx ).
Proof. Let Tr = (T , K) be the tree representation of int(C, Tx ) derived from Tx .
For the sake of contradiction, assume that Tr is not a clique tree, so that K is not
a one-to-one map between the nodes of T and the maximal cliques of int(C, Tx ).
Then there must be nodes u and v of T such that K(v) ⊆ K(u). Let v 0 be the
7
(ab, χ1 )
(ac, χ2 )
(cd, χ1 )
(de, χ3 )
(bde, χ2 )
(ab, χ3 )
Fig. 3. The tree representation Tr of int(C, Tx ) derived from Tx in Figure 1.a. The
triangulation int(C, Tx ) of int(C) is depicted in Figure 1.b; note that any two subtrees
of Tr intersect precisely when the corresponding vertices of int(C, Tx ) are adjacent.
Observe that Tr is not a clique tree; for example, the two left-most nodes map to nonmaximal cliques. Additionally, there are two nodes that map to the maximal clique
{(cd, χ1 ), (bde, χ2 ), (de, χ3 )}. Obtaining a clique tree from a tree representation is described in [8].
closest node to v between u and v in T , allowing the possibility that v 0 = u. Note
that each vertex in K(v) is also a vertex of K(v 0 ) by convexity, so K(v) ⊆ K(v 0 ).
Now, vv 0 is distinguished by some χ ∈ C 0 , so there are distinct cells A and A0
of χ such that v is a node of Tx (A) and v 0 is a node of Tx (A0 ). By Observation
1, v is a node of Tr (A, χ), so (A, χ) ∈ K(v) and by containment (A, χ) ∈ K(v 0 ).
Further, (A0 , χ) ∈ K(v 0 ), so (A, χ)(A0 , χ) is a fill edge of int(C, Tx ) and it breaks
χ. This contradicts the assumption that Tx displays C 0 , so Tr must be a clique
tree of int(C, Tx ).
t
u
Lemma 3. Let C be a set of partial characters on X and Tx be an X−tree
that displays C 0 ⊆ C. Suppose that Tx is free, ternary, and each edge of Tx is
distinguished by C 0 . Then int(C, Tx ) is uniquely representable.
Proof. To prove that int(C, Tx ) is uniquely representable, we use Theorem 8 and
show that there are no containment relationships between the minimal separators
of int(C, Tx ). Working towards a contradiction, assume that S ⊂ S 0 are minimal
separators of int(C, Tx ), and let Tr = (T , K) be the tree representation derived
from Tx . Then Tr is a clique tree by Lemma 2, and there are edges uv and
u0 v 0 of T such that S = K(u) ∩ K(v) and S 0 = K(u0 ) ∩ K(v 0 ) by Theorem 7.
Without loss of generality, assume that the path from v to v 0 does not contain
either u or u0 (perhaps v = v 0 ). Let w be the node on this path adjacent to v
if v 6= v 0 , otherwise let w = u0 . In either case, K(w) contains S: if w = u0 , then
S ⊆ K(w). Otherwise, each (A, χ) in S is an element of both K(u) and K(u0 ),
so (A, χ) ∈ K(w) by convexity, and S ⊆ K(w).
To complete the proof, we will obtain a contradiction by showing that S
has a vertex not in K(w). There is a character χ0 in C 0 that distinguishes wv,
and distinct cells A0 and A00 of χ0 such that v is a node of Tx (A0 ) and w is a
8
node of Tx (A00 ). Now, v has at least two neighbors, and because Tx is ternary,
v must have degree three. Also, v is not mapped to by φ because Tx is free,
so in order for v to be a node of Tx (A0 ), there must be at least two nodes
of Tx in φ(A0 ) that are not v, and the path between these two nodes must
contain v. This path must also contain two of v’s neighbors, and neither of
these vertices can be w, because w is not a node of Tx (A0 ). Thus u must be
on this path, so it is a node of Tx (A0 ). Both K(u) and K(v) contain (A0 , χ0 ) by
Observation 1, so it must be a vertex of S = K(u)∩K(v). Further, (A0 , χ0 ) ∈
/ K(w)
because w is not a node of Tx (A0 ) = Tr (A0 , χ0 ). This is impossible because we
have shown that both (A0 , χ0 ) ∈ S − K(w) and S ⊆ K(w). Thus there are
no containment relationships between the minimal separators of int(C, Tx ), and
int(C, Tx ) is uniquely representable by Theorem 8.
t
u
Lemma 4. [3] Let C be a set of partial characters on X, Tx an X−tree, and
C 0 be the subset of C displayed by Tx . Then int(C, Tx ) is a triangulation of int(C)
in which the displayed characters are C 0 .
The previous three lemmas can be summarized as follows.
Theorem 9. Let C be a set of partial characters on X and Tx be an X−tree
that displays C 0 ⊆ C. Suppose that Tx is free, ternary, and each edge of Tx is
distinguished by C 0 . Then int(C, Tx ) is a uniquely representable chordal graph,
and the tree representation derived from Tx is its unique clique tree. Further, the
displayed characters of int(C, Tx ) are exactly C 0 .
Now suppose that Tr = (T , K) is a clique tree of a triangulation H of int(C),
with the goal of defining an X−tree Tx . The discussion we provide here is standard, e.g. see [3,25]. Construct a map φ from X to T by defining, for each a ∈ X,
φ(a) = v if and only if K(v) contains every vertex of int(C) whose cell contains
a. These vertices form a clique because a is contained in each of their cells. Because we have only added fill edges to obtain H, this clique a subset of a maximal
clique of H, and hence v exists. There may be more than one choice for v, each
of which we call a candidate node for a. Let u be a leaf of T with neighbor w.
Then K(u) is a maximal clique that contains a vertex (A, χ) of int(C) that is
not found in K(w). By convexity, u is the only node of T whose corresponding
maximal clique contains (A, χ). Each a0 ∈ A has u as its unique candidate node,
and hence every leaf of T is a unique candidate node for at least one element of
X. Thus each leaf of T must be labeled by φ. To finish constructing an X−tree,
obtain T 0 by suppressing any unlabeled nodes of T that have degree two. The
result is an X−tree Tx = (T 0 , φ), and we say that Tr induces Tx and Tx is induced
by Tr . We emphasize that the underlying tree of Tx need not be the same as the
underlying tree of Tr . Note that, because an element of X may have multiple
candidate nodes, Tr may induce multiple X−trees. Next, we show that when
H is a minimal triangulation, much of Tx ’s structure is described by Tr . The
following lemma will be useful.
Lemma 5. [17, 22] Let G be a graph and H be a minimal triangulation of G.
If uv is a fill edge of H, then there is a minimal separator of H that contains
both u and v.
9
ab
c
de
(a) Tr0
(b) Tx0
Fig. 4. A clique tree Tr0 of int(C, Tx ) from Figure 1.b and an X−tree Tx0 induced by
Tr0 . Note that int(C, Tx0 ) = int(C, Tx ), where Tx is the X−tree from Fig. 1.a.
Though not stated in this form, the following lemma follows from the proof
of Lemma 2.4 and the statement of Corollary 2.5 in [24].
Lemma 6. Let H be a minimal triangulation of int(C), and suppose Tx is induced by a clique tree of H. Then H = int(C, Tx ).
Lemma 7. Let H be a minimal triangulation of int(C), Tr = (T , K) be a clique
tree of H, and suppose Tr induces Tx . Then the underlying tree of Tx is T .
Proof. We have already seen that every leaf of T is the unique candidate node
of some element of X. In addition to this, it was also shown in [13] that if u is a
node of Tr of degree two, then u is the unique candidate node of some element
of X. This was done by showing that K(u) contains either:
1. a vertex (A1 , χ1 ) of int(C) that is not contained in K(w) for any other node
w 6= u of T ; or
2. an edge (A2 , χ2 )(A3 , χ3 ) of int(C), whose incident vertices have cells with
non-empty intersection, and are not both contained in K(w) for any other
node w 6= u of T .
For completeness, we outline a proof here. Using convexity and the fact that
u has degree two, it follows that either a unique vertex or unique pair of vertices
are contained in K(u). It remains to show that (A2 , χ2 )(A3 , χ3 ) is actually an
edge of int(C) (so A2 ∩ A3 is non-empty). If not, then by Lemma 5 there is a
minimal separator S of H containing both (A2 , χ2 ) and (A3 , χ3 ). By Theorem
7, there is an edge u1 u2 of T such that S = K(u1 ) ∩ K(u2 ). But this contradicts
case 2, so it must be that (A2 , χ2 )(A3 , χ3 ) is an edge of int(C).
In both cases, u is the unique candidate node of some element in X (this
element is either a ∈ A1 or a ∈ A2 ∩ A3 ), so every degree two node of T is
labeled by φ, and there are no nodes of T that need to be suppressed. Hence the
underlying tree of Tx is T .
t
u
10
Lemma 8. Let H be a minimal triangulation of int(C), Tr = (T , K) be a clique
tree of H, and suppose Tr induces Tx . Then for each vertex (A, χ) of int(C),
Tx (A) = Tr (A, χ).
Proof. Let (A, χ) be a vertex of int(C) and consider a node v of Tx (A). Either
v = φ(a) for some a ∈ A or v lies between φ(a1 ) and φ(a2 ) for some a1 , a2 ∈ A.
In the first case, (A, χ) ∈ K(v) because v is a candidate node for a. Similarly,
in the second case, (A, χ) is an element of both K(φ(a1 )) and K(φ(a2 )), and
therefore (A, χ) ∈ K(v) by convexity. In both cases v is a node of Tr (A, χ), so
Tx (A) ⊆ Tr (A, χ).
To finish proving equality, suppose that Tx (A) ⊂ Tr (A, χ). Define a tree
representation Tr0 = (T 0 , K0 ) of a graph H 0 as follows: set T 0 = T , and define
subtrees Tr0 (A0 , χ0 ) of T for each vertex (A0 , χ0 ) of int(C) as follows:
1. Tr0 (A0 , χ0 ) = Tr (A0 , χ0 ) if (A0 , χ0 ) 6= (A, χ), and
2. Tr0 (A0 , χ0 ) = Tx (A) if (A0 , χ0 ) = (A, χ).
We have already seen that Tr0 (A0 , χ0 ) ⊆ Tr (A0 , χ0 ) for every vertex (A0 , χ0 ) of
int(C), so the edge set of H 0 is a subset of the edge set of H. If (A1 , χ1 )(A2 , χ2 )
is an edge of int(C), then A1 and A2 have at least one element a in common,
and thus Tx (A1 ) and Tx (A2 ) intersect at φ(a). Further, Tx (A0 ) ⊆ Tr0 (A0 , χ0 ) for
each vertex (A0 , χ0 ) of int(C), so Tr0 (A1 , χ1 ) and Tr0 (A2 , χ2 ) also intersect at φ(a).
Therefore (A1 , χ1 )(A2 , χ2 ) is an edge of H 0 , and H 0 is chordal by Theorem 3, so
it is a triangulation of int(C).
To complete the proof, we show that H 0 must have an edge that does not
exist in H. Because Tx (A) ⊂ Tr (A, χ), there is a node u of Tr (A, χ) − Tx (A) that
is adjacent to a node w of Tx (A) = Tr0 (A, χ). By maximality, there is a vertex
(A0 , χ0 ) ∈ K(u) − K(w), and because u is a node of Tr (A, χ), (A, χ) ∈ K(u).
Therefore (A0 , χ0 )(A, χ) is an edge of H. The situation in H 0 is different: if
(A0 , χ0 )(A, χ) is an edge of H 0 , then Tr0 (A, χ) and Tr0 (A0 , χ0 ) intersect at a node
v 0 . But if v 0 is a node of Tr0 (A, χ), then w is on the path from v 0 to u, and by
convexity this would imply that (A0 , χ0 ) ∈ K(w). Hence (A0 , χ0 )(A, χ) is not an
edge of H 0 , so the edge set of H 0 is a proper subset of the edge set of H. This
is impossible because H is a minimal triangulation of int(C), so it must be that
Tx (A) = Tr (A, χ) for each vertex (A, χ) of int(C).
t
u
Lemmas 6, 7, and 8 are summarized below.
Theorem 10. Let H be a minimal triangulation of int(C), Tr be a clique tree of
H, and suppose Tr induces Tx . Then the underlying tree of Tx is the underlying
tree of Tr , and for each vertex (A, χ) of int(C), Tx (A) = Tr (A, χ). Further,
H = int(C, Tx ).
4
Maximal Defining Subsets of Characters
This section is devoted to the proof of Theorem 6. Its proof will follow mainly
from Propositions 1 and 2. Recall that, for a graph H and a subset U of its
11
vertices, the graph H − U is obtained by removing the vertices in U and edges
of H incident to one or more vertices of U .
Lemma 9. Let C be a set of partial characters and C 0 ⊆ C. Suppose H 0 is a
minimal triangulation of int(C 0 ), and let U be the vertices of int(C) that are not
vertices of int(C 0 ). Then there is a minimal triangulation H of int(C) such that
H0 = H − U .
Proof. Let H 0 be a minimal triangulation of int(C 0 ) and H ∗ be the graph obtained by adding the following fill edges to int(C):
1. the fill edges of H 0 ; and
2. fill edges of the form (A, χ)(A0 , χ0 ) where (A, χ) is a vertex of U , and (A0 , χ0 )
is any vertex of int(C).
First we prove that H ∗ is chordal, and then we will use it to construct H, a
minimal triangulation of int(C) such that H − U = H 0 .
Let (A1 , χ1 ), (A2 , χ2 ), . . . , (Ak , χk ) be a cycle in H ∗ . If χi ∈ C 0 for all i =
1, 2, . . . , k, then this cycle is also a cycle of H 0 , and therefore has a chord that is
an edge of H 0 . Each edge of H 0 is also an edge of H ∗ , so this cycle has a chord
in H ∗ . Otherwise, without loss of generality, χ1 ∈ C − C 0 and (A1 , χ1 ) ∈ U so
either (A1 , χ1 )(A3 , χ3 ) is an edge of int(C) or is a fill edge of H ∗ of type 2. In
either case, this cycle has a chord, so H ∗ is chordal.
Now let H be any minimal triangulation of int(C) such that the edge set of
H is a subset of the edge set of H ∗ . Every edge of H − U is either an edge of
int(C 0 ) or is an edge of H 0 by the construction of H ∗ and H. Therefore the edge
set of H − U is a subset of the edge set of H 0 . Further, H − U is chordal because
any cycle of H − U is a cycle of H (i.e. chordality is inherited [23]), so it is a
triangulation of int(C 0 ). By minimality of H 0 , it must be that the edge set of
H − U is equal to the edge set of H 0 , so H 0 = H − U .
t
u
Lemma 10. [12] see also [3] Let C be a set of partial characters on X and
suppose that C 0 is a compatible subset of C. Then there is a minimal triangulation
of int(C) whose displayed characters are at least C 0 .
Though not stated in this form, the following lemma is a direct result of
Lemma 5.1 in [3] and its proof.
Lemma 11. Let C be a set of partial characters on X. Suppose H is a triangulation of int(C) with displayed characters C 0 . Then if Tx is induced by a clique
tree of H, it is a perfect phylogeny for C 0 .
Proposition 1. Let C be a set of partial characters on X and C 0 ⊆ C. Suppose
that the following conditions hold:
(i) int(C) has a unique minimal triangulation H that has C 0 as its displayed
characters, and no other minimal triangulation of int(C) has at least C 0 as
its displayed character set;
(ii) H is a ternary ur-chordal graph with leafage |X|; and
12
(iii) each edge of H’s unique clique tree is incontractable with respect to C 0 .
Then C 0 is a maximal defining subset of C.
Proof. We begin by showing that C 0 has a unique perfect phylogeny using Theorem 2, and finish the proof showing that no superset of C 0 has a unique perfect
phylogeny.
Throughout the proof, H will denote the unique minimal triangulation of
int(C) given by (i) whose displayed character set is C 0 . By Lemma 11, C 0 has a
perfect phylogeny so it is compatible. There is a proper triangulation of int(C 0 )
by Theorem 1, so int(C 0 ) has a proper minimal triangulation as well. To see that
condition (a) of Theorem 2 holds, suppose that H10 and H20 are proper minimal
triangulations of int(C 0 ) given by Theorem 1, and let U be the set of vertices of
int(C) not in int(C 0 ). By Lemma 9, there are minimal triangulations H1 and H2
of int(C) that satisfy H10 = H1 −U and H20 = H2 −U . The displayed characters of
H1 and H2 must be at least C 0 , because any fill edge that breaks a character of C 0
would also appear in H10 or H20 , and both H10 and H20 are proper triangulations
of int(C 0 ). By (i), we have H1 = H = H2 . Therefore H10 = H − U = H20 , so
condition (a) of Theorem 2 is satisfied with respect to C 0 .
Now we show condition (b) of Theorem 2 holds. Let Tr be the unique clique
tree of H given by (ii), and suppose Tx is the X−tree induced by Tr . Tx displays
C 0 by Lemma 11, and by Theorem 10 the underlying tree T of Tr is also the
underlying tree of Tx . Further, T is ternary and has |X| leaves by (ii), so Tx
must be free and ternary. To see that Tx is distinguished by C 0 , consider an edge
uv of T . By (iii) uv is incontractible with respect to C 0 , so there is a character
χ ∈ C 0 and distinct cells A and A0 of χ such that u is a node of Tr (A, χ) and v is
a node of Tr (A0 , χ). But Tx (A) = Tr (A, χ) and Tx (A0 ) = Tr (A0 , χ) by Theorem
10, so χ distinguishes uv. Hence condition (b) of Theorem 2 also holds with
respect to C 0 , so C 0 defines Tx .
Last, we show that no proper superset of C 0 also defines an X−tree. If any
superset C ∗ of C 0 was compatible, then by Lemma 10 some minimal triangulation
of int(C) has at least C ∗ as its displayed character set. This would contradict (i),
so no such superset can exist. This completes the proof.
t
u
Lemma 12. Suppose C 0 is a maximal defining subset of C, and H, H 0 are minimal triangulations of int(C) with C 0 as its displayed character set. Then H = H 0 .
Proof. Let Tx be an X−tree induced by a clique tree Tr of H and Tx0 be an
X−tree induced by a clique tree Tr0 of H 0 . By Lemma 11, Tr and Tr0 are perfect
phylogenies for C 0 , and because C 0 defines an X−tree it must be that Tx ∼
= Tx0 via
isomorphism ψ. Additionally, for each vertex (A, χ) of int(C) we have Tr (A, χ) =
Tx (A) and Tr0 (A, χ) = Tx0 (A) by Theorem 10.
To prove that H = H 0 , it suffices to show that their fill edge sets are the same.
Suppose that (A1 , χ1 )(A2 , χ2 ) is a fill edge of H. By the edge coverage property
of clique trees, Tr (A1 , χ1 ) and Tr (A2 , χ2 ) intersect at a node v of T . We will
show that Tr0 (A1 , χ1 ) and Tr0 (A2 , χ2 ) intersect at ψ(v). If there is an a ∈ A1 such
that φ(a) = v, then ψ(v) = φ0 (a) is a node of Tx0 (A1 ) = Tr0 (A1 , χ1 ). Otherwise
13
there are a1 , a2 ∈ A1 and v is an internal node on the path from φ(a1 ) = v1
to φ(a2 ) = v2 . Because ψ is a graph isomorphism, ψ(v) is an internal node on
the path from ψ(v1 ) to ψ(v2 ). Further, ψ(v1 ) and ψ(v2 ) are nodes of Tx0 (A1 ) =
Tr0 (A1 , χ1 ), so ψ(v) is a node of Tr0 (A1 , χ1 ) as well. In both cases ψ(v) is a node
of Tr0 (A1 , χ1 ), and a similar argument shows that ψ(v) is a node of Tr0 (A2 , χ2 ).
Therefore Tr0 (A1 , χ1 ) and Tr0 (A2 , χ2 ) intersect at ψ(v), so (A1 , χ1 )(A2 , χ2 ) is a
fill edge of H 0 , and the fill edge set of H is a subset of the fill edge set of H 0 . A
symmetric argument shows that the fill edge set of H is a subset of the fill edge
set of H 0 , so these fill edge sets must be equal, completing the proof.
t
u
Proposition 2. Let C be a set of partial characters on X and C 0 be a maximal
defining subset of C. Then the following conditions hold:
(i) int(C) has a unique minimal triangulation H that has C 0 as its displayed
characters, and no other minimal triangulation of int(C) has at least C 0 as
its displayed character set;
(ii) H is a ternary ur-chordal graph with leafage |X|; and
(iii) each edge of H’s unique clique tree is incontractable with respect to C 0 .
Further, if C 0 defines Tx , then int(C, Tx ) = H.
Proof. To see that (i) holds, first observe that C 0 is compatible by definition.
There is a minimal triangulation H1 of int(C) with at least C 0 as its displayed
character set by Lemma 10. Because C 0 is a maximal defining subset of C, there
is no C 0 ⊂ C ∗ ⊆ C that is compatible. By Lemma 11, the displayed characters of
H1 are compatible, so this set must be exactly C 0 . This is true of any minimal
triangulation of int(C) that has at least C 0 as its displayed characters. If H2 is
such a minimal triangulation, then H1 = H2 by Lemma 12, so there is a unique
minimal triangulation of int(C) that has at least C 0 as its displayed characters.
We will refer to this unique minimal triangulation as H in the remainder of the
proof.
Now we show that (ii) holds. Let Tx be the X−tree induced by a clique tree
of H. By Lemma 11, Tx is a perfect phylogeny for C 0 , and since C 0 is a maximal
defining subset it must be that C 0 defines Tx . Recall that Tx is free, ternary,
and distinguished by C 0 according to Theorem 2. By Theorem 9, int(C, Tx ) is urchordal. On the other hand, H = int(C, Tx ) by Theorem 10, so H is ur-chordal
as well. By the same theorem, H’s unique clique tree has the same underlying
tree as Tx . Since Tx is ternary, this clique tree must also be ternary, so H is a
ternary ur-chordal graph. This proves statement (ii).
Now consider condition (iii), and let uv be an edge of T . By Theorem 2
the edge uv is distinguished by C 0 , so there is a character χ ∈ C 0 that has cells
A 6= A0 and u is a node of Tx (A) and v is a node of Tx (A0 ). From Theorem 10 we
see that Tx (A) = Tr (A, χ) and Tx (A0 ) = Tr (A0 , χ), so uv is incontractable with
respect to C 0 . Hence Tr is incontractable with respect to C 0 .
The remainder of the theorem was shown while proving (ii) holds.
t
u
Proof of Theorem 6. Propositions 1 and 2 show that C 0 is a maximal defining
subset of C if and only if conditions (a) – (c) hold. The fact that Tx is free,
14
ternary, and distinguished by C 0 follows by Theorem 2. Finally, int(C, Tx ) = H
due to Proposition 2.
t
u
Proof of Theorem 4. Use Theorem 6, taking C 0 = C.
5
t
u
Discussion
We conclude with a brief discussion on the role minimal separators play in minimal triangulation theory [15], and how our characterization may contribute towards constructing an algorithm that sometimes finds a maximal defining subset
of characters when one exists. Minimal triangulations have been characterized
by their minimal separators, which happen to be minimal separators of the
triangulated graph as well [17, 22]. Further, a minimal separator of a minimal
triangulation has connected components (and full components) that are identical
in the graph that has been triangulated [15].
Bouchitté and Todinca [4, 5] used minimal separators and potential maximal cliques, the maximal cliques of minimal triangulations, to create a dynamic
programming algorithm to solve the treewidth and minimum-fill problems in
time polynomial in the number of minimal separators of a graph. This approach
was extended to create a dynamic programming algorithm that solves a variety perfect phylogeny problems in [11], including the unique perfect phylogeny
problem.
Our results elucidate the structure of minimal separators of triangulations
associated with maximal defining subsets of characters. This structure is retained in the partition intersection graph, and is closely related to the structure
of potential maximal cliques, because the connected components obtained by
removing the vertices in a potential maximal clique have neighborhoods that
are minimal separators [4]. This may allow for the computation of a ternary
ur-chordal minimal triangulation in time polynomial in the number of minimal
separators of int(C) (or asserting that no ternary ur-chordal minimal triangulations exist), yielding a candidate subset C 0 of C that may be a maximal subset
of characters. The number of minimal separators of int(C 0 ) is bounded by the
number of minimal separators of int(C) (this is a specific example of a more
general fact; see Corollary 4 in [5]). Therefore if it is computationally feasible to
find C 0 due to int(C) having a small number of minimal separators, checking if
C 0 defines a perfect phylogeny using the method from [11] may also be feasible.
6
Acknowledgements
This research was partially supported by NSF grants IIS-0803564 and CCF1017580.
15
References
1. Anne Berry and Romain Pogorelcnik. A simple algorithm to generate the minimal
separators and the maximal cliques of a chordal graph. Information Processing
Letters, 111(11):508–511, 2011.
2. J.R.S. Blair and B.W. Peyton. An introduction to chordal graphs and clique
trees. In J.A. George, J.R. Gilbert, and J.W-H. Liu, editors, Graph Theory and
Sparse Matrix Computations, volume 56 of IMA Volumes in Mathematics and its
Applications, pages 1–27. Springer–Verlag.
3. M. Bordewich, K.T. Huber, and C. Semple. Identifying phylogenetic trees. Discrete
Mathematics, 300(1–3):30–43, 2005.
4. V. Bouchitté and I. Todinca. Treewidth and minimum fill-in: grouping the minimal
separators. SIAM Journal on Computing, 31(1):212–232, 2001.
5. V. Bouchitté and I. Todinca. Listing all potential maximal cliques of a graph.
Theoretical Computer Science, 276(1-2):17–32, 2002.
6. P. Buneman. A characterisation of rigid circuit graphs. Discrete Mathematics,
9(3):205–212, 1974.
7. G.A. Dirac. On rigid circuit graphs. Abhandlungen aus dem Mathematischen
Seminar der Universität Hamburg, 25(1–2):71–76, 1961.
8. F. Gavril. The intersection graphs of subtrees in trees are exactly the chordal
graphs. Journal of Combinatorial Theory, 16(1):47–56, 1974.
9. M.C. Golumbic. Algorithmic graph theory and perfect graphs. Elsevier Science,
2nd edition edition, 2004.
10. Dan Gusfield. The multi–state perfect phylogeny problem with missing and removable data: solutions via integer–programming and chordal graph theory. Journal
of Computational Biology, 17(3):383–399, 2010.
11. R. Gysel. Potential maximal clique algorithms for perfect phylogeny problems.
Pre-print: arXiv, 1303.3931.
12. R. Gysel and D. Gusfield. Extensions and improvements to the chordal graph
approach to the multistate perfect phylogeny problem. IEEE/ACM Transactions
on Computational Biology and Bioinformatics, 8(4):912–917, 2011.
13. R. Gysel, F. Lam, and D. Gusfield. Constructing perfect phylogenies and proper
triangulations for three–state characters. Algorithms for Molecular Biology, 7:26,
2012.
14. H. Hara and A. Takemura. Boundary cliques, clique trees and perfect sequences of
maximal cliques of a chordal graph. Technical Report METR 2006-41, Department
of Mathematical Informatics, University of Tokyo, 2006.
15. P. Heggernes. Minimal triangulations of graphs: a survey. Discrete Mathematics,
306(3):297–317, 2006.
16. C. Ho and R.C.T. Lee. Counting clique trees and computing perfect elimination
schemes in parallel. Information Processing Letters, 31(2):61–68, 1989.
17. T. Kloks, D. Kratsch, and J. Spinrad. On treewidth and minimum fill-in of asteroidal triple-free graphs. Theoretical Computer Science, 175(2):309–335, 1997.
18. P.S. Kumar and C.E.V. Madhavan. Clique tree generalization and new subclasses
of chordal graphs. Discrete Applied Mathematics, 117(1–3):109–131, 2002.
19. I.J. Lin, T.A. McKee, and D.B. West. The leafage of a chordal graph. Discussiones
Mathematicae Graph Theory, 18:23–48, 1998.
20. T.A. McKee and F.R. McMorris. Topics in Intersection Graph Theory. Number 2
in SIAM Monographs on Discrete Mathematics and Applications. 1999.
16
21. C.A. Meacham. Theoretical and computational considerations of the compatibility
of qualitative taxonomic characters. In J. Felsenstein, editor, Numerical Taxonomy,
volume 1 of NATO ASI Series G, pages 304–314. Springer–Verlag, 1983.
22. A. Parra and P. Scheffler. Characterizations and algorithmic applications of chordal
graph embeddings. Discrete Applied Mathematics, 79(1–3):171–188, 1997.
23. D.J. Rose. Triangulated graphs and the elimination process. Journal of Mathematical Analysis and Applications, 32(3):597–609, 1970.
24. C. Semple and M. Steel. A characterization for a set of partial partitions to define
an X -tree. Discrete Mathematics, 247(1–3):169–186, 2002.
25. C. Semple and M. Steel. Phylogenetics. Oxford Lecture Series in Mathematics and
Its Applications. Oxford University Press, 2003.
26. M. Steel. The complexity of reconstructing trees from qualitative characters and
subtrees. Journal of Classification, 9(1):91–116, 1992.
27. J.R. Walter. Representations of chordal graphs as subtrees of a tree. Journal of
Graph Theory, 2(3):265–267, 1978.
17
| 5cs.CE
|
arXiv:1604.01551v1 [math.AC] 6 Apr 2016
Stably co-tame polynomial automorphisms
over commutative rings
Shigeru Kuroda∗
Abstract
We say that a polynomial automorphism φ in n variables is stably
co-tame if the tame subgroup in n variables is contained in the subgroup generated by φ and affine automorphisms in n + 1 variables. In
this paper, we give conditions for stably co-tameness of polynomial
automorphisms.
1
Introduction
Let R be a commutative ring of characteristic p 6= 1, R[x] := R[x1 , . . . , xn ]
the polynomial ring in n variables over R, and GAn (R) := AutR R[x] the
automorphism group of the R-algebra R[x]. We identify each φ ∈ GAn (R)
with the n-tuple (φ(x1 ), . . . , φ(xn )) of elements of R[x]. The composition is
defined by
φ ◦ ψ = (g1 (f1 , . . . , fn ), . . . , gn (f1 , . . . , fn ))
(1.1)
for φ = (f1 , . . . , fn ), ψ = (g1 , . . . , gn ) ∈ R[x]n . For each r ≥ 1, we regard
GAn (R) as a subgroup of GAn+r (R) by identifying each φ ∈ GAn (R) with
the unique extension φ̃ ∈ GAn+r (R) of φ defined by φ̃(xn+i ) = xn+i for
i = 1, . . . , r.
We say that φ ∈ GAn (R) is affine if φ = (x1 , . . . , xn )A + b for some
A ∈ GLn (R) and b ∈ Rn , and set Aff n (R) := {φ ∈ GAn (R) | φ is affine}. We
define
ǫ(f ) := (x1 + f, x2 , . . . , xn ) ∈ GAn (R) for each f ∈ R[x̂] := R[x2 , . . . , xn ],
and set En (R) := {ǫ(f ) | f ∈ R[x̂]}. We call Tn (R) := hAff n (R), En (R)i
the tame subgroup, and elements of Tn (R) are said to be tame. Here,
∗
Partly supported by JSPS KAKENHI Grant Number 15K04826
1
for subsets S1 , . . . , Sr and elements g1 , . . . , gs of a group
S G, we denote by
hS1 , . . . , Sr , g1 , . . . , gs i the subgroup of G generated by ri=1 Si ∪ {g1 , . . . , gs }.
If n ≥ 3 and R contains Q, then Tn (R) = hAff n (R), σi holds for σ = ǫ(x22 )
by Derksen (cf. [Es, Thm. 5.2.1]). We remark that Derksen’s theorem requires
that R is generated by R∗ as a Q-vector space, but this assumption is in
fact unnecessary (cf. §3). When R is a field and p = 0, Bodnarchuk [B]
proved a similar result for more general σ’s. The situation is different if p is
a prime. Maubach-Willems [MW] showed that T3 (F2 ) 6= hAff 3 (F2 ), ǫ(x22 )i,
and conjectured that, if n ≥ 3 and R is a finite field, then no finite subset E
of Tn (R) satisfies Tn (R) = hAff n (R), Ei.
Edo [Ed] found a class of φ ∈ GAn (R) for which hAff n (R), φi contains
(R).
Such φ is said to be co-tame. If R is a field, no element of GA2 (R)
Tn
is co-tame thanks to Jung [J] and van der Kulk [K]. For n ≥ 3, it is easy to
find elements of GAn (R) \ Aff n (R) which are not co-tame if R is not a field
(cf. (4.1)), or if p is a prime (cf. [EK, §1 (d)]). In the case where n ≥ 3, R is
a field and p = 0, the first example of such an automorphism was found by
Edo-Lewis [EL] for n = 3.
Recall that φ ∈ GAn (R) is said to be stably tame if φ belongs to Tn+1 (R).
It is known that some non-tame automorphisms are stably tame (cf. [BEW],
[N], [SU], [S]). The following is an analogue to the stably tame automorphisms.
Definition 1.1. We say that φ ∈ GAn (R) is stably co-tame if hAff n+1 (R), φi
contains Tn (R), or equivalently hAff n+1 (R), φi contains En (R).
Clearly, co-tame automorphisms are stably co-tame. When R is a field,
there exist elements of T3 (R) which are not co-tame but stably co-tame in
both cases p = 0 and p > 0 (cf. §6.1). The purpose of this paper is to
study when elements of GAn (R) are stably co-tame or not. If R contains
an infinite field, we have a necessary and sufficient condition for stably cotameness (Corollary 2.4).
This paper is organized as follows. The main results are stated in Section 2, and three key results are proved in Sections 3, 4 and 5. In Section 6,
we study stably co-tameness of the example of Edo-Lewis. We also discuss
a technique which is useful when R does not contain an infinite field.
2
Main results
Since T1 (R) = Aff 1 (R), we always assume that n ≥ 2 unless otherwise
stated. Take any φ ∈ GAn (R). We define Mφ to be the R-submodule of
2
R[x] generated by 1, xi and η(φ(xi )) for i = 1, . . . , n and η ∈ Aff n (R). By
definition, we have
η(Mφ ) ⊂ Mφ
for each η ∈ Aff n (R).
(2.1)
The following theorem holds for any commutative ring R of characteristic
p 6= 1.
Theorem 2.1. φ ∈ GAn (R) is stably co-tame in the following four cases:
(a) Mφ contains xi xj for some 1 ≤ i < j ≤ n.
(b) 2 is a unit of R, and Mφ contains x2i for some 1 ≤ i ≤ n.
(c) n = p = 2 and Mφ contains x21 x2 or x1 x22 .
(d) n = p = 2, Mφ contains x31 or x32 , and there exists ξ ∈ R∗ satisfying
ξ + 1 ∈ R∗ .
Next, assume that p is zero or a prime. For a ∈ R and t1 , . . . , tn ≥ 0, we
call axt11 · · · xtnn a good monomial in the following five cases:
(I) p = 0 and t1 + · · · + tn ≥ 2.
(II) p ≥ 2 and there exist 1 ≤ i < j ≤ n such that ti , tj ≡ 1 (mod p).
(III) p ≥ 3 and there exists 1 ≤ i ≤ n such that ti 6≡ 0, 1 (mod p).
(IV) n = p = 2 and there exist i, j ∈ {1, 2} such that ti ≡ 1 and tj ≡ 2
(mod 4).
(V) n = p = 2 and there exists 1 ≤ i ≤ 2 such that ti ≡ 3 (mod 4).
For each f ∈ R[x], let Cf denote the set of the coefficients of good
monomials
appearing in f . We define Iφ to be the ideal of R generated
Sn
by i=1 Cφ(xi ) .
Theorem 2.2. Assume that p is zero or a prime. If φ ∈ GAn (R) satisfies
Iφ 6= R, then φ is not stably co-tame.
Throughout this paper, let k be a field. When R is a commutative kalgebra, we say that f ∈ R[x] satisfies the degree condition if
degsxi f ≤ #k − 2 for i = 1, . . . , n.
(2.2)
Here, degsxi f denotes the separable degree of f if p > 0, and the standard
degree degxi f of f if p = 0, as a polynomial in xi . The separable degree of
a polynomial g(x) ∈ R[x] in one variable is defined as the degree of h(x) ∈
e
R[x] \ R[xp ] such that h(xp ) = g(x) for some e ≥ 0. We say that φ ∈
GAn (R) satisfies the degree condition if φ(xi ) satisfies the degree condition
for i = 1, . . . , n.
Now,
φ to be the ideal of R generated by the union of Cf for
P we define JP
f ∈ ni=1 Rφ(xi ) + ni=1 Rxi satisfying the degree condition. Since no good
3
monomial is linear, Jφ is contained in Iφ . If φ satisfies the degree condition,
then Jφ is equal to Iφ . Hence, (ii) of the following theorem is a consequence
of (i).
Theorem 2.3. Let k be a field, R a commutative k-algebra, and φ ∈ GAn (R).
(i) If Jφ = R, then φ is stably co-tame.
(ii) If φ satisfies the degree condition and if Iφ = R, then φ is stably co-tame.
From Theorems 2.2 and 2.3 (ii), we obtain the following corollary.
Corollary 2.4. Let k be an infinite field, and R a commutative k-algebra.
Then, φ ∈ GAn (R) is stably co-tame if and only if Iφ = R.
In particular, if R = k is an infinite field, then φ ∈ GAn (k) is stably
co-tame if and only if a good monomial appears in φ(xi ) for some i. When
p = 0, this is the same as φ is non-affine.
In the following three sections, we prove Theorems 2.1, 2.2 and 2.3 (i).
3
Proof of Theorem 2.1
Let R be any commutative ring, and Γ a subgroup of GAn+1 (R) containing
Aff n+1 (R), where n ≥ 2. First, we study properties of Γ. Define
ǫ̂(f ) := (x1 , . . . , xn , xn+1 + f ) ∈ GAn+1 (R) for each f ∈ R[x].
We identify each permutation σ ∈ Sn+1 with (xσ(1) , . . . , xσ(n+1) ) ∈ Aff n+1 (R),
and write φσ := σ −1 ◦ φ ◦ σ for φ ∈ GAn+1 (R). Then, we have the following:
(A) If φ = (f1 , . . . , fn ) ∈ GAn (R) belongs to Γ, then ǫ̂(fi ) = φ ◦ ǫ̂(xi ) ◦ φ−1
belongs to Γ for each 1 ≤ i ≤ n.
(B) If f, g ∈ R[x̂] satisfy ǫ(f ), ǫ(gxn+1 ) ∈ Γ, then we have ǫ̂(f ) = ǫ(f )(1,n+1) ∈
Γ, and so
ǫ(f g) = ǫ̂(f ) ◦ ǫ(gxn+1 ) ◦ ǫ̂(f )−1 ◦ ǫ(gxn+1 )−1 ∈ Γ.
In particular, ǫ(f ) ∈ Γ implies ǫ(af ) ∈ Γ for each f ∈ R[x̂] and a ∈ R, since
ǫ(axn+1 ) belongs to Aff n+1 (R), and hence to Γ.
When Q ⊂ R and n ≥ 3, Derksen showed that Tn (R) is generated by
Aff n (R) and {ǫ(ax22 ) | a ∈ R}, and Tn (R) = hAff n (R), ǫ(x22 )i holds if R is
generated by R∗ as a Q-vector space (cf. [Es, Thm. 5.2.1]). We remark that,
since n ≥ 3, the first statement and (B) imply that Tn (R) = hAff n (R), ǫ(x22 )i
whenever Q ⊂ R.
4
Lemma 3.1. If Γ contains ǫ̂(x1 x2 ), or if n = 2 and Γ contains ǫ̂(x1 x22 ), then
Γ contains Tn (R).
Proof. We show that Γ contains ǫ(f ) for each f ∈ R[x̂]. Since ǫ(g + h) =
ǫ(g)◦ǫ(h) holds for each g, h ∈ R[x̂], we may assume that f is a monomial. In
the former case, Γ contains ǫ̂(x1 x2 )(1,n+1)(2,i) = ǫ(xi xn+1 ) for each 2 ≤ i ≤ n.
By (B), it follows that ǫ(f ) ∈ Γ implies ǫ(xi f ) ∈ Γ. Since Γ contains ǫ(a) for
each a ∈ R, the assertion follows by induction on deg f . In the latter case,
Γ contains ǫ̂(x1 x22 )(1,3) = ǫ(x22 x3 ). Hence, ǫ(axl2 ) ∈ Γ implies ǫ(axl+2
2 ) ∈ Γ for
each a ∈ R and l ≥ 0 similarly. Since Γ contains ǫ(a) and ǫ(ax2 ), it follows
that Γ contains ǫ(axl2 ) for all l ≥ 0.
The following two implications hold for the conditions listed in Theorem 2.1.
Lemma 3.2. (b) implies (a), and (d) implies (c).
Proof. By (2.1), (b) implies (x1 + x2 )(x1 − x2 ) = x21 − x22 ∈ Mφ . Since 2 is a
unit of R, we have (x1 +x2 , x1 −x2 , x3 , . . . , xn ) ∈ Aff n (R). Hence, Mφ contains
x1 x2 by (2.1). Similarly, (d) implies x21 x2 + x1 x22 = x31 + x32 + (x1 + x2 )3 ∈ Mφ .
Hence, Mφ contains ((ξx1 )2 x2 + ξx1 x22 ) + ξ 2 (x21 x2 + x1 x22 ) = ξ(ξ + 1)x1 x22 .
Since ξ(ξ + 1) is a unit of R, this implies that Mφ contains x1 x22 .
Now, we prove Theorem 2.1. Thanks to Lemmas 3.1 and 3.2 and (2.1),
it suffices to show that Γ := hAff n+1 (R), φi contains ǫ̂(f ) for each f ∈ Mφ .
Write φ = (f1 , . . . , fn ) and
f = h+
n X
r
X
i=1 j=1
ai,j ηj (fi ), where h ∈
n
X
Rxi + R, ai,j ∈ R and ηj ∈ Aff n (R).
i=1
Then, ǫ̂(f ) is the product of ǫ̂(h) and ǫ̂(ηj (ai,j fi )) for i = 1, . . . , n and j =
1, . . . , r. Since Γ contains ǫ̂(fi ) for each i by (A), Γ contains ǫ̂(ai,j fi ) for each
i, j by (B). Hence, Γ contains ηj ◦ ǫ̂(ai,j fi ) ◦ ηj−1 = ǫ̂(ηj (ai,j fi )) for each i, j.
Since ǫ̂(h) is affine, it follows that Γ contains ǫ̂(f ). This completes the proof
of Theorem 2.1.
4
Proof of Theorem 2.2
Assume that p is zero or a prime. We define NGn (R) to be the set of φ ∈
GAn (R) such that no good monomial appears in φ(x1 ), . . . , φ(xn ). If p = 0,
then we have NGn (R) = Aff n (R). In this case, the following theorem is
obvious.
5
Theorem 4.1. NGn (R) is a subgroup of GAn (R), and no element of NGn (R)
is stably co-tame. In fact, En (R) 6⊂ hAff n+r (R), NGn (R)i holds for any r ≥ 1.
To prove Theorem 4.1, we need a lemma, where
L p need not be zero or a
prime. Consider the standard grading R[x]
L = l>0 R[x]l . An R-submodule
V of R[x] is said to be graded if V =
l>0 (V ∩ R[x]l ). If V is generated
by monomials, then V is graded. Recall that each (f1 , . . . , fn ) ∈ R[x]n is
identified with the substitution map R[x] → R[x] defined by xi 7→ fi for
i = 1, . . . , n, and R[x]n forms a monoid for the composition defined in (1.1).
Note that V n is closed under this operation if and only if φ(V ) ⊂ V holds
for each φ ∈ V n .
Lemma 4.2. Let V be a graded R-submodule of R[x] such that V n is closed
under composition. If x1 , . . . , xn ∈ V , then V n ∩ GAn (R) is a subgroup of
GAn (R).
Proof. Since V n ∩ GAn (R) contains (x1 , . . . , xn ), and is closed under composition, we show that φ−1 belongs to V n for each φ ∈ V n ∩ GAn (R).
L There
exists η ∈ V n ∩ Aff n (R) for which ψ := φ ◦ η satisfies ψ(xi ) ∈ xi + l>2 R[x]l
for i = 1, . . . , n. Since φ−1 = η ◦ ψ −1 , and V n is closed under composition,
it suffices to verify that ψ −1 belongs to V n . Suppose to theP
contrary that
f := ψ −1 (xi ) does not belong to V for some i. Write f = l>0 fl , where
fl ∈ R[x]
holds for some d ≥ 0. Take the minimal d, and set
P l . Then, fd 6∈ V P
f ′ := l<d fl and f ′′ := l>d fl . Let h, h′ and h′′ be the homogeneous components of ψ(fd ), ψ(f ′ ) and ψ(f ′′ ) of degree d, respectively. Then, h + h′ + h′′
belongs to V , since ψ(fd ) + ψ(f ′ ) + ψ(f ′′ ) = ψ(f ) = xi belongs to V , and
V is graded by assumption. By the minimality of d, we have f ′ ∈ V . Since
ψ = φ ◦ η belongs to V n , it follows that ψ(f ′ ) belongs to V . This implies that
h′ belongs to L
V as before. As for h and h′′ , we have h = fd and h′′ = 0, since
ψ(xi ) ∈ xi + l>2 R[x]l holds for i = 1, . . . , n. Thus, h + h′ + h′′ = fd + h′
does not belong to V , a contradiction. Therefore, ψ −1 belongs to V n .
P
We remark that, if V = s∈Σ As for some Σ ⊂ R[x] and an R-subalgebra
A of R[x], and if φ(A) ⊂ A and φ(Σ) ⊂ V hold for each φ ∈ V n , then
V n is closed under composition. For example, assume that R has prime
l
l
l
characteristic p, and define R[xp ] := R[xp1 , . . . , xpn ] for each l ≥ 0. Set
Zd := {a ∈ Z | a ≥ d} for each d ∈ Z. Let d, e ∈ Z0 and ∅ =
6 N ⊂ Z0 be
such that d ≤ e, and each u, v ∈ N satisfy u + v ∈ N ∪ Zd . Then, we define
a graded R-submodule of R[x] by
d
V := V (d, e, N) := R[xp ] +
n X
X
i=1 u∈N
6
e
u
R[xp ]xpi .
Let us prove that V n is closed under composition using the remark for A :=
u
e
d
R[xp ] and Σ := R[xp ] ∪ {xpi | 1 ≤ i ≤ n, u ∈ N}. First, note that f ∈ V
w
u
v
v
implies f p ∈ V for any v ∈ N, since each u, v ∈ N satisfy (xpi )p ∈ {xpi |
u
d
d
e
w ∈ N} ∪ R[xpi ], and R[xp ] contains R[xp ]. Hence, φ(xpi ) belongs to V for
l
l
each φ ∈ V n , 1 ≤ i ≤ n and u ∈ N. Clearly, φ(R[xp ]) ⊂ R[xp ] holds for any
φ ∈ R[x]n and l ≥ 0. Therefore, we have φ(A) ⊂ A and φ(Σ) ⊂ V for each
φ ∈ V n.
Now, let us prove Theorem 4.1 when p is a prime. Clearly, u, v ∈ {0}
implies u + v ∈ {0} ∪ Z1 . So, we define Vn := V (1, 1, {0}) and Wn :=
V (1, 2, {0}), i.e.,
Vn = R[xp ] + R[xp ]x1 + · · · + R[xp ]xn
2
2
Wn = R[xp ] + R[xp ]x1 + · · · + R[xp ]xn .
If n = p = 2, then t1 , t2 ≥ 0 satisfy xt11 xt22 ∈ W2 if and only if t1 , t2 ≡ 0
(mod 2), or ti ≡ 1, tj ≡ 0 (mod 4) for some i, j ∈ {1, 2}. Hence, we have
xt11 xt22 6∈ W2 if and only if xt11 xt22 is a good monomial. Similarly, when (n, p) 6=
(2, 2), a nonzero monomial m is good if and only if m does not belong to Vn .
Thus, NGn (R) is equal to W22 ∩ GA2 (R) if n = p = 2, and to Vnn ∩ GAn (R)
otherwise. Therefore, NGn (R) is a subgroup of GAn (R) by Lemma 4.2. Note
n+r
n+r
that hAff n+r (R), NGn (R)i is contained in Wn+r
if n = p = 2, and in Vn+r
n+r
otherwise. Since En (R) \ Vn+r
contains ǫ(x2 x3 ) if n ≥ 3, and ǫ(x22 ) if p ≥ 3,
n+r
while En (R) \ Wn+r contains ǫ(xp+1
2 ) for any n, p ≥ 2, we get the last part of
Theorem 4.1. This completes the proof.
If I is a proper ideal of R, then each φ ∈ GAn+r (R) induces an element
φI of GAn+r (R/I). Since En (R) ∋ τ 7→ τI ∈ En (R/I) is surjective, (σ ◦ τ )I =
σI ◦τI for each σ, τ ∈ GAn+r (R), and τ ∈ Aff n+r (R) implies τI ∈ Aff n+r (R/I),
we see that
En (R) ⊂ hAff n+r (R), φi implies En (R/I) ⊂ hAff n+r (R/I), φI i
(4.1)
for any r ≥ 0 and φ ∈ GAn (R). Hence, Theorem 4.1 implies the following
corollary.
Corollary 4.3. Let I be a proper ideal of R such that char(R/I) is zero
or a prime. If φ ∈ GAn (R) satisfies φI ∈ NGn (R/I), then En (R) 6⊂
hAff n+r (R), φi holds for any r ≥ 1. Here, char(R/I) is the characteristic
of R/I.
Now, we can prove Theorem 2.2. By assumption, I := Iφ is a proper ideal
of R. If p is a prime, then R contains Fp , and so char(R/I) = p. Hence, φI
belongs to NGn (R/I) by the definition of I. Thus, φ is not stably co-tame
7
by Corollary 4.3. If p = 0, then φI is affine. By (4.1), this implies that φ is
not stably co-tame.
Finally, we remark that, if p = 2, then R[x2 ] is contained in Vn . Hence,
we have E2 (R) ⊂ V22 , and so T2 (R) ⊂ V22 . Thus, for any φ ∈ T2 (R), no
monomial xt11 xt22 with t1 and t2 odd appears in φ(x1 ) and φ(x2 ). If R is a
domain, the same holds for any φ ∈ GA2 (R), since T2 (K) = GA2 (K) for any
field K (cf. [J], [K]).
Lemma 4.4. Assume that p = 2, φ ∈ GA2 (R), 1 ≤ i ≤ 2 and t1 , t2 ≥ 0
are odd. If a ∈ R is the coefficient of xt11 xt22 in φ(xi ), then a belongs to the
nilradical of R.
Proof. Let p be any prime ideal of R. Since R/p is a domain of characteristic
two, xt11 xt22 does not appear in φp (xi ) as mentioned. Hence, a belongs to p.
5
Proof of Theorem 2.3 (i)
Assume that R is a commutative k-algebra. To prove Theorem 2.3 (i), we
verify that one of (a), (b) and (c) of Theorem 2.1 holds when Jφ = R. If
Jφ = R, a good monomial appears in a polynomial satisfying the degree
condition. This implies that #k > 2, and so there always exists ξ ∈ R∗
satisfying ξ + 1 ∈ R∗ .
P
e
Take any f (x) ∈ R[x] and write f (x) = di=0 ui xip with d, e ≥ 0 and
ui ∈ R, where we regard pe = 1 if p = 0. If k contains d + 1 distinct
e
elements ξ0 , . . . , ξd , then ui xip can be written as a k-linear combination of
f (ξ0 x), . . . , f (ξd x) for each 0 ≤ i ≤ d by linear algebra. This remark is used
to prove the following lemma.
Lemma 5.1. If f ∈ R[x] satisfies the degree condition, each monomial appearing in f is written as a k-linear combination of f (ξ1 x1 , . . . , ξn xn ) for
ξ1 , . . . , ξn ∈ k ∗ .
Proof. We prove the lemma by P
induction on n. The case n = 0 is clear. Ase
sume that n ≥ 1, and write f = di=0 fi xip
1 , where d := #k−2, fi ∈ R[x̂] and
e ≥ 0. Take any monomial m appearing in f . Then, there exists i such that
m appears in fi xi1 . Since #k = d + 2, we may find distinct α0 , . . . , αd ∈ k ∗ .
Then, fi xi1 is written as a k-linear combination of f (αl x1 , x2 , . . . , xn ) for
l = 0, . . . , d as remarked. By the choice of i, we have m = m′ xi1 for some
monomial m′ appearing in fi . Note that degsxj fi ≤ degsxj f ≤ d for all
j. Hence, by induction assumption, m′ is written as a k-linear combination of fi (β2 x2 , . . . , βn xn ) for β2 , . . . , βn ∈ k ∗ . Thus, m = m′ xi1 is written
as a k-linear combination of fi (β2 x2 , . . . , βn xn )xi1 ’s, and therefore that of
f (αl x1 , β2 x2 , . . . , βn xn ) for l = 0, . . . , d and β2 , . . . , βn ∈ k ∗ .
8
Next, we define λ ∈ Aff n (R) by λ(xi ) = xi + 1 for i = 1, . . . , n.
Lemma 5.2. If xt11 · · · xtnn is a good monomial, there appears in λ(xt11 · · · xtnn )
the monomial uxi xj with u ∈ k ∗ in the cases (I), (II) and (III), xi x2j in the
case (IV), and x3i in the case (V), where i, j ∈ {1, . . . , n}, and where i 6= j if
(II) or (IV).
′
Proof. In the case (IV), we have λ(xt11 xt22 ) = (xi +1)(x2j +1)(x4i +1)ti (x4j +1)tj
for some i, j ∈ {1, 2} with i 6= j, and t′1 , t′2 ≥ 0. Then, it is easy to see that the
monomial xi x2j appears in λ(xt11 xt22 ). Other cases are verified similarly.
′
Now, let us prove Theorem 2.3 (i).
by assumption, we may
PSince Jφ = RP
find a1 , . . . , ar ∈ R and g1 , . . . , gr ∈ ni=1 Rφ(xi ) + ni=1 Rxi with r ≥ 1 as
follows:
(1) For
Pr each 1 ≤ l ≤ r, a good monomial ml appears in gl with coefficient al .
(2) l=1 al = 1, and g1 , . . . , gr satisfy the degree condition.
Since gl satisfies the degree condition for each l, so does ml , and hence
so does λ(ml ). Clearly, gl belongs to Mφ . Hence, ml belongs to Mφ by
Lemma 5.1 and (2.1). Then, λ(ml ) belongs to Mφ , so each monomial appearing in λ(ml ) belongs to Mφ similarly. By Lemma 5.2, there appears in
λ(ml ) the monomial nl := ul al xil xjl with ul ∈ k ∗ in the cases (I), (II) and
(III), nl := al xil x2jl in the case (IV), and nl := al x3il in the case (V), where
il , jl ∈ {1, . . . , n}, and where il 6= jl if (II) or (IV). Since nl belongs to Mφ ,
we have al x1 x2 ∈ Mφ in the case (II). The proof of Lemma 3.2 shows that
al x1 x22 ∈ Mφ in the case (V) as well as in the case (IV).
If n ≥P3 and p = 2, then ml must be of type (II) for all l. Hence, Mφ
contains rl=1 al x1 x2 = x1 x2 . Therefore, (a) of Theorem 2.1 holds.
If n = p = 2, then ml is of type (II) or (IV)
Pror (V) for each l. In the
case (II),
l=1 al = 1, it follows that
P al is nilpotent by Lemma 4.4. Since
P
u := ′ al is a unit of R, where the sum ′ isPtaken over l such that ml is
of type (IV) or (V). Hence, Mφ contains u−1 ′ al x1 x22 = x1 x22 . Therefore,
(c) of Theorem 2.1 holds.
If p 6= 2, then each ml is of type (I) or (II) or (III), so Mφ contains
ul al xil xjl . By (2.1), Mφ contains al x21 if il = jl . If il 6= jl , then Mφ contains
2
a
l x1 x2 , and hence contains al x1 (x2 + x1 ) − al x1 x2 = al x1 . Thus, Mφ contains
P
r
2
2
l=1 al x1 = x1 . Therefore, (b) of Theorem 2.1 holds. This completes the
proof of Theorem 2.3 (i).
9
6
Remarks
6.1. When R = k, Edo-Lewis [EL] showed that θN := (β ◦ π)−N ◦ π ◦ (β ◦ π)N
is neither affine nor co-tame for each N ≥ 3, where
β := (x1 + x22 (x2 + x23 )2 , x2 + x23 , x3 ) and π := (x2 , x1 , x3 ).
Here, we mention that the expression of θN above is slightly different from the
original one, due to the difference in the definitions of composition (cf. §2.2
of [EL] and (1.1)). Since R = k and θN is not affine, θN is stably co-tame
if p = 0 by the remark after Corollary 2.4. If p = 2, then θN belongs to
NG3 (k), and hence is not stably co-tame by Theorem 4.1. When p ≥ 3, we
have the following theorem.
Theorem 6.1. If p ≥ 3, N ≥ 1 and #k ≥ 42N −1 + 2, then θN is stably
co-tame.
Proof. Observe that θN is written as σ1 ◦ τ1 ◦ · · · ◦ σN ◦ τN ◦ σ, where
σi ∈ {π ◦ β ◦ π, π ◦ β −1 ◦ π},
τi ∈ {β, β −1 } and σ ∈ {id, π}
with σ1 = π ◦ β −1 ◦ π =: (h1 , h2 , x3 ). For each w = (w1 , w2 , w3 ) ∈ Z3 , we
define the w-weighted grading on R[x] by degw xi := wi for i = 1, 2, 3. For
f ∈ R[x] \ {0}, we denote by f w the highest w-homogeneous part of f . Now,
′
set θN
:= θN ◦ σ. We claim that, if degw h2 is greater than (1/4) degw h1 and
2w3 , then
2N−i
′
4
θN
(xi )w = (σ1 ◦ τ1 ◦ · · · ◦ σN ◦ τN )(xi )w = ±(hw
2 )
for i = 1, 2.
In fact, the case N = 1 is checked directly, and the case N ≥ 2 follows by
induction on N, since (σi ◦ τi )(x2 ) = x2 ± x21 (x1 ± x23 )2 ± x23 and
(σi ◦τi )(x1 ) = x1 ±x23 ±(x2 ±x21 (x1 ±x23 )2 )2 (x2 ±x21 (x1 ±x23 )2 ±x23 )2
for each i.
Since h1 = x1 − x23 and h2 = x2 − x21 h21 , the assumption of the claim holds
for w = (1, 0, 0), (0, 1, 0), (0, 0, 1), (2, 0, 1), for which
4
2 4
2
2 2
hw
2 = −x1 , x2 , −x1 x3 , −x1 (x1 − x3 ) ,
′
respectively. The first three of these imply degxi θN
(x2 ) ≤ 4 · 42N −2 ≤
′
#k − 2 for i = 1, 2, 3. Hence, θN
(x2 ) satisfies the degree condition. Since
′
(2,0,1)
2 t
θN (x2 )
= ±(x1 (x1 − x3 )) with t := 2 · 42N −2 similarly, the monomial
′
±(x1 x23 )t appears in θN
(x2 ). Since p ≥ 3, we see that xt1 x2t
3 is a good mono′
mial of type (III). Thus, we get JθN = JθN = R. Therefore, θN is stably
co-tame by Theorem 2.3 (i).
10
6.2. When k is a finite field, Theorem 2.3 might not be useful due to the
degree condition. In such a case, the following technique may be effective.
Take any η ∈ Aff n (R), and set δ := η − id. Then, δ(Mφ ) ⊂ Mφ holds for
each φ ∈ GAn (R) by (2.1). Since φ(xj ) belongs to Mφ for each j, we get the
following lemma.
Lemma 6.2. If φ ∈ GAn (R) and h ∈ R[x] satisfy the following condition,
then h belongs to Mφ : There exist 1 ≤ j ≤ n and η1 , . . . , ηl ∈ Aff n (R) with
l ≥ 1 such that (δ1 ◦ · · · ◦ δl ◦ φ)(xj ) belongs to R∗ h + Mφ , where δi := ηi − id
for each i.
Assume that p is a prime. For i = 1, . . . , n, take any ai ∈ R∗ , define
ηi ∈ Aff n (R) by ηi (xi ) = xi + ai and ηi (xj ) = xj for j 6= i, and put δi :=
ηi − id. Then, we have δip = 0 since ηip = id, degxi δi (xdi ) < d for d ≥ 0, and
P
Rxji if p ∤ d − l for all 0 ≤ l < t. Note that ker δi
δit (xdi ) ∈ R∗ xd−t
+ d−t−1
i
j=0
is an R-subalgebra of R[x], and δi is ker δi -linear. Also, ker δi contains the
R-subalgebra Ai generated by qi := xpi − ap−1
xi and xj for j 6= i. We claim
i
that ker δi = Ai . In fact, each f ∈ ker δi \ {0} satisfies p | degxi f =: lp,
hence degxi (f − hqil ) < lp holds for some h ∈ R[{xj | j 6= i}]. Thus, we get
f − hqil ∈ Ai by induction on l, and so f ∈ Ai .
Now, fix l = (l1 , . . . , ln ) ∈ {0, . . . , p − 1}n , and set δ l := δ1l1 ◦ · · · ◦ δnln and
Ml :=
n
X
i=0
l
Rxi x +
li −1
n X
X
Ai xji , where x0 := 1 and xl := xl11 · · · xlnn .
i=1 j=0
Then, since δi ’s commute
P with each other, and are Ai -linear, we see that
δ l (Ml ) is contained in ni=0 Rxi . Hence, δ l (Ml ) ⊂ Mφ holds for every φ ∈
GAn (R). Using this, we can give sufficient conditions for stably co-tameness
of φ. For P
example, assume that l1 , l2 ≤ p − 2. Then, δ l (x1 x2 xl ) belongs to
R∗ x1 x2 + 2i=0 Rxi . Hence, if φ satisfies φ(xj ) ∈ R∗ x1 x2 xl + Ml for some
1 ≤ j ≤ n, then (δ l ◦ φ)(xj ) belongs to R∗ x1 x2 + Mφ . This implies that x1 x2
belongs to Mφ by Lemma 6.2. Therefore, φ is stably co-tame by Theorem 2.1.
Similarly, if p ≥ 3, l1 ≤ p − 3, and φ(xj ) belongs to R∗ x21 xl + Ml for some
1 ≤ j ≤ n, then φ is stably co-tame.
6.3. When n ≥ 3, Motoki Kuroda [MK] showed that ǫ(x22 ) is stably cotame for R = Fp with p ≥ 3, and found it hard to decide whether ǫ(x52 ) is
stably co-tame for R = F3 . Note that ǫ(x52 ) is stably co-tame for R = F9 by
Theorem 2.3 (ii).
6.4. The Tame Generators Problem asks if GAn (R) = Tn (R). In the context of stably co-tame automorphisms, NGn (R) is a natural generalization of
Aff n (R). So, the following problem is of interest, as a generalization of the
11
Tame Generators Problem in prime characteristic (see also [EK] for other
generalizations).
Problem. When p is a prime, does it hold that GAn (R) = hNGn (R), Tn (R)i?
References
[BEW] J. Berson, A. van den Essen and D. Wright, Stable tameness of
two-dimensional polynomial automorphisms over a regular ring, Adv.
Math. 230 (2012), no. 4-6, 2176–2197.
[B]
Y. Bodnarchuk, On generators of the tame invertible polynomial maps
group, Internat. J. Algebra Comput. 15 (2005), no. 5-6, 851–867.
[Ed]
E. Edo, Coordinates of R[x, y]: constructions and classifications,
Comm. Algebra 41 (2013), no. 12, 4694–4710.
[EK]
E. Edo and S. Kuroda, Generalisations of the tame automorphisms
over a domain of positive characteristic, Transform. Groups 20
(2015), no. 1, 65–81.
[EL]
E. Edo and D. Lewis, The affine automorphism group of A3 is not a
maximal subgroup of the tame automorphism group, Michigan Math.
J. 64 (2015), no. 3, 555–568.
[Es]
A. van den Essen, Polynomial automorphisms and the Jacobian
conjecture, Progress in Mathematics, Vol. 190, Birkhäuser, Basel,
Boston, Berlin, 2000.
[J]
H. Jung, Über ganze birationale Transformationen der Ebene, J.
Reine Angew. Math. 184 (1942), 161–174.
[K]
W. van der Kulk, On polynomial rings in two variables, Nieuw Arch.
Wisk. (3) 1 (1953), 33–41.
[MK]
M. Kuroda, Stably Derksen polynomial automorphisms over finite
fields (Japanese), Master’s Thesis, Tokyo Metropolitan University,
January 2016.
[MW] S. Maubach and R. Willems, Polynomial automorphisms over finite
fields: mimicking tame maps by the Derksen group, Serdica Math. J.
37 (2011), no. 4, 305–322 (2012).
12
[N]
M. Nagata, On Automorphism Group of k[x, y], Lectures in Mathematics, Department of Mathematics, Kyoto University, Vol. 5, Kinokuniya Book-Store Co. Ltd., Tokyo, 1972.
[SU]
I. Shestakov and U. Umirbaev, The tame and the wild automorphisms
of polynomial rings in three variables, J. Amer. Math. Soc. 17 (2004),
197–227.
[S]
M. K. Smith, Stably tame automorphisms, J. Pure Appl. Algebra 58
(1989), 209–212.
Department of Mathematics and Information Sciences
Tokyo Metropolitan University
1-1 Minami-Osawa, Hachioji
Tokyo 192-0397, Japan
[email protected]
13
| 4math.GR
|
Generalized Value Iteration Networks:
Life Beyond Lattices
Sufeng Niu∗† Siheng Chen∗‡ , Hanyu Guo∗† , Colin Targonski† , Melissa C. Smith† , Jelena Kovačević‡
†
arXiv:1706.02416v2 [cs.LG] 26 Oct 2017
‡
Clemson University, 433 Calhoun Dr., Clemson, SC 29634, USA
Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA
Abstract
In this paper, we introduce a generalized value iteration network (GVIN), which is an end-to-end neural network planning
module. GVIN emulates the value iteration algorithm by using
a novel graph convolution operator, which enables GVIN to
learn and plan on irregular spatial graphs. We propose three
novel differentiable kernels as graph convolution operators
and show that the embedding-based kernel achieves the best
performance. Furthermore, we present episodic Q-learning,
an improvement upon traditional n-step Q-learning that stabilizes training for VIN and GVIN. Lastly, we evaluate GVIN
on planning problems in 2D mazes, irregular graphs, and realworld street networks, showing that GVIN generalizes well
for both arbitrary graphs and unseen graphs of larger scale
and outperforms a naive generalization of VIN (discretizing a
spatial graph into a 2D image).
Introduction
Reinforcement learning (RL) is a technique that solves
sequential decision making problems that lacks explicit rules
and labels (Sutton and Barto 1998). Recent developments
in Deep Reinforcement Learning(DRL) have lead to enormous progress in autonomous driving (Bojarski et al. 2016),
innovation in robot control (Levine et al. 2015), and humanlevel performance in both Atari games (Mnih et al. 2013;
Guo et al. 2014) and the board game Go (Silver et al. 2016a).
Given a reinforcement learning task, the agent explores the
underlying Markov Decision Process (MDP) (Bellman 1957;
Bertsekas et al. 1995) and attempts to learn a mapping
of high-dimensional state space data to an optimal policy
that maximizes the expected return. Reinforcement learning can be categorized as model-free (Lillicrap et al. 2015;
Mnih et al. 2016; 2013) and model-based approaches (Sutton and Barto 1998; Deisenroth and Rasmussen 2011;
Schmidhuber 1990). Model-free approaches learn the policy
directly by trial-and-error and attempt to avoid bias caused
by a suboptimal environment model (Sutton and Barto 1998);
the majority of recent architectures for DRL follow the
model-free approach (Lillicrap et al. 2015; Mnih et al. 2016;
2013). Model-based approaches, on the other hand, allow
for an agent to explicitly learn the mechanisms of an environment, which can lead to strong generalization abilities. A
∗
Equal contribution.
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
recent work, the value iteration networks (VIN) (Tamar et al.
2016) combines recurrent convolutional neural networks and
max-pooling to emulate the process of value iteration (Bellman 1957; Bertsekas et al. 1995). As VIN learns an environment, it can plan shortest paths for unseen mazes.
The input data fed into deep learning systems is usually associated with regular structures. For example, speech
signals and natural language have an underlying 1D sequential structure; images have an underlying 2D lattice
structure. To take advantage of this regularly structured
data, deep learning uses a series of basic operations defined for the regular domain, such as convolution and uniform pooling. However, not all data is contained in regular structures. In urban science, traffic information is associated with road networks; in neuroscience, brain activity
is associated with brain connectivity networks; in social
sciences, users’ profile information is associated with social networks. To learn from data with irregular structure,
some recent works have extended the lattice structure to general graphs (Defferrard, Bresson, and Vandergheynst 2016;
Kipf and Welling 2016) and redefined convolution and pooling operations on graphs; however, most works only evaluate
data that has both a fixed and given graph. In addition, most
lack the ability to generalize to new, unseen environments.
In this paper, we aim to enable an agent to self-learn and
plan the optimal path in new, unseen spatial graphs by using model-based DRL and graph-based techniques. This task
is relevant to many real-world applications, such as route
planning of self-driving cars and web crawling/navigation.
The proposed method is more general than classical DRL,
extending for irregular structures. Furthermore, the proposed
method is scalable (computational complexity is proportional
to the number of edges in the testing graph), handles various
edge weight settings and adaptively learns the environment
model. Note that the optimal path can be self-defined, and is
not necessarily the shortest one. Additionally, the proposed
work differs from conventional planning algorithms; for example, Dijkstra’s algorithm requires a known model, while
GVIN aims to learn a general model via trial and error, then
apply said model to new, unseen irregular graphs.
To create GVIN, we generalize VIN in two aspects. First,
to work for irregular graphs, we propose a graph convolution
operator that generalizes the original 2D convolution operator. With the new graph convolution operator, the proposed
network captures the basic concepts in spatial graphs, such
as direction, distance and edge weight. It also is able to transfer knowledge learned from one graph to others. Second, to
improve reinforcement learning on irregular graphs, we propose a reinforcement learning algorithm, episodic Q-learning,
which stabilizes the training for VIN and GVIN. The original VIN is trained through either imitation learning, which
requires a large number of ground-truth labels, or reinforcement learning, whose performance is relatively poor. With
the proposed episodic Q-learning, the new network performs
significantly better than VIN in the reinforcement learning
mode. Since the proposed network generalizes the original
VIN model, we call it the generalized value iteration network
(GVIN).
The main contributions of this paper are:
• The proposed architecture, GVIN, generalizes the VIN
(Tamar et al. 2016) to handle both regular structures and irregular structures. GVIN offers an end-to-end architecture
trained via reinforcement learning (no ground-truth labels);
see Section Framework;
• The proposed graph convolution operator generalizes 2D
convolution learns the concepts of direction and distance,
which enables GVIN to transfer knowledge from one graph
to another; see Section Graph Convolution;
• The proposed reinforcement learning algorithm, episodic
Q-learning, extends the classical n-step Q-learning as Monte
Carlo control and significantly improves the performance of
reinforcement learning for irregular graphs; see Section Training via Reinforcement Learning; and
• Through intensive experiments we demonstrate the generalization ability of GVIN within imitation learning and
episodic Q-learning for various datasets, including synthetic
2D maze data, irregular graphs, and real-world maps (Minnesota highway and New York street maps); we show that
GVIN significantly outperforms VIN with discretization input on irregular structures; See Section Experimental Results.
Background
Markov Decision Process. We consider an environment
defined as an MDP that contains a set of states s ∈ S, a
set of actions a ∈ A, a reward function Rs,a , and a series
of transition probabilities Ps0 ,s,a , the probability of moving
from the current state s to the next state s0 given an action a.
The goal of an MDP is to find a policy that maximizes
the
P∞
expected return (accumulated rewards) Rt = k=0 γ k rt+k ,
where rt+k is the immediate reward at the (t + k)th time
stamp and γ ∈ (0, 1] is the discount rate. A policy πa,s is
the probability of taking action a when in state s. The value
of state s under a policy π, vsπ , is the expected return when
starting in s and following π; that is, vsπ = E[Rt |St = s].
The value of taking action a in state s under a policy π, qπ (a)
s ,
is the expected return when starting in s, taking the action
a and following π; that is, qπ (a)
= E[Rt |St = s, At = a].
s
There is at least one policy that is better than or equal to all
other policies, called an optimal policy π ∗ ; that is, the optimal
policy is π ∗ = arg maxπ vsπ , the optimal state-value function
is vs∗ = maxπ vsπ , and the optimal action-value function
∗
∗
is q∗ (a)
= maxπ qπ (a)
s
s . To obtain π and v , we usually
consider solving the Bellman equation. Value iteration is a
popular algorithm used to solve the Bellman equation in the
discrete
Pstate space; that is, we iteratively compute vs ←
maxa s0 Ps0 ,s,a (Rs,a +γvs0 ) until convergence.
Differentiable planning module. VIN employs an embedded differentiable planning architecture, trained end-toend via imitation learning (Tamar et al. 2016). In VIN, the
Bellman equation is encoded within the convolutional neural networks, and the policy can be obtained through backpropagation. However, VIN is limited to regular lattices; it
requires imitation learning for maximum performance and is
trained separately with a reactive policy. A more recent work
Memory Augmented Control Network (MACN) (Khan et al.
2017) combines the VIN model with a memory augmented
controller, which can then backtrack through the history of
previous trajectories. However, as we shown later in Table 2,
GVIN outperform MACN on both performance and problem scales. A different model-based work, Predictron, uses a
learning and planning model that simulates a Markov reward
process (Silver et al. 2016b). The architecture unrolls the
"imagined" plan via a predictron core. However, Predictron
is limited to the Markov rewards process and is relatively
computationally expensive compared to VIN.
Deep Learning with Graphs. A number of recent works
consider using neural networks to handle signals supported
on graphs (Niepert, Ahmed, and Kutzkov 2016; Duvenaud
et al. 2015; Henaff, Bruna, and LeCun 2015). The principal
idea is to generalize basic operations in the regular domain,
such as filtering and pooling, to the graph domain based
on spectral graph theory. For example, (Bruna et al. 2013;
Henaff, Bruna, and LeCun 2015) introduce hierarchical clustering on graphs and the spectrum of the graph Laplacian to
neural networks; (Defferrard, Bresson, and Vandergheynst
2016) generalizes classical convolutional neural networks by
using graph coarsening and localized convolutional graph
filtering; (Kipf and Welling 2016) considers semi-supervised
learning with graphs by using graph-based convolutional
neural networks; (Li et al. 2015) investigate learning graph
structure through gated recurrent unit; (Gilmer et al. 2017)
considers a message passing framework that unifies previous
work, see some recent overviews in (Bronstein et al. 2016).
Methodology
We propose a new model-based DRL framework, GVIN,
that takes a general graph with a starting node and a goal
node as inputs and outputs the designed plan. The goal of
GVIN is to learn an underlying MDP that summarizes the
optimal planning policy applied for arbitrary graphs, which
requires GVIN to capture general knowledge about planning
that is structure and transition invariant and does not depend
on any specific graph structure. A key component of an MDP
is the transition matrix, which is needed to solve the Bellman
equation. To train a general transition matrix that works for
arbitrary graphs, similar to the VIN, we treat it as a graph
convolution operator and parameterize it by using graphbased kernel functions, each of which represents a unique
action pattern. We train the parameters in GVIN by using
episodic Q-learning, which makes reinforcement learning on
irregular graphs practical.
Figure 1: Architecture of GVIN. The left module emulates value iteration and obtains the state values; the right module is
responsible for selecting an action based on an -greedy policy (for training) or a greed policy (for testing). We emphasize our
contributions, including graph convolution operator and episodic Q-learning, in the blue blocks.
Framework
The input of GVIN is a graph with a starting node and a
goal node. In the training phase, GVIN trains the parameters
by trial-and-error on various graphs; during the testing phase,
GVIN plans the optimal path based on the trained parameters.
The framework includes the planning module (left) and the
action module (right), shown in Figure 1. The planning module emulates value iteration by iteratively operating the graph
convolution and max-pooling. The action module takes the
greedy action according to the value function.
Mathematically, we consider a directed, weighted spatial
graph G = (V, X, E, A), where V = {v1 , ..., vN } is the
node set, X ∈ RN ×2 are the node embeddings with the ith
row Xi ∈ R2 being the embedding of the ith node (here we
consider 2D spatial graphs, but the method is generalizable),
E = {e1 , ..., eM } is the edge set, and A ∈ RN ×N is the
adjacency matrix, with the (i, j)th element Ai,j representing
the edge weight between the ith and jth nodes. We consider
a graph signal as a mapping from the nodes to real values.
We use a graph signal g ∈ {0, 1}N to encode the goal node,
where g is one-sparse and only activates the goal node. Let
r ∈ RN , v ∈ RN , and q ∈ RN be the reward graph signal,
the state-value graph signal, and the action-value graph signal,
respectively. We represent the entire process in a matrixvector form as follows,
P
r
=
fR (g; wr ),
(1)
(a)
=
fP (G; wP(a) ),
(2)
(a)
qn+1
vn+1
(a)
=
P
(r + γvn ) ,
=
(a)
max qn+1 .
a
(3)
(4)
In the feature-extraction step (1), g is encoded to become
the robust reward r via the feature-extract function fR (·),
which is a convolutional neural network in the case of regular graphs, but is the identity function when operating on
irregular graphs; in step (2), where P(a) is the graph convolution operator in the ath channel, a set of graph convolution
operators is trained based on the graph G, which is further
described in Section Graph Convolution; in (3) and (4), value
iteration is emulated by using graph convolution to obtain
the action-value graph signal q(a) in the ath channel and
max-pooling to obtain the state-value graph signal v. wr and
wP(a) are training parameters to parameterize r and P(a) ,
respectively. As shown in Figure 1, we repeat the graph convolution operation (3) and max-pooling (4) for K iterations
b . When G is a 2D
to obtain the final state-value graph signal v
lattice, the planning module of GVIN degenerates to VIN.
In the training phase, we feed the final state-value graph
b to the action module. The original VIN extracts
signal v
the action values from step (3) and trains the final action
probabilities for eight directions; however, this is problematic
for irregular graphs, as the number of actions (neighbors) at
b
each node varies. To solve this, we consider converting v
b ∈ RN , whose sth
to a pseudo action-value graph signal, q
bs = maxs0 ∈Nei(s) v
bs0 , representing the action
element is q
value moving from s to one of its neighbors. The advantages
of this approach come from the following three aspects: (1)
the final state value of each node is obtained by using the
maximum action values across all the channels, which is
robust to small variations; (2) the pseudo action-value graph
signal considers a unique action for each node and does not
depend on the number of actions; that is, at each node, the
agent queries the state values of its neighbors and always
moves to the one with the highest value; and (3) the pseudo
action-value graph signal considers local graph structure,
because the next state is always chosen from one of the
neighbors of the current state.
The pseudo action-value graph signal is used through
episodic Q-learning, which learns from trial-and-error experience and backpropagates to update all of the training parameters. In episodic Q-learning, each episode is obtained as
follows: for each given starting node s0 , the agent will move
sequentially from st to st+1 by the -greedy strategy; that
bs0
is, with probability (1 − ), st+1 = arg maxs0 ∈Nei(st ) v
and with probability , st+1 is randomly selected from
one of the neighbors of st . An episode terminates when
st+1 is the goal state or the maximum step threshold is
reached. For each episode, we consider the loss function
PT
bst )2 , where q
bst is a function of
as, L(w) = t=1 (Rt − q
the training parameters w = [wr , wP(a) ] in GVIN, T is the
episode length and Rt is the expected return at time stamp t,
defined as Rt = (rt+1 + γRt+1 ), where γ is the discount factor, and rt is the immediate return at time stamp t. Additional
details of the algorithm will be discussed in Section Training
via Reinforcement Learning. In the testing phase, we obtain
the action by greedily selecting the maximal state value; that
bs0 .
is, st+1 = arg maxs0 ∈Nei(st ) v
Graph Convolution
The conventional CNN takes an image as input, which is
a 2D lattice graph. Each node is a pixel and has the same
local structure, sitting on a grid and connecting to its eight
neighbors. In this case, the convolution operator is easy to
obtain. In irregular graphs, however, nodes form diverse local
structures, making it challenging to obtain a structured and
translation invariant operator that transfers knowledge from
one graph to another. The fundamental problem here is to
find a convolution operator that works for arbitrary local
structures. We solve this through learning a 2D spatial kernel
function that provides a transition probability distribution in
the 2D space, and according to which we evaluate the weight
of each edge and obtain a graph convolution operator.
The 2D spatial kernel function assigns a value to each
position in the 2D space, which reflects the possibility to
transit to the corresponding position. Mathematically, the
transition probability from a starting position x ∈ R2 to
another position y ∈ R2 is K(x, y), where K(·, ·) is a 2D
spatial kernel function, which will be specified later.
Definition 1. A 2D spatial kernel function K(·, ·) is shift
invariant when it satisfies K(x, y) = K(x + t, y + t), for
all x, y, t ∈ R2 .
The shift invariance requires that the transition probability
depend on the relative position, which is the key for transfer learning; in other words, no matter where the starting
position is, the transition probability distribution is invariant. Based on a shift-invariant 2D spatial kernel function
and the graph adjacency matrix, we obtain the graph convolution operator P = fP (G; wP ) ∈ RN ×N , where each
element Pi,j = Ai,j ·KwP (Xi , Xj ), where the kernel function KwP (·, ·) is parameterized by wP and Xi , Xj ∈ R2 are
the embeddings of the ith and jth node. The graph convolution operator follows from (1) graph connectivity and (2) 2D
spatial kernel function. With the shift-invariant property, the
2D spatial kernel function leads to the same local transition
distribution at each node; the graph adjacency matrix works
as a modulator to select activations in the graph convolution
operator. When there is no edge between i and j, we have
Ai,j = 0 and Pi,j = 0; when there is an edge between i and
j, Pi,j is high when KwP (Xi , Xj ) is high; in other words,
when the transition probability from the ith node to the jth
node is higher, the edge weight Pi,j is high and the influence
from the ith node to the jth node is bigger during the graph
convolution. Note that P is a sparse matrix and its sparsity
pattern is the same with its corresponding adjacency matrix,
which ensures cheap computation.
As shown in (2), the graph convolution is a matrix-vector
multiplication between the graph convolution operator P and
the graph signal r + γvn ; see Figure 2. Note that when we
work with a lattice graph and an appropriate kernel function,
this graph convolution operator P is nothing but a matrix representation of the conventional convolution (LeCun, Bengio,
and others ); in other words, VIN is a special case of GVIN
when the underlying graph is a 2D lattice; see more details in
Supplementary Graph-based Kernel Functions.
We consider three types of shift-invariant 2D spatial kernel
functions: the directional kernel, the spatial kernel, and the
embedding kernel.
Figure 2: Matrix-vector multiplication as graph convolution.
Through a graph convolution operator P, r + γv diffuses
over the graph to obtain the action-value graph signal q.
Directional Kernel. The directional kernel is embedded
with the direction information. The (i, j)th element in the
graph convolution operator models the probability of following the edge from i to j ; that is,
Pi,j = Ai,j ·
L
X
(t,θ` )
w` Kd
(θij ) ,
(5)
`=1
where
(t,θ )
Kd `
(θ) =
1 + cos(θ − θ` )
2
t
,
w` is kernel coefficient, θij is the direction of the edge connecting the ith and thejth nodes, which can be computed
(t,θ )
through the node embeddings Xi , Xj ∈ R2 , and Kd ` (θ)
is the directional kernel with order t and reference direction
θ` , reflecting the center of the activation. The hyperparameters include the number of directional kernels L and the order
t, reflecting the directional resolution (a larger t indicates
more focus in one direction); see Figure 7. The kernel coefficient w` and the reference direction θ` are the training
parameters, which is wP in (2).
Spatial Kernel. We next consider both direction and distance. The (i, j)th element in the graph convolution operator
is then,
L
X
Pi,j = Ai,j ·
w` Ks(d` ,t,θ` ) (dij , θij ) ,
(6)
`=1
where Ks(d` ,t,θ` ) (d, θ) = I|d−d` |≤
1 + cos(θ − θ` )
2
and dij is the distance between the ith and the jth nodes,
which can be computed through the node embeddings
(d ,t,θ )
Xi , Xj ∈ R2 , Ks ` ` (d, θ) is the spatial kernel with reference distance d` and reference direction θ` and the indicator
function I|d−d` |≤ = 1 when |d − d` | ≤ and 0, otherwise.
The hyperparameters include the number of directional kernels L, the order t, the reference distance d` and the distance
threshold . The kernel coefficient w` and the reference direction θ` are training parameters (wP in (2)).
t
,
Embedding-based Kernel. In the directional kernel and
spatial kernel, we manually design the kernel and provide
hints for GVIN to learn useful direction-distance patterns.
Now we directly feed the node embeddings and allow GVIN
to automatically learn implicit hidden factors for general planning. The (i, j)th element in the graph convolution operator
is then,
(Ii=j + Ai,j )
·Kemb (Xi , Xj ) ,
P
k (1 + Ak,j )
k (1 + Ai,k )
(7)
where the indicator function Ii=j = 1 when i = j, and
0, otherwise, and the embedding-based kernel function is
Kemb (Xi , Xj ) = mnnet ( [Xi − Xj ]), with mnnet(·) is a
standard multi-layer neural network. The training parameters
wP in (2) are the weights in the multi-layer neural network.
In practice, when the graph is weighted, we may also include
the graph adjacency matrix Ai,j as the input of the multilayer neural network.
Pi,j = pP
Theorem 1. The proposed three kernel functions, the directional kernel, the spatial kernel and the embedding-based
kernel, are shift invariant.
The proof follows from the fact that those kernels use only
the direction, distance and the difference between two node
embeddings, which only depend on the relative position.
Training via Reinforcement Learning
We train GVIN through episodic Q-learning, a modified version of n-step Q-learning. The difference between
episodic Q-learning and the n-step Q-learning is that the nstep Q-learning has a fixed episode duration and updates the
training weights after n steps; while in episodic Q-learning,
each episodic terminates when the agent reaches the goal
or the maximum step threshold is reached, and we update
the trainable weights after the entire episode. During experiments, we found that for both regular and irregular graphs, the
policy planned by the original Q-learning keeps changing and
does not converge due to the frequent updates. Similar to the
Monte Carlo algorithms (Sutton and Barto 1998), episodic
Q-learning first selects actions by using its exploration policy until the goal is reached. Afterwards, we accumulate
the gradients during the entire episode and then update the
trainable weights, allowing the agent to use a stable plan
to complete an entire episode. This simple change greatly
improves the performance (see Section Revisting 2D Mazes).
The pseudocode for the algorithm is presented in Algorithm 1
(Supplementary Episodic Q-learning).
Experimental Results
In this section, we evaluate the proposed method on three
types of graphs: 2D mazes, synthesized irregular graphs and
real road networks. We first validate that the proposed GVIN
is comparable to the original VIN for 2D mazes, which have
regular lattice structure. We next show that the proposed
GVIN automatically learns the concepts of direction and
distance in synthesized irregular graphs through the reinforcement learning setting (without using any ground-truth
labels). Finally, we use the pre-trained GVIN model to plan
paths for the Minnesota road network and Manhattan street
network. Additional experiment parameter settings are listed
in the Supplementary Experiment Settings.
Revisting 2D Mazes
Given a starting point and a goal location, we consider
planning the shortest paths for 2D mazes; see Figure 8(a)
(Supplementary) as an example. We generate 22, 467 2D
mazes (16 × 16) using the same scripts1 that VIN used. We
use the same configuration as VIN (6/7 data for training and
1/7 data for testing). Here we consider four comparisons:
VIN vs. GVIN, action-value based imitating learning vs. statevalue based imitating learning, direction-guided GVIN vs.
unguided GVIN, and reinforcement learning.
Four metrics are used to quantify the planning performance, including prediction accuracy—the probability of
taking the ground-truth action at each state (higher means
better); success rate—the probability of successfully arriving
at the goal from the start state without hitting any obstacles
(higher means better); path difference—the average length
difference between the predicted path and the ground-truth
path (lower means better); and expected reward—the average accumulated reward (higher means better). The overall
testing results are summarized in Table 1.
VIN vs. GVIN. GVIN performs competitively with VIN
(Table 1), especially when GVIN uses direction-aware actionvalue based imitation learning (4th column in Table 1), which
outperforms the others for all four metrics. Figure 8(b) (Supplementary) shows the value map learned from GVIN with
direction-unaware state-value based imitation learning. We
see negative values (in blue) at obstacles and positive values
(in red) around the goal, which is similar to the value map
that VIN reported in (Tamar et al. 2016).
Action-value vs. State-value. VIN with action-value imitation learning slightly outperforms VIN with state-value imitation learning. Similarly, GVIN with action-value based imitation learning slightly outperforms GVIN with state-value
based imitation learning. The results suggest that our action
approximation method (Section Framework) does not impact
the performance while maintaining the ability to be extended
to irregular graphs.
Direction-aware GVIN vs. Unaware GVIN. Directionaware GVIN slightly outperforms direction-unaware GVIN,
which is reasonable because the fixed eight directions are
ground truth for regular 2D mazes. It remains encouraging
that the GVIN is able to find the ground-truth directions
through imitation learning. As shown later, direction-unaware
GVIN outperforms direction-aware GVIN in irregular graphs.
Figures 3(a) and (b) show that the planning performance
improves as the kernel exponential t in (8) increases due to
the resolution in the reference direction being low when t is
small. Figure 7 (in Supplementary) compares the kernel with
the same reference direction, but two different kernel orders.
When t = 5, the kernel activates wide-range directions; when
t = 100, the kernel focuses on a small-range directions and
has a higher resolution.
1
https://github.com/avivt/VIN
VIN
Action-value State-value
Prediction accuracy
Success rate
Path difference
Expected reward
95.00%
99.30%
0.089
0.963
95.00%
99.78%
0.010
0.962
GVIN
Action-value
State-value
dir-aware unaware dir-aware unaware
95.20%
99.91%
0.004
0.965
92.90%
98.60%
0.019
0.939
94.40%
99.57%
0.013
0.958
94.80%
99.68%
0.015
0.960
Table 1: 2D Maze performance comparison for VIN and GVIN. GVIN achieves similar performance with VIN for 2D mazes
(16 × 16); state-value imitation learning achieves similar performance with action-value imitation learning.
(a) Prediction accuracy
in 2D mazes.
(b) Success rate
in 2D mazes.
(c) Prediction accuracy
in irregular graphs
(d) Success rate
in irregular graphs.
Figure 3: Kernel direction order influences the planning performance in both regular and irregular graphs.
based kernel, direction-aware vs. direction-unaware, scale
generalization, and reinforcement learning vs. imitation learning. We use the same performance metrics as the previously
discussed 2D maze experiments.
(a) Expected rewards.
(b) Success rate.
Figure 4: Q- vs. episodic Q-learning on 16 × 16 Maze.
Reinforcement Learning. We also examine the performance of episodic Q-learning (Section Training via Reinforcement Learning) in VIN. Table 4 (Supplementary) shows
that the episodic Q-learning algorithm outperforms the training method used in VIN (TRPO + curriculum learning). For
the results reported in Table 4, we were able to train the VIN
using our algorithm (episodic Q-learning) in just 200 epochs,
while TRPO and curriculum learning took 1000 epochs to
train VIN, as reported in (Tamar et al. 2016) (both algorithms
used the same settings). As shown in Figure 4, the episodic
Q-learning algorithm shows faster convergence and better
overall performance when compared with Q-learning.
Exploring Irregular Graphs
We consider four comparisons in the following experiments: Directional kernel vs. Spatial kernel vs. Embedding-
Directional Kernel vs. Spatial Kernel vs. Embeddingbased Kernel. We first train the GVIN via imitation learning.
Table 2 shows that the embedding-based kernel outperforms
the other kernel methods in terms of both action prediction
and path difference (5th column in Table 2), indicating that
the embedding-based kernel captures the edge weight information (distance) within the neural network weights better than the other methods. The spatial kernel demonstrates
higher accuracy and success rate when compared with the
directional kernel, which suggests the effectiveness of using
bin sampling. The direction-unaware method shows slightly
better results for the spatial kernel, but has a larger success
rate gain for the directional kernel. Figure 8(d) (Supplementary) shows the visualization of the learned value map which
shares similar properties with the regular graph value map.
We also train VIN (1st column) by converting graph to 2D
image. As shown in the Table, VIN fails significantly (See
Supplementary Experiment Settings).
Figures 3(c) and (d) show the planning performance for
the irregular domain as the kernel order t in 8 increases.
The results show that a larger t in the irregular domain has
the opposite effect when compared with the regular domain.
The observation is reasonable: in the irregular domain, the
direction of each neighbor is extremely variable and a larger
kernel order creates a narrower direction range (as seen in
Prediction acc.
Success rate
Path diff.
Expected reward
VIN
MACN
(36 nodes)
26.57%
10.29%
0.992
−0.905
78%
89.4%
-
Directional Kernel
dir-aware unaware
Spatial Kernel
dir-aware unaware
41.50%
34.75%
0.175
0.266
57.45%
96.56%
0.082
0.911
41.51%
65.30%
0.141
0.599
57.90%
97.17%
0.082
0.917
Embedding-based Kernel
train 100 (IL) train 10 (IL) train 10 (RL)
58.90%
97.34%
0.079
0.922
56.14%
6.73%
0.041
−0.03
50.90%
100%
0.148
0.943
Table 2: The performance comparison amongst VIN and three different kernels of GVIN. All experiments except MACN (Khan
et al. 2017) are tested on 100-node irregular graphs. Note the last column is trained using episodic Q-learning. IL and RL stands
for imitate learning and reinforcement learning, respectively. Under similar experimental settings, MACN achieves an 89.4%
success rate for 36-node graphs, while GVIN achieves a 97.34% success rate for 100-node graphs. The details about training
VIN on irregular graphs sees Section Irregular Graphs in the supplementary material.
Prediction Accuracy
Success rate
Path difference
Expected reward
Optimal
Minnesota
|V| = 100
|V| = 10
Optimal
100%
100%
0.0000
0.96043
78.37%
100%
0.1069
0.95063
78.15%
100%
0.1025
0.95069
100%
100%
0.0000
0.97279
New York City
|V| = 100 |V| = 10
78.66%
100%
0.03540
0.97110
79.11%
100%
0.0353
0.97136
Table 3: Performance comparison on Minnesota and New York City street map data using GVIN. |V| = 100 is trained on
100-node graphs and |V| = 10 is trained on 10-node graphs.
(a) Expected rewards
(b) Success rate
Figure 5: Q- vs. Episodic Q-learning on irregular graphs.
Figure7), thus resulting in information loss.
Reinforcement Learning. We then train the GVIN using
episodic Q-learning to compare with imitation learning. As
a baseline, we also train GVIN by using standard deep Qlearning techniques, including using an experience replay
buffer and a target network. Both networks use the same kernel function (embedding-based kernel) and configurations.
Figure 5 shows the comparison of the two algorithms’ success rate and expected rewards during the training. Clearly,
episodic Q-learning converges to both a high success rate and
a high expected rewards, but the standard deep Q-learning
techniques fail to achieve reasonable results.
Scale Generalization. We also examine the scale generalization by training on 10-node graphs and then testing on
100-node graphs using the embedding-based kernel. When
GVIN is trained on 10-node graphs via imitation learning, the
performance is significantly hindered as shown in Table 2 (6th
column). When GVIN is trained using episodic Q-learning,
Table 2 (7th column) shows excellent generalization abilities that outperform all imitation learning based results for
success rate and expected rewards. Compared with imitation
learning, we also observe the performance decreases for path
differences and action prediction.
Graph with Edge Weights. We also test how GVIN handles edge weights. We set the true weighted shortest path to be
Xi − Xj
Wij , where Xi − Xj is the distance between two nodes
and Wij is the edge weight. As shown in Table 5, imitation
learning is trained on 100-node graphs, while reinforcement
learning is trained on 10-node. We also examine the GVIN
by excluding edge weights from the input to see if there are
any effects on performance. Table 5 (Supplementary) shows
that for reinforcement learning, edge weights slightly help
the agent find a more suitable policy; for imitation learning,
input edge weights cause a significant failure.
Validating Real Road Networks
To demonstrate the generalization capabilities of GVIN,
we evaluate two real-world maps: the Minnesota highway
map, which contains 2642 nodes representing intersections
and 6606 edges representing roads, and the New York City
street map, which contains 5069 nodes representing intersections and 13368 edges representing roads. We use the
same models trained on the graphs containing |V| = 100 and
|V| = 10 nodes with the embedding-based kernel and using
episodic Q-learning in Section Exploring Irregular Graphs,
separately. We normalize the data coordinates between 0 and
1, and we set recurrence parameter to K = 200. We randomly pick start points and goal points 1000 different times.
We use the A* algorithm as a baseline. Table 3 shows that
both |V| = 100 and |V| = 10 generalize well on large scale
data. The policy could reach the goal position with 100%
in the experiments. One sample planned path is shown in
Supplementary (Figures 9 and 10).
Conclusions
We have introduced GVIN, a differentiable, novel planning
module capable of both regular and irregular graph navigation and impressive scale generalization. We also introduced
episodic Q-learning that is designed to stabilize the training
process of VIN and GVIN. The proposed graph convolution may be applied to many other graph-based applications,
such as navigation, 3D point cloud processing and molecular
analysis, which is left for future works.
References
Bellman, R. 1957. Dynamic programming. Princeton, USA:
Princeton University Press 1(2):3.
Bertsekas, D. P.; Bertsekas, D. P.; Bertsekas, D. P.; and Bertsekas, D. P. 1995. Dynamic programming and optimal
control, volume 1. Athena Scientific Belmont, MA.
Bojarski, M.; Del Testa, D.; Dworakowski, D.; Firner, B.;
Flepp, B.; Goyal, P.; Jackel, L. D.; Monfort, M.; Muller, U.;
Zhang, J.; et al. 2016. End to end learning for self-driving
cars. arXiv preprint arXiv:1604.07316.
Bresenham, J. 1977. A linear algorithm for incremental
digital display of circular arcs. Communications of the ACM
20(2):100–106.
Bronstein, M. M.; Bruna, J.; LeCun, Y.; Szlam, A.; and Vandergheynst, P. 2016. Geometric deep learning: going beyond
euclidean data. arXiv preprint arXiv:1611.08097.
Bruna, J.; Zaremba, W.; Szlam, A.; and LeCun, Y. 2013.
Spectral networks and locally connected networks on graphs.
arXiv preprint arXiv:1312.6203.
Defferrard, M.; Bresson, X.; and Vandergheynst, P. 2016.
Convolutional neural networks on graphs with fast localized
spectral filtering. In Advances in Neural Information Processing Systems, 3837–3845.
Deisenroth, M., and Rasmussen, C. E. 2011. Pilco: A modelbased and data-efficient approach to policy search. In Proceedings of the 28th International Conference on machine
learning (ICML-11), 465–472.
Duvenaud, D. K.; Maclaurin, D.; Iparraguirre, J.; Bombarell,
R.; Hirzel, T.; Aspuru-Guzik, A.; and Adams, R. P. 2015.
Convolutional networks on graphs for learning molecular
fingerprints. In Advances in neural information processing
systems, 2224–2232.
Gilmer, J.; Schoenholz, S. S.; Riley, P. F.; Vinyals, O.; and
Dahl, G. E. 2017. Neural message passing for quantum
chemistry. arXiv preprint arXiv:1704.01212.
Guo, X.; Singh, S.; Lee, H.; Lewis, R. L.; and Wang, X.
2014. Deep learning for real-time atari game play using
offline monte-carlo tree search planning. In Advances in
neural information processing systems, 3338–3346.
Henaff, M.; Bruna, J.; and LeCun, Y. 2015. Deep convolutional networks on graph-structured data. arXiv preprint
arXiv:1506.05163.
Khan, A.; Zhang, C.; Atanasov, N.; Karydis, K.; Kumar, V.;
and Lee, D. D. 2017. Memory augmented control networks.
arXiv preprint arXiv:1709.05706.
Kipf, T. N., and Welling, M. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint
arXiv:1609.02907.
LeCun, Y.; Bengio, Y.; et al. Convolutional networks for
images, speech, and time series.
Levine, S.; Finn, C.; Darrell, T.; and Abbeel, P. 2015. Endto-end training of deep visuomotor policies. arXiv preprint
arXiv:1504.00702.
Li, Y.; Tarlow, D.; Brockschmidt, M.; and Zemel, R. 2015.
Gated graph sequence neural networks. arXiv preprint
arXiv:1511.05493.
Lillicrap, T. P.; Hunt, J. J.; Pritzel, A.; Heess, N.; Erez, T.;
Tassa, Y.; Silver, D.; and Wierstra, D. 2015. Continuous
control with deep reinforcement learning. arXiv preprint
arXiv:1509.02971.
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.;
Antonoglou, I.; Wierstra, D.; and Riedmiller, M. 2013. Playing atari with deep reinforcement learning. arXiv preprint
arXiv:1312.5602.
Mnih, V.; Badia, A. P.; Mirza, M.; Graves, A.; Lillicrap, T.;
Harley, T.; Silver, D.; and Kavukcuoglu, K. 2016. Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning, 1928–1937.
Niepert, M.; Ahmed, M.; and Kutzkov, K. 2016. Learning
convolutional neural networks for graphs. In Proceedings of
the 33rd annual international conference on machine learning. ACM.
Schmidhuber, J. 1990. An on-line algorithm for dynamic
reinforcement learning and planning in reactive environments.
In Neural Networks, 1990., 1990 IJCNN International Joint
Conference on, 253–258. IEEE.
Silver, D.; Huang, A.; Maddison, C. J.; Guez, A.; Sifre, L.;
Van Den Driessche, G.; Schrittwieser, J.; Antonoglou, I.;
Panneershelvam, V.; Lanctot, M.; et al. 2016a. Mastering
the game of go with deep neural networks and tree search.
Nature 529(7587):484–489.
Silver, D.; van Hasselt, H.; Hessel, M.; Schaul, T.; Guez, A.;
Harley, T.; Dulac-Arnold, G.; Reichert, D.; Rabinowitz, N.;
Barreto, A.; et al. 2016b. The predictron: End-to-end learning
and planning. arXiv preprint arXiv:1612.08810.
Sutton, R. S., and Barto, A. G. 1998. Reinforcement learning:
An introduction, volume 1. MIT press Cambridge.
Tamar, A.; Levine, S.; Abbeel, P.; WU, Y.; and Thomas, G.
2016. Value iteration networks. In Advances in Neural
Information Processing Systems, 2146–2154.
Tieleman, T., and Hinton, G. 2012. Lecture 6.5-rmsprop:
Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning
4(2).
Appendix
is,
Computational Complexity
Let the input graph G have |V| nodes and |E| edges.
In the testing phase, since the input graph are commonly
sparse, the computational complexities of (1), (3) and (4)
are O(|V|), O(|E|) and O(|V|) based on sparse computation,
respectively. Therefore, the total computational complexity is
O(|V| + K(|E| + |V|)), where K is number of iterations. For
a spatial graph, the number of edges is usually proportional
to the number of nodes; thus the computational complexity
is O(K|V|), which is scalable to huge graphs.
Episodic Q-learning
We highlight the differences between episodic Q-learning
and the original n-step Q-learning in blue, including the
initial expected return, the termination condition and the
timing of updating the gradient.
Algorithm 1 Episodic Q-learning
1: input graph G and the goal sg
2: initialize global step counter T = 0
3: initialize GVIN parameters w = [wr , wP(a) ]
4: initialize parameter gradients ∆w
5: repeat (one episode)
6:
clear gradients ∆w ← 0
7:
t=0
8:
randomly pick a start node st
9:
repeat (one action)
take action at according to the -greedy policy
10:
(a)
based on qst
11:
receive reward rt and new state st+1
12:
t←t+1
13:
until terminal st = sg or t > tmax
14:
R=0
15:
for i = t : −1 : 0 do
16:
R ← ri + γR
17:
accumulate gradients wrt w : ∆w ← ∆w +
2
∂(R−q(a)
st )
∂w
18:
end for
19:
w ← w − ∆w
20:
T =T +1
21: until T > Tmax
Graph-based Kernel Functions
Directional Kernel. We first consider direction. When we
face several roads at an intersection, it is straightforward
to pick the one whose direction points to the goal. We aim
to use the directional kernel to capture edge direction and
parameterize the graph convolution operation.
The (i, j)th element in the graph convolution operator
models the probability of following the edge from i to j ; that
Pi,j = Ai,j ·
L
X
(t,θ` )
w` Kd
(θij ) ,
`=1
(t,θ` )
where Kd
(θ) =
1 + cos(θ − θ` )
2
(8)
t
.
where w` is kernel coefficient, θij is the direction of the edge
connecting the ith and thejth nodes, which can be computed
through the node embeddings Xi ∈ R2 and Xj ∈ R2 , and
(t,θ )
Kd ` (θ) is the directional kernel with order t and reference direction θ` , reflecting the center of the activation. The
hyperparameters include the number of directional kernels L
and the order t, reflecting the directional resolution (a larger
t indicates more focus in one direction); see Figure 7. The
kernel coefficient w` and the reference direction θ` are the
training parameters, which is wP in (2). Note that the graph
convolution operator P ∈ RN ×N is a sparse matrix and its
sparsity pattern is the same with the input adjacency matrix,
which ensures that the computation is cheap.
(a) t = 5
(b) t = 100
Figure 6: The directional kernel function activates the areas
around the reference direction θ` in the 2D spatial domain.
The activated area is more concentrated when t increases.
The intuition behind (8) is that each graph convolution
operator represents a unique direction pattern. An edge is
a 2D vector sampled from the 2D spatial plane. When the
direction of the edge connecting i and j matches one or some
of the L reference directions, we have a higher probability to
follow the edge from i to j. In GVIN, we consider several
channels. In each channel, we obtain an action value for
each node, which represents the matching coefficient of a
direction pattern. The max-pooling operation then selects the
most matching direction pattern for each node.
Spatial Kernel. We next consider both direction and distance. When all the roads at the current intersection opposite
to the goal, it is straightforward to try the shortest one first.
We thus include the edge length into the consideration. The
(i, j)th element in the graph convolution operator is then,
Pi,j = Ai,j ·
L
X
w` Ks(d` ,t,θ` ) (dij , θij ) ,
`=1
where
Ks(d` ,t,θ` )
(d, θ) = I|d−d` |≤
1 + cos(θ − θ` )
2
t
,
(9)
where dij is the distance between the ith and the jth nodes,
which can be computed through the node embeddings Xi ∈
(d ,t,θ )
R2 and Xj ∈ R2 , Ks ` ` (d, θ) is the spatial kernel with
reference distance d` and reference direction θ` and the indicator function I|d−d` |≤ = 1 when |d − d` | ≤ and 0,
otherwise. The hyperparameters include the number of directional kernels L, the order t, the reference distance d` and
the distance threshold . The kernel coefficient w` and the
reference direction θ` are training parameters, which is wP
in (2).
(a) t = 5.
(b) t = 100.
Figure 7: The spatial kernel function activates the areas
around the reference direction θ` and reference distance d`
in the 2D spatial domain.
Compared to the directional kernel, the spatial kernel adds
another dimension, distance; in other words, the directional
kernel is a special case of the spatial kernel when we ignore
the distance. Each spatial kernel activates a localized area
in the direction-distance plane. With the spatial kernel, the
graph convolution operator (9) represents a unique directiondistance pattern; that is, if the direction/distance of the edge
connecting i and j matches one or some of the L reference
directions and distances, we have a higher probability to
follow the edge from i to j.
Embedding-based Kernel. In the directional kernel and
spatial kernel, we manually design the kernel and hint GVIN
to learn useful direction-distance patterns. Now we directly
feed the node embeddings and allow GVIN to automatically
learn implicit hidden factors for general planning. The (i, j)th
element in the graph convolution operator is then,
(Ii=j + Ai,j )
·Kemb (Xi , Xj ) ,
P
k (1 + Ak,j )
k (1 + Ai,k )
(10)
where the indicator function Ii=j = 1 when i = j, and
0, otherwise, and the embedding-based kernel function is
Kemb (Xi , Xj ) = mnnet ( [Aij , Xi − Xj ]), with mnnet(·)
is a standard multi-layer neural network. The training parameters wP in (2) are the weights in the multi-layer neural network. Note that the graph convolution operator P ∈ RN ×N
is still a sparse matrix and its sparsity pattern is the same with
the input adjacency matrix plus the identity matrix.
In the directional kernel and spatial kernel, we implicitly
discretize the space based on the reference direction and
distance; that is, for each input pair of given direction and
distance, the kernel function outputs the response based on
its closed reference direction and distance. In the embeddingPi,j = pP
Success rate
No. of Epochs
Path difference
Expected reward
TRPO
EQL2
82.50%
1000
N/A
N/A
98.67%
200
0.1617
0.9451
Table 4: Performance comparison using different training
algorithms on the VIN model. The first column is VIN trained
by TRPO with curriculum learning reported in (Tamar et al.
2016), the second column is VIN trained by episodic Qlearning.
based kernel, we do not set the reference direction and distance to discretize the space; instead, we use the multi-layer
neural network to directly regress from an arbitrary edge
(with edge weight and embedding representation) to a response value. The embedding-based kernel is thus more flexible than the directional kernel and spatial kernel and may
learn hidden factors.
Experiment Settings
Our implementation is based on Tensorflow with GPUenabled platform. All experiments use the standard centered RMSProp algorithm as the optimizer with learning rate
η = 0.001 (Tieleman and Hinton 2012). All reinforcement
learning experiments use a discount of γ = 0.99, RMSProp
decay factor of α = 0.999, and exploration rate annealed
linearly from 0.2 to 0.001 over the first 200 epochs.
2D Mazes The experiments are set up as follows. We consider the rules as follows: the agent receives a +1 reward
when reaching the goal, receives a −1 reward when hitting
an obstacle, and each movement gets a −0.01 reward. To
preprocess the input data, we use the same two-layer CNN
for both VIN and GVIN, where the first layer involves 150
kernels with size 3 × 3 and the second layer involves a kernel
with size 3 × 3 for output. The transition probability matrix
is parameterized by 10 convolution kernels with size 3 × 3 in
both VIN and GVIN. In GVIN, we use the directional kernel
based method as shown in Equations 8 and 9 and we set
` = 8 to represent the eight reference directions. We consider
two approaches to initialize the directions θ` . In directionaware approach, we fix θ` as [0, π/4, π/2, ..., 7π/4]. In the
direction-unaware approach, we set θ` to be weights and train
them via backpropagation. We set the recurrence K in GVIN
to be 20 for 16 × 16 2D mazes. In the regular domain, we set
the kernel order t = 100 to be the default.
Irregular Graphs We evaluate our proposed methods in
Section Graph Convolution for the irregular domain. Our
experimental domain is a synthetic data which consists of
N = 10000 irregular graphs, in which each graph contains
100 nodes. We follow the standard rules of random geometric graphs to generate our irregular graphs. Specifically, we
generate |V| vertices with random coordinates in the box
[0, 1]2 and connects a pair of two vertices with an edge when
the distance between two vertices is smaller than a certain
threshold. For each node in the graph, we define the coordi-
(a) Input map
(2D mazes).
(b) Value map
(2D mazes).
(c) Input map
(irregular graph).
(d) Value map
(irregular graph).
Figure 8: value map visualization on regular and irregular graph.
Imitation Learning
w edge weight w/o edge weight
Prediction accuracy
Success rate
Path difference
Expected reward
62.24%
3.00%
0.020
−0.612
43.12%
97.60%
0.763
0.771
Reinforcement Learning
w edge weight w/o edge weight
50.38%
100%
0.146
0.943
50.14%
100%
0.151
0.940
Table 5: Performance comparison for testing weighted graphs. Imitation learning is trained on 100-node irregular graphs while
reinforcement learning is trained on 10-node irregular graphs.
nates which represent its spatial position ranged between 0
and 1. The dataset is split into 7672 graphs for training and
1428 graphs for testing. Additionally, to exam whether GVIN
could handle weighted graphs, we also generated a synthetic
dataset consisting of 100 node graphs and partitioned 42857
graphs for training and 7143 graphs for testing.
For the directional kernel and spatial kernel, we set the
number of reference directions to ` = 8 and kernel order t = 20 to be default values for Equations 8 and 9.
We also set θ` to be 0 to 2π with an interval of π/4 for
direction-aware mode and we set θ` to be trainable weights
for direction-unaware mode. For the spatial kernel function,
we set the number of bins d` to be 10 in Equation 9. In the
embedding-based kernel, we use three layers of fully connected neural networks (32-64-1), where each layer uses
ReLU (·) = max(0, ·) as its activation function. The neural
network are initialized with zero-mean and 0.01 derivation.
or all three kernel methods, we set the graph convolution
channel number to be 10 and K = 40 for recurrence.
Training VIN on Irregular Graphs. To show the strong
generalization of GVIN, we evaluate the VIN on irregular
graph by converting graph data format to 2D image. Each
testing set contains reward map and obstacle map that sizes
100 × 100 pixels. We use pre-trained weights from 28 × 28
maze with all parameters tuned to be highest performance.
To make the training and testing consistent, we set the rewards map and obstacle map the same settings as 2D maze:
vertices and edges are marked as free path (value set to be
0), while the other area are marked as obstacles (value set
to be 1). The edge path is generated via Bresenham’s line
algorithm (Bresenham 1977). We set recurrence K to be 200
so that the value iteration could cover the whole map.
(a) Ground-Truth
(b) GVIN prediction
Figure 9: Sample planning trajectories on Minnesota highway map.
(a) Ground-Truth
(b) GVIN prediction
Figure 10: Sample planning trajectories on New York City street map.
| 2cs.AI
|
1
Theory and Application on Adaptive-Robust
Control of Euler-Lagrange Systems with Linearly
Parametrizable Uncertainty Bound
arXiv:1708.01442v1 [cs.SY] 4 Aug 2017
Spandan Roy, Sayan Basu Roy and Indra Narayan Kar, Senior Member, IEEE
Abstract—This work proposes a new adaptive-robust control
(ARC) architecture for a class of uncertain Euler-Lagrange (EL)
systems where the upper bound of the uncertainty satisfies linear
in parameters (LIP) structure. Conventional ARC strategies
either require structural knowledge of the system or presume
that the overall uncertainties or its time derivative is norm
bounded by a constant. Due to unmodelled dynamics and
modelling imperfection, true structural knowledge of the system
is not always available. Further, for the class of systems under
consideration, prior assumption regarding the uncertainties (or
its time derivative) being upper bounded by a constant, puts
a restriction on states beforehand. Conventional ARC laws
invite overestimation-underestimation problem of switching gain.
Towards this front, Adaptive Switching-gain based Robust Control (ASRC) is proposed which alleviates the overestimationunderestimation problem of switching gain. Moreover, ASRC
avoids any presumption of constant upper bound on the overall
uncertainties and can negotiate uncertainties regardless of being
linear or nonlinear in parameters. Experimental results of
ASRC using a wheeled mobile robot notes improved control
performance in comparison to adaptive sliding mode control.
Index Terms—Adaptive-robust control, Euler-Lagrange systems, Wheeled mobile robot, Uncertainty.
I. I NTRODUCTION
A. Background
HE controller design aspect for nonlinear systems subjected to parametric and nonparametric uncertainties has
always been a challenging task. Adaptive control and Robust
control are the two popular control strategies to deal with
uncertain nonlinear systems. In case of adaptive control, online
computation of the unknown system parameters and controller
gains for complex systems is significantly intensive [1]. On
the other front, robust control reduces computation burden
for complex systems compared to adaptive control, while
requiring a predefined upper bound on the uncertainties. However, in practice it is not always possible to estimate a prior
uncertainty bound due to the effect of unmodelled dynamics.
Again, to increase the operating region of the controller, often
higher uncertainty bounds are assumed. This in turn leads to
overestimation of switching gain and high control effort [2].
Considering the individual constraints of adaptive and robust control, recently global research is reoriented towards
adaptive-robust control (ARC). The series of publications [1],
T
S. Roy, S. Basu Roy and I. N. Kar are with the Department of Electrical
Engineering, Indian Institute of Technology-Delhi, New Delhi, India e-mail:
([email protected], [email protected], [email protected]).
[3]-[11] regarding ARC, estimate the individual uncertain
system parameters through adaptive law and robust control
is utilized to negate the effect of external disturbances. These
works utilize the projection operator in their respective adaptive laws which necessitate the knowledge of lower and upper
bound of individual uncertain system parameters. Adaptive
sliding mode control (ASMC) is designed in [12] for parameter
identification of mechanical servo systems with LuGre friction
considering the uncertainties to be linear in parameters (LIP).
In contrast, the controllers [13]-[20] assume that the overall
uncertainty (or its time derivative) is bounded by some constant. Thereafter, that constant term is estimated by adaptive
law, rather estimating individual uncertain system parameters.
The adaptive laws in [17]-[18] involve a predefined threshold
value; as a matter of fact, until the threshold value is achieved,
the switching gain may still be increasing (resp. decreasing)
even if the tracking error decreases (resp. increases) and
thus creates overestimation (resp. underestimation) problem
of switching gain.While the underestimation problem compromises the controller accuracy by applying lower switching
gain than the required amount, the overestimation problem
causes larger gain and high control input requirement. The
adaptive law reported in [19] requires predefined bound on
the time derivative of the uncertainties. As observed in [20],
the method in [19] also requires frequency characteristics of
the perturbation to design the filter for equivalent control.
However, the work in [20] assumed that the time derivative of
the uncertainties are bounded by an unknown constant.
B. Motivation
Let us consider the following system representing a chemostat operating under Monod kinetic [21]:
ż1 = f1 (z1 , z2 ) − Dz1 , ż2 = f2 (z1 , z2 ) + S0 − Dz2 ,
δ1 z1 z2
−δ3 z1 z2
where f1 (z1 , z2 ) =
, f2 (z1 , z2 ) =
.
δ2 + z2
δ2 + z2
(1)
Here z1 ≥ 0, z2 ≥ 0 ∀t ≥ 0 are states; δ1 , δ2 , δ3 are uncertain
positive parameters; S0 is a known constant and D is the
control input. For the system (1), the following relations hold:
|f1 | ≤ y1 f (z), |f2 | ≤ y2 f (z),
(2)
where y1 = |δ1 |/|δ2 |, y2 = |δ3 |/|δ2 |, f (z) = |z1 ||z2 |.
Inspection of the uncertainties f1 and f2 from (1)-(2) reveal
that though f1 , f2 are nonlinear in parameters (NLIP) but their
upper bounds are LIP.
2
Similarly Euler-Lagrange (EL) systems can have uncertainties with LIP or NLIP (e.g. system with nonlinear friction
([22])) structure. However, the upper bound of the overall (or
lumped) uncertainty for such systems has LIP property [23].
EL systems, in general, represent a large class of real world
systems like robotic manipulators [24]-[25], mobile robots
[26], ship dynamics, aircraft, pneumatic muscles [27] etc.
These systems have immense applications in various domains
such as defence, automation industry, surveillance, space missions etc. The controllers [13]-[18] assume that the overall
uncertainties are upper bounded by some constant, while [19][20] assume the time derivative of the overall uncertainty to
be bounded by some constant. Hence, for the aforementioned
class of systems, consideration of such constant bound (known
or unknown) restricts the system state a priori. Further, the
switching gain in [17]-[18] suffers from over- and underestimation problems. In practice, it is also not always possible
to have prior knowledge of bounds for system parameters as
required in [1], [3]-[11] for projection operator.
where q ∈ Rn denotes system state, τ ∈ Rn denotes vector
of generalized control input forces, M (q) ∈ Rn×n represents
mass/inertia matrix, C(q, q̇) ∈ Rn×n denotes Coriolis, centripetal terms, g(q) ∈ Rn denotes gravity vector, f (q̇) ∈ Rn
represents the vector of slip, damping and friction forces and
ds (t) denotes the bounded external disturbances. The system
(3) possesses the following properties [23]:
Property 1: The matrix (Ṁ − 2C) is skew symmetric.
Property 2: ∃gb , fb , d¯ ∈ R+ such that ||g(q)|| ≤ gb , ||f (q̇)|| ≤
¯
fb ||q̇|| and ||ds (t)|| ≤ d.
Property 3: The matrix M (q) is uniformly positive definite
and there exist two positive constants µ1 , µ2 such that
0 < µ1 I ≤ M (q) ≤ µ2 I.
Property 4: ∃Cb ∈ R such that ||C(q, q̇)|| ≤ Cb ||q̇||.
Let q d (t) is the desired trajectory to be tracked and it is
selected such that q d , q̇ d , q̈ d ∈ L∞ . Let e(t) , q(t) − q d (t) be
the tracking error and ef be the filtered tracking error:
ef , ė + Ωe ⇒ ef = Γξ,
T
C. Contribution
In view of the above discussion and the importance of EL
systems in real-life scenarios, it is imperative to formulate a
dedicated ARC framework for uncertain EL systems. Towards
this front, Adaptive Switching-gain based Robust Control
(ASRC) is presented in this paper for tracking control of
uncertain EL systems. The formulation of ASRC is insensitive
towards the nature of the uncertainties, i.e., it can negotiate
uncertainties that can be either LIP or NLIP. ASRC utilizes
LIP structure of the upper bound of uncertainty and does
not presume the overall uncertainty (or its time derivative) to
be upper bounded by a constant. The adaptive law of ASRC
prevents the switching gain from becoming a monotonically
increasing function by allowing the switching gain to decrease
within a finite time when tracking error decreases. Moreover,
ASRC alleviates the overestimation-underestimation problem
of switching gain. To realize the effectiveness, the performance
of ASRC is compared with ASMC [17]-[18] experimentally
using PIONEER 3 wheeled mobile robot (WMR).
D. Organization and Notations
The remainder of the article is organized as follows: The
proposed ASRC framework for second order EL systems is
detailed in Section II. This is followed by the experimental
results of ASRC and its comparison with ASMC [17]-[18] in
Section III. Section IV presents concluding remarks.
The following notations are used in this paper: λmin (•) and
|| • || represent minimum eigenvalue and Euclidean norm of
(•) respectively; I denotes identity matrix with appropriate
dimension; R+ denotes the set of positive real numbers.
(4)
+
T T
(5)
n×n
where Γ , [Ω I], ξ , [e ė ] and Ω ∈ R
is a positive
definite matrix. Multiplying the time derivative of (5) by M
and using (3) yields
M ėf = M (q̈ − q̈ d + Ωė) = τ − C(q, q̇)ef + σ,
where σ , −(C(q, q̇)q̇ + g(q) + f (q̇) + ds + M q̈ −
M Ωė − C(q, q̇)ef ) represents the overall uncertainty. Further
ξ = [eT ėT ]T implies ||ξ|| ≥ ||e||, ||ξ|| ≥ ||ė||.
Characterization of the upper bound of σ: Relation (5)
and system Property 4 yields
||Cef − C q̇|| = ||C(ė + Ωe) − C q̇|| = || − C q̇ d + CΩe||
≤ Cb ||q̇||||q̇ d || + Cb ||q̇||||Ω||||e||
≤ Cb ||ė + q̇ d ||||q̇ d || + Cb ||ė + q̇ d ||||Ω||||ξ||
≤ Cb {||ξ||||q̇ d || + ||q̇ d ||2 + ||ξ||2 ||Ω|| + ||q̇ d ||||Ω||||ξ||}. (7)
Further, system Properties 2 and 3 provide the following:
||g(q) + f (q̇) + ds + M q̈ d − M Ωė||
≤ gb + fb ||q̇|| + d¯ + µ2 ||q̈ d || + µ2 ||Ω||||ė||
≤ gb + fb ||ė + q̇ d || + d¯ + µ2 ||q̈ d || + µ2 ||Ω||||ξ||
≤ gb + fb ||ξ|| + fb ||q̇ d || + d¯ + µ2 ||q̈ d || + µ2 ||Ω||||ξ||. (8)
Since q d , q̇ d , q̈ d ∈ L∞ , it can be verified using (7)-(8) that
∃θi∗ ∈ R+ i = 0, 1, 2 such that the upper bound of σ holds
the following LIP structure [23]:
||σ|| ≤ θ0∗ + θ1∗ ||ξ|| + θ2∗ ||ξ||2 , Y (ξ)T Θ∗ ,
A. Problem Formulation
In general, an EL system with second order dynamics can
be written as
M (q)q̈ + C(q, q̇)q̇ + g(q) + f (q̇) + ds = τ,
(3)
(9)
where Y (ξ) = [1 ||ξ|| ||ξ||2 ]T and Θ∗ = [θ0∗ θ1∗ θ2∗ ]T .
Let Θ̄ , {Θ ∈ R3 : θi ≥ θi∗ ∀i = 0, 1, 2} such that the
following condition always holds from (9):
||σ|| ≤ Y (ξ)T Θ, ∀Θ ∈ Θ̄.
II. C ONTROLLER D ESIGN
(6)
d
(10)
A robust controller for the system (3) can be designed as [23]
( e
ρ ||eff || if ||ef || ≥ $
(11)
τ = −e − Gef − ∆τ, ∆τ =
e
ρ $f
if ||ef || < $,
ρ = Y (ξ)T Θ,
(12)
3
where ∆τ provides robustness against σ through switching
gain ρ; $ ∈ R+ is a small scalar used for chattering removal;
G ∈ Rn×n is a positive definite matrix.
Evaluation of Switching Gain: Evaluation of ρ like (12)
is conservative in nature and evidently requires the knowledge
of Θ∗ , which is not always possible in the face of uncertain
parametric variations and external disturbances. The control
laws developed in [13]-[18] and [19]-[20] assume that σ
and σ̇ is upper bounded by constant, respectively. Exploring
the structure of ||σ|| from (9) it can be easily inferred that
such constant bound assumption on the uncertainties, whether
known or unknwon, puts a restriction on the states a priori. Moreover, the switching gain in [17]-[18] suffers from
overestimation-underestimation problem.
B. Adaptive Switching-gain based Robust Control (ASRC)
The major aims of the proposed ASRC framework are:
• To compensate the uncertainties that can be either LIP
or NLIP. However, the upper bound of the uncertainties
satisfies the LIP property (9).
• To alleviate the overestimation-underestimation problem
of switching gain.
The control input of the proposed ASRC is designed as
( e
ρ̂ ||eff || if ||ef || ≥ $
(13)
τ = −e − Gef − ∆τ, ∆τ =
e
if ||ef || < $,
ρ̂ $f
ρ̂ = θ̂0 + θ̂1 ||ξ|| + θ̂2 ||ξ||2 + γ , Y (ξ)T Θ̂ + γ,
(14)
where ∆τ provides robustness against σ through ρ̂; Θ̂ =
[θ̂0 θ̂1 θ̂2 ]T is the estimate of Θ; γ is an auxiliary gain.
The importance of γ will be explained later. The gains γ, θ̂i ,
i = 0, 1, 2 are evaluated using the following adaptive laws:
(i) for ||ef || ≥ $
S2
i
T
αi ||ξ|| ||ef || if {e ė > 0} ∪ { i=0 θ̂i ≤ 0}
˙
θ̂i =
∪{γ ≤ β}
i
−αi ||ξ|| ||ef || otherwise,
S2
T
α3 ||ef || if {e ė > 0} ∪ { i=0 θ̂i ≤ 0}
γ̇ =
∪{γ ≤ β}
−ςα3 ||ξ||4 otherwise,
(15)
with θ̂i (t0 ) > 0, i = 0, 1, 2, γ(t0 ) > β.
Remark 2. For ||ef || < $, the tracking error remains
bounded inside the ball B$ , {||Γξ|| < $} using the relation
ef = Γξ. This implies that the switching gains are sufficient
enough to keep the error within B$ . Hence, the gains are
kept unchanged for ||ef || < $. One can choose small $ to
improve tracking accuracy (as B$ gets reduced) as long as
the value of $ does not invite chattering.
Remark 3. The initial condition of the gains are selected
as θ̂i (t0 ) > 0, γ(t0 ) > β. Further, for ||ef || ≥ $, the
adaptive laws (15)-(16) force the gains to increase if either
of the gains attempt
S2 to breach their respective lower bounds
(governed by { i=0 θ̂i ≤ 0} ∪ {γ ≤ β}). This ensures that
γ(t) ≮ β, θ̂i (t) ≮ 0 ∀i = 0, 1, 2 when ||ef || ≥ $. Again,
gains remain unchanged for ||ef || < $. Hence, combination
of the conditions mentioned above implies
θ̂i (t) ≥ 0 ∀i = 0, 1, 2 and γ(t) ≥ β ∀t ≥ t0 .
(20)
The condition (20) is later exploited in stability analysis.
To guarantee the alleviation of the overestimation problem
of switching gain, it is necessary that θ̂i , γ decrease within a
finite time. This is shown through Theorem 1.
Theorem 1. Let t = tin be any time instant when gains start
increasing. Then there exist finite times t1 , t2 , t3 , δt such that
the gains θ̂0 , θ̂1 , θ̂2 , γ decrease for t ≥ tin + T where T ≤
t̄ + δt, t̄ = max{t1 , t2 , t3 }. These times are obtained as
θ∗ ||Γ||
θ∗ ||Γ||2
θ0∗
, t2 ≤ 1 2 , t3 ≤ 2 3 ,
(α0 + α3 )$
α1 $
α2 $
2
δt ≤ (1/%)ln{2V (t̄)/(||e(t̄)|| )},
t1 ≤
where V = 21 eTf M ef + 21 eT e, % ,
(21)
(22)
min{λmin (G),λmin (Ω)}
.
max{µ2 ,1}
(17)
Proof. Here, tin can be any time when gains start increasing
and it is solely used for analysis. The objective of Theorem
1 is to find when the gains start to decrease. Further, it is to
be noted from the laws (15)-(17) that the gains increase only
when ||ef || ≥ $. So, it is sufficient to investigate the condition
when all the gains increase and ||ef || ≥ $. Moreover, using
ef = Γξ from (5) one has
(18)
$ ≤ ||ef || ≤ ||Γ||||ξ|| ⇒ ||ξ|| ≥ ($/||Γ||).
(16)
(ii) for ||ef || < $
˙
θ̂i = 0, γ̇ = 0,
the proposed law certainly does not make the switching gain
a monotonically increasing function and thus alleviates the
overestimation problem.
(23)
+
Here t0 is the initial time and β, ς, α0 , α1 , α2 , α3 ∈ R are
user defined scalars. Substituting (13) into (6), the closed loop
system is formed as:
M ėf = −e − Gef − ∆τ − Cef + σ.
(19)
Remark 1. For ||ef || ≥ $, it can be noticed from the
adaptive laws (15)-(16) that the gains θ̂i , γ increase if error
trajectories move away from ||e|| = 0 (governed by eT ė > 0)
and decrease if error trajectories do not move away from
||e|| = 0 (governed by the ’otherwise’
condition in (15)-(16)
S2
which implies {eT ė ≤ 0} ∩ { i=0 θ̂i > 0} ∩ {γ > β}). Hence,
So, the first laws of (15), (16) and the condition (23) yields
˙
˙
˙
θ̂0 ≥ α0 $, θ̂1 ≥ (α1 $2 )/||Γ||, θ̂2 ≥ (α2 $3 )/||Γ||2 , γ̇ ≥ α3 $.
(24)
Let V be a Lypaunov function. Using (19) and the relation
eT ė = eT (ef − Ωe) (from (5)), the time derivative of V yields
V̇ = eTf M ėf + (1/2)eTf Ṁ ef + eT ė
= eTf (−e − Gef − ∆τ + σ) + (1/2)eTf (Ṁ − 2C)ef
+ eT (ef − Ωe).
(25)
4
Further, substituting (13) into (25) and using Property 1
(implying eTf (Ṁ − 2C)ef = 0), V̇ is simplified as
C. Stability Analysis
Exploring the structures of the adaptive laws (15)-(17), three
˙
possible
scenarios are identified: Case (1): θ̂i , γ̇ increase and
V̇ = −eTf Gef − eT Ωe + eTf (−ρ̂(ef /||ef ||) + σ)
˙
||ef || ≥ $; Case (2): θ̂i , γ̇ decrease and ||ef || ≥ $; Case (3)
≤ −eTf Gef − eT Ωe − (Y (ξ)T Θ̂ + γ)||ef || + Y (ξ)T Θ∗ ||ef || ˙
θ̂i = 0, γ̇ = 0 when ||ef || < $ ∀i = 0, 1, 2.
≤ −λmin (G)||ef ||2 − λmin (Ω)||e||2 − {(θ̂0 + γ − θ0∗ )
Theorem 2. The closed loop system (19) with control input
+ (θ̂1 − θ1∗ )||ξ|| + (θ̂2 − θ2∗ )||ξ||2 }||ef ||.
(26) (13)-(17) guarantees e(t), ef (t), θ̃i (t), γ(t) to be Uniformly
Ultimately Bounded (UUB) where θ̃i , (θ̂i − θi∗ ), i = 0, 1, 2.
Thus, the sufficient condition to achieve V̇ < 0 would be
Proof. The stability analysis of the overall system is carried
∗
∗
∗
θ̂0 + γ ≥ θ0 , θ̂1 ≥ θ1 and θ̂2 ≥ θ2 .
(27) out for the three cases mentioned above using the following
common Lyapunov function:
Let the system (3) does not posses finite time escape [28].
2
X
1 2
1 2
Then integrating both sides of the inequalities in (24) and
θ̃i +
γ .
(33)
V1 = V +
2α
2α
i
3
using those results in (27) lead to the expressions of t1 , t2 , t3
i=0
in (21). So, for t ≥ tin + t̄
˙
Case (1): θ̂i , γ̇ increase ∀i = 0, 1, 2 and ||ef || ≥ $.
P
˙
2
V̇ ≤ −λmin (G)||ef ||2 − λmin (Ω)||e||2
Note that i=0 α1i θ̃i θ̂i = Y (ξ)T (Θ̂ − Θ∗ )||ef ||. Then using
≤ −%m (||ef ||2 + ||e||2 ),
(28) (15)-(16) and following the procedure in (26) one obtains
where %m , min{λmin (G), λmin (Ω)}. Further, the definition
of V yields
2
2
V ≤ %M (||ef || + ||e|| ),
(29)
V̇ ≤ −%V ⇒ V (t) ≤ V (tin + t̄)e−%(t−t̄) ∀t ≥ tin + t̄, (30)
where % , %m /%M . Here θ̂i > 0, γ > β as gains were
T
increasing.
S2 So, to ensure the ‘otherwise’ condition (i.e.T{e ė ≤
0} ∩ { i=0 θ̂i > 0} ∩ {γ > β}), the condition e ė ≤ 0
(i.e. ||e(t)|| does not increase) needs to take place. From the
definition of V , the upper bound of e follows
p
V (t) ≥ (1/2)||e(t)||2 ⇒ ||e(t)|| ≤ 2V (t) ∀t ≥ t0 . (31)
Let ||e(tin + t̄)|| = ψ which implies V (tin + t̄) ≥ ψ 2 /2 from
(31). As V (t) decreases exponentially ∀t ≥ tin + t̄ following
(30), there exist a finite time δt = t − (tin + t̄) such that
V (tin + t̄ + δt) =Sψ 2 /2 implying ||e(tin + t̄ + δt)|| ≤ ψ.
2
So, {eT ė ≤ 0} ∩ { i=0 θ̂i > 0} ∩ {γ > β} would occur at
t ≥ tin +T where T ≤ t̄+δt and θ̂i , γ would start decreasing.
The time δt is found from (30):
∀t ≥ tin + t̄
2
⇒ δt ≤ (1/%)ln{2V (tin + t̄)/ψ }.
+ Y (ξ)T (Θ̂ − Θ∗ )||ef || + γ||ef ||
≤ −eTf Gef − eT Ωe − (Y (ξ)T Θ̂ + γ)||ef || + Y (ξ)T Θ∗ ||ef ||
where %M , max{µ2 , 1}. Substituting (29) into (28) and
using the comparison Lemma [29] yields
ψ 2 ≤ 2V (tin + t̄)e−%δt ,
V̇1 ≤ −eTf Gef − eT Ωe + eTf (−ρ̂(ef /||ef ||) + σ)
(32)
+ Y (ξ)T (Θ̂ − Θ∗ )||ef || + γ||ef ||
≤ −λmin (G)||ef ||2 − λmin (Ω)||e||2 ≤ 0.
From (34) it can be inferred that V1 (t) ∈ L∞ . Thus, the definition of V1 yields γ(t), θ̃i (t), e(t), ef (t) ∈ L∞ ⇒ ξ(t), θ̂i ∈
L∞ . Thus, for this case the closed-loop system remains stable.
The gains γ, θ̂i , i = 0, 1, 2 remain bounded for Case (1),
decrease for Case (2) and remain constant for Case (3). Hence,
∃ γ̄, θ̄i ∈ R+ such that
θ̂0 (t) ≤ θ̄0 , θ̂1 (t) ≤ θ̄1 , θ̂2 (t) ≤ θ̄2 , γ(t) ≤ γ̄ ∀t ≥ t0 . (35)
˙
Case (2): θ̂i , γ̇ decrease ∀i = 0, 1, 2 and ||ef || ≥ $.
Using γ ≥ β (from (20)) and ||ef || ≤ ||Γ||||ξ|| yields:
V̇1 ≤ −eTf Gef − eT Ωe + eTf (−ρ̂(ef /||ef ||) + σ)
− Y (ξ)T (Θ̂ − Θ∗ )||ef || − γς||ξ||4
≤ −eTf Gef − eT Ωe − (Y (ξ)T Θ̂ + γ)||ef || + Y (ξ)T Θ∗ ||ef ||
− ςβ||ξ||4 − Y (ξ)T (Θ̂ − Θ∗ )||ef ||
≤ −ςβ||ξ||4 + 2||Γ||{θ0∗ + θ1∗ ||ξ|| + θ2∗ ||ξ||2 }||ξ||
− λmin (G)||ef ||2 − λmin (Ω)||e||2 .
(36)
Since 0 ≤ θ̂i (t) ≤ θ̄i , β ≤ γ ≤ γ̄ (from (20) and (35)), the
definition of V1 in (33) yields
V1 ≤ %M (||ef ||2 + ||e||2 ) + ζ,
where ζ ,
Remark 4. The increment and decrement of θ̂i , γ can occur
several times depending on the error incurred by the system.
However, time interval ∆t between two successive decrement
will always satisfy ∆t ≤ t̄ + δt. Moreover, high values of
α0 , α1 , α2 , α3 help to reduce t̄ and achieve faster adaptation.
(34)
P2
1
∗2
i=0 α i (θi
+ θ̄i2 ) +
1 2
α3 γ̄ .
(37)
Thus using (37)
−λmin (G)||ef ||2 − λmin (Ω)||e||2 ≤ −%V1 + %ζ.
(38)
Substitution of (38) into (36) yields
V̇1 ≤ −%V1 + fp (||ξ||),
(39)
5
where fp (||ξ||) = −ςβ||ξ||4 + 2||Γ||{θ0∗ ||ξ|| + θ1∗ ||ξ||2 +
+ %ζ. Applying Descarte’s rule of sign change [31],
one can verify that fp (||ξ||) has maximum one positive real
root. Further, it is to be noticed that fp (||ξ|| = 0) = %ζ ∈ R+
and fp (||ξ||) → −∞ as ||ξ|| → ∞. Hence, according to
Bolzano’s Intermediate Value Theorem [32], fp (||ξ||) will have
at least one positive real root. So, combination of the Intermediate Value Theorem and Descarte’s rule of sign change reveals
that fp (||ξ||) has exactly one positive real root. Therefore, the
nature of roots of fp (||ξ||) will be either (i) one positive real
root and three negative real roots or (ii) one positive real root,
one negative real root and a pair of complex conjugate roots.
θ2∗ ||ξ||3 }
Let ι ∈ R+ be the positive real root of fp (||ξ||). Figure
1 depicts the nature of fp (||ξ||) depending on the various
combination of roots [33]. It is to be noted that the actual
graph and values of the roots of the polynomial fp (||ξ||) would
depend on the values of the coefficients of fp (||ξ||). However,
θ0∗ , θ1∗ , θ2∗ and ζ are unknown here. Nevertheless, to study the
stability of the system, it is sufficient to analyse the nature of
fp (||ξ||) (i.e. the instances when fp (||ξ||) > 0 or fp (||ξ||) ≤ 0)
rather than determining the values of the roots. The nature of
any polynomial can be understood from the occurrence of its
real roots [33]. Moreover, the leading coefficient of fp (||ξ||)
(the coefficient of the highest degree term ||ξ||4 ) is negative
(as ς, β ∈ R+ ). As a matter of fact, one can notice from
Fig. 1 that fp (||ξ||) ≤ 0 when ||ξ|| ≥ ι. Hence, the overall
system would be UUB for this case. From the point of view
of controller design, it is important to reduce ι such that better
tracking accuracy can be achieved and this can be obtained by
increasing ς.
Similar to Case (1), V̇1 can be simplified for this case as
V̇1 ≤ −eTf Gef − eT Ωe + eTf (−ρ̂(ef /$) + σ)
≤ −eTf Gef − eT Ωe − (Y (ξ)T Θ̂ + γ)(||ef ||2 /$)
+ Y (ξ)T Θ∗ ||ef ||
≤ −λmin (G)||ef ||2 − λmin (Ω)||e||2 + Y (ξ)T Θ∗ ||ef ||.
(40)
For ||ef || < $, the system remains bounded inside the ball
B$ , {||Γξ|| < $} as ef = Γξ. This implies that Y (||ξ||) ∈
L∞ . Hence, for ||ef || < $, ∃ϑ ∈ R+ such that
||Y (ξ)T Θ∗ ||||ef || ≤ $ϑ.
(41)
Let us define a scalar z as 0 < z < λmin (G). Then using (37)
and (41), (40) is modified as
V̇1 ≤ −λmin (Ω)||e||2 − {λmin (G) − z}||ef ||2 − z||ef ||2 + $ϑ
≤ −%1 V1 − z||ef ||2 + %1 ζ + $ϑ,
(42)
where %1 , {min{(λmin (G) − z), λmin (Ω)}}/%M . Hence, the
system would be UUB when
p
(43)
||ef || = ||Γξ|| ≥ (%1 ζ + $ϑ)/z.
Since the closed loop system remains UUB for both the cases
||ef || ≥ $ and ||ef || < $ using the common Lyapunov
function (33), the overall closed loop system also remains
UUB [34].
Remark 5. It is noteworthy that the condition (20) is necessary for stability of the system. Moreover, high values of ς
helps to reduce ι which consequently can improve controller
accuracy. However, one needs to be careful that too high
value of ς may excite the condition γ ≤ β leading to the
increment in all the gains γ, θ̂i , i = 0, 1, 2. Further, the scalar
terms z, ϑ, ψ, µ2 , ζ, θ̄i and γ̄ are only used for the purpose of
analysis and not used to design control law.
Remark 6. The importance of the auxiliary gain γ can be
realized from Theorems 1 and 2. It can be observed from (21)
that t1 gets reduced due to the presence of α3 (contributed
by γ̇ > 0) which leads to faster adaptation. Moreover, the
negative fourth degree term −ςβ||ξ||4 in fp (||ξ||) (contributed
by γ̇ < 0) ensures system stability for Case (2) by making
fp (||ξ||) ≤ 0 for ||ξ|| ≥ ι. This also indicates the reason
for selecting β > 0 while lower bounds of other gains θ̂i
i = 0, 1, 2 are selected as zero.
Figure 1.
fp (||ξ||) with (a) one positive real root and three negative real
roots (b) one positive real root, one negative real root and a pair of complex
conjugate roots.
Hence, the overall system would be UUB [29] for this case.
˙
Case 3: θ̂i = 0, γ̇ = 0, ∀i = 0, 1, 2 when ||ef || < $.
Special Case: The quadratic term ||ξ||2 in uncertainty
bound (9) is contributed by the matrix C(q, q̇) (through
property 4 in (7)). EL systems such as robotic manipulator,
underwater vehicles, ship dynamics etc. includes C(q, q̇).
However, there also exist second order EL system (e.g. reduced
order WMR system) which does not have the term C(q, q̇).
For such systems, the following LIP structure would hold:
||σ|| ≤ θ0∗ + θ1∗ ||ξ|| , Y (ξ)T Θ∗ ,
(44)
6
where Y (ξ) = [1 ||ξ||]T and Θ∗ = [θ0∗ θ1∗ ]T . Hence,
following the switching gain laws (14)-(17), the control laws
for uncertainty structure (44) are modified as
ρ̂ = θ̂0 + θ̂1 ||ξ|| + γ , Y (ξ)T Θ̂ + γ,
(45)
(i) for ||ef || ≥ $
S1
i
T
αi ||ξ|| ||ef || if {e ė > 0} ∪ { i=0 θ̂i ≤ 0}
˙
θ̂i =
∪{γ ≤ β}
−αi ||ξ||i ||ef || otherwise,
(46)
S1
T
α3 ||ef || if {e ė > 0} ∪ { i=0 θ̂i ≤ 0}
γ̇ =
(47)
∪{γ ≤ β}
3
−ςα3 ||ξ|| otherwise,
(ii) for ||ef || < $
˙
θ̂i = 0, γ̇ = 0,
insufficient and error will increase again. However, K will not
increase (rather it keeps on decreasing) until ||s|| > , which
creates underestimation problem. Low (resp. High) value of
may force K to increase (resp. decrease) for longer duration
when ||s|| ≥ (resp. ||s|| < ) resulting in escalation of the
overestimation (resp. underestimation) problem of ASMC.
Whereas, ASRC allows its gains to decrease when error trajectories move towards ||e|| = 0 and ||ef || ≥ $
(overcoming overestimation problem) and keeps the gains
unchanged when they are sufficient to keep the error within
the ball B$ (overcoming underestimation problem). Since
the overestimation-underestimation problems are alleviated by
ASRC for any $, one can in fact reduce $ for better tracking
accuracy as long as chattering does not occur.
III. A PPLICATION : N ONHOLONOMIC WMR
(48)
with θ̂i (t0 ) > 0, i = 0, 1, γ(t0 ) > β.
(49)
System stability employing (45)-(48) can be analysed exactly
like Theorem 2 using the following Lyapunov function:
V1 = V +
1
X
(1/2αi )θ̃i2 + (1/2α3 )γ 2 .
(50)
i=0
One can verify that the cubic polynomial 2||Γ||{θ0∗ + θ1∗ ||ξ|| +
θ2∗ ||ξ||2 }||ξ|| in fp (||ξ||) of Case (2) would be modified as
quadratic polynomial 2||Γ||{θ0∗ + θ1∗ ||ξ||}||ξ|| using (44) and
(50). Hence, following the argument in Remark 6, it can be
noticed that a cubic term −ςα3 ||ξ||3 is selected in the adaptive
law (48) for system stability.
Thus, with EL system (3), only two structures are possible
for ||σ||: (i) Y (ξ) = [1 ||ξ|| ||ξ||2 ]T , Θ∗ = [θ0∗ θ1∗ θ2∗ ]T and
(ii) Y (ξ) = [1 ||ξ||]T , Θ∗ = [θ0∗ θ1∗ ]T . Both these situations
are covered here. For better inference, the ASRC algorithm is
summarized in Table I for various system structures.
Table I
ASRC A LGORITHM FOR VARIOUS S YSTEM S TRUCTURES
System Structure
(3)
C(q, q̇) 6= 0
C(q, q̇) = 0
LIP structure of ||σ||
Control law
(9)
(44)
(13) - (18)
(13), (45) - (49)
Comparison with existing Adaptive-Robust law: To gain
further insight into the advantage of the proposed adaptive law,
the following adaptive law of ASMC [17]-[18] for switching
gain K is provided:
(
K̄||s||sgn(||s|| − ), if K > β
K̇(t) =
(51)
β
if K ≤ β,
where , K̄ ∈ R+ are user defined scalars and s is the sliding
surface. It can be observed from (51) that when ||s|| ≥
the switching gain K increases monotonically even if error
trajectories move close to ||s|| = 0. This gives rise to the
overestimation problem of switching gain. Again, even if K
is sufficient to keep ||s|| within , it decreases monotonically
when ||s|| < . So, at certain time, K would become
Figure 2. Schematic of a WMR.
Nonholonomic WMR, which has vast applications in transportation, planetary exploration, surveillance, security, humanmachine-interfaces etc., provides a unique platform to test the
proposed control law. Hence, the performance of the proposed
ASRC is verified using a commercially available ’PIONEER
3’ WMR in comparison to Adaptive Sliding Mode Control
(ASMC) [17]-[18]. The ASMC law is detailed in [17]-[18]
while it follows the adaptive law (51).
The Euler-Lagrangian formulation of a nonholonomic
WMR (Fig. 2) is given as [35]:
M (q)q̈ + C(q, q̇)q̇ = Lτ − A∗ T λ∗ ,
m
0
mdsinϕ
0
0
m
−mdcosϕ
0
¯
mdsinϕ
−mdcosϕ
I
0
where M =
0
0
0
Iw
0
0
0
0
2
0 0
mdϕ̇ cosϕ
0 0
mdϕ̇2 sinϕ
, C(q, q̇)q̇ =
, τ = τr .
0
0
0
L =
τl
1 0
0
0 1
0
(52)
0
0
0
,
0
Iw
7
Here q ∈ R5 = {xc , yc , ϕ, θr , θl } is the generalized coordinate
vector of the system; (xc , yc ) are the coordinates of the center
of mass (CM) of the system and ϕ is the heading angle;
(θr , θl ) and (τr , τl ) are rotation and torque inputs of the right
¯ Iw , rw and b represent the
and left wheels respectively; m, I,
system mass, system inertia, wheel inertia, wheel radius and
robot width respectively; d is the distance to the CM from
the center of the line joining the two wheel axis; A∗ and λ∗
represent the constraint matrix and vector of constraint forces
(Lagrange multipliers) respectively. Expressions of A∗ and I¯
can be found in [35].
It is noteworthy that the system (52) has only two control
inputs although having five generalized coordinates. In fact,
for WMR, one can only directly control wheel positions
(θr , θl ) rather than (xc , yc , ϕ). So, the system dynamics can
be represented as a combination of a reduced order dynamics
and kinematic model for efficient controller design as in [30],
[35]:
specifically in the face of unmodeled dynamics and timevarying uncertainty. During the experiment, the payload of the
system may be varied due to addition or removal of sensors
according to the application requirement; this causes variations
in overall system mass, center of mass, inertia etc. Further, the
original systems dynamics (53) is formed based on the pure
rolling assumption. This assumption is not satisfied in practice
due to the friction effect between wheel and surface; this is
denoted by f (q̇R ) in the WMR dynamics (57). Apart from this,
there are also effects of external disturbances ds . However,
the evaluation of switching gain for robust controller like [26]
requires prior knowledge of the bound of the uncertainties.
This implies the designer needs to have the knowledge of
θ0∗ , θ1∗ for WMR (due to the absence of Coriolis term in the
WMR dynamics). This further means that the designer should
have the knowledge of the parametric variations in systems
as well as bound of f (q̇R ) and ds . This demands tedious
modelling job which is also not always accurate.
The benefit, applicability and efficacy of the proposed
MR q̈R + CR q̇R = τ,
(53) ASRC can be realized in this context. ASRC does not require
any knowledge of the systems dynamics terms MR , f (q̇R ) and
rw b
rw b
d
s of WMR system (57) (and for the matter of fact any EL sysb
2 cos(ϕ) − dsin(ϕ)
b
2 cos(ϕ) + dsin(ϕ)
rw b
tem
representing dynamics (3)). Further, while implementing
rw b sin(ϕ) + dcos(ϕ)
b 2
b
2 sin(ϕ) − dcos(ϕ)
the
control
law, it does not need any knowledge of θ0∗ , θ1∗ and
q̇ =
rw /b
−rw /b
q̇R ,
rather adapts these terms by the adaptive law (46)-(49) (since
1
0
Coriolis component is zero, the ASRC algorithm applied to
0
1
|
{z
}
WMR is based on the control laws (13), (45)-(49)). Hence,
S(q)
ASRC eliminates any effort to model the system as well as
(54) avoids any need to characterize the time-varying uncertain
parameters and disturbances. It is to be noted that S(q) is only
used
for coordinate transformation and WMR pose (xc , yc , ϕ)
k k2
where MR = S T M S = 1
,
(55) representation and, not for control law design.
k2 k1
Hence, the objective is to apply ASRC and ASMC to the
k1 = Iw + {I¯ + m(b2 /4 − d2 )}(rw 2 /b2 ),
d
reduced
order WMR system (57) to track a desired qR
(t)
¯ w 2 /b2 ),
k2 = {m(b2 /4 + d2 ) − I}(r
d
which
in
effect
track
a
desired
q
(t)
through
(54).
To
illustrate
0 0
CR = S T (M Ṡ + CS) =
, qR = [θr θl ]T .
(56) the fact: one can direct a WMR to move in a specified circular
0 0
path by designing two suitable different and fixed wheel
velocities or in a Lawn-Mower path by applying approximated
As WMR moves on ground, the gravity vector g(q) and the
square-wave velocity profile to the wheels [26].
potential function would certainly be zero which implies that
MR , CR satisfies the Properties 1 and 3 [30], [36]. The main
implication of system Property 1 is to hold eTf (Ṁ −2C)ef = 0 A. Experimental Scenario
and this can be easily verified from (55)-(56). The WMR
The WMR is directed to follow a circular path using the
dynamics (52) is based on rolling without slipping condition following desired trajectories:
and hence the term f (q̇R ) is omitted. However, in practical
θrd = (4t + π/10)rad, θld = (3t + π/10)rad.
circumstances a WMR is always subjected to uncertainties like
friction, slip, skid, external disturbance etc. Hence, incorpoPIONEER 3 uses two quadrature incremental encoders (500
rating (56), the system dynamics (53) is modified as
ppr) and always starts from θr (t0 ) = θl (t0 ) = 0 and the
initial
wheel position error (π/10, π/10) rad helps to realize
MR q̈R + f (q̇R ) + ds = τ,
(57)
the error convergence ability of the controllers. The desired
where f (q̇R ) and ds are considered to be the unmodelled dy- WMR pose (xdc , ycd , ϕd ) and actual WMR pose (xc , yc , ϕ)
namics and disturbance respectively. Often, simple controllers can be determined from (54) using (θ̇rd , θ̇ld ) and (θ̇r , θ̇l )
such as open loop control (OLC), PID controller are used (obtained from encoder) respectively with rw = 0.097m, b =
in practice for their simplicity. However, the various works 0.381m, d = 0.02m (supplied by the manufacturer). The
such as [26] and the references of [26] (e.g., reference [15] control laws for both ASRC and ASMC are written in VC++
in [26]) have discussed the need to formulate advanced robust environment. Considering the hardware response time, the
tracking controllers for WMR compared to conventional open sampling interval is selected as 20ms for all the controllers.
loop control or PID control to improve tracking accuracy, Further, to create a dynamic payload variation, a 3.5kg payload
8
is added (kept for 5 sec) and removed (for 5 sec) periodically
on the robotic platform at different places.
The controller parameters for ASRC are selected as: G =
Ω = I, $ = 0.5, θ̂i (t0 ) = γ(t0 ) = 20, αi = α3 = 10 ∀i =
0, 1, β = 0.1, ς = 10. Further, the controller parameters for
ASMC are selected as s = ef , K̄ = 10, K(t0 ) = 35, = 0.5.
B. Experimental Results and Comparison
The path tracking performance of ASRC is depicted in
Fig. 3 while following the desired circular path. The tracking
performance comparison of ASRC with ASMC is illustrated
in Fig. 4 in terms of Ep ( defined by the Euclidean distance
in xc , yc error) and Eτ (defined as ||τ ||). ASMC framework is
built on the assumption that uncertainties are upper bounded
by an unknown constant (i.e. θ1∗ = θ2∗ = 0 for general EL
systems and θ1∗ = 0 for this particular WMR based experiment
as CR = 0). This assumption is restrictive in nature for EL
systems and the switching gain is thus insufficient to provide
the necessary robustness. As a matter of fact, ASRC provides
better tracking accuracy over ASMC.
To evaluate the benefit of the proposed adaptive-robust law,
the evaluation of switching gain for ASMC and ASRC are
provided in Fig. 5 and 6 respectively. Figure 5 reveals that
K, the switching gain of ASMC, increases even when ||s||
approaches towards ||s|| = 0 during the time t=0−1.2 sec.
This is due to the fact that K does not decrease unless ||s|| <
and this gives rise to the overestimation problem. On the other
hand for ASRC, it can be seen from Fig. 6 that all the gains
d
) decreases
(i.e. γ, θ̂0 , θ̂1 ) decrease when ||eR || (eR = qR − qR
during t=0−1 sec when ||efR || ≥ $ (||efR || = ėR + ΩeR ).
So, ASRC overcomes the overestimation problem which is
encountered in ASMC. Further, K decreases monotonically
for time durations t=1.2−38.5 sec, when ||s|| < . This
monotonic decrement makes K insufficient to tackle uncertainties at certain time creating underestimation problem. As
a result, ||s|| starts increasing again for t > 38.5 sec leading to
poor tracking accuracy and K increases again when ||s|| ≥ .
Gains of ASRC, on the contrary, stay unchanged for t > 1 sec
when the gains are sufficient to keep ||efR || < $ avoiding any
underestimation problem. Further the evaluation ρ̂ of ASRC is
shown in Fig. 7. It is to be noted that ρ̂ = θ̂0 + θ̂1 ||ξ|| + γ for
WMR. Hence, though θ̂0 , θ̂1 and γ remain constant for t > 1
sec, ρ̂ is not constant for t > 1 due to the presence of ξ. This
is shown by magnifying ρ̂ for time durations t = 1−5 sec and
t = 20 − 25 sec in Fig. 7. While reduction of would cause
more overestimation problem for ASMC, Table II shows how
tracking accuracy of ASRC improves with reduced $ (other
control parameters are kept unchanged) while chattering is not
observed in control input.
$ = 0.5
0.053
89.78
$ = 0.3
0.0421
75.46
Figure 4. Performance comparison between ASRC and ASMC.
Figure 5. Switching gain evaluation of ASMC.
Table II
P ERFORMANCE OF ASRC FOR VARIOUS $
RMS (root mean squared) Ep (m)
RMS ||τ || (Nm)
Figure 3. Circular path tracking with ASRC.
$ = 0.1
0.0362
67.13
It can be noticed from Fig. 7 that initials gains are high
enough such that ||eR || decreases from the beginning and
so do the gains γ, γ̂i . Hence, it would be prudent to verify
the capability of ASRC in alleviating the overestimation-
9
Figure 8. Performance comparison between various controllers.
Figure 6. Switching gain evaluation of ASRC.
Figure 9. Switching gain evaluation of ASRC2.
of the gains.
Figure 7.
Evaluation of ρ̂ of ASRC.
underestimation problem while starting with relatively low
gains. Therefore, the same experiment for ASRC is repeated
with much lower initial value of the gains θ̂i , i = 0, 1 and γ.
Previously, the initial values were γ(t0 ) = θ̂i (t0 ) = 20. This
time they are selected to be γ(t0 ) = θ̂i (t0 ) = 10. The tracking
performance and evaluation of the switching gain for the later
case is shown in Fig. 8 and 9 respectively. Here ASRC1
denotes when γ(t0 ) = θ̂i (t0 ) = 20 and ASRC2 denotes when
γ(t0 ) = θ̂i (t0 ) = 10.
It can be noticed from Fig. 8 and 9 that initially the tracking
error is high for ASRC2 compared to ASRC1 and ASMC
(initial gain K(t0 ) = 35) due to low initial gains. However, at
t ≥ 5sec tracking accuracy of ASRC2 begins to improve as the
gains became sufficient enough to negotiate the uncertainties
and eventually the tracking performance of ASRC2 is found
to be similar to ASRC1 from t ≥ 12sec and much improved
compared to ASMC. This proves that the proposed adaptive
law can perform satisfactorily even with low initial conditions
Another important aspect to verify is whether ASRC2
can alleviate the over- and under-estimation issue similar to
ASRC1. It can verified from Fig. 9 that when ||efR || > $, the
gains follow the pattern of ||eR || according to the adaptive laws
(46)-(47). Due to the initial low values, ||eR || increases and so
do the gains; similarly at t ≥ 5sec the gains decrease as ||eR ||
decreases. Further, at t ≥ 8sec the gains remain unchanged
when they were sufficient to keep the filtered tracking error
||efR || ≤ $, according to the law (48), thus overcoming
the underestimation problem. Moreover, that gains do not
increase during t = 3.78 − 8sec when ||eR || decrease sand
thus avoids the overestimation problem. Hence, low initial
gain conditions do not affect the capability ASRC2 in
alleviating the over- and under-estimation problem.
Moreover, it can be noticed from the first conditions of (46)(47) that the rate of increment of θ̂0 and γ are same; hence they
have similar value in Fig. 9 when ||eR || increases during t =
0 − 3.7sec (please note that we have selected α0 = α3 = 10
in the experiment) . However, their rate of falling are different
(second conditions of (46)-(47)); thus, they exhibit different
falling pattern in Fig. 9.
10
IV. C ONCLUSIONS
A novel ASRC law is proposed for a class of uncertain EL
systems where the upper bound of uncertainty possesses a LIP
structure. The benefit of the ASRC lies in the fact that it is
independent of the nature of the uncertainty and can negotiate
uncertainties which can be LIP or NLIP. ASRC does not
presume the overall uncertainty to be bounded by a constant
and avoids putting any prior restriction on states. Moreover, the proposed adaptive law alleviates the overestimationunderestimation problem of switching gain. The experimental
results validate the efficacy of the proposed control law in
comparison with the existing adaptive sliding mode control.
The future work would be to extend the ASRC law for systems
with unmatched disturbances.
R EFERENCES
[1] 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.
[2] B. Bandyopadhayay, S. Janardhanan and S. K. Spurgeon, Advances in
Sliding Mode Control, Springer-Verlag, New York, 2013.
[3] 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.
[4] 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.
[5] G. Zhang, J. Chen and Z. Lee, Adaptive robust control of servo
mechanisms with partially known states via dynamic surface control
approach, IEEE Transactions on Control System Technology, vol. 18(3),
pp. 723-731, 2010.
[6] C.Guan and S. Pan, Nonlinear adaptive robust Control of single-rod
electro-hydraulic actuator with unknown nonlinear parameters, IEEE
Transactions on Control System Technology, vol. 16(3), pp. 434-445,
2008.
[7] C. Hu, B. Yao, Z. Chen, and Q. Wang, Adaptive robust repetitive control
of an industrial biaxial precision gantry for contouring tasks, IEEE
Transactions on Control System Technology, vol. 19(6), pp. 1559-1568,
2011.
[8] 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.
[9] 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.
[10] 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.
[11] 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.
[12] Q. Chen, L. Tao, Y. Nan and X. Ren, Adaptive nonlinear sliding mode
control of mechanical servo system with lugre friction compensation,
ASME Journal of Dynamic Systems, Measurement, and Control, vol.
138, DOI: 10.1115/1.4032068, 2016.
[13] 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.
[14] Q. Meng, T. Zhang, X. Gao and J. Y. Song, Adaptive sliding mode
fault-tolerant control of the uncertain stewart platform based on offline
multibody Dynamics, IEEE/ASME Transactions on Mechatronics, vol.
19(3), pp. 882-894, 2014.
[15] S. Liu, H. Zhou, X. Luo and J. Xiao, Adaptive sliding fault tolerant
control for nonlinear uncertain active suspension systems, Journal of
The Franklin Institute, DOI 10.1016/j.jfranklin.2015.11.002, 2015.
[16] S. Mobayen, An adaptive chattering-free PID sliding mode control based
on dynamic sliding manifolds for a class of uncertain nonlinear systems,
Nonlinear Dynamics, DOI 10.1007/s11071-015-2137-7, 2015.
[17] 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.
[18] 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.
[19] V. Utkin and A. Poznyak, Adaptive sliding mode control with application
to super-twist algorithm: equivalent control method, Automatica, vol. 49,
pp. 39-47, 2013.
[20] J. A. Moreno, D. Y. Negrete, V. Torres-Gonzalez and L. Fridman, Adaptive continuous twisting algorithm, International Journal of Control,
DOI: 10.1080/00207179.2015.1116713, 2015.
[21] V. Adetola, M. Guay and D. Lehrer, Adaptive estimation for a class
of nonlinearly parametrized dynamical systems, IEEE Transactions on
Automatic Control, vol. 59(10), pp. 2818-2824, 2014.
[22] A. M. Annaswamy, F. P. Skantze and A-P Loh, Adaptive control of continuous time systems with convex/concave Parametrization, Automatica,
vol. 34(1), pp. 33-49, 1998.
[23] M. W. Spong, S. Hutchinson and M. Vidyasagar, Robot Dynamics and
Control, John Wiley and Sons, New York, 2004.
[24] W. Shang and S. Cong, Motion Control of parallel manipulators using
acceleration feedback, IEEE Transactions on Control System Technology, vol. 20(1), pp. 314-320, 2014.
[25] T. Y. Choi, B. S. Choi and K. H. Seo, Position and compliance control
of a pneumatic muscle actuated manipulator for enhanced safety, IEEE
Transactions on Control System Technology, vol. 19(4), pp. 832-842,
2011.
[26] 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. 28862892, 2014.
[27] J. Wu, J. Huang, Y. Wang and K. Xing, Nonlinear disturbance observerbased dynamic surface control for trajectory tracking of pneumatic
muscle system, IEEE Transactions on Control System Technology, vol.
22(2), pp. 440-455, 2014.
[28] W. M. Haddad and V. Chellaboina, Nonlinear dynamical systems and
control: a Lyapunov-based approach, Princeton University Press, 2008.
[29] H. Khalil, Nonlinear systems, 3rd ed. Prentice Hall, 2002.
[30] T. Das and I. N. Kar, Design and implementation of an adaptive fuzzy
logic-based controller for wheeled mobile robots, IEEE Transactions on
Control System Technology, vol. 14(3), pp. 501-510, 2006.
[31] B. Anderson, J. Jackson and M. Sitharam, Descartes’ Rule of Signs
Revisited, The American Mathematical Monthly, vol. 150, pp. 447-451,
1998.
[32] S. B Russ, A Translation of Bolzano’s Paper on the Intermediate Value
Theorem, Historia Mathematica, vol. 2, pp. 156-185, 1980.
[33] K. T. Leung, I. A. C Mok and S. N. Suen Polynomials and functions,
Honk Kong University Press, 1992.
[34] D. Liberzon, Switching in systems and control, Birkhäuser, 2003.
[35] P. Coelho and U. Nunes, Path-following control of mobile robots in presence of uncertainties, IEEE Transactions on Robotics and Automation,
vol. 21, no. 2, pp. 252-261, 2005.
[36] G. Campion, B. d’Andrea-Novel and G. Bastin, Modelling and state
feedback control of nonholonomic mechanical systems, 30th IEEE
Conference on Decision and Control, pp. 1184-1189, 1991.
| 3cs.SY
|
How fast can we make interpreted Python?
Russell Power
Alex Rubinsteyn
arXiv:1306.6047v2 [cs.PL] 13 Aug 2013
New York University
{power,alexr}@cs.nyu.edu
Abstract
Python is a popular dynamic language with a large part of its
appeal coming from powerful libraries and extension modules. These augment the language and make it a productive
environment for a wide variety of tasks, ranging from web
development (Django) to numerical analysis (NumPy).
Unfortunately, Python’s performance is quite poor when
compared to modern implementations of languages such as
Lua and JavaScript. Why does Python lag so far behind
these other languages? As we show, the very same API and
extension libraries that make Python a powerful language
also make it very difficult to efficiently execute.
Given that we want to retain access to the great extension
libraries that already exist for Python, how fast can we make
it? To evaluate this, we designed and implemented Falcon, a
high-performance bytecode interpreter fully compatible with
the standard CPython interpreter. Falcon applies a number
of well known optimizations and introduces several new
techniques to speed up execution of Python bytecode. In our
evaluation, we found Falcon an average of 25% faster than
the standard Python interpreter on most benchmarks and in
some cases about 2.5X faster.
1. Introduction
Python is popular programming language, with a long history and an active development community. A major driver
of Python’s popularity is the diverse ecosystem of libraries
and extension modules which make it easy to do almost
anything, from writing web-servers to numerical computing. But despite significant effort by Python’s developers,
the performance of Python’s interpreter still lags far behind
implementations of languages such as Lua and JavaScript.
What differentiates Python from “faster” languages? Obviously, implementation choices (JIT vs. interpreter) can
have a dramatic effect on performance. In the case of Python,
however, the landscape of choices is severely constrained by
the very same API that makes it easy to extend. The standard
interpreter for Python (called CPython) exposes a low-level
API (the Python C API [4]) which allows for building extension libraries and for embedding the interpreter in other
programs. The Python C API allows access to almost every aspect of the interpreter, including inspecting the current
interpreter state (what threads are running, function stacks,
etc..) and pulling apart the representation of various object
types (integer, float, list, dictionary, or user-defined). For performance reasons, many Python libraries are written in C or
another compiled language, and interface to Python via the
C API.
In an ideal world, any implementation of Python’s semiformal specification [19] would be interchangeable with the
CPython implementation. Unfortunately, to avoid breaking
libraries, an alternative implementation must also support
the full C API. While the size of the C API (∼700 functions) is burdensome, what really makes it problematic is
the degree to which it exposes the internal memory layout
and behavior of Python objects. As a result of this many
Python extensions have become intimately coupled to the
current implementation of the CPython interpreter. For instance, modifying the layout of the basic object format (for
example, to use less memory) breaks even source level compatibility with existing extensions.
The value of these extensions to Python is hard to overstate. Python already has a fast JIT compiler in the form of
PyPy [9], but is has not seen widespread adoption. This is, to
a large extent, due to the lack of support for existing CPython
extension libraries. Replacing these libraries is not a simple undertaking; NumPy [16] alone consists of almost 100k
lines of C source, and the SciPy libraries which build upon
it are another 500k lines.
To evaluate how much we can improve over CPython
(without breaking existing extension modules), we developed an alternative Python virtual machine called Falcon.
Falcon converts CPython’s stack-oriented bytecode into a
register-based format and then performs optimizations to remove unnecessary operations and occasionally elide type
checks. Falcon’s register bytecode is then executed by a
threaded interpreter, which attempts to accelerate common
code patterns using attribute lookup caching and register tagging.
Overall, we found that:
• A combination of register conversion, simple bytecode
optimization, and a virtual machine with threaded dispatch result in an average 25% speedup over CPython.
On certain benchmarks, Falcon was up to 2.5X faster.
the array of local values? This is the essence of a register
bytecode. Every instruction explicitly labels which registers
(local variable slots) it reads from and to which register it
writes its result.
Translated into register code, the above example would
look like:
r2 = B I N A R Y _ A D D( r0 , r1 )
R E T U R N _ V A L U E r2
• Just-in-time register conversion and optimization are fast
enough to run online for every function. This means that
a system like Falcon can accelerate any code without the
need for profiling and compilation heuristics.
Figure 3. Register code for adding two inputs
When converted into register code, the local variables
and z are represented by the registers r0, r1 and r2.
Since the source and destination registers are part of each
instruction, it is possible to express this function using only
two instructions.
The downside to Falcon’s register code format (like any
register code) is that each instruction must be larger to make
room for register arguments. There are two advantages to
register code which make the space increase worthwhile.
The first is a potential for reducing the time spent in virtual machine dispatch by reducing the number of instructions that must be executed. Previous research has verified
that switching from a stack-based virtual machine to a register machines can improve performance [12, 18].
Additionally, it is much easier to write optimizations for a
register code. The reason for this is that when every instruction implicitly affects a stack, program analyses and optimizations must track these side effects in some form of virtual stack. A register code, on the other hand, admits the
expression of more compact optimization implementations
(section 3.2), since instructions only depend on each other
through explicit flows of data along named registers.
x, y
2. Overview
Falcon does not replace the standard CPython interpreter,
but rather runs inside of it. A user-defined function can be
marked for execution by Falcon with the decorator @falcon.
When a decorated function is called, Falcon translates that
function’s stack bytecode into Falcon’s more compact register code. This register code is then optimized to remove redundant computations and decrease the number of registers
needed by a function. The optimized register code is then
passed on to Falcon’s virtual machine for evaluation.
For example, consider executing the following function,
which adds its two inputs, assigns their sum to a local variable, and then returns that variable:
def add (x , y ):
z = x + y
return z
Figure 1. Python function that adds two inputs
When Python first encounters this code, it is compiled
into the following bytecode.
3. Compiler
LOAD_FAST
LOAD_FAST
BINARY_ADD
STORE_FAST
LOAD_FAST
RETURN_VALUE
(x)
(y)
(z)
(z)
Figure 2. Python stack code that adds two inputs
Each operation in Python’s bytecode implicitly interacts
with a value stack. Python first pushes the values of the
local variables x and y onto the stack. The instruction BINARY_ADD then takes these two values, adds them, and pushes this
new result value onto the stack. Where did the values of
the local variables come from? In addition to the stack,
Python’s virtual machine maintains a distinct array of values
for named local variables. The numerical arguments attached
to the LOAD_FAST and STORE_FAST instructions indicate which
local variable is being loaded or stored.
Even from this simple example, we can see that Python
bytecode burdens the virtual machine with a great deal of
wasteful stack manipulations. Could we get better performance if we did away with the stack and instead only used
The Falcon compiler is structured as a series of passes, each
of which modifies the register code in some way. To illustrate
the behavior of each pass we will use a simple example
function called count_threshold - which counts the number of
elements in a list below a given threshold:
def c o u n t _ t h r e s h o l d(x , t ):
return sum ([ xi < t for xi in x ])
For count_threshold Python generates the following stack
bytecode (Figure 4):
Before we can do anything else, we need to convert our
original stack machine bytecode to the equivalent register
code.
3.1 Stack-to-register conversion
To convert from Python stack code to register code Falcon
uses abstract interpretation [10]. In Falcon this takes the
form of a “virtual stack” which stores register names instead
of values. Falcon steps through a function’s stack operations,
evaluates the effect each has on the virtual stack, and emit an
equivalent register machine operation.
LOAD_GLOBAL
BUILD_LIST
LOAD_FAST
GET_ITER
10: FOR_ITER
STORE_FAST
LOAD_FAST
LOAD_FAST
COMPARE_OP
LIST_APPEND
JUMP_ABSOLUTE
31: C A L L _ F U N C T I O N
RETURN_VALUE
( sum )
(x)
( to 31)
( xi )
( xi )
(t)
( <)
Python
Falcon
BUILD_LIST 0
r5 = BUILD_LIST 0
Stack
h4i → h5, 4i
instructions directly, we can alias these variables to specially
designated register names, which simplifies our code and
reduces the number of instructions needed.
10
Python
Falcon
LOAD_FAST 0 (x)
Stack
h5, 4i → h1, 5, 4i
Figure 4. Python stack machine bytecode
Handling control flow
For straight-line code this process is fairly easy; most Python
instructions have a fairly straightforward effect on the stack.
But what happens when we encounter a branch? We need
to properly simulate both execution paths. To handle this
situation, we must make a copy of our virtual stack, and
evaluate both sides of the branch.
With branches come merge points; places where two or
more branches of execution come together. Each thread of
control flow might have assigned different register names to
each stack position. To handle this situation Falcon inserts
rename instructions before merge points, ensuring that all
incoming register stacks are compatible with each other.
(This is the same mechanism employed by compilers which
use static single assignment form (SSA)[11] to resolve φnodes.)
Register r1 is aliased to the local variable x. Therefore
for the LOAD_FAST operation here, we don’t need to generate a
Falcon instruction, and can instead simply push r1 onto our
virtual stack.
GET_ITER pops a sequence off of the stack and pushes back
an iterator for the sequence.
Python
Falcon
GET_ITER
r6 = GET_ITER(r1)
Stack
h1, 5, 4i → h6, 5, 4i
FOR_ITER is a branch instruction. It either pushes the next
element in the iterator onto the stack and falls-through to the
next instruction, or pops the iterator off the stack and jumps
to the other side of the loop.
Python
Falcon
FOR_ITER
r7 = FOR_ITER(r6)
Stack
h6, 5, 4i → h7, 6, 5, 4i
or h5, 4i
Example conversion
Let’s walk through how this works for the example stack
code above (figure 4).
First we find the value of the function “sum” using the
LOAD_GLOBAL instruction. In the CPython interpreter, LOAD_GLOBAL
looks up a particular name in the dictionary of global values
and pushes that value onto the stack. Since the set of literal
names used in a function is known at compile time, the instruction can simply reference the index of the string “sum”
in a table of constant names. The equivalent register machine
instruction assigns the global value to a fresh register (in this
case r4). For brevity, the “stack” column in the listings below
will show just the register number for each instruction.
Python
Falcon
LOAD_GLOBAL 0
r4 = LOAD_GLOBAL 0
Stack
hi → h4i
The effect of this operation on the virtual stack is to
push the register r4 on top. When a later operation consumes
inputs off the stack, it will be correctly wired to use r4 as an
argument.
BUILD_LIST constructs an empty list to contain the results.
We create a new register r5 and push it onto the stack.
Python has special operations to load and store local
variables and to load constants. Rather than implement these
One branch of the FOR_ITER instruction takes us into inner
loop, which continues until the iterator is exhausted:
Python
Falcon
STORE_FAST (xi)
r3 = r7
LOAD_FAST (xi)
LOAD_FAST (t)
COMPARE_OP
r8 = r3 > r2
LIST_APPEND
APPEND(r5, r8)
JUMP_ABSOLUTE
JUMP_ABSOLUTE
Stack
h7, 6, 5, 4i → h6, 5, 4i
h6, 5, 4i → h3, 6, 5, 4i
h3, 6, 5, 4i → h2, 3, 6, 5, 4i
h2, 3, 6, 5, 4i → h8, 6, 5, 4i
h8, 6, 5, 4i → h6, 5, 4i
h6, 5, 4i
The behavior of the LIST_APPEND instruction here might
look somewhat surprising; it appears to “peek into” the stack
to find r5. This special behavior is unique to the LIST_APPEND
instruction, and likely is a result of past performance tuning
in the CPython interpreter (building lists is a very common
operation in Python).
And the other branch takes us to our function’s epilogue:
Python
Falcon
CALL_FUNCTION (sum)
r9 = sum(r4)
RETURN_VALUE
RETURN_VALUE(r9)
Stack
h5, 4i → h6i
h6i → hi
Operations with dynamic stack effects
In the above example, the effect of each instruction on the
stack was known statically. It turns out that this is the case
for almost all Python instructions. In fact, only one operation (END_FINALLY) has a stack effect that must be determined
at runtime. This instruction appears in functions which have
a try...finally block, and determines whether a caught exception should be re-raised. While it is possible to handle
such an instruction dynamically (by inserting branches in
the generated code for each possible stack effect), we chose
a much simpler option - we simply do not compile functions
containing this instruction. Instead these functions are evaluated using the existing Python interpreter. As this instruction
is relatively rare, (occurring in only 4% of the functions in
the Python standard library), and is almost never found in
performance sensitive code, the cost of not supporting it is
minimal.
stead use the original source (r7). In the code from Figure 5,
copy propagation changes “r8 = COMPARE_OP(r3, r2)” into “r8 =
COMPARE_OP(r7, r2)”. By itself, this optimization will not improve performance. Instead, it enables other optimizations
to remove useless instructions and to reuse unoccupied registers.
Dead Code Elimination
If the value contained in a register is never used in a program
(likely to occur after copy propagation), it may be possible to
delete the instruction which “created” that register. Instructions which lack side effects (simple moves between registers) are safe to delete, whereas instructions which may run
user-defined code (such as BINARY_ADD) must be preserved even
if their result goes unused. Once copy propagation is applied
to the code above, the register r3 is never used. Thus, the
move instruction “r3 = r7” gets deleted by dead code elimination.
3.2 Bytecode Optimizations
After the stack to register pass, the bytecode for
count_threshold now looks like Figure 5.
bb_0 :
r4 = L O A D _ G L O B A L ( sum )
r5 = B U I L D _ L I S T
r6 = GET_ITER ( r1 ) -> bb_10
bb_10 :
r7 = FOR_ITER ( r6 ) -> bb_13 , bb_31
bb_13 :
r3 = r7
r8 = C O M P A R E _ O P( r3 , r2 )
L I S T _ A P P E N D( r5 , r8 )
J U M P _ A B S O L U T E() -> bb_10
bb_31 :
r9 = C A L L _ F U N C T I O N( r5 , r4 )
R E T U R N _ V A L U E( r9 )
Figure 5. Unoptimized register code
Note that rather than using positions in the code for jump
targets, Falcon splits up code into basic blocks (indicated
with the bb_* prefix). This change has no effect on the code
that ultimately gets run by the virtual machine but greatly
simplifies the implementation of optimization passes.
The register machine bytecode emitted from the stack-toregister pass tends to be sub-optimal: it uses too many registers and often contains redundant loads used to emulate the
effect of stack operations. To improve performance, we perform a number of optimizations that remove these redundant
operations and improve register usage.
The advantage of switching to a register code becomes
clear at this point; we can apply known optimization techniques to our unoptimized register code with almost no modification. The optimizations used in Falcon are common to
most compilers; we briefly describe them and their application to Falcon here.
Copy Propagation
Whenever a value is copied between registers (e.g. r3 = r7), it
is possible to change later uses of the target register (r3) to in-
Register Renaming
Even if a register is used at some point in the program,
it might not necessarily be “alive” for the entire duration
of a function’s execution. When two registers have nonoverlapping live ranges, it may be possible to keep just one
of them and replace all uses of the other register. This reduces the total number of registers needed to run a function,
saving memory and giving a slight performance boost.
Register code after optimization
After Falcon’s optimizations are applied to the bytecode in 5,
extraneous store instructions (such as r3 = r7) are removed.
Furthermore, register renaming causes the registers r7 and r4
to be used repeatedly in place of several other registers.
The optimized register bytecode achieves a greater instruction density, compared with the original stack code. Optimization in general reduces the number of instructions by
30%, with a similar improvement in performance.
bb_0 :
r4 = L O A D _ G L O B A L( sum )
r5 = B U I L D _ L I S T()
r6 = GET_ITER ( r1 ) -> bb_10
bb_10 :
r7 = FOR_ITER ( r6 ) -> bb_13 , bb_31
bb_13 :
r7 = C O M P A R E _ O P( r7 , r2 )
L I S T _ A P P E N D( r5 , r7 )
J U M P _ A B S O L U T E() -> bb_10
bb_31 :
r4 = C A L L _ F U N C T I O N [1]( r5 , r4 )
R E T U R N _ V A L U E( r4 ) ->
Figure 6. Optimized register code
Difficulty of Optimizing Python Code
It would be desirable to run even more compiler optimizations such as invariant code motion and common subexpression elimination. Unfortunately these are not valid
when applied to Python bytecode. The reason these optimizations are invalid is that almost any Python operation
might trigger the execution of user-defined code with unrestricted side effects. For example, it might be tempting
to treat the second BINARY_ADD in the following example as
redundant.
r3 = B I N A R Y _ A D D r1 , r2
r4 = B I N A R Y _ A D D r1 , r2
However, due to the possibility of encountering an overloaded __add__ method, no assumptions can be made about
the behavior of BINARY_ADD. In the general absence of type
information, almost every instruction must be treated as
INVOKE_ARBITRARY_METHOD.
4. Virtual Machine
After compilation, the register code is passed over to the
virtual machine to evaluate. Falcon uses a 3 main techniques
(token-threading, tagged-registers and lookup hints) to try
and improve the dispatch performance, which we cover in
this section.
Token-threading
The common, straightforward approach (used by Python
2.*) to writing a bytecode interpreter is to use a switch and a
while loop:
Code * ip = i n s t r u c t i o n s;
while (1) {
switch ( ip - > opcode ) {
case B I N A R Y _ A D D:
Add ( ip - > reg [0] , ip - > reg [1] , ip - > reg [2]);
break ;
case B I N A R Y _ S U B T R A C T:
Sub ( ip - > reg [0] , ip - > reg [1] , ip - > reg [2]);
break ;
}
++ ip ;
}
Figure 7. Switch dispatch
Most compilers will generate an efficient jump-table
based dispatch for this switch statement. The problem with
this style of dispatch is that it will not make effective use
of the branch prediction unit on a CPU. Since every instruction is dispatched from the top of the switch statement, the
CPU is unable to effectively determine which instructions
tend to follow others. This leads to pipeline stalls and poor
performance. Fortunately, there is an easy way to improve
on this.
Token-threading is a technique for improving the performance of switch based interpreters. The basic idea is to “inline” the behavior of our switch statement at the end of every
instruction. This requires a compiler which supports labels
as values [2] (available in most C compilers, with the Microsoft C compiler being a notable exception).
By inlining the jump table lookup we replace the single difficult to predict branch with many, more predictable
j u m p _ t a b l e = { && BINARY_ADD ,
&& BINARY_SUBTRACT , ... };
B I N A R Y _ A D D:
Add ( ip - > reg [0] , ip - > reg [1] , ip - > reg [2]);
goto j u m p _ t a b l e [(++ ip ) - > opcode ];
B I N A R Y _ S U B T R A C T:
Sub ( ip - > reg [0] , ip - > reg [1] , ip - > reg [2]);
goto j u m p _ t a b l e [(++ ip ) - > opcode ];
Figure 8. Token threading
branches. For example, if BINARY_ADD is always followed
by BINARY_SUBSCR in a certain loop, the processor will be
able to accurately predict the branch and avoid stalling.
Token threading was recently added to the Python 3.1
interpreter[3].
We can go one step further, and modify our bytecode to
contain the actual address of the handler for each instruction.
This results in direct-threading [7].
foreach ( instr in function ) {
instr . handler = j u m p _ t a b l e[ ip . opcode ]
}
B I N A R Y _ A D D:
Add ( ip - > reg [0] , ip - > reg [1] , ip - > reg [2]);
goto (++ ip ) - > handler ;
Figure 9. Direct threading
Direct threading increases the size of each instruction
while removing a lookup into the jump table. We implemented both token and direct threading for Falcon. Token
threading seems to provide a modest (∼5%) performance improvement over switch based dispatch.
Tagged Registers
The default Python object format (PyObject) is inefficient.
A simple word-size integer requires 3 words of space, and
must be dereferenced to get the actual value. For numerically
intensive code, the cost of this indirection can dominate the
interpreter runtime.
This overhead can be reduced by using a more efficient
object format. In general, changing the object format would
break compatibility with existing code, but here the Falcon’s
use of registers proves to be very convenient. As long as a
value is in a register, we can store it in whatever format is
most efficient. Only when a value has to be handed to the
Python API (or other external call) do we need to convert it
back to the normal Python object format.
For our tests, we chose a simple tagged integer format.
Integer tagging takes advantage of the fact that object pointers are always aligned in memory to word boundaries; the
least significant 2 bits are always zero. We can therefore use
the least significant bit of a register to indicate whether it is
storing an integer value or a pointer. If it is storing an integer, we shift the register right one bit to obtain the value.
Boland [8] provides detailed descriptions of different tagged
representations and their cost/benefits.
A simplified example of how tagged registers are implemented is show in Figure 10 .
struct Register {
union {
int int_val ;
PyObject * py_val ;
};
bool is_int () { return int_val & 1; }
int as_int () { return int_val >> 1; }
// Mask bottom 2 bits out
PyObject * as_obj () {
return py_val & 0 x f f f f f f f c;
}
};
Figure 10. Tagged register format
Lookup Hints
Attribute lookups (e.g. myobj.foo) are handled by the LOAD_ATTR
instruction and account for a significant portion of the time
spent in the Python interpreter. Just-in-time compilers for
many languages accelerate method lookups using polymorphic inline caching [15] (PIC) and shadow classes. In many
instances, these can replace expensive dictionary (hash-map)
with direct pointer offsets (making them effectively no more
costly than a struct in C). Unfortunately, this technique is
difficult to employ in the context of Falcon for a few reasons:
nately, this would increase the complexity of Falcon by an
order of magnitude; while it may be an appropriate choice
in the future, we were interested in determining whether a
simpler approach might be effective.
Falcon uses a variant on PIC, which we call “lookup
hints”. As the name suggests, these provide a guess for
where an attribute can be found. A hint records the location
where an attribute was found the last time the instruction was
run. A hint can indicate that the attribute was in the instance
dictionary an object, or as an attribute of parent class. When
a hint is found, the location specified by the hint is checked
first before the normal lookup traversal is performed. If the
hint matches, the resulting value is returned immediately;
otherwise the full lookup procedure is performed and a new
hint generated.
The benefit of hints depends greatly on the application being executed. Code that references lots of attributes in a consistent manner can see a large improvement in performance;
in general we observed a ∼5% improvement for most of our
benchmarks.
void LoadAttr ( RegOp op , E v a l u a t o r* eval ) {
// Load the hint for this i n s t r u c t i o n.
L o o k u p H i n t h = eval - > hint [ op . hint_pos ];
PyObject * o b j _ k l a s s = g e t _ c l a s s( op . reg [0]);
PyObject * obj_dict = get_dict ( op . reg [0]);
PyObject * key = op . reg [1];
// If we p r e v i o u s l y found our a t t r i b u t e in
// our instance dictionary , look there again .
if ( h . dict == obj_dict &&
h . d i c t _ s i z e == obj_dict - > size &&
h . klass == o b j _ k l a s s &&
obj_dict . keys [ h . offset ] == key ) {
return obj_dict . values [ h . offset ];
}
// no hint , normal path
...
• Fixed object format. Shadow classes require control
over how objects are laid out in memory, which would
break our compatibility goal.
• Complex lookup behavior. Python provides a great deal
of flexibility to application programmers in choosing how
attribute lookup is performed. The builtin behavior for resolving lookups is similar to that found in most languages
(first check the object dictionary, then the class dictionary, then parent classes. . . ). In addition to this, Python
offers an unusual degree of flexibility for programmers
in the form of accessor methods. These allow specifying what happens if an attribute is not found (using the
__getattr__ method) or even completely override the normal lookup process (the __getattribute__ method). What’s
more, these methods can be added to a class after objects have been created, which creates complications for
mechanisms like shadow-classes, which expect lookup
behavior to remain consistent.
• Attribute lookup is hidden by the bytecode. Making
matters worse is that the bytecode generated by Python
does not explicitly check whether methods such as __getattr__ have been defined. Instead, the LOAD_ATTR instruction is expected to implicitly perform checks for the various accessor functions for every lookup.
One general way of handling such complex behavior is to
use a traces within a just-in-time compiler [6]. Unfortu-
}
Figure 11. Simplified implementation lookup hints
5. Implementation
Falcon is implemented in C++ as a Python extension module and is compatible with Python 2.6 through 3.3. Building
on top of the existing Python interpreter means that Falcon
takes a relatively small amount of code to implement; the entire package is only 3000 lines of code, evenly split between
the compiler/optimizer and the interpreter.
The method used to implement the Falcon VM is somewhat unusual. A typical VM interpreter is structured as a
single method containing the code to handle each type of
instruction. The approach we took with Falcon was implemented as a number of C++ classes, one for each Python
opcode. We use the compiler function attributes to force the
inlining of the code for each opcode into the main interpreter
dispatch loop. We have found this technique to be very effective, allowing a clean separation of code without sacrificing
speed.
Like the CPython interpreter, Falcon overlays Python
function calls onto the C execution stack (each Python call
corresponds to a C function call).
Python exceptions are emulated using C++ exceptions.
This allows Falcon to leverage the builtin C++ exception
handling facility, and greatly simplifies the main interpreter
code. (Proper handling of exceptions and return values is
a significant source of complexity for the mainline Python
interpreter).
100%
80%
60%
40%
20%
• Compile times are negligible. All of our benchmark re-
sults include time taken to compile from stack to register
code and run the optimization passes. Despite our best efforts at making an inefficient compiler (copies used many
places where references might suffice, multiple passes
over the code), the time taken to convert and optimize
h
uc
fan
nk
ld
tre
e
ho
ta
thr
es
fas
ou
nt
cry
pto
ult
rdc
wo
60%
40%
20%
h
uc
nk
fan
ld
tre
e
ho
es
thr
ta
fas
nt
cry
pto
ou
rdc
wo
ma
tm
ult
0%
qu
Figure 13 shows the runtime performance of Falcon relative to the runtime of CPython interpeter. The three bars represent the time taken by (1) unoptimized Falcon code using
untagged (ordinary PyObject) registers, (2) optimized code
with untagged registers and (3) optimized code with tagged.
registers. We were surprised by the inconsistency of benefit from using tagged registers. For some benchmarks, such
as matrix multiplication, the performance improvement from
switching to tagged registers was quite dramatic. Most of
the other benchmarks saw either little improvement or even
some slowdown from the switch.
We also looked at the change in the number of instructions used after converting to register code, and after optimizations have been run. (figure 14) As expected, the register code version of each benchmark requires significantly
fewer instructions to express the same computation, using on
average 45% fewer instructions.
A few interesting observations can be made:
tm
so
ma
ick
qu
80%
ick
Figure 12. Benchmark Descriptions
Unoptimized
Optimized
100%
rt
Description
Multiply square matrices
Recursive control flow
# of distinct words
AES encrypt+decrypt
Classic sorting algorithm
Random string generation
# values > threshold
Count permutations
Figure 13. Falcon performance relative to Python
so
We evaluated the runtime performance of Falcon on a variety
of benchmarks. All tests were performed on a machine with
8GB of memory and a Xeon W3520 processor. For most
benchmarks, Falcon provides a small performance benefit
over the CPython interpreter. For benchmarks that are bound
by loop or interpreter dispatch overhead, Falcon is over twice
as fast as CPython.
rt
0%
6. Evaluation
Benchmark
Matrix Multiplication
Decision Tree
Wordcount
Crypto
Quicksort
Fasta
Count threshold
Fannkuch[5]
Unoptimized
Optimized
Optimized+tagged
Figure 14. Effect of compiler optimizations on number of
opcodes, relative to number of Python stack operations
functions is very small, varying from 0.1ms for simple
functions to 1.1ms for the most complex function in our
benchmark set (AES encryption). This implies that is
profitable to simply convert everything to register code,
rather than relying on profile-based techniques (such as
those used by Psyco [17]) to determine whether it is
worthwhile.
• Optimization is important. For some benchmarks, the
compiler optimizations result in a 30% improvement over
the unoptimized code; in some cases changing Falcon
from being slower than CPython to being significantly
faster. The register code is more amenable to optimization, but the register machine instructions are slower and
more expensive to dispatch than simple stack machine
operations.
• Bit tagging registers yields mixed results. Switching
to a more compact and efficient internal representation
seemed like it would be a straightforward win, but this
is not always the case. The potential benefit of using a
tagged inline integer value must be weighed against the
potential cost of converting these integers into Python
objects whenever they must be passed into API functions.
In functions that are dominated by arithmetic and logic
operations, tagged registers are a performance win. In
other functions, however, unpacking an integer value to
be stored directly in a register is simply wasted work.
7. Related Work
Many different projects have sought to speed up the performance of Python programs using a variety of techniques.
Nuitka, Cython [1], and ShedSkin reduce the runtime
overhead of opcode dispatch by statically compiling Python
programs into C API calls. This approach, if combined with
aggressive optimization, can also remove some redundant
runtime checks. The disadvantage of this approach is that
it requires an explicit (and sometimes lengthy) compilation
step, which is at odds with the usual programming style in a
dynamic language like Python.
The currently most popular approach to accelerating dynamic language is tracing just-in-time (JIT) compilation
[13], which has proven particularly effective for JavaScript [14].
One of the primary ways a JIT is able to achieve good performance is by using unboxed representations for data, which
is incompatible with a native API that exposes the internal
representation of data. Unfortunately, this is the case with
Python. The only currently active JIT project for Python
is PyPy [9]. Although PyPy is able to achieve impressive
performance gains, it does so at the expense of breaking C
API compatibility. This is particularly problematic for scientific Python libraries, which act largely as wrappers over
pre-compiled C or Fortran code and are often written with
particular expectations about the Python object layout.
Psyco [17] was an older (now abandoned) just-in-time
compiler for Python. By coupling intimately with the Python
interpreter and switching between efficient (unboxed) representations and externally compatible boxed representations,
Psyco was able to avoid breaking C extensions. Unfortunately, this compatibility required a great deal of conceptual
and implementation complexity, which eventually drove the
developers to abandon the project in favor of PyPy.
8. Conclusion
To investigate how fast we could make a binary compatible
interpreter for Python, we built Falcon, a fast register based
compiler and virtual machine for Python. Falcon combines
many well-known techniques and few new ones in order to
achieve a significant speedup over regular Python.
What did we learn from the experience?
Stack and register bytecodes aren’t too different. Our
register based interpreter proved to be ∼25% faster then the
basic Python stack interpreter for most tasks. While this is
a nice improvement, much larger gains could be made if we
had the ability to change the object format.
Tagged object formats are important. The performance
improvement of using an inline tagged format (integer or
NaN tagging) for primitive types is worth the extra effort;
for any sort of performance sensitive code, it easily means
the difference between an interpreter that is 5 times slower
then C and one that is 100 times slower. If this type of
object format could be used uniformly within the CPython
interpreter, it would greatly improve the performance for
almost every task.
API design. The most important lesson we can draw from
our experience is that interpreter APIs should be designed
with care. In particular, an API which exposes how the internals of an interpreter work may be convenient for gaining a
quick performance boost (i.e. use a macro instead of a function), but in the long-term, exposing these internal surfaces
makes it nearly impossible to change and improve performance in the future. For Python, it is not the size of the C
API that is the problem, but rather its insistence on a particular object format.
The assumption made by an API are not always obvious.
For instance, when writing an API for an interpreter, it may
be tempting to have functions which directly take and return
object pointers. This simple decision has unexpected consequences; it prevents the use of a copying garbage collector.
9. Future Work
One of our goals with Falcon was to build a platform that
would simplify writing new experiments. The use of register
code and a pass based compiler format makes trying out new
optimization techniques on Python bytecode easy. Particular
ideas we would like to explore in the future include:
• Type specialization. At compile time, type propagation
can be performed to determine the types of registers.
Unboxed, type specific bytecode can then be generated
to leverage this information.
• Container specialization. The performance benefit of
tagged registers is primarily limited by the need to convert to and from the CPython object format whenever an
API call is made. This is almost always due to a register being stored into a Python list or dictionary object.
We can improve on this by creating specialized versions
of lists and dictionaries for each primitive type. These
specialized objects would support the standard list/dictionary interface and convert to and from the Python object format on demand (thus allowing them to be used in
external code); internally they would store objects in an
efficient tagged format.
• Improving attribute hints. The current Falcon hinting
mechanism improves performance slightly, but is very
limited in its application. Better results could be obtained
by making lookup more explicit in the bytecode (first
check for accessor functions, then look up the actual
name).
F RANZ , M. Trace-based just-in-time type specialization for
dynamic languages. In PLDI (2009), pp. 465–478.
The source code for Falcon is available online at:
http://github.com/rjpower/falcon/; we encourage
anyone who is interested to try it out and provide feedback.
[15] H ÖLZLE , U., C HAMBERS , C., AND U NGAR , D. Optimizing
dynamically-typed object-oriented languages with polymorphic inline caches. In ECOOP’91 European Conference on
Object-Oriented Programming (1991), Springer, pp. 21–38.
References
[16] O LIPHANT, T. E. Python for scientific computing. Computing in Science & Engineering 9, 3 (2007), 10–20.
[1] Cython: C-Extensions for Python. http://cython.org/.
[2] GCC
documentation:
Labels
as
Values.
gcc.gnu.org/onlinedocs/gcc/Labels-as-Values.html.
Accessed: August 13th 2013.
[3] Python 3 token threading. bugs.python.org/issue4753.
Accessed: August 13th 2013.
[4] Python/C
API
Reference
Manual.
docs.python.org/2/c-api/.
Accessed: August 13th
2013.
[5] A NDERSON , K. R., AND R ETTIG , D. Performing Lisp analysis of the FANNKUCH benchmark.
[6] BALA , V., D UESTERWALD , E., AND BANERJIA , S. Dynamo: a transparent dynamic optimization system. In PLDI
2000 (2000), vol. 35, ACM, pp. 1–12.
[7] B ELL , J. R. Threaded code. Commun. ACM 16, 6 (June
1973), 370–372.
[8] B OLAND , C. Memory Allocation and Access Patterns in
Dynamic Languages. PhD thesis, Heinrich Heine University
Dsseldorf, 2012.
[9] B OLZ , C. F., C UNI , A., F IJALKOWSKI , M., AND R IGO , A.
Tracing the meta-level: PyPy’s tracing JIT compiler. In Proceedings of the 4th workshop on the Implementation, Compilation, Optimization of Object-Oriented Languages and Programming Systems (2009), ACM, pp. 18–25.
[10] C OUSOT, P., AND C OUSOT, R. Abstract interpretation: a
unified lattice model for static analysis of programs by construction or approximation of fixpoints. In Proceedings of the
4th ACM SIGACT-SIGPLAN symposium on Principles of programming languages (1977), ACM, pp. 238–252.
[11] C YTRON , R., F ERRANTE , J., ROSEN , B. K., W EGMAN ,
M. N., AND Z ADECK , F. K. Efficiently computing static
single assignment form and the control dependence graph.
ACM Transactions on Programming Languages and Systems
13, 4 (Oct 1991), 451–490.
[12] DAVIS , B., B EATTY, A., C ASEY, K., G REGG , D., AND
WALDRON , J. The case for virtual register machines. In
In Interpreters, Virtual Machines and Emulators (IVME ’03)
(2003), ACM Press, pp. 41–49.
[13] G AL , A., B EBENITA , M., AND F RANZ , M. One method
at a time is quite a waste of time. In Proceedings of the
Second ECOOP Workshop on Implementation, Compilation,
Optimization of Object-Oriented Languages, Programs and
Systems (2007).
[14] G AL , A., E ICH , B., S HAVER , M., A NDERSON , D., M AN DELIN , D., H AGHIGHAT, M. R., K APLAN , B., H OARE , G.,
Z BARSKY, B., O RENDORFF , J., RUDERMAN , J., S MITH ,
E. W., R EITMAIER , R., B EBENITA , M., C HANG , M., AND
[17] R IGO , A. Representation-based just-in-time specialization
and the Psyco prototype for Python. In Proceedings of
the 2004 ACM SIGPLAN symposium on Partial evaluation
and semantics-based program manipulation (2004), ACM,
pp. 15–26.
[18] S HI , Y., C ASEY, K., E RTL , M. A., AND G REGG , D. Virtual machine showdown: Stack versus registers. ACM Trans.
Archit. Code Optim. 4, 4 (Jan. 2008), 2:1–2:36.
[19] VAN ROSSUM , G.,
1994.
ET AL .
Python programming language,
| 6cs.PL
|
Logic Programming for an Introductory Computer
Science Course for High School Students
Timothy Yuen1, Maritz Reyes2, Yuanlin Zhang3
1
The University of Texas at San Antonio, USA
2 The University of Texas at Austin, USA
3 Texas Tech University, USA
[email protected], [email protected], [email protected]
Abstract. This paper investigates how high school students approach computing
through an introductory computer science course situated in the Logic Programming (LP) paradigm. This study shows how novice students operate within the
LP paradigm while engaging in foundational computing concepts and skills, and
presents a case for LP as a viable paradigm choice for introductory CS courses.
Keywords: CS education, high school CS, declarative programming, logic programming, answer set programming
1
Introduction
The debate in paradigm selection in CS0 and CS1 type courses is often split between
object-oriented (OO) and procedural paradigms, which also leads to discussions on programming language choice. Selection is a difficult task as there are no standard languages or paradigms to use in the field. Though both these approaches have been successful in introductory CS courses [1], some research has also shown only minimal
differences when comparing the outcomes between paradigms [2, 3]. At the same time,
research on the teaching of introductory CS courses also reveals the limitations of those
approaches [4, 5] (e.g., the suitability of Java, C, C++ for education is debatable [4] as
well as the objects-first). However, the Logic Programming (LP) based approach is
largely ignored by the CS0/CS1 community [4] despite continuing efforts in both teaching and research on LP.
1.1
Logic Programming and CS Education
Prolog may be the most well-known instance in which LP was used to teach CS. In
particular, Prolog was used to teach children in the 1980s [6, 7]. As a pioneer, Kowalski
[6] focused on the declarative aspects of Prolog and restricted the use of procedural
aspects of Prolog to a minimum. Later, researchers and practitioners found that the procedural aspects of Prolog have been the main source of misconceptions and difficulties
[8], while the benefits of its declarative aspects were acknowledged. In the last two
2
decades, Prolog has made some appearances: listed in high school curriculum [9], and
taught in gifted and talented high school students [10] and in general high school/undergraduate students [11], though it does not enjoy the same attention as procedural and
OO paradigms.
1.2
Advantages of Logic Programming
From the existing work and practice in LP and general programming [4, 5, 9, 12], the
main advantages of LP as related to computing education are: simple syntax and intuitive/declarative semantics, natural connection to abstraction, logic reasoning and
knowledge representation which form the foundation of computing and other disciplines, involvement of non-trivial, interesting and challenging problems (e.g., Sudoku
problem) early, and the mathematical flavor of LP. LP also provides a rich context for
embedded essential computing concepts and skills. As for industrial relevance, LP and
Declarative Programming (DP), in general, have seen their profound application and
impact in database query languages, problem (formal) specification languages, and domain specification languages including popular web application development languages such as HTML, CSS and XSLT etc. A breakthrough in LP research in the last
two decades is the establishment of the Answer Set Programming (ASP) [13, 14] paradigm which has inherited the declarative nature of Prolog while fully getting rid of its
procedural features. Currently, ASP is a dominating LP formalism in knowledge representation.
1.3
Research Questions
Although there are some movements [e.g., 15] in expanding LP, research on how ASP,
with full declarative nature, facilitates student learning in introductory computer science is still absent. The research question for this study is: How does adopting ASP to
teach an introductory CS course for high school students impact their understanding
of computer science and computing?
2
Methods
This study adopts a qualitative approach in exploring, in-depth, how high students’ understanding of CS and computing through LP, specifically ASP.
2.1
Setting and Participants
Participants were recruited from the TexPREP program at Texas Tech University during Summer 2015. TexPREP is a summer program for 6th-12th graders and offers full
day courses in science, engineering, math, and computer science over several weeks.
Courses are typically taught by local K-12th grade teachers as well as university faculty
members and students. In the case of this study, the class was taught by the third author
with some TAs (the second author was also a TA). TexPREP is a selective program:
3
students had to have an A/B+ average, letters of recommendations, a written, and a
personal interview for consideration. Prospective participants were in their fourth year
of TexPREP taking Computer Science 4 (CS4). They have already taken its CS2
(Scratch [17]) and CS3 (Alice [18]) courses in the CS sequence. The TexPREP CS4
course teaches the foundations of computing using ASP [14]. This course emphasized
problem solving and knowledge representation through the ASP paradigm. The course
met Monday through Friday from 1:00pm to 1:50pm.
Informed consent was obtained from both the participants and their parents prior to
the start of class. There were sixteen (N=16) participants for the study. There were
seven females and nine males. Participants were students at high school in the Texas
Tech area in Lubbock, TX. In terms of rising grades, there were eleven 10th graders,
three 11th graders, and two 12th graders.
2.2
Data Collection and Procedures
Surveys. Participants responded to surveys with open-ended questions at the beginning
(pre) and end (post) of the course regarding their experiences with computing and declarative programming. The purpose of the surveys was to provide some background
information on participants' CS knowledge and to triangulate the findings from the clinical interviews. The survey refers to declarative programming (DP), rather than LP or
ASP even though this study focused on the latter two paradigms. The rationale for referring to the more general DP is because: ASP belongs to the LP paradigm and LP is
one paradigm of DP and knowing other DP paradigms (such as functional programming) may give an edge to the students in learning ASP.
The pre-survey asked about participants' general knowledge of computer science and
declarative programming. The pre-survey questions were:
• What is computer science?
• Describe what you know about declarative programming.
The post-survey asked about participants' knowledge of computer science and the
topics of this course. The post-survey questions specifically asked what participants
learned in this course and how they understood the concepts; that is, rather than asking
participants directly to explain declarative programming, these questions asked how
they perceived declarative programming.
• What is computer science?
• What have you learned, if anything?
• After having taken this course do you feel that you understood the subject of
this course and the tasks that were asked of you? Why or why not?
Clinical Interviews. Clinical interviews are task-based activities in which the researcher attempts to explore subjects’ understanding and cognitive process through active questioning and probing [19]. During the last week of class, participants were asked
to work out one class-based problem while thinking aloud. Since this was a class assignment during a limited time with few researchers, participants only worked on part
of the assignment. Clinical interviews were conducted one-on-one by the research team
and the participant. Researchers asked probing questions in order to prompt participants
4
to explain and clarify their thought processes. All researchers received the same training
from the first and third author on conducting the clinical interviews. Clinical interviews
lasted 30 minutes or when the participant completed their tasks—whichever came first.
Clinical interviews were video recorded and transcribed.
In the interview task, each student was asked to write a SPARC program to represent
the knowledge about family relationships. Some relations, such as father(X, Y), denoting X is the father of Y, and mother(X, Y) are given. A program stub including the
declarations of these relations is also given. The students are asked to represent the
following knowledge: 1) Jon is the father of Matthew, and 2) about the following relations: grandparent(X, Y), son(X, Y), aunt(X, Y), and descendant(X, Y). SPARC is a programming language and system instance of ASP paradigm [20]. It offers a type system
to overcome some challenging syntax restrictions in existing ASP systems, such as
DLV [21] or Clingo [22] and help discover some programming errors early.
The results of the surveys and interviews are presented in the following sections.
3
Survey Results
Survey responses were transcribed and were analyzed by the researchers using an inductive coding approach. The main unit of analysis was a sentence. Longer sentences
with representing multiple ideas were sometimes separated for further coding. Similarly, some consecutive sentences were grouped together for analysis when they continued the same line of thought. Through this analysis, researchers found common
themes across participants. Responses were grouped by common themes, which are
presented below. Participants were given pseudonyms in the presentation of the results
in the form of P#.
3.1
What LP Students Know about Computer Science
Since this course was the third in the pre-engineering program sequence, participants
already had two courses in computing with other science and engineering courses. Most
participants’ definition of computer science in the pre-survey centered on the creation
of a program that does some task with the programming language being the means
through which those tasks are completed [e.g., The science behind how a computer
functions and performs tasks (P1), Computer science is using different computer “languages” to get computers to perform tasks (P13), and Understanding and being able to
reproduce the code and programs that make a computer do its task (P4)]. There was a
strong connection between computer science and programming with ten participants
mentioning programming or the creation of software in their responses. Some participants acknowledged the importance of computers and technologies in the advancement
of society [e.g., The study on how to develop computers to help the world be an easier
and more efficient place (P10); and It has become extremely prominent in our daily
lives and must always be improved for technology to come (P11)].
At the end of the course, participants were asked the same question. In the post-survey, only three participants mentioned programming in their definition of computer
5
science. The participants’ responses could be categorized as defining computer science
as problem solving [e.g., Computer science is a study that involves mathematics and
helps us understand the way a computer thinks (P3); Computer science is the study on
how to make a computer work/think like a “human” (P10); It is how we get computers
to solve problems and how we can use them for simplifying, numbers, solutions, or
problems themselves (P11)] and the study of technologies [e.g., The science behind a
computer and how it functions (P1); Computer science is the study of how and why
computers work (P16)]. Thus, by the end of the course, the participants’ definition of
CS had broadened a bit away from just programming; rather, it become more holistic
and more on emphasizing problem solving in order to complete tasks.
3.2
What LP Students Know about Declarative Programming
In the pre-survey, most participants did not know what DP was; however, many were
familiar with the more “traditional” procedural programming and/or had prior programming experience. Eight participants said that they did not know what declarative programming was. Some participants offered an explanation of what declarative programming could be [e.g., I would guess that it is a branch of computer science that has to
do with making statements and receiving some form of output from the statements
(P12); or I believe that you give a description to of an object then the program runs it
to find the item similar to what was described (P8)] or explanation that they have done
programming before [e.g., In school, I have learned a little about Java and how Java
is a declarative language, Other than that, I don’t know much else about declarative
programming (P3)]. In such explanations, participants were attempting to make connections to previous programming experiences, even though those were incorrect.
The remaining five participants did not explicitly say they did not know what DP
was. Rather, they offered some incorrect or incomplete definitions: Declarative programming is stating a statement and being able to program that statement. Giving the
computer a “command” and programming it to accomplish that particular command
(P5); It is a program where it does the things you type out (P10); Declarative programming is more of an elaboration off of algorithmic programming as it is developed for a
broader use compared to Algorithmic programming. It is much more elaborates and is
well suited for situations with similarities (P12); and It is the type of programming
needs to be defined before any output action can be achieved (P15). Participant 16 was
the closest to having a correct answer: I know that declarative programming works like
telling a computer rules for what something is then having the computer rules for what
something is then having the computer go find all the somethings [sic] in a set. Thus,
most participants did not have experience with DP though have had some experience
in other paradigms, which included object-oriented, procedural, and imperative.
3.3
What Students Have Learned in this LP Course
The data showed that most students reported learning how to program with the SPARC
language within the context of their lab assignments. Though most did not explicitly
articulate their understanding of LP, as they did with procedural-type languages from
6
prior CS experiences, they were able to report how they solved problems using ASP
methodologies and tools. Although the participants emphasized learning how to program with SPARC, eleven responses implied that the programming language provided
the structure that required students to think about their problems through a deeper, more
thorough perspective, which lends itself to declarative programming. P12’s response
was representative of how learning the programming language helped guide declarative
understanding: I’ve learned how to reevaluate my thinking process and consider the
basis of my thought (so I can program). I’ve learned how to ‘translate’ – by writing my
goals/rules/knowledge in English, then transferring and modifying it to fit SPARC code.
Similar responses include: I learned how to look at a problem in a different way (P2);
I have learned how to sort objects, define relationships with predicates, use multiple
variables, use multiple predicates/ rules, use tex*** rules with variables, and use if
when for rules to help simplify solutions (P11); I have learned a new coding language
to go along with the others I know, which is very valuable to me. I have also learned a
new way to look [at] relations between objects (P15); and New syntax - A new way to
specify relations (P9). These examples show that the rules and structure (i.e., syntax
and semantics) that SPARC required and the associated methodologies helped guide
participants in their problem solving. Four participants emphasized learning about the
syntax and coding style. P8 also mentioned that this was the first course in which she
had to type out her programming rather than through a drag-and-drop interface.
3.4
How Students Understood the LP Course Topics and Tasks
Eight participants stated that they understood the subject of this course within the confines of the course activities and three were indifferent. They understood that there was
more to learn with respect to the programming language [e.g., I think if given more time
and work on the computer, I could use this language in the future (P3); The prompt and
things that were asked were clear and easy to understand. They were easy to pick up
on and gave a general understanding – making it relatively easy to complete a given
task (P5); and I understand the subject because it was easy. I’m pretty sure that if you
wanted to program more complex things, it would be harder (P13).
Two participants stated they were able to understand the subject of the course, but
were unsure how it was applicable in a real world or everyday setting: I understand
how, but not why. So, I don’t understand the purpose of this type of program and how
it will be beneficial to the world, other than artificial intelligence, but I feel as though
this program requires more to get to that point (P9) and I do feel I understood the
subject of the course as I did actually learn something. However, I did not always understand exactly what I was supposed to do and I was occasionally completely lost (P4).
Therefore, most participants knew that there was more to learn about LP beyond the
contexts with which they were presented. The two participants (P10 and P14) that stated
they did not understand the course topics mainly due to time management in the course.
P10 felt that the course was too repetitive with the simple topics and was not challenging enough. Similarly, P14 felt the course was also repetitive with instructor teaching
topics already found in their notes.
7
4
Interview Results
A grounded theory methodology [16] was adopted to generate a theory that explained
how students understood and approached computing through ASP activities based on
the interview data. This qualitative approach generates a theory or explanation through
systematic analysis of data [16]. Grounded theory goes beyond describing and categorizing by finding the processes and relationships in the phenomenon being studied [16].
This approach was conducted on the clinical interview data and served as the main
analysis for this study. The unit of analysis was an utterance, which consisted of a complete thought or phrases spoken by the participant. In cases where utterances were rich
with many potential units, they were broken down into smaller utterances. Utterances
made by the researcher provided context to those made by the participant, but were not
analyzed. There were 1,452 utterances used for analysis.
The first author analyzed the data through an open-coding process [16] coding each
utterance with a descriptive label. This process required several iterations to ensure that
coding was consistent across the entire dataset. Then, the third author coded the data
using the set of labels created by the first author. Through that process, the researchers
debated the labels and coding of the utterances, which led to several iterations of recoding the data and revision of the initial labels. Memos were kept during this analysis
process regarding the rationale behind labels and the formation of coding [16]. Since
the third author was the expert in the field of ASP and DP and taught the course, his
coding was used for the remaining analysis.
The next step was creating categories that represented larger constructs that was happening the data. Axial coding analyzed the way these categories were related based on
their properties [16] and formed larger categories. In the case of this study, many of the
merged labels were found to relate to each other as well with infrequent labels. The five
major categories were: abstraction, representation, reasoning, revision, and procedural.
The first author led the first stage of axial coding which led to these five major categories. Then, these were also debated and discussed by the researchers. The surveys, particularly the post-surveys, were used to guide the axial coding and the creation of the
categories. The definitions of the categories also went through several revisions. The
next sections define each category and give example quotes.
4.1
Abstraction
This category describes instances where the participants understand the problem space
in more general or abstract terms and are thinking on a higher level or in real world
terms and within the context of the code they are writing. Participants are generally
describing relationships with respect to these real-world terms and/or integrating their
prior knowledge of the problem space.
Example quotes included: (P8) The descendant part because the descendants can be
anything that's…that was before that person, well before X in this case.; (P10) So, father of person and person, ok, parent, sister, same parent, gender; (P12) I am thinking
of like a family tree I guess sort of and so I am thinking of X being or I am thinking of
Y being so I am thinking of Y being a person, whose descendants we are looking at…
8
4.2
Representation.
This category refers to the process in which the participants are trying to understand the
problem space in coding terms; that is, how their abstraction or understanding of the
problem translates into the code. This category includes references to both the conceptual and coding as well as the transition process between the two.
Example quotes included: (P1) You can’t say, X is the parent of Y if X, if Y is the son,
because it’s flipped. I hope I am explaining it right. So then I’m saying X is the son of
Y if Y is the parent of X, and I’m putting gender because if you are a son you have to
be male. And if X is a male. Yes, Ok. All right I am going to move on to aunt.; (P14) A
descendant is like it’s the same as a child it’s umm like the child of two parents. That
is their descendant that’s their heir so that is what I am trying to say but I don’t know
how to say it.; (P15) I put X is the son of Y. I put that in the predicates. This is it. The
son is a person. And I know that we have the example here. I can put Matthew is the
son of Jon…
4.3
Reasoning
This category represents the problem solving process. It describes instances in which
participants were actively thinking about the problem, apply problem-solving strategies
associated with ASP, and adapting existing code to solve the current problem. Reasoning was selected for the category name as that the most frequent related label.
Example quotes included: (P2) So grandchild Isaac of George, grandchild Joseph,
George, grandchild Susie, George. Grandchild, those are the children, so, grandchild,
grandchild, X is a grandchild of Y, if Y is grandfather, if Y is the grandfather of X and
X is the, do we define gender? Yes, gender of X is male, because grandfather has to be
a male; (P5) I have like separate ones like I have sister and I have like aunt and I have
parent; (P11) Nephew, so then you have Isaac, and actually you can create a rule...
4.4
Revision
Revision encapsulates a lot of the overall debugging process, but emphasizes the importance of participants asking questions in SPARC to see if their code was correct.
This category includes the participants' process of questioning to see if they completed
the tasks correctly, which was done through querying to see if their solutions were right.
Example quotes included: (P1): So I’ll write aunt and then I’ll write Rose comma =
Susie, and then I’ll have to add a question mark, and then have to press execute; (P14):
Ok and now I need to write queries to see if my program works; (P12): I am going to
ask who are descendants of George because George is like the patriarch of the family…
4.5
Procedural
This category represents a large part of the data in which participants were either restating or interpreting facts or instructions. In some cases, participants were asking if
9
they have the correct interpretation of the assignment instructions. In this category, participants frequently mentioned specific rules and strategies associated with declarative
programing. Often, these utterances were co-labeled with other more descriptive categories such as representation and reasoning.
Example quotes included: (P1): Whenever I am talking about the son I forgot to include that it has- and he has to be male, and that X has to be male.; (P8): So now I am
at son. Oh forgot to add the periods. Ok. Got it, got it, and got it. Now I need it aunt.;
(P16): Jon is the father of Matthew, so that rule that I wrote works.
4.6
Order of Category Occurrences
As part of axial coding, there was an analysis of when these categories occurred during
the clinical interview to learn more about the categories themselves and how they related to the other categories. A color-coded graph plotted out the occurrence of each
category in the order they appear
with each participant. Plots varied in length due to the variance
in number of coded utterances.
Figure 1 shows an example of
this graph. When an utterance
had more than one category associated with it, plots are stacked.
Fig. 1. Sample of Order of Category Occurrence
A visual analysis of all participants’ graphs was conducted to
describe when each category occurred and the relationships between those occurrences.
Across all participants, interviewers had them all read the instructions once at the beginning, but these utterances were not coded. Since participants were asked to work on
their assignments, some participants had already started on them prior to the interview.
Thus, those graphs immediately started with non-procedural labels. The visual analysis
found the following:
Revision. Ten participants were engaged in revision during the last half the interview,
which was expected as they spent the first half identifying the objects, relationships,
knowledge, and coding before testing. P10 did not get as far during the interview to be
engaged in revision. Other participants had already started their assignments before the
interview, so they were quicker to be in revisions.
Abstraction. Evidence of abstraction mostly came from the beginning to the middle
for most participants. For several participants, it occurred alongside procedural (7 participants) as well as with representation (6 participants). Three participants had abstraction after revision, which may suggest that the questioning process may support part of
abstraction abilities. Another three participants showed no evidence of abstraction during the interview. However, the interviewer reported that P2 did not get a chance to test
10
their program and P3 had technical issues with the laptop. P13 was reported to feel
confident about what they were doing, and started revision early.
Reasoning. Reasoning came throughout interview for most participants, similar to the
procedural utterances, which may suggest that it is an important recurring process. Reasoning occurred with and between instances of revision (P1, P16) and representation
(P2, P13), abstraction (P14, P15). Thus, reasoning is also intertwined with other parts
of the ASP process. There was no evidence reasoning from P6; however, there were
not many coded utterances from that participant in general.
Representation. Similar to reasoning and procedural, representation also happened
throughout the interview. For the most part, representation started before clusters of
revision for 11 participants. However, for three of these participants, there were more
revision afterward. For P13, representation was intertwined into revision throughout
the interview. Most of P14's interview consisted of utterances labeled as representation
as he was mostly listing out what had to be represented in the knowledge base.
Procedural. All participants had instances of procedural utterances and distributed
throughout the interview as well as simultaneously with other codes. An interesting
observation was that there were procedural moments throughout the interview, but not
just at the beginning. That is, participants kept going back to the instructions and/or the
facts presented in the assignment.
Although the procedural category did not seem to provide much insight into how the
participants approached computing and computer science, at first. After all, these utterances were of participants reading and re-reading the rules and asking if they understood the instructions and information correctly. Procedural utterances were made
throughout the entire overall computing process. In selective coding where analysis
examines a category that could be connected to the other categories [16], procedural
was found as that central theme. Identifying objects and relations in the problem,
through English definitions, by carefully reading the problem description and integrating one's common sense knowledge was a major component of the explicit methodology for solving LP problems. Thus, there was deep interaction between the procedural
and other processes. Students had also mentioned in their post surveys that they had
learned this methodology to solve the problems. Although that may not be surprising,
it showed that they were able to adopt those strategies, which in turn, led to more reasoning and representation processes.
11
4.7
Model of How Participants Approach LP
Based on the grounded theory analysis, supported by the visual analysis
of ordering of categories, model was
constructed that may explain how
participants approach computing
through LP as based on the data (see
Figure 2).
This figure highlights the founFig. 2. Model of How Participants Approach LP
dational importance of the following the strategies of declarative programming. In the case of Answer Set Programming, it provided an explicit methodology, which guided students towards abstracting
the concepts, representing it in code, and reasoning in their problem solving. Abstraction, representation, and reasoning often happened together and are grouped separately
from revision, which mostly came toward the end of the task. Similarly, abstraction and
representation happened more sequentially with abstraction coming before representation. Reasoning was observed throughout the interview, which is indicated by a larger
rectangle.
The model showed the importance of procedures; in this case, students were able to
follow the explicit methodology in LP. The students also demonstrated the general
problem solving skill of iterative refinement. There are two distinguishable big loop
components: knowledge specification and revision. Knowledge specification is again
an iterative refinement using interleaving steps: abstraction, representation and reasoning. As an example, when students try to specify their knowledge about aunt, they try
to figure out the objects (individual persons) and relations (e.g., who is whose sister, or
who is whose parent) and then try to define their knowledge about aunt using these
objects and relations. Participants tried to make sure the definition captures their intended meaning, which requires them to reason with and understand the relations better.
In fact, this model reflects the intended skills most researchers would like the students
to obtain in an introductory course: the skills should be general but important in computing (and beyond). The skills include the procedure (a general methodology for problem solving), iterative refinement in problem solving, programming (revision in our
category name), abstraction, rigorous high level (logic) representation of knowledge
and (logic based) reasoning with the knowledge.
LP can be thought of as introducing an explicit problem solving methodology: problem understanding and precise representation and reasoning with knowledge. This
methodology is almost universal in problem solving. It will also particularly be useful
as LP lays the foundation for problem understanding and knowledge representation.
5
Discussions
The students had done very well in their lab assignments and final interview tasks,
which, together with the pre and post surveys, which shows that LP is easy to learn [8].
12
The main reason is the simple syntax and intuitive/natural semantics of ASP, which
allow the students to focus less on language specific feature and more on the problem
solving skills. The clinical interview data shows further that the students were able to
learn and apply intensively the important concepts underlying computing. The model
(Figure 2) shows that they were able to apply explicit LP methodologies in problem
solving and iterative refinement in both big steps (knowledge specification and revision) and small steps (abstraction, representation and reasoning), when solving the
problems. Participants applied abstraction, representation and reasoning heavily, in an
interwoven manner, in understanding, extracting and defining the knowledge needed in
solving a given problem. Once they identified/defined the knowledge needed, they are
able to carry out the standard programming tasks (labeled as “revision” in our categories) from coding to debugging the program. The model generated showed that students
were able to engage in computing skills, such as abstraction, problem solving (representation), and debugging (revision). Indeed, they were able to explain computing concepts within the context of this course and tasks. Students were able to operate within
the LP paradigm driven by the explicit methodology for LP; however, they may not
immediately see how it is applicable outside the context of the assignments. They understood the LP concepts and processes and were successful in completing their assignments. More work needs to be done to make connections to how they see CS in the real
world through LP. Students understood the problem solving nature of CS and believed
LP was one way of solving problems.
5.1
Limitations
Participants were above average students and were pre-screened for this program, and
had some experience with Scratch and Alice. However, participants stated no experience with LP or DP. Though every student initially volunteered for this study, there
were still some participants who stopped attending or did not complete an assignment.
Lastly, this course met for four weeks, which provided limited exposure to LP as compared to regular K-12 or university courses. Future studies on LP in introductory CS
courses could be conducted on full semester courses rather than short courses.
6
Conclusions
This paper asserts that CS educators should take a closer look at using LP during the
introductory courses. As suggested in [12], future computer scientists should be
equipped with foundational programming language principles involving logic and formal specification to design and implement complex software systems needed by the
society. Our results show that the students were able to focus on the key concepts of
computing including abstraction, representation and reasoning when solving problems.
The simple syntax and intuitive semantics of ASP allows them to put less attention on
the language specific features. These findings support LP as a viable option to teach an
introductory CS course using ASP.
13
Acknowledgments. The authors acknowledge Cynthia Perez and Rocky Upchurch
for their contributions to this project, and thank Michael Gelfond for sharing his
teaching materials. This work is partially supported by NSF grant CNS-1359359.
References
1. S. Draft, "Computer Science Curricula 2013," ed, 2012.
2. T.-Y. Chen, A. Monge, and B. Simon, "Relationship of early programming language to
novice generated design," ACM SIGCSE Bulletin, vol. 38, pp. 495-499, 2006.
3. T. Vilner, E. Zur, and J. Gal-Ezer, "Fundamental concepts of CS1: Procedural vs. object
oriented paradigm-a case study," in ACM SIGCSE Bulletin, 2007, pp. 171-175.
4. A. Pears, S. Seidman, L. Malmi, L. Mannila, E. Adams, J. Bennedsen, et al., "A survey of
literature on the teaching of introductory programming," ACM SIGCSE Bulletin, vol. 39, pp.
204-223, 2007.
5. M. M. Chakravarty and G. Keller, "The risks and benefits of teaching purely functional
programming in first year," Journal of Functional Programming, vol. 14, pp. 113-123, 2004.
6. B. Kowalski, Logic as a computer language for children: Wiley-Interscience, 1987.
7. J. Nichol, J. Dean, and J. Briggs, Prolog, children and students: Nichols Publishing Co.,
1988.
8. P. Mendelsohn, T. Green, and P. Brna, "Programming languages in education: The search
for an easy start," Psychology of programming, pp. 175-200, 1990.
9. Z. Scherz and B. Haberman, "Logic programming based curriculum for high school
students: the use of abstract data types," in ACM SIGCSE Bulletin, 1995, pp. 331-335.
10. J. Stutterheim, W. Swierstra, and D. Swierstra, "Forty hours of declarative programming:
Teaching Prolog at the Junior College Utrecht," arXiv preprint arXiv:1301.5077, 2013.
11. S. Beux, D. Briola, A. Corradi, G. Delzanno, A. Ferrando, F. Frassetto, et al.,
"Computational Thinking for Beginners: A Successful Experience using Prolog."
12. T. Ball and B. Zorn, "Teach foundational language principles," Communications of the
ACM, vol. 58, pp. 30-31, 2015.
13. R. Kowalski, "Logic Programming," in Computational Logic. vol. 9, J. Siekmann, Ed., ed:
Elsevier, 2014.
14. M. Gelfond and Y. Kahl, Knowledge representation, reasoning, and the design of intelligent
agents: The answer-set programming approach: Cambridge University Press, 2014.
15. A. Dovier, P. Benoli, M. C. Brocato, L. Dereani, and F. Tabacco, "Reasoning in High
Schools: do it with ASP!," in 31st Italian Conference on Computational Logic, 2016.
16. A. Strauss and J. Corbin, Basics of Qualitative Research: Techniques and Procedures for
Developing Grounded Theory. Thousand Oaks, CA: SAGE Publications, Inc., 1998.
17. MIT Media Lab, "Scratch," ed.
18. Carnegie Mellon University, "Alice," ed.
19. H. P. Ginsburg, Entering the Child's Mind: The Clinical Interview in Psychological
Research and Practice. New York, Ny: Cambridge University Press, 1997.
20. E. Balai, M. Gelfond, and Y. Zhang, "Towards answer set programming with sorts," in
International Conference on Logic Programming and Nonmonotonic Reasoning, 2013, pp.
135-147.
21. M. Alviano, W. Faber, N. Leone, S. Perri, G. Pfeifer, and G. Terracina, "The disjunctive
datalog system DLV," in Datalog Reloaded, ed: Springer, 2011, pp. 282-301.
22. M. Gebser, B. Kaufmann, R. Kaminski, M. Ostrowski, T. Schaub, and M. Schneider,
"Potassco: The Potsdam answer set solving collection," AI Communications, vol. 24, pp.
107-124, 2011.
| 2cs.AI
|
arXiv:1701.01393v5 [math.AC] 24 Aug 2017
On the n-th row of the graded Betti table of an
n-dimensional toric variety
Alexander Lemmens
Abstract
We prove an explicit formula for the first non-zero entry in the n-th
row of the graded Betti table of an n-dimensional projective toric variety
associated to a normal polytope with at least one interior lattice point.
This applies to Veronese embeddings of Pn . We also prove an explicit
formula for the entire n-th row when the interior of the polytope is onedimensional. All results are valid over an arbitrary field k.
Index terms— syzygies, toric varieties, lattice polytopes, Koszul cohomology
Contents
1 Introduction
1
2 Toric varieties and graded Betti tables
4
3 Combinatorial proof
8
1
Introduction
Let k be a field. In this article we study syzygies of projectively embedded
toric varieties X/k. More precisely, we give explicit formulas in terms of the
combinatorics of the defining polytope for certain graded Betti numbers, which
appear in the minimal free resolution of the homogeneous coordinate ring of
X as a graded module, obtained by repeatedly taking syzygies. These Betti
numbers are typically gathered in the graded Betti table:
0
1
2
3
..
.
0
1
1
0
0 κ1,1
0 κ1,2
0 κ1,3
..
..
.
.
2
0
3
0
4
0
κ2,1
κ2,2
κ2,3
..
.
κ3,1
κ3,2
κ3,3
..
.
κ4,1
κ4,2
κ4,3
..
.
...
...
...
...
...
Here κp,q is the number of degree p + q summands of the p-th module in
the resolution. One alternatively defines κp,q as the dimension of the Koszul
1 The
final
publication
is
http://dx.doi.org/10.1007/s10801-017-0786-y
available
1
at
springer
via
cohomology space Kp,q (X, O(1)). The graded Betti table is expected to contain
a wealth of geometric information and is the subject of several important open
problems and conjectures. But the vast part of it is poorly understood.
For a number of entries an explicit formula in terms of the defining lattice
polytope was known before. Examples of this can be found in [3, 11]. But for
this paper the most relevant result is that of Schenck, who proved [16] that for
projective toric surfaces coming from a lattice polygon with b lattice boundary
points κp,2 = 0 for all p ≤ b − 3. Hering proved in [11, Theorem IV.20] using
a theorem of Gallego-Purnaprajna [9, Theorem 1.3] that the next entry κb−2,2
is not zero. Both results were already known in the case this polygon is equal
to the triangle with vertices (0, d), (d, 0), (0, 0). This polygon gives the d-fold
Veronese embedding of the projective plane, for which Loose [12] proved that
the number of zeroes in the quadratic strand equals 3d−3 (not counting κ0,2 = 0
as a zero). This result was independently rediscovered by Ottaviani and Paoletti
[14], and they generalized this to the following conjecture:
Conjecture 1. For the d-fold Veronese embedding of n-dimensional projective
space κp,q = 0 whenever p ≤ 3d − 3 and q ≥ 2.
This is known as property Np with p = 3d − 3. For d > n this is generalized
by the following conjecture [7, p. 643, Conjecture 7.6] which the authors already
proved for q = n:
Conjecture 2. If we take a minimal free resolution of the line bundle OPn (b)
on the Veronese embedding of Pn of degree d with d ≥ b + n + 1 then κp,q = 0
for all 1 ≤ q ≤ n and
d+q
d−b−1
p<
−
− q.
q
q
Syzygies of Veronese embeddings are still an active area of research [2, 6, 10,
15, 13]. For a short introduction to syzygies and to toric varieties we refer the
reader to the next section.
We will not prove this conjecture, but we will prove an explicit formula for
κ(d+n)−(d−b−1)−n,n which is the first non-zero entry on the n-th row. We also
n
n
prove a formula for the first non-zero entry in the n-th row of the Betti table
of any projectively normal toric variety of dimension n, if this row is not zero.
Note that the n-th row is the last non-zero row if it is not zero. We will work
over an arbitrary field k. For a convex lattice polytope ∆ we denote by ∆(1)
the convex hull of the lattice points in the topological interior of ∆.
∆(1)
We are now ready to formulate our main result:
2
Theorem 3. Let X be a toric variety coming from an n-dimensional normal
polytope ∆ ⊆ Rn . Let ∆(1) be the interior polytope of ∆. Let S = ∆ ∩ Zn ,
T = ∆(1) ∩ Zn and N = #S, N (1) = #T .
• If p < N − N (1) − n then κp,n = 0.
(1)
−2
• If ∆(1) is not 1-dimensional then κN −N (1) −n,n = t+N
where t is the
N (1) −1
number of translations of T that are contained in S. (If N (1) = 0 then
κN −N (1) −n,n = 0.)
N −ℓ
• If ∆(1) is 1-dimensional then ∀p ≥ 0: κp+N −N (1) −n,n = (p + 1) N (1)
−p−1
where ℓ is the number of lines parallel to ∆(1) that are not disjoint with
S.
The first statement actually follows from Green’s linear syzygy theorem [8,
Theorem 7.1] combined with Koszul duality. The second statement already appeared for n = 2 as a conjectural formula in [3], where more information on
Koszul cohomology of toric surfaces can be found. Recall that κp,q = 0 whenever q > n, and note that if ∆(1) = ∅ then κp,q = 0 whenever q ≥ n as follows
from [11, Proposition IV.5 p. 17-18].
Theorem 4. In the context of conjecture 2 the first non-zero entry on the n-th
row equals
b+2n+1
+ d−b−1
−2
n
n
κ(d+n)−(d−b−1)−n,n =
.
d−b−1
n
d
−1
n
These two theorems will be proved at the end of section 2 using results from
section 3.
Corollary 5. For toric surfaces coming from polygons of lattice width two, we
know the entire Betti table explicitly.
N −3
N −1
N −3
κp,2 = max(p−N +N (1) +3, 0)
, κp,1 = κp−1,2 +p
−2A
p
p+1
p−1
where A = N/2 + N (1) /2 − 1 is the area of ∆. Of course κ0,0 = 1 and everything
else is zero.
The second formula comes from [3, lemma 1.3], the first follows directly from
our theorem 3.
Using [4, Theorem 1.3] one can deduce the following formula for the graded
Betti table of the canonical model of a tetragonal curve in a toric surface:
X
2
g−3
g−3
max(p − bi − 1, 0)
,
κg−p−2,1 = κp,2 = (g − p − 2)
+
p
p−2
i=1
where b1 , b2 are the tetragonal invariants introduced by Schreyer in [17, p. 127],
and g is the genus. Actually this formula is true for all tetragonal curves as
follows from the explicit minimal free graded resolution in Schreyer’s article.
3
We include this explicit formula because it is not easy to find in the literature.
In section 2 we explain toric varieties, syzygies, Koszul cohomology and we
prove these theorems using results from section 3. We use Koszul duality [1, p.
21] which expresses Betti numbers on the n-th row in terms of Betti numbers
on the first row of the Betti table of the Serre dual line bundle.
The core of the article is section 3 where we construct an explicit basis for
the last non-zero entry
L on the first row of the graded Betti table of any graded
module of the form q≥0 H 0 (qL + M ) for line bunldes L, M with H 0 (M ) = 0,
H 0 (L) 6= 0, H 0 (L + M ) 6= 0 on any normal projective toric variety. This comes
down to constructing a basis of the kernel of the map
p
^
H 0 (L) ⊗ H 0 (L + M ) →
p−1
^
H 0 (L) ⊗ H 0 (2L + M )
with p = dim H 0 (L + M )− 1. Theorems 3 and 4 can then be proved from results
from section 3, namely theorem 9 (which actually also follows from Green’s linear
syzygy theorem [8, Theorem 7.1]), corollary 21 and theorem 22.
Acknowledgements
This article is part of my Ph.D thesis which is funded by the Research Foundation Flanders (FWO). It was my colleagues Wouter Castryck and Filip Cools
who noticed patterns in Betti tables of certain toric surfaces, which motivated
me to find an explicit basis. I am also grateful to the referee for carefully reading
my article and making many useful suggestions. I also want to thank Milena
Hering for bringing the article [7] to my attention.
2
Toric varieties and graded Betti tables
Projectively normal toric varieties
We work over an arbitrary field k. By lattice points we mean points of Zn .
Projective toric varieties are built out of polytopes ∆ ⊆ Rn that are the convex
hull of a finite set of lattice points. This works as follows. Suppose ∆ is ndimensional and let P1 , . . . , PN∆ be a list of all lattice points of ∆, we define an
embedding
φ∆ : (A1 \{0})n → PN∆ −1 : (λ1 , . . . , λn ) 7→
n
Y
i=1
P
λi 1,i : . . . :
n
Y
i=1
PN∆ ,i
λi
,
where Pj,i is the i-th coordinate of Pj . Let X∆ be the closure of the image of
φ∆ . If it happens that a∆ ∩ Zn + b∆ ∩ Zn = (a + b)∆ ∩ Zn for all positive
integers a, b, then the polytope is called normal. In this case the projective toric
variety corresponding to ∆ is just X∆ and it is projectively normal.
Example.
The d-fold Veronese embedding of projective space is given by a polytope of the
following form:
∆ = {(x1 , . . . , xn ) ∈ (R≥0 )n |x1 + . . . + xn ≤ d}
4
This gives the Veronese embedding of Pn into PN −1 where N = #∆ ∩ Zn =
n+d
d . For instance if n = 2 and d = 2 we get the embedding
A1 \{0} × A1 \{0} → P5 : (x, y) 7→ (x2 : xy : x : y 2 : y : 1).
The monomials x2 , xy, x, y 2 , y, 1 correspond to the lattice points of the triangle
∆ with vertices (2, 0), (0, 2), (0, 0).
y2
y
xy
1
x
x2
When taking the Zariski closure of the image this corresponds to the standard
Veronese embedding
P2 → P5 : (x : y : z) 7→ (x2 : xy : xz : y 2 : yz : z 2 ).
If ∆ is not normal then one can still take integer multiples q∆, q ≥ 1, which
will be normal for sufficiently large q. One can then associate to ∆ the toric
variety Xq∆ where q is large enough so that q∆ is normal. This variety does not
depend on q (but its embedding does). However for simplicity we will restrict
to the case when ∆ is normal. The homogeneous coordinate ring of X∆ is given
by
M
M
Vq∆ =
H 0 (X, qL),
q≥0
q≥0
where by Vq∆ we mean the vector space spanned by the monomials (possibly
with negative exponents) xi11 . . . xinn corresponding to lattice points (i1 , . . . , in ) ∈
q∆. By L we mean the very ample line bundle coming from the embedding into
projective space.
Graded Betti tables
Given any projective variety X with homogeneous coordinate ring R =
H 0 (X, qL) we can consider the graded Tor modules
ToriS ∗ H 0 (X,L) (R, k).
L
q≥0
Note that R is a graded module over the symmetric algebra S ∗ H 0 (X, L). These
graded Tor modules can be computed either by taking a graded free resolution
of R (syzygies) or by taking a graded free resolution of k (Koszul cohomology).
We will mainly work with the latter. The graded Betti table is a table of nonnegative integers κp,q , in the p-th column and the q-th row, where p, q ≥ 0. They
are defined as the dimension over k of the degree p+q part of TorpS ∗ H 0 (X,L) (R, k).
In general the table has the following shape:
0
1
2
3
..
.
0
1
1
0
0 κ1,1
0 κ1,2
0 κ1,3
..
..
.
.
2
0
3
0
4
0
κ2,1
κ2,2
κ2,3
..
.
κ3,1
κ3,2
κ3,3
..
.
κ4,1
κ4,2
κ4,3
..
.
5
...
...
...
...
...
Example.
When ∆ is the convex hull of (2, 0), (0, 2), (0, 0) we have the minimal graded
free resolution of R:
d
d
d
d
1
2
3
0
S ∗ V∆∩Z2 −→
F1 −→
F2 −→
0 −→ F3 −→
R.
Here V∆∩Z2 is the vector space spanned by the monomials x2 , xy, x, y 2 , y, 1 corresponding to the lattice points of ∆. The symmetric algebra S ∗ V∆∩Z2 is the
polynomial ring in 6 variables x(2,0) , x(1,1) , x(1,0) , x(0,2) , x(0,1) , x(0,0) and is the
homogeneous coordinate ring of P5 . The image of d0 corresponds to the ideal
cutting out the Veronese surface. This ideal is generated by six elements:
x(2,0) x(0,2) − x2(1,1) ,
x(0,2) x(0,0) − x2(0,1) ,
x(2,0) x(0,0) − x2(1,0) ,
x(2,0) x(0,1) − x(1,0) x(1,1) , x(1,0) x(0,2) − x(0,1) x(1,1) , x(1,1) x(0,0) − x(1,0) x(0,1) .
These constitute a minimal set of generators of the ideal. So F1 is a free graded
module of rank six over the polynomial ring S ∗ V∆∩Z2 where the basis elements
all have degree two and are mapped by d0 to the generators of the ideal. This
makes sure that d0 is a degree-preserving morphism of modules. This means
that κ1,1 = 6.
The image of d1 consists of the relations between these generators, called syzygies. It turns out that there is a minimal generating set of eight syzygies of
degree 3, so that F2 is a rank 8 graded free module where the basis elements
have degree 3. So κ2,1 = 8. It turns out that F3 has rank 3 where the basis
elements have degree 4. This gives the graded Betti table:
0 1
0 1 0
1 0 6
2 0 0
.. .. ..
. . .
2
0
8
0
..
.
3
0
3
0
..
.
4
0
0
0
..
.
...
...
...
...
Note that κ0,0 = 1 because the polynomial ring S ∗ V∆∩Z2 is a free module of
rank one over itself with the monomial 1 as a generator.
Koszul cohomology
Let L, N be line bundles on a complete variety
S ∗ V = S ∗ H 0 (X, L) be
L X. Let
0
0
the symmetric algebra over H (X, L), then
q≥0 H (X, qL + N ) is a graded
module over S ∗ V . We define the Koszul cohomology space Kp,q (X, N, L) as the
homology of the following sequence:
p+1
^
δp+1
V ⊗ H 0 (X, (q − 1)L + N ) −→
δp
−→
p
^
V ⊗ H 0 (X, qL + N )
p−1
^
V ⊗ H 0 (X, (q + 1)L + N )
Pp
where δp (v1 ∧ . . . ∧ vp ⊗ w) = i=1 (−1)i v1 ∧ . . . ∧ vˆi ∧ . . . ∧ vp ⊗ (vi w). The
vˆi indicates that vi is removed from the wedge product. When N = 0 we
write Kp,q (X, L) = Kp,q (X, 0, L). We denote the dimension of Kp,q (X, N, L)
6
(resp. Xp,q (X, L)) by κp,q (X, N, L) (resp. κp,q (X, L)). If L is the very ample
line bundle coming from a projective embedding this agrees with our earlier
definition of κp,q using syzygies.
Example.
For our Veronese example with n = d = 2 we will construct an explicit element
of the cohomology space K2,1 (X, L):
y 2 ∧ xy ⊗ xz − y 2 ∧ yz ⊗ x2 + xy ∧ yz ⊗ yx ∈
2
^
V∆ ⊗ V∆ ,
which is in the kernel of δ2 , so it defines an element of K2,1 (X, L).
We now turn to the proof of theorem 1 and corollary 2. To any n-dimensional
convex lattice polytope ∆ ⊆ Rn one can associate the inner normal fan Σ [5,
p. 321] whose rays ρ are in one-to-one correspondence with the facets of ∆ and
the torus-invariant
prime divisors Dρ . In general for any torus-invariant divisor
P
D = ρ aρ Dρ the vector space H 0 (X, D) has a basis that naturally corresponds
to {P ∈ Zn |∀ρ ∈ Σ(1) : hP, ρi ≥ −aρ } where Σ(1) is the set of rays of the
fan Σ. Multiplication of these global sections corresponds to coordinatewise
addition of lattice points. The divisor whose global sections correspond to ∆
gives the very ample line bundle of our embedding into projective space. Note
that in this setting nothing changes when extending the field k. In the next
proof we assume k algebraically closed. We also use in the following proof
that taking the pull-back of a line bundle through a birational morphism of
projective normal
varieties preserves global
P
P sections. We will also use the fact
that adding ρ∈Σ(1) −Dρ to a divisor ρ aρ Dρ amounts to taking the interior
of the corresponding polytope (∗).
Proof of theorem 3 using results from the next section. This will rely on Koszul
duality which requires smoothness, so letP
X ′ → X be a toric resolution of singularities as in [5, p. 515-519] and let K = ρ∈Σ′ (1) −Dρ be the canonical divisor
of X ′ . Let L be the line bundle on X coming from its projective embedding.
The pull-back of L to X ′ (which we also denote by L) is globally generated on
X ′ and hence nef. By Demazure vanishing [5, p. 410] H i (X ′ , qL) = 0, ∀i, q ≥ 1.
By Koszul duality [1, p. 21] we have
κN −1−n−p,n (X, L) = κN −1−n−p,n (X ′ , L) = κp,1 (X ′ , K, L), ∀p ≥ 0.
The first equality follows because H 0 (X, qL) = H 0 (X ′ , qL), ∀q ≥ 0 as taking
the pull-back of L through X ′ → X preserves global sections. We claim that
κp,1 (X ′ , K, L) is the dimension of the kernel of the following map:
δ:
p
^
VS ⊗ VT →
p−1
^
VS ⊗ V(2∆)(1) ∩Zn
where by VS (resp. VT ) we mean a vector space with S (resp. T ) as a basis.
This is because H 0 (X ′ , K) = 0 and H 0 (X ′ , L+K) corresponds to T = ∆(1) ∩ Zn
Vp−1
VS ⊗VS+T .
which we know by (∗). Note that the image of δ is contained in
Now all the results of the theorem follow from theorem 9, corollary 21 and
theorem 22, except when p = 0 and N (1) = 1, but then δ = 0 and the result is
easy. For the case where ∆(1) is one-dimensional, note that ℓ (as in theorem 3)
equals N − #X, with X as in theorem 22.
7
By the same duality as in the proof of theorem 3 it follows that κp,q = 0
whenever q > n.
Proof of theorem 4. Denote by κp,q (b; d) the dimension of the Kuszul cohomology of the graded module
M
H 0 (OPn (b + id)),
i≥0
Let N =
n+d
n
be the number of lattice points in dΣ where
Σ = {(x1 , . . . , xn ) ∈ R≥0 |x1 + . . . + xn ≤ d}
is the standard simplex of dimension n. As in the proof of theorem 3 we have
duality:
κp,n (b; d) = κN −n−p−1,1 (−b − n − 1; d).
Strictly speaking, we cannot apply theorem [1, p. 21], but the proof obviously
generalizes and the vanishing condition will be satisfied because H 1 , . . . , H n−1
of any line bundle on Pn will vanish. This κN −n−p−1,1 (−b − n − 1; d) is the
dimension of the kernel of
N −n−p−1
^
S⊗T →
N −n−p−2
^
S ⊗ S + T,
where T = (d − b − n − 1)Σ ∩ Zn and S = dΣ ∩ Zn so that X = (b + n + 1)Σ ∩ Zn.
Applying theorem 9 we find that κp,n (b; d) = 0 whenever N − n − p − 1 ≥ #T =
d−b−1
, or equivalently p < n+d
− d−b−1
− n. The formula for κp,n (b; d)
n
n
n
d−b−1
when p = n+d
−
−
n
follows
from
corollary
21.
n
n
3
Combinatorial proof
In this section we do not require our polytopes to be normal. From now on instead of working with spaces of monomials VS , VT etcetera, we replace monomials xi11 . . . xinn by the corresponding lattice point (i1 , . . . , in ). Let S and T ⊆ Zn
with T finite, let p ≥ 0. We abusively write S (resp. T ) for a vector space with
S (resp. T ) as a basis. We are interested in the kernel of the following map:
δ:
p
^
P1 ∧ . . . ∧ Pp ⊗ Q 7−→
S⊗T →
p
X
p−1
^
S ⊗ (S + T ) :
(−1)i P1 ∧ . . . ∧ P̂i ∧ . . . ∧ Pp ⊗ (Q + Pi ),
i=1
where
in Zn .
V−1
of a vector space is zero. By Q+Pi we mean coordinate-wise addition
Definition
6. Let S and T be finite subsets of Zn and p ≥ 0 an integer. If
Vp
S⊗T then x can be uniquely written (up to order) as a linear combination
x∈
(with non-zero coefficients) of expressions of the form P1 ∧ . . . ∧ Pp ⊗ Q with
P1 , . . . , Pp ∈ S, Q ∈ T and P1 > . . . > Pp for some total order on S. We define
the support of x, denoted supp(x), as the convex hull of the set of points Pi
occurring in the wedge part of the terms of x.
8
Definition 7. A lattice pre-order on Zn is a reflexive transitive relation ≤
on Zn such that ∀P1 , P2 ∈ Zn : P1 ≤ P2 or P2 ≤ P1 and ∀P1 , P2 , P3 ∈ Zn : if
P1 ≤ P2 then P1 +P3 ≤ P2 +P3 . We call ≤ a lattice order if it is anti-symmetric.
One way to obtain a lattice pre-order is to take a linear map L : Rn → R
and set P1 ≤ P2 if L(P1 ) ≤ L(P2 ). If the coefficients defining L are linearly
independent over Q then it defines a lattice order. We write P1 < P2 if P1 ≤
P2 and not P2 ≤ P1 , and we write P1 ∼ P2 if P1 ≤ P2 and P2 ≤ P1 . In
the proof of the following lemma we will use the property that for any points
P, PM , Q, Q′ ∈ Zn and any pre-order on Zn , if P ≤ PM and Q ≤ Q′ then either
P + Q < PM + Q′ or PM ∼ P and Q′ ∼ Q.
Lemma 8. Let S and T be finite subsets of Zn , p ≥ 1. Let x ∈ ker δ be non-zero
with δ as above. Let ≤ be a lattice pre-order such that supp(x) has a unique
maximum PM . Let S ′ = supp(x)\{PM } and T ′ the set of non-maximal points
Vp−2 ′
Vp−1 ′
S ⊗(S ′ +T ′) be defined analogously
S ⊗T ′ →
Q ∈ T . Finally, let δ ′ :
′
to δ. Then there exists a non-zero y ∈ ker δ such that
x = PM ∧ y + terms not containing PM in the ∧ part.
Proof. Write
X
λi PM ∧ P2,i ∧ . . . ∧ Pp,i ⊗ Qi + terms not containing PM in the ∧ part,
x=
i
without any redundant terms. Then define
X
λi P2,i ∧ . . . ∧ Pp,i ⊗ Qi .
y=
i
All we have to prove now is that y ∈ ker(δ ′ ). It is clear that supp(y) ∩ Zn ⊆
S ′ . Let us prove that every Qi in this expression is in T ′ . We know that
Qi ∈ T . If it is not in T ′ then it is maximal. But then, applying δ, the term
−P2,i ∧. . .∧Pp,i ⊗ (PM + Qi ) of δ(x) has nothing to cancel against, contradicting
the fact that δ(x) = 0. The reason that it has nothing to cancel against is that
all terms in δ(x) end with ⊗(P + Q) with P ≤ PM and Q ≤ Qi so that
P + Q < PM + Qi unless P ∼ PM , and Q ∼ Qi . As PM is the unique maximum
of supp(x) P ∼ PM implies P = PM . But if Q 6= Qi and P = PM then we
still have P + Q 6= PM + Qi , so that we have only one term of δ(x) ending on
⊗(PM + Qi ), which has nothing to cancel against.
So y is in the domain of δ ′ . We now prove that δ ′ (y) = 0:
0 = δ(x) = −PM ∧ δ ′ (y) + terms not containing PM in the ∧ part.
Because terms of PM ∧ δ ′ (y) cannot cancel against terms without PM in the ∧
part, δ ′ (y) must be zero.
Example.
Our Veronese example of n = d = 2 becomes S = T = {(2, 0), (1, 1), (1, 0),
V2
S⊗S
(0, 2), (0, 1), (0, 0)} and in the new notation the explicit cochain in
becomes
x = (0, 2) ∧ (1, 1) ⊗ (1, 0) − (0, 2) ∧ (0, 1) ⊗ (2, 0) + (1, 1) ∧ (0, 1) ⊗ (1, 1).
9
If we take the pre-order coming from the linear map L(x1 , x2 ) = x2 then PM =
(0, 2) is the unique maximum of supp(x) = {(0, 2), (1, 1), (0, 1)} and the lemma
gives y = (1, 1)⊗(1, 0)−(0, 1)⊗(2, 0). Indeed, x = (0, 2)∧y+(1, 1)∧(0, 1)⊗(1, 1)
and the last term does not have (0, 2) in the wedge part.
x2
2
x2
PM
2
supp(x)
1
1
S
T′
1
2
x1
1
Theorem 9. Notation as above. Suppose #T ≤ p, then δ :
Vp−1
S ⊗ (S + T ) is injective.
2
Vp
x1
S⊗T →
Proof. By induction on p. The case p = 0 is trivial as the domain of δ is zero
(because T = ∅).
Let p ≥ 1, and take a non-zero element x of the kernel. Let ≤ be a lattice order.
Now apply the construction of lemma 8 to obtain S ′ and T ′ and a non-zero
y ∈ ker δ ′ . Since #T ′ < #T , we have #T ′ ≤ p − 1. Applying the induction
hypothesis we get a contradiction.
Note that this also follows fromL
Green’s Linear syzygy theorem [8,L
Theorem 7.1]
applied to the graded module q≥0 qS + T over the graded ring q≥0 qS. We
give this direct proof because we rely on the same technique later.
We now want to construct elements of the kernel of δ when p = #T − 1. To
this end we do the following construction: let X = {P ∈ Zn |P +QT ⊆ S} and
consider the elements of X as variables. To any monomial A = P ∈X P aP of
degree p with variables in X we will associate an element xA of the kernel of δ.
Let P1 , . . . , Pp be a list of points in X such that each point P occurs aP
times in the list. Let Q1 , . . . , Qp+1 be a list of all points of T . Now we define
xA =
Y 1 X
sgn(σ)(P1 + Qσ(1) ) ∧ . . . ∧ (Pp + Qσ(p) ) ⊗ Qσ(p+1) .
aP !
P
σ∈Sp+1
Up to sign, this will be independent of the choice of lists P1 , . . . , Pp and
Q1 , . . . , Qp+1 .
Lemma 10. The xA we just constructed has integer coefficients and is in the
kernel of δ.
Proof. Consider permutations σ, σ ′ ∈ Sp+1 as in the previous definition such
that σ({i|Pi = P }) = σ ′ ({i|Pi = P }) for all P ∈ X. We claim that the terms
in xA corresponding to σ and σ ′ will be equal. This is because in the wedge
product the only thing that changes is the order, and the change in sign caused
by changing the order is compensated by the change in sgn(σ).
′
Now the number of bijections σQ
with the property that σ({i|Pi = P }) =
′
σ ({i|Pi = P }), ∀P is equal to P aP !, hence the expression will have integer coefficients.
We now prove that xA is in the kernel of δ. Obviously the sums Pi + Qσ(i) are
10
all in S. We claim that when applying δ everything cancels. Let C be the set
of all ordered pairs (σ, i) where σ ∈ Sp+1 and i ∈ {1, . . . , p}. Then
Y
aP ! δ(xA )
P
=
X
(−1)i sgn(σ)(P1 + Qσ(1) ) ∧ . . . ∧ (Pi \
+ Qσ(i) ) ∧ . . . ∧ (Pp + Qσ(p) ))
(σ,i)∈C
⊗ (Qσ(p+1) + Qσ(i) + Pi ).
We now partition C into (unordered) pairs: (σ, i) and (σ ′ , i′ ) belong to the same
pair if either they are equal or the following conditions are met:
• i = i′
• σ(j) = σ ′ (j), for all j ∈
/ {i, p + 1}
• σ(i) = σ ′ (p + 1) and σ ′ (i) = σ(p + 1).
These conditions imply that σ ′−1 ◦ σ is a transposition, and hence has sign -1.
One now easily sees that pairs in C yield terms that cancel.
Example.
Suppose S = {(2, 0), (1, 1), (1, 0), (0, 2), (0, 1), (0, 0)} and T = {(1, 0), (0, 1), (0, 0)}
and p = 2. Then X = {(1, 0), (0, 1), (0, 0)}. Take for example the monomial
A = (1, 0)(0, 1). If we take the lists (1, 0), (0, 1) and (1, 0), (0, 1), (0, 0) we get:
xA =(1, 0) ∧ (1, 1) ⊗ (0, 1) − (1, 0) ∧ (0, 2) ⊗ (1, 0) + (2, 0) ∧ (0, 2) ⊗ (0, 0)
−(2, 0) ∧ (0, 1) ⊗ (0, 1) + (1, 1) ∧ (0, 1) ⊗ (1, 0) − (1, 1) ∧ (1, 1) ⊗ (0, 0).
Of course the last term is zero. Note that each term is of the form P ∧ Q ⊗
((2, 2) − P − Q) where P belongs to the lower right blue triangle and Q to the
upper left one.
(0,2)
S
(0,1)
(0,1) (1,1)
T
(0,0) (1,0) (2,0)
(0,0)
(1,0)
Example.
Suppose S = {(0, 0), (1, 0), (2, 0), (3, 0), (0, 1), (1, 1), (2, 1), (3, 1)} and T =
{(0, 0), (1, 0), (0, 1), (1, 1)} so that X = {(0, 0), (1, 0), (2, 0)}, p = 3 and consider
the monomial A = (0, 0)2 (2, 0), then we get
xA = − (0, 0) ∧ (1, 0) ∧ (3, 1) ⊗ (0, 1) + (0, 0) ∧ (1, 0) ∧ (2, 1) ⊗ (1, 1)
− (0, 0) ∧ (1, 1) ∧ (2, 1) ⊗ (1, 0) + (0, 0) ∧ (1, 1) ∧ (3, 0) ⊗ (0, 1)
− (0, 0) ∧ (0, 1) ∧ (3, 0) ⊗ (1, 1) + (0, 0) ∧ (0, 1) ∧ (3, 1) ⊗ (1, 0)
− (1, 0) ∧ (0, 1) ∧ (3, 1) ⊗ (0, 0) + (1, 0) ∧ (0, 1) ∧ (2, 0) ⊗ (1, 1)
− (1, 0) ∧ (1, 1) ∧ (2, 0) ⊗ (0, 1) + (1, 0) ∧ (1, 1) ∧ (2, 1) ⊗ (0, 0)
− (0, 1) ∧ (1, 1) ∧ (3, 0) ⊗ (0, 0) + (0, 1) ∧ (1, 1) ∧ (2, 0) ⊗ (1, 0)
11
In this case the first two wedge factors in each term are from the left blue square
and the third wedge factor is from the right blue square. In the definition there
are 24 terms but each term occurs twice and we divide by two so only twelve
terms are left.
S
(0,1)
(1,1)
(0,1) (1,1) (2,1) (3,1)
T
(0,0) (1,0) (2,0) (3,0)
(0,0)
(1,0)
Q
One way to get rid of the factor P a1P ! is to only sum over one element of
each equivalence class, where two permutations σ, σ ′ are equivalent if σ ′ ({i|Pi =
P }) = σ ′ ({i|Pi = P }) for all P ∈ X. So the construction works over any field.
Proposition 11. The xA for distinct monomials A are linearly independent
and the support of any linear combination of them is the convex hull of the
union of the supp(xA ) occurring with non-zero coefficient.
Proof. By induction on p. In the case p = 0 there is only one monomial namely
the constant monomial 1. The corresponding xA is ∧(∅) ⊗ Q where Q is the
unique
point of T . So the statement is obvious. So suppose p ≥ 1. Let x =
P
λ
x
j
A
j with the Aj distinct. We have to prove that
j
[
supp(xAj ) 6= ∅.
supp(x) = conv
j
To prove this equality, it is enough to prove that every linear map L : Rn → R
attains the same maximum on both sides of the equation. It is enough to show
this when L|Zn is injective (as these L are dense). Given such an L, let ≤ be
the order it induces on Zn . Let QM be the maximum of T for this order and
PM ∈ X the greatest point occurring as a variable in some monomial Aj . We
will prove that PM + QM is the maximum of both sides of the equation, proving
that L attains the same maximum on both, and that both sides are non-empty.
Obviously nothing greater than PM + QM can possibly occur in any supp(xAj ).
We have
X
X
λj xAj .
λj xAj +
x=
j
PM |Aj
j
PM does not occur in Aj
For any Aj containing PM we define Bj as the monomial Aj /PM of degree
p − 1. Using T ′ = T \{QM } we can associate to any Bj an element xBj so
that xAj = ±(PM + QM ) ∧ xBj plus terms where everything in the ∧ part is
smaller than PM + QM . By induction xBj 6= 0 so PM + QM is the maximum of
supp(xAj ). Finally
X
± λj xBj + terms without PM + QM in the ∧ part.
x = (PM + QM ) ∧
j
PM |Aj
By induction the linear combination of the xBj is not zero so PM + QM is the
maximum of supp(x).
12
So far, we have studied the kernel of the map
δ:
p
^
P1 ∧ . . . ∧ Pp ⊗ Q 7−→
S⊗T →
p
X
p−1
^
S ⊗ (S + T ) :
(−1)i P1 ∧ . . . ∧ P̂i ∧ . . . ∧ Pp ⊗ (Q + Pi ).
i=1
We now introduce the following maps
δi : S ⊗p ⊗ T → S ⊗(p−1) ⊗ (S + T ) :
P1 ⊗ . . . ⊗ Pp ⊗ Q 7→ P1 ⊗ . . . ⊗ P̂i ⊗ . . . ⊗ Pp ⊗ (Pi + Q).
This time we look at the intersection
of the kernels of the δi . If p = 0 there is
T
nothing to intersect so we put 0i=1 ker δi = T .
We introduce this new machinery because it helps us prove our main result.
Example. Let S = T = {(1, 1), (1, 0), (0, 1), (0, 0)} then
x =(1, 0) ⊗ (0, 1) ⊗ (0, 0) − (1, 0) ⊗ (0, 0) ⊗ (0, 1)
−(0, 0) ⊗ (0, 1) ⊗ (1, 0) + (0, 0) ⊗ (0, 0) ⊗ (1, 1)
is in ker δ1 ∩ ker δ2 .
Proposition 12. There is an injective map
ι:
p
^
S ⊗ T → S ⊗p ⊗ T : P1 ∧ . . . ∧ Pp ⊗ Q 7→
X
sgn(σ)Pσ(1) ⊗ . . . ⊗ Pσ(p) ⊗ Q
σ∈Sp
and ι(ker δ) ⊆
Tp
i=1
ker δi
Proof. Note that the definition of ι does not depend on any choices. It is injective
as cancellation is impossible. Let us prove the last assertion. We define
g:
p−1
^
S ⊗ (S + T ) → S ⊗(p−1) ⊗ (S + T )
analogously to T
ι. If we can prove that δi ◦ ι = (−1)i g ◦ δ for all i then it follows
that ι(ker δ) ⊆ i ker δi . Let x = P1 ∧ . . . ∧ Pp ⊗ Q, we compute
X
[
δi (ι(x)) =
sgn(σ)Pσ(1) ⊗ . . . ⊗ P
σ(i) ⊗ . . . ⊗ Pσ(p) ⊗ (Pσ(i) + Q)
σ∈Sp
g(δ(x)) =
p
X
X
(−1)j sgn(τ )Pτ ′ (1) ⊗ . . . ⊗ Pτ ′ (p−1) ⊗ (Pj + Q),
j=1 τ ∈Sp−1
where τ ′ = (j j + 1 . . . p) ◦ τ . Here (j j + 1 . . . p) maps every number from j
up to p − 1 to itself plus one and everything smaller than j to itself, and p is
mapped to j. For every τ ∈ Sp−1 we formally put τ (p) = p so that Sp−1 ⊆ Sp .
There is a bijection from Sp to {1, . . . , p} × Sp−1 mapping σ to (j, τ ) where
σ ◦ (i i + 1 . . . p) = (j j + 1 . . . p) ◦ τ.
Note that sgn(σ)(−1)p−i = sgn(τ )(−1)p−j . Using this bijection one sees that
δi (ι(x)) = (−1)i g(δ(x)). This proves that δi ◦ ι = (−1)i g ◦ δ and hence that
ι(ker δ) ⊆ ker δi .
13
For anyT
sequence P1 , . . . , Pp of points in X = {P |P + T ⊆ S} one defines an
p
element of i=1 ker δi :
X
xP1 ,...,Pp :=
sgn(σ)(P1 + Qσ(1) ) ⊗ . . . ⊗ (Pp + Qσ(p) ) ⊗ Qσ(p+1)
σ∈Sp+1
where Q1 , . . . , Qp+1 is a list of all the points of T . Whenever we use this notation
we assume that #T = p + 1. Just as the xT
A are intended to be a basis of ker δ,
p
the xP1 ,...,Pp are intended to be a basis of i=1 ker δi .
Lemma 13. Consider the right group action of Sp on the set of sequences
P1 , . . . , Pp ∈ X by permutation. So P1 , . . . , Pp · σ = Pσ(1) , . . . , Pσ(p) . For a
given such sequence let A be the monomial P1 . . . Pp , then
X
xPσ(1) ,...,Pσ(p) .
ι(xA ) =
σ̄∈Stab(P1 ,...,Pp )\Sp
(We sum over the right cosets of the stabilizer.)
Proof. It is enough to prove this equality in characteristic zero. We have
Y
ι(xA )
aP !
P
X
=ι
sgn(σ)(P1 + Qσ(1) ) ∧ . . . ∧ (Pp + Qσ(p) ) ⊗ Qσ(p+1)
σ∈Sp+1
=
X
X
sgn(σ ◦ τ )(Pτ (1) + Qσ(τ (1)) ) ⊗ . . . ⊗ (Pτ (p) + Qσ(τ (p)) ) ⊗ Qσ(p+1)
τ ∈Sp σ∈Sp+1
=
X
X
sgn(σ ′ )(Pτ (1) + Qσ′ (1) ) ⊗ . . . ⊗ (Pτ (p) + Qσ′ (p) ) ⊗ Qσ′ (p+1)
τ ∈Sp σ′ ∈Sp+1
=
X
xPτ (1) ,...,Pτ (p)
τ ∈Sp
=
Y
P
aP !
X
xPτ (1) ,...,Pτ (p)
τ̄ ∈Stab(P1 ,...,Pp )\Sp
The last equality follows because theQorder of the stabilizer is
result follows by removing the factor P aP !.
Q
P
aP !. The
ι(ker δ) is generated by
Lemma 14. The span of the xP1 ,...,Pp intersected with T
p
the ι(xA ). In particular if the xP1 ,...,Pp are a basis of i=1 ker δi then the xA
are a basis of ker δ.
Proof. We have a right group action of Sp on S ⊗p ⊗ T , restricting to one on
T
p
i=1 ker δi : any σ ∈ Sp maps P1 ⊗ . . . ⊗ Pp ⊗ Q to sgn(σ)Pσ(1) ⊗ . . . ⊗ Pσ(p) ⊗ Q.
Clearly any element of ι(ker δ) is fixed by this action. The action of Sp on the
set of sequences P1 , . . . , Pp in X from the previous lemma is compatible with
the action on S ⊗p ⊗ T in the sense that xP1 ,...,Pp · σ = xPσ(1) ,...,Pσ(p) . Choose an
element Rj = P1,j , . . . , Pp,j out of each orbit of the action on sequences.
Consider an element x of ι(ker δ) that is a linear combination of the xP1 ,...,Pp .
We prove that it is generated by the ι(xA ). Write it as a linear combination of
the xP1 ,...,Pp :
X
X
λj,σ̄ xPσ(1),j ,...,Pσ(p),j .
x=
j
σ̄∈Stab(Rj )\Sp
14
Applying the action on S ⊗p ⊗ T to this expression permutes the λj,σ̄ , for each
j. Since x is fixed by the action of Sp and the xP1 ,...,Pp are linearly independent
(by lemma 18 below), λj,σ̄ doesn’t depend on σ. So x is a linear combination of
the
X
xPσ(1),j ,...,Pσ(p),j = ι(xA ),
σ̄∈Stab(Rj )\Sp
where A = P1,j . . . Pp,j . Note that we used the previous lemma in the last
equality. This proves the first assertion. The second assertion follows from the
first, the injectivity of ι and proposition 11.
Vp
Having established a connection between the
S ⊗T context and the S ⊗p ⊗
T context, we now focus on the latter.
Tp
Definition 15. For any x ∈ i=1 ker δi and i ∈ {1, . . . , p} we define suppi (x)
to be the convex hull of the set of lattice points occurring in the i-th factor of
some term of x.
The following lemma is analogous to lemma 8.
T
Lemma 16. Let x ∈ pi=1 ker δi and let ≤ be a lattice pre-order on Zn . Fix an
i and suppose P ∈ suppi (x) is strictly greater than any other point of suppi (x).
′
Let T ′ be the set of non-maximal points in T . Let δ1′ , . . . , δp−1
: S ⊗(p−1) ⊗ T ′ →
S ⊗(p−2) ⊗ (S + T ′ ) be defined analogously to δ1 , . . . , δp . Write
X
λj P1,j ⊗. . .⊗P (place i)⊗. . .⊗Pp,j ⊗Qj +terms not having P at place i
x=
j
y=
X
λj P1,j ⊗ . . . ⊗ Pp,j ⊗ Qj (with P removed from place i).
j
Then y ∈
Tp−1
j=1
ker δj′ .
We omit the proof since it is analogous to that of lemma 8.
Example. Let S = T = {(1, 1), (1, 0), (0, 1), (0, 0)} then
x =(1, 0) ⊗ (0, 1) ⊗ (0, 0) − (1, 0) ⊗ (0, 0) ⊗ (0, 1)
−(0, 0) ⊗ (0, 1) ⊗ (1, 0) + (0, 0) ⊗ (0, 0) ⊗ (1, 1)
is in ker δ1 ∩ ker δ2 . If we take the order coming from the linear map L(x1 , x2 ) =
x2 then PM = (0, 1) is the unique maximum of supp2 (x) = {(0, 1), (0, 0)}.
Applying the lemma with i = 2 we get y = (1, 0) ⊗ (0, 0) − (0, 0) ⊗ (1, 0).
x2
x2
PM
supp2 (x)
S
supp1 (x)
Lemma 17. If p ≥ #T then
S
x1
x1
Tp
i=1
ker δi = 0.
Again the proof is analogous to that of theorem 9.
15
LemmaP
18. Let δi : S ⊗p ⊗T → S ⊗(p−1) ⊗(S+T ) be the usual maps, p = #T −1.
combination with non-zero coefficients then
Let x = j λj xP1,j ,...,Pp,j be a linear
S
suppi (x) is the convex hull of j suppi (xP1,j ,...,Pp,j ). In particular the xP1 ,...,Pp
are linearly independent.
One can prove this with the same technique as proposition 11.
Lemma 19. Let ∆ be a lattice polytope of dimension at least two with p + 1
lattice points and let δi T
: (Zn )⊗p ⊗ (∆ ∩ Zn ) → (Zn )⊗(p−1) ⊗ Zn be the usual
p
maps. Then for all x ∈ i=1 ker δi \{0} and i ∈ {1, . . . , p} we have
∆ − ∆ ⊆ suppi (x) − suppi (x).
Note that the lemma can fail if ∆ is one-dimensional. If n = 1, ∆ = [0, 2]
and x = v0 ⊗v0 ⊗v2 −v0 ⊗v1 ⊗v1 −v1 ⊗v0 ⊗v1 +v1 ⊗v1 ⊗v0 then supp1 (x) = [0, 1]
and the conclusion of the lemma fails.
Proof. We prove this
Tpby induction on p. We can suppose i = 1 without loss of
generality. Let x ∈ i=1 ker δi \{0} and take v ∈ (∆ − ∆)\(supp1 (x) − supp1 (x))
with integer coordinates. Let P1 , P2 ∈ ∆∩Zn with P2 −P1 = v. Using a unimodular transformation we can suppose P1 = (0, 0, . . . , 0) and P2 = v = (d, 0, . . . , 0)
for some d > 0.
Case 1: ∆\[P1 , P2 ] contains more than one lattice point.
Take a linear map L : Rn → R that does not attain a maximum at P1 or
P2 on ∆. We take L to attain its maximum on supp2 (x) at only one point.
This induces a lattice pre-order ≤.
Tp−1We apply lemma 16 for place 2 to obtain S ′ , T ′ and a non-zero y ∈ j=1 ker(δj′ ) with supp1 (y) ⊆ supp1 (x), so
[P1 , P2 ] − [P1 , P2 ] * supp1 (y) − supp1 (y). If #T ′ = p we get a contradiction
with the induction hypothesis and if #T ′ < p we get a contradiction with lemma
17. We needed the fact that ∆\[P1 , P2 ] contains more than one lattice point
to ensure that T ′ is of dimension at least two, so we can apply the induction
hypothesis.
Case 2: ∆\[P1 , P2 ] contains only one lattice point.
We can suppose this lattice point is (0, −1, 0, . . . , 0). Note that p = #T −
1 = d + 1. Define L1 , L2 : Zn → Z as follows: L1 ((x1 , x2 , . . .)) = −x1 ,
L2 ((x1 , x2 , . . .)) = x1 −dx2 . We claim that L1 (resp. L2 ) attains its maximum on
supp1 (x) at more than one lattice point of supp1 (x). Indeed, applying lemma 16
with L1 (resp. L2 ) and place 1 we get T ′ = {(1, 0, 0, . . . , 0), . . . , (d, 0, 0, . . . , 0)}
(resp. {(0, 0, . . . , 0), . . . , (d − 1, 0, 0, . . . , 0)}). In each case the y we obtain leads
to a contradiction with lemma 17, unless L1 (resp. L2 ) does not attain its maximum on supp1 (x) at a unique point of supp1 (x) (in which case we cannot apply
lemma 16). Therefore L1 (resp. L2 ) attains its maximum at more than one
lattice point of supp1 (x).
Case 2a: n = 2.
Let (x1 , y1 ) and (x1 , y1 + 1) be points of supp1 (x) on which L1 reaches its maximum and let (x2 , y2 ) and (x2 + d, y2 + 1) be points of supp1 (x) on which L2
reaches its maximum. We know by maximality that x2 ≥ x1 and L2 ((x1 , y1 )) ≤
L2 ((x2 , y2 )). If y2 ≤ y1 then [(x2 , y2 + 1), (x2 + d, y2 + 1)] ⊆ supp1 (x) and
similarly if y2 ≥ y1 then [(x1 , y1 + 1), (x1 + d, y1 + 1)] ⊆ supp1 (x). In any case
we get a contradiction with the fact that v = (d, 0) ∈
/ supp1 (x) − supp1 (x).
16
(x1 , y1 + 1)
∇L1
(x1 , y1 )
(x2 + d, y2 + 1)
(x2 , y2 )
∇L2
Case 2b: n ≥ 3.
Let π : Zn → Zn−2 bePthe projection that deletes the first two coordinates. So
π(∆) = 0. Now δ1 = P ∈Zn−2 δP where δP maps Q1 ⊗ . . . ⊗ Qp ⊗ Q to
Q2 ⊗ . . . ⊗ Qp ⊗ (Q + Q1 ) if π(Q1 ) = P and to zero otherwise.
As π(Q + Q1 ) = π(Q) + π(Q1 ) = P , there can’t be any cancellation between
δP (x) for different P . Therefore δP (x) = 0 for all P ∈ Zn−2 .
For any P ∈ Zn−2 and Q0 ∈ π −1 (0) we define a linear automorphism
αP,Q0 : (Zn )⊗p ⊗ T −→ (Zn )⊗p ⊗ T
(
(Q1 + Q0 ) ⊗ . . . ⊗ Qp ⊗ Q
: Q1 ⊗ . . . ⊗ Qp ⊗ Q 7→
Q1 ⊗ . . . ⊗ Qp ⊗ Q
if π(Q1 ) = P
else.
(Recall that by Zn (resp. T ) we mean the vector space with Zn (resp. T ) as a
basis. So we define the linear map on basis elements and linearly extend them
over the base field.) For any P, P ′ ∈ Zn−2 and Q0 we define
α′P,P ′ ,Q0 : (Zn )⊗(p−1) ⊗ T −→ (Zn )⊗(p−1) ⊗ T
(
Q1 ⊗ . . . ⊗ Qp−1 ⊗ (Q + Q0 ) if P = P ′
: Q1 ⊗ . . . ⊗ Qp−1 ⊗ Q 7→
Q1 ⊗ . . . ⊗ Qp−1 ⊗ Q
else.
Then α′P,P ′ ,Q0 ◦ δP ′ = δP ′ ◦ αP,Q0 , from which it follows that αP,Q0 (x) is in
ker(δP ′ ) for all P ′ ∈ Zn−2 and Q0 ∈ π −1 (0). So αP,Q0 (x) ∈ ker δ1 and if we
define
α′′P,Q0 : (Zn )⊗p−1 ⊗ T −→ (Zn )⊗p−1 ⊗ T
(
(Q1 + Q0 ) ⊗ . . . ⊗ Qp−1 ⊗ Q
: Q1 ⊗ . . . ⊗ Qp−1 ⊗ Q 7→
Q1 ⊗ . . . ⊗ Qp−1 ⊗ Q
if π(Q1 ) = P
else,
then α′′P,Q0 ◦ δi = δi ◦ αP,Q0 for all i = 2, . . . , p. Therefore αP,Q0 (x) ∈
Tp
i=1
ker δi .
suppP (αP,Q (x))
0
suppP (x)
supp1 (αP,Q0 (x))
supp1 (x)
For any P ∈ Zn−2 we define suppP (x) as the convex hull of all points
in π −1 (P ) that occur in the first factor of some term of x. Of course this
is at most two-dimensional. If we can prove that whenever this support is
17
non-empty it contains more than 1 point where L1 (resp. L2 ) attains its maximum, then we can perform the same reasoning as in case 2a on suppP (x)
to obtain a contradiction. If we choose a Q0 with L1 (Q0 ) (resp. L2 (Q0 ))
high enough, then L1 (resp. L2 ) will attain its maximum on supp1 (αP,Q0 (x))
only at points of suppP (αP,Q0 (x)). This means that there are at least two
points of suppP (αP,Q0 (x)) where L1 (resp. L2 ) attains its maximum. Since
suppP (αP,Q0 (x)) = Q0 + suppP (x) the same is true for suppP (x), so we are
done.
Theorem 20. If S = ∆′ ∩ Zn , T = ∆ ∩ Zn and p = #T − 1 with ∆, ∆′
convex and ∆ a bounded lattice polytope of dimension greater than one,
Tpthen the
expressions xP1 ,...,Pp with Pi ∈ X := {P |P + T ⊆ S} are a basis of i=1 ker δi
and hence the xA for monomials A of degree p with variables in X are a basis
of ker δ.
Proof. Let H be the set of all bounded convex lattice polytopes in Zn that are
either of dimension greater than one or have just two lattice points. By lemma
14 we only have to prove
Tpthe first statement. In fact we only have to prove that
the xP1 ,...,Pp generate i=1 ker δi by lemma 18. We prove it for all ∆ ∈ H by
induction on p = #∆ ∩ Zn − 1.
Suppose first p = 1, then T has just two points and we have to show that
the kernel of δ1 : S ⊗ T → S + T is generated by expressions of the form
(P + Q1 ) ⊗ Q2 − (P + Q2 ) ⊗ Q1 where T = {Q1 , Q2 } and P ∈ X. Consider
the map f : S × T → S + T of sets given by addition of lattice points, then
every point P ′ of S + T is reached by at most two elements of S × T namely
(P + Q1 , Q2 ) and (P + Q2 , Q1 ) with P = P ′ − Q1 − Q2 . We can write S ⊗ T as
the direct sum of the linear span of each f −1 (P ) with P ∈ S + T . The kernel of
δ1 is the direct sum of the kernels of δ1 restricted to each span of f −1 (P ). The
result easily follows.
Now for the induction step suppose p ≥ 2. Let QM be any extreme point of ∆
such that conv(∆ ∩ Zn \{QM }) ∈ H. Using some unimodular transformations
one can squeeze ∆ into (R≥0 )n in such a way that QM = (xM , 0, . . . , 0) and all
other points of ∆ have first coordinate smaller than xM . We can also make sure
that the smallest first coordinate in ∆ is zero.
(One can do all this as follows: first one chooses a linear form L : Zn → Z
that attains its maximum on ∆ only at QM . One can choose L with integer coefficients with no prime factors that they all share. One then chooses
∼
=
a unimodular transformation U1 : Zn → Zn whose first component is L.
Then U1 (QM ) has its first coordinate greater than that of any other point
of U1 (∆). Next one chooses a unimodular transformation U2 of the form
(x1 , . . . , xn ) 7→ (x1 , x2 − a2 x1 , . . . , xn − an x1 ) with a2 , . . . , an large enough so
that all the other coordinates of U2 (U1 (QM )) are smaller than those of the
other points of U2 (U1 (∆)). Finally one uses a translation to map U2 (U1 (QM ))
to (xM , 0, . . . , 0) where xM is the greatest first coordinate on U2 (U1 (∆)) minus
the smallest.)
Claim: It is enough to prove the statement in the case where ∆′ = Rn .
Suppose it is true for ∆′ = Rn , we prove it for arbitrary ∆′ . If x ∈
Proof.
Tp
ker
δi then it is a linear combination of some xP1 ,...,Pp . By lemma 18 their
i=1
supports are contained in the supports of x, hence in ∆′ .
Henceforth we assume ∆′ = Rn . We put the lexicographical ordering on (Z≥0 )n ,
18
meaning (x1 , . . . , xn ) < (x′1 , . . . , x′n ) if for the smallest i with xi 6= x′i we have
xi < x′i .
Tp
So suppose there exists an x ∈ i=1 ker δi that is not a linear combination of
the xP1 ,...,Pp . We can translate the first factor so that supp1 (x) ⊆ (R≥0 )n . We
take x so that the lexicographic maximum of supp1 (x) is minimal. (We can do
this because there are no lexicographic infinite descents in (Z≥0 )n .) We will
′
find a contradiction. Let PM
be the maximum of supp1 (x) and e its first coorn
dinate. Let Qm ∈ ∆ ∩ Z be some point with first coordinate zero. By lemma
19 QM − Qm ∈ ∆ − ∆ ⊆ supp1 (x) − supp1 (x). It follows that e ≥ xM . So
′
PM := PM
− QM ∈ (R≥0 )n because its first coordinate e − xM is ≥ 0 and all
′
the other coordinates are equal to those of PM
.
x2
x2
′
PM
supp1 (x)
Qm
e
∆
QM
xM
x1
x1
T
n ⊗(p−1)
′
′
⊗
We now apply lemma 16 to x to obtain y ∈ p−1
i=1 ker δi where δi : (Z )
′
n ⊗(p−2)
n
′
T → (Z )
⊗ Z are the usual maps and where T = T \{QM }. This y
satisfies
′
′
x = PM
⊗ y plus terms whose first factor is < PM
.
By induction
y=
X
λj yP1,j ,...,Pp−1,j , for some Pi,j ∈ Zn .
j
Using the fact that xPM ,P1,j ,...,Pp−1,j = (PM + QM ) ⊗ yP1,j ,...,Pp−1,j plus terms
′
′
whose first factor is smaller than PM + QM = PM
we see that PM
is the
′
maximum of supp1 (x ) where
X
λj xPM ,P1,j ,...,Pp−1,j .
x′ =
j
′
In x − x′ the terms with PM
cancel so the maximum of supp1 (x − x′ ) is smaller
′
than PM , contradicting the minimal choice of x. (The fact that PM ∈ (R≥0 )n
is important because it ensures that supp1 (x − x′ ) ⊆ (R≥0 )n .)
The last assertion of the theorem follows from lemma 14.
Corollary 21. Let ∆ and ∆′ be convex lattice polytopes with ∆ of dimension
greater than 1 and T = ∆ ∩ Zn , S = ∆′ ∩ Zn and p = #T − 1. If
δ:
p
^
S⊗T →
p−1
^
S ⊗ (S + T )
is the usual map then the dimension of ker δ is
S}.
p+#X−1
p
where X = {P |P +T ⊆
This follows
because the number of degree p monomials with variables in X
is p+#X−1
.
We
end with the case when ∆ is 1-dimensional. This time the
p
formula works for all p.
19
Theorem 22. Let ∆ = conv((0, 0, . . . , 0), (d, 0, . . . , 0)) with d ≥ 0 and let
∆′ ⊆ Zn a bounded convex lattice polygon, then for all p ≤ d + 1 the dimension of the kernel of the usual map δ is (d − p + 1) #X
where X =
p
n
′
{P |P + {(0, . . . , 0), (1, 0, . . . , 0)} ⊆ S}, S = ∆ ∩ Z .
Proof. Let T = ∆ ∩ Zn . Put a lattice order ≤ on Zn such that (1, 0, . . . , 0) >
(0, . . . , 0). Let I = {(0, 0, . . . , 0), (1, 0, . . . , 0)}. For any P1 < . . . < Pp in X and
Q ∈ T with first coordinate in {p, . . . , d} we define
X
(1)
(−1)i1 +...+ip (P1 + i1 ) ∧ . . . ∧ (Pp + ip ) ⊗ (Q − i1 − . . . − ip ),
i1 ,...,ip ∈I
where we abusively write (−1)i1 +...+ip for the power of −1 whose exponent is
the first coordinate of i1 + . . . + ip . These expressions are in the kernel of δ. We
will prove that they are a basis of the kernel which proves the theorem because
there are exactly (d − p + 1) #X
of these. We will do so by induction on #S.
p
The case where p = 0 is easy as the domain and kernel of δ are both just T
and have the points in T as a basis. So suppose p ≥ 1 and let x ∈ ker δ, we
will show that it is a linear combination of expressions like (1). Let PM be
the maximum of S. If PM ∈
/ supp(x) we apply the induction hypothesis to
S ′ = S\{PM } and we are done. So assume PM ∈ supp(x), then by lemma 8
we can
x = PM ∧ y plus terms not containing PM in the ∧ part. Here
Vp write
S ′ ⊗ T ′ where T ′ = {(0, 0, . . . , 0), . . . , (d − 1, 0, . . . , 0)}. Note also that
y∈
PM − (1, 0, . . . , 0) ∈ S as otherwise the terms in δ(x) where PM is removed from
the ∧ would have nothing to cancel against. This is because these would be the
only terms of δ(x) where the point after the ⊗ agrees with PM in all but the
first coordinate. Applying the induction hypothesis to y we get
X
X
λj
(−1)i1 +...+ip−1
y=
j
i1 ,...,ip−1 ∈I
(P1,j + i1 ) ∧ . . . ∧ (Pp−1,j + ip−1 ) ⊗ (Qj − i1 − . . . − ip−1 ).
Therefore x can be written as (−1)p x′ plus terms not containing PM in the ∧
part where
X
X
λj
(−1)i1 +...+ip
x′ =
j
i1 ,...,ip ∈I
(P1,j + i1 ) ∧ . . . ∧ (Pp−1,j + ip−1 ) ∧ (PM − (1, 0, . . . , 0) + ip )
⊗ (Qj + (1, 0, . . . , 0) − i1 − . . . − ip ).
So we can apply the induction hypothesis to x − (−1)p x′ to conclude that x is
a linear combination of expressions like (1).
Finally, we show linear independence of the expressions, P
again by induction
on #S. the case p = 0 is again trivial, let p ≥ 1. Let
i λi xi be a linear
combination of our expressions that yields zero. Each xi containing PM in its
support can be written as PM ∧yi plus terms not containing PM . Then up to sign
yi is an expression like (1) but with p − 1 instead of p and with the set S\{PM }
in stead of S. By the induction hypothesis the yi are linearly independent.
But then it follows that PM cannot occur at all in the wedge part of any xi ,
otherwise the linear combination could not yield zero. And then one again
applies the induction hypothesis with S\{PM } to obtain a contradiction.
20
References
[1] Marian Aprodu and Jan Nagel, Koszul cohomology and algebraic geometry, University Lecture Series 52, American Mathematical Society, 125 pp.
(2010)
[2] Winfried Bruns, Aldo Conca, Tim Römer, Koszul homology and syzygies of
Veronese subalgebras, Mathematische Annalen 351(4), pp. 761-779 (2011)
[3] Wouter Castryck, Filip Cools, Jeroen Demeyer, Alexander Lemmens, Computing graded Betti tables of toric surfaces, preprint
[4] Wouter Castryck, Filip Cools, Alexander Lemmens, Canonical syzygies of
smooth curves in toric surfaces, preprint
[5] David Cox, John Little and Hal Schenck, Toric varieties, Graduate Studies
in Mathematics 124, American Mathematical Society (2011)
[6] Lawrence Ein, Daniel Erman, Robert Lazarsfeld, A quick proof of nonvanishing for asymptotic syzygies, Algebraic Geometry 3(2), pp. 211-222
(2016)
[7] Lawrence Ein, Robert Lazarsfeld, Asymptotic Syzygies of Algebraic Varieties, Inventiones Mathematicae 190, pp. 603-646 (2012)
[8] David Eisenbud, The geometry of syzygies. A second course in commutative algebra and algebraic geometry, Graduate Texts in Mathematics 229,
Springer-Verlag, New York (2005)
[9] Francisco J. Gallego and Bangere P. Purnaprajna, Some results on rational surfaces and Fano varieties, Journal für die Reine und Angewandte
Mathematik 538, pp. 25-55 (2001)
[10] Ornella Greco and Ivan Martino, Syzygies of the Veronese modules, Communications in Algebra 44(9), pp. 3890-3906
[11] Milena Hering, Syzygies of toric varieties, Ph.D. thesis, University of Michigan (2006)
[12] Frank Loose, On the graded Betti numbers of plane algebraic curves,
Manuscripta Mathematica 64(4), pp. 503-514 (1989)
[13] Elena Rubei, A result on resolutions of Veronese embeddings, Annali
dell’Università di Ferrara Sezione VII. Scienze Matematiche 50, pp. 151-165
(2004)
[14] Giorgio Ottaviani, Raffaela Paoletti, Syzygies of Veronese embeddings,
Compositio Mathematica 125(1), pp. 31-37 (2001)
[15] Euisung Park, On syzygies of Veronese embedding of arbitrary projective
varieties, Journal of Algebra 322(1), pp. 108-121 (2009)
[16] Hal Schenck, Lattice polygons and Green’s theorem, Proceedings of the
American Mathematical Society 132(12), pp. 3509-3512 (2004)
[17] Frank-Olaf Schreyer, Syzygies of canonical curves and special linear series,
Mathematische Annalen 275, pp. 105-138 (1986)
21
| 0math.AC
|
Acting Thoughts: Towards a Mobile Robotic Service Assistant for
Users with Limited Communication Skills
arXiv:1707.06633v3 [cs.AI] 30 Nov 2017
F. Burget*
L.D.J. Fiederer* D. Kuhner* M. Völker* J. Aldinger* R.T. Schirrmeister
J. Boedecker B. Nebel T. Ball W. Burgard
Abstract— As autonomous service robots become more affordable and thus available also for the general public, there
is a growing need for user friendly interfaces to control the
robotic system. Currently available control modalities typically
expect users to be able to express their desire through either
touch, speech or gesture commands. While this requirement
is fulfilled for the majority of users, paralyzed users may not
be able to use such systems. In this paper, we present a novel
framework, that allows these users to interact with a robotic
service assistant in a closed-loop fashion, using only thoughts.
The brain-computer interface (BCI) system is composed of
several interacting components, i.e., non-invasive neuronal signal recording and decoding, high-level task planning, motion
and manipulation planning as well as environment perception.
In various experiments, we demonstrate its applicability and
robustness in real world scenarios, considering fetch-and-carry
tasks and tasks involving human-robot interaction. As our
results demonstrate, our system is capable of adapting to
frequent changes in the environment and reliably completing
given tasks within a reasonable amount of time. Combined with
high-level planning and autonomous robotic systems, interesting
new perspectives open up for non-invasive BCI-based humanrobot interactions.
I. I NTRODUCTION
For patients with heavily impaired communication capabilities, such as severly paralyzed patients, their condition
forces them to constantly rely on the help of human caretakers. Robotic service assistants can re-establish some autonomy for these patients, if they offer adequate interfaces
and possess a sufficient level of intelligence. Generally, an
intelligent system requires adaptive task and motion planning
modules to determine appropriate task plans and motion
trajectories for the robot, that implement a task in the real
world. Moreover, it requires a perception component, e.g.,
to detect objects of interest or to avoid accidental collisions
with obstacles. Typically used interfaces, such as haptic
(buttons), audio (speech) or visual (gesture) interfaces, to
command the robotic system are intuitive and easy options
for healthy users, but difficult to impossible to use for
paralyzed individuals.
In this paper, we present a novel framework, schematically depicted in Fig. 1, that allows closed-loop interaction
between users with minimal communication capabilities and
C. Do
Motion and Manipulation Planning
Motion Execution
High-Level Planner
Knowledge
Base
EEG-based User Interaction
Object Detection
Fig. 1. Framework unifying decoding of neuronal signals, high-level task
planning, low-level motion and manipulation planning, scene perception
with a centralized knowledge base at its core. Intuitive goal selection is
provided through an adaptive graphical user interface.
a robotic service assistant. To do so, we record neuronal
activity elicited in the human brain, the common origin of
all types of communication, with an electroencephalography
(EEG) system. Furthermore, we adopt a convolutional neural
network approach for online decoding of neuronal activity,
in order to allow users to navigate through a graphical user
interface (GUI) provided by a high-level task planner. The
set of feasible actions displayed in the GUI, depends in turn
on the current state of the world, which is stored in a central
knowledge base and continuously updated with information
provided by the robot and a camera perception system. Once
a task has been selected, it is decomposed into a sequence of
atomic actions by the high-level planner. Subsequently, each
action is resolved to a motion for the mobile manipulator using low-level motion and manipulation planning techniques.
In the following, the individual components shown in Fig. 1
will be described in detail, before presenting a quantitative
evaluation of the overall system regarding its performance.
II. R ELATED W ORK
* These authors contributed equally to the work. Authors are with the
Department of Computer Science and Faculty of Medicine, University
of Freiburg, Germany. {burgetf, kuhnerd, do, aldinger, jboedeck, nebel,
burgard}@informatik.uni-freiburg.de and {lukas.fiederer, martin.voelker,
robin.schirrmeister, tonio.ball}@uniklinik-freiburg.de. This research was
supported by the German Research Foundation (DFG, grant number EXC
1086) and grant BMI-Bot by the Baden-Württemberg Stiftung.
Multiple previous studies have focused on robotic systems
assisting people with disabilities. For example, Park et al. [1]
implemented a system for the autonomous feeding of yogurt
to a person. Chung et al. [2] focused on autonomous drinking
which involved locating the drink, picking it up and bringing
it to the person’s mouth. Using a hybrid BCI and head
movement control, Achic et al. [3] studied a setup with a
moving wheelchair and an attached robotic arm. None of
these systems used pure BCI control. In contrast, Wang et
al. [4] used a motor imagery BCI with three classes to
achieve low-level control of a robotic arm. More relevant,
Schröer et al. [5] developed a robotic system which receives
a BCI command from a user and autonomously assists
the user in drinking from a cup. However, this approach
only considers a single object and a fixed-base manipulator.
More recently, Muelling et al. [6] presented a shared-control
approach to assistive robotics, albeit focused on invasive
BCIs. Nonetheless, their approach could be combined with
the high-level planning approach presented in our work.
In these applications, robust decoding of brain signals is
required. Inspired by the successes of deep convolutional
neural networks (ConvNets) in computer vision [7], [8] and
speech recognition [9], [10], deep ConvNets have recently
been applied more frequently to EEG brain-signal decoding.
Deep ConvNets were already applied to decoding tasks
useful for building brain-computer interfaces. Lawhern et
al. [11] used a deep ConvNet to decode P300 oddball signals,
feedback error-related negativity and two movement-related
tasks. When evaluated cross-subject, i.e., trained on some
subjects and evaluated on others, their ConvNet yields competitive accuracies compared with widely-used traditional
brain-signal decoding algorithms. Tabar and Halici [12] used
a ConvNet combined with a convolutional stacked autoencoder to decode motor imagery within-subject, yielding
better accuracies than several non-ConvNet decoding algorithms. Schirrmeister et al. [13] used a shallow and a deep
ConvNet to decode both motor imagery and motor execution
within-subject, reaching or slightly surpassing the accuracies
of the widely used EEG motor-decoding algorithm filter bank
common spatial patterns [14]. Bashivan et al. [15] used a
ConvNet trained on fourier-transformed inputs to estimate
mental workload. In addition to the work on evaluating
ConvNet decoding accuracies, ConvNet visualization methods allow us to get a sense of what brain-signal features
the network is using [13], [15], [16]. Taken together, these
advances make deep ConvNets a viable alternative for brainsignal decoding in brain-computer interfaces. Still, to our
knowledge, online control with deep ConvNets has not yet
been reported for an EEG-based brain-computer interface.
III. O NLINE D ECODING OF N EURONAL S IGNALS
The system at hand is developed to control more complex scenarios than the ones considered in previous work.
Particularly, we consider scenarios involving manipulation
of objects as well as human-robot interaction. Feasible goals
are determined by our GUI which is controlled by directional
commands. As reliable classification of brain signals into
navigation directions cannot yet be achieved directly with
non-invasive BCIs, we used a deep ConvNet approach for
decoding of multiple mental tasks from EEG (Schirrmeister
et al. [13]). This approach introduces a hybrid network,
combining a deep ConvNet with a shallower ConvNet ar-
chitecture. The deep part consists of 4 convolution-pooling
blocks using exponential linear units (ELU) [17] and max
pooling, whereas the shallow part uses a single convolutionpooling block with squaring non-linearities and mean pooling. Both parts use a final convolution with ELU to produce
output features. These features are then concatenated and
fed to a final classification layer. We trained the ConvNet
to decode five mental tasks: right hand finger and both feet
toe movements, object rotation, word generation and rest.
These mental tasks evoke discernible brain patterns and are
used as surrogate signals to control the GUI. Offline training
was done with a cropped training strategy using shifted time
windows within the trials as input data [13].
From our experience it is important to train the BCI
decoder and subjects in an environment that is as close
as possible to the real application environment to avoid
pronounced performance drops. Therefore, we designed a
gradual training paradigm within the high-level planner GUI
where the displayed environment, timing and actions are
identical to those of the real control task. The training
paradigm proceeds as follows: We first train each subject
offline using simulated feedback. Subjects are aware of not
being in control of the GUI. The mental tasks are cued
using grayscale images presented for 0.5 s in the center
of the display. At all times a fixation circle is displayed
at the center of the GUI and the subject is instructed to
fixate on it to minimize eye movements. After a random
time interval of 1-7 s the fixation circle is switched to a
disk for 0.2 s, which indicates the end of the mental task.
At the same time the GUI action (go up, go down, select,
go back, do nothing) corresponding to the cued mental task
is performed to update the GUI. To keep training realistic
we include a 20 % error rate, i.e., on average every fifth
action is erroneous. We instruct the subjects to count the
error occurrences to assert their vigilancy. This offline data
is used to train the individual deep ConvNets. Then, the
subjects do online training by performing the decoded mental
tasks in the GUI. Finally, we stop cueing the mental tasks.
To evaluate the performance of the BCI control, we let
the subjects create instructed high-level plans in the GUI.
These tasks are then executed by a simulated robot or the
real mobile manipulator, when available. To provide more
control over the mobile manipulator and enhance the feeling
of agency, subjects have to confirm the execution of every
planned action and can interrupt the chain of actions at any
moment during their execution. BCI decoding accuracies
for the label-less instructed tasks are assessed by manually
rating each decoding based on the instructed task steps.
Statistical significance of the decoding accuracies were tested
using a conventional permutation test with 100 k random
permutations of the labels (i.e., p-value is the fraction of
label permutations that would have led to better or equal
accuracies than the accuracy for the original labels).
IV. H IGH -L EVEL G OAL F ORMULATION P LANNING
We use domain independent planning to derive the required steps for reaching a desired high-level goal in a
complex task. The user can formulate a high-level goal
without knowledge of the internal representation of objects
in the planning system and the exact capabilities of the robot.
This is achieved by an intuitive graphical user interface,
where the object parameters of the goal are specified by
incrementally refining the objects by referring to their type,
e.g., “cup” or attributes, e.g., “content = apple-juice”.
Domain independent planning identifies a sequence of
actions that transforms the current world state into a state
satisfying a goal condition. A planning task consists of: (i)
a planning domain describing static components such as the
object type hierarchy and the available actions and (ii) a
problem instance describing the objects present in the world
and their current state, as well as a goal description. While
the current state of the objects can be extracted from the
knowledge base, the goal has to be chosen in the GUI.
A restricted vocabulary is shared between the user and the
planning system. Objects or sets of objects are identified by
creating referring expressions to them composed of shared
references built on this vocabluary [18]. We briefly describe
the relevant aspects of our previous work in this area [19].
In general, a referring expression φ is a logical formula with
a single free variable. φ refers to an object o if φ(o) is valid.
E.g., the reference φ(x) ≡ cup(x) ∧ contains(x, water)
refers to all cups containing water. We restrict ourselves to
references that are simple conjunctions of facts, which is not
only preferable for computational reasons, but also allows us
to incrementally refine references by adding constraints. For
example, adding contains(x, water) to cup(x), restricts the
set of all cups to the set of cups containing water.
We distinguish between three types of fundamental object
references: individual references, typename references and
relational references. Individual references are identified by
name, such as the “omniRob” robot. Typename references
can be identified by the name of their type. While we cannot
refer to the cups in our scenario directly, we can refer to an
unspecific cup. Relational references are encountered when
objects can be referred to via predicates in which they occur
as an argument. The relations in our scenario are object
attributes. For example, the content of the cup is used to
clarify which cup is meant. These object references are
used to create references to goals. We start defining goals
with the action that achieves it as we found that this is
most natural to the user, e.g., put(x, y) ∧ cup(x) ∧ shelf (y).
After the initial selection of a goal type (e.g., drop) it is
necessary to determine the objects for all parameters of the
goal predicate or action. These parameters are refined by
constraining the previous choice until the argument is either
determined uniquely (i.e., it is impossible to constrain the
argument further) or the user declares that any remaining
option is acceptable. We exclude unreachable goals, but we
allow for goals that can only be achieved after a sequence
of preceding actions (e.g., drinking water could require to
fetch a cup, bring it to the patient, fetch a bottle and pour
the water into the cup in order to be executed). The goal
that is determined by the selection process of the GUI is
then passed to a custom domain independent planner.
V. ROBOT M OTION G ENERATION
For generating paths for the mobile base, we apply the
sampling-based planning framework BI2 RRT* [20]. Given a
pair of terminal configurations, it performs a bidirectional
search using uniform sampling in the configuration space
until an initial sub-optimal solution path is found. This path is
subsequently refined for the remaining planning time, adopting an informed sampling strategy, which yields a higher
rate of convergence towards the optimal solution. Execution
of paths is implemented via a closed-loop joint trajectory
tracking algorithm using robot localization feedback.
To realize pick, place, pour and drink motions efficiently,
we adopt a probabilistic roadmap planner approach [21]. The
planner uses a graph of randomly sampled task poses (endeffector poses), which are connected by edges. To find a plan
between two poses, the planner connects both poses with the
roadmap graph and uses the A* algorithm to find an optimal
path between the start and goal pose. The execution of the
plan maps the task space velocity commands to joint velocity
commands by employing a task space motion controller. We
sample random poses around the object to determine grasp
motions. For dropping objects we extract horizontal planes
from the camera’s point cloud and sample poses above those
planes to find a suitable drop location.
VI. I MPLEMENTATION D ETAILS
Implementation of our framework in the real world requires several components, such as neuronal signal decoding,
scene perception, knowledge base operations as well as
symbolic and motion planning, to run in parallel. Therefore, we distributed the computation across a network of
7 computers, communicating among each other via ROS.
The decoding of neuronal signals has four components. EEG
measurements are performed using Waveguard EEG caps
with 64 electrodes and a NeurOne amplifier in AC mode.
Additionally, vertical and horizontal EOGs, EMGs of the
four extremities and ECG’s are recorded. For recording and
online-preprocessing, we used BCI2000 and Matlab. We
then transferred the data to a GPU server where our deep
ConvNet classified the data into 5 classes. The high-level
planner GUI consists of a back- and front-end. The backend of the GUI uses the Fast Downward planner [22] to
iteratively build goal references and to find symbolic plans
for the selected goal. As the planning time is not crucial
for the performance of our system, we used Fast Downward
with a basic configuration in our experiments. The central
knowledge base is implemented as a ROS node, which is
able to store objects with arbitrary attribute information. All
changes in the knowledge base automatically trigger updates
of the front-end, unexpected ones interrupt the current motion
trajectory execution. Finally, we used SimTrack [23] for
object pose detection and tracking.
VII. E XPERIMENTS
To evaluate our framework, we consider the environment
schematically depicted in Fig. 2, containing two shelves and
a table as potential locations for manipulation actions. The
TABLE I
AGGREGATED MEAN±STD RESULTS FOR 52 BCI CONTROL RUNS
(E XP. VII-A), * P - VALUE < 10−6
Runs Accuracy*
#
[%]
Shelf
Shelf
S1
S2
S3
S4
Robot
User
GUI
Table
EEG
Camera
Object Location
Fig. 2.
Experimental environment: Two shelves and a table can be
considered by the robot for performing manipulation actions. Five RGBD
sensors observe the environment. A human operator selects a goal using
EEG control and the high-level planner GUI.
user sits in a wheelchair in front of a screen, displaying
the graphical interface of the high-level planner. The robot
used in the experiments is the omniRob omni-directional
mobile manipulator platform by KUKA Robotics, which is
composed of 10 degrees of freedom (DOF), i.e., 3 DOF
for the mobile base and 7 DOF for the manipulator. Additionally, the Dexterous Hand 2.0 by Schunk is attached
to the manipulator’s flange and used to perform grasping
and manipulation actions. The tasks we considered in our
experiments required the robotic system to autonomously
perform the following actions: drive from one location to
another, pick up an object, drop an object (on a shelf or
table), pour liquid from a bottle into a cup, supply a user with
a drink. Moreover, we use a perception system composed of
five RGBD cameras. Three of them are statically mounted at
the shelves and the table, in order to observe the scene and
to report captured information to the knowledge base. The
other two cameras are carried by the robot on-board. The
first one is located at the mobile base and used to perform
collision checks in manipulation planning. The second camera is mounted at the robot’s end-effector and used for tasks
involving physical human-robot interaction. A demonstration
of our framework can be found in the accompanying video:
http://www.informatik.uni-freiburg.de/~burgetf/ecmr17/.
A. Online Decoding of Neuronal Signals
We evaluated the BCI control setup with four healthy subjects (S1-4, all right-handed, three females, aged 26.75±5.9).
At the time of writing the validation, S4 was still in
progress and no validation with the mobile manipulator
was performed. In total, 52 runs have been recorded (20
with the real robot) where the subjects executed various
instructed high-level plans. For 32 runs, we used simulated
feedback from the GUI in order to generate a significant
amount of data for the evaluation. The performance of the
BCI decoding during these runs was assessed using video
Time
[s]
Steps
#
Path Optimality Time/Step
[%]
[s]
18
14
17
3
84.1±6.1
76.8±14.1
82.0±7.4
63.8±15.6
125±84
150±32
200±159
176±102
13.0±7.8
10.1±2.8
17.6±11.4
26.3±11.2
70.1±22.3
91.3±12.0
65.7±28.9
34.5±1.2
9±2
9±3
11±4
6±2
52
76.7±9.1
148±50
16.7±7.1
65.4±23.4
9±2
recordings of interactions with the GUI. We rated GUI
actions as correct if they correspond to the instructed path
and incorrect otherwise. Actions which are necessary to
remediate a previous error are interpreted as correct if the
correction is intentionally clear. Finally, we rated rest actions
as correct during the (simulated) robot executions, incorrect
if the next robotic action had to be initialized and ignored
them during high-level plan creation. For evaluation, five
metrics have been extracted from the video recordings: (i)
the accuracy of the control, (ii) the time it took the subjects
to execute a high-level plan, (iii) the number of steps used
to execute a high-level plan, (iv) the path optimality, i.e., the
ratio of the steps used to the minimally possible number of
steps, and (v) the average time per step. We summarized the
results in Table I. In total, a 76.67 % correct BCI control
was achieved, which required 9 s per step. Selecting a plan
using the GUI took on average 148 s and required the user
to perform on average 16.74 steps in the GUI of the highlevel planner. The path formed by these steps is on average
34.6 % away from the optimal path. The decoding accuracy
of every subject is significantly above chance (p < 10−6 ).
The subject-averaged EEG data used to train the hybrid
ConvNets and the decoding results of the train/test transfer
are visualized in Fig. 3. In Fig. 3(a) we show the signal-tonoise ratio (SNR) of all 5 classes C of the labeled datasets.
We define the SNR for a given frequency f , time t and
electrode e as
SNRf,t,e =
IQR ({median (Mi )})
median ({IQR (Mi )})
i ∈ C,
where Mi corresponds to the set of values at position (f, t, e)
of the i-th task, with |Mi | being the number of repetitions.
median(·) and IQR(·) is the median and interquartile range
(IQR), respectively. The upper part describes the variance of
the class medians, i.e., a large variance means more distinguishable class clusters and a higher SNR. The denominator
describes the variance of values in each class, i.e., a lower
variance of values results in a higher SNR. The low SNR in
EMG channels shows that the subjects did not move during
the tasks.
The decoding accuracies achieved on the test data after
initial training of the ConvNets are visualized in Fig. 3(b). To
further support the neural origin of the BCI control signals,
Fig. 3(c) shows physiologically plausible input-perturbation
network-prediction correlation results (see [13] for methods).
TABLE II
AGGREGATED RESULTS FOR 10 RUNS (E XP. VII-B)
(a)
Actions
SNR
0.7
# Executions
(# Scheduled)
Success
Executions [%]
Grasp
Drop
Approach
10 (10)
9 (10)
19 (20)
90.0
89.0
100.00
37.56
34.13
33.05
4.62
5.75
18.48
Total
38 (40)
94.74
34.42
14.02
TABLE III
AGGREGATED RESULTS FOR 10 RUNS (E XP. VII-C)
0
Actions
(b)
Grasp
Drop
Approach
Pour
Drink
(c)
Hand (R)
Feet
Predictions
Runtime [s]
Mean
Std
Hand (R)
Rotation
Feet
Words
Rotation
Rest
Words
Sensitivity
Targets
Rest
Correlation x10-2
-0.4 0 0.4
Fig. 3. EEG data and decoding results. (a) SNR of the first 4 s of data
used to train the hybrid ConvNet. Highest SNR can be observed in the
alpha (7-14 Hz) and lower beta (16-26 Hz) bands. These frequency bands
are robust markers of task related mental tasks. Note that the non-EEG
channels (top row) were withheld from the ConvNets at any time and
are displayed as negative control. Not all channels are displayed because
of space constraints. (b) Confusion matrix of decoding accuracies for the
train/test transfer. Accuracies are well above the theoretical chance level of
20 %. (c) Topographically plausible input-perturbation network-prediction
correlation maps in the alpha (7-13 Hz) frequency band. For details on the
visualization technique we refer the reader to [13].
Total
# Executions
(# Scheduled)
34
30
80
10
13
(30)
(30)
(80)
(10)
(10)
167 (160)
Success
Executions [%]
Runtime [s]
Mean
Std
91.0
97.0
100.0
100.0
77.0
40.42
37.59
20.91
62.90
57.10
10.31
4.83
7.68
7.19
8.20
95.86
32.46
15.51
might differ for two reasons. A number of executed calls,
lower than the scheduled ones, indicates that a previous
action step has failed to succeed and plan recovery was not
possible. On the other hand, a higher number of executed
calls indicates that the user was able to achieve plan recovery
by commanding a repetition of the failed action. Moreover,
we recorded the largest standard deviation for the approach
action, which can be attributed to the diverse complexity of
the planning problem for the mobile base and the distance to
travel between the selected grasp and drop location. In total,
our system achieved a success rate of 80% for the entire task.
Planning and execution required on average 140.63±36.7 s.
Errors were mainly caused by object detection issues, i.e.,
the system was not able to detect the object or the detection
was not precise enough to be able to successfully grasp or
drop an object.
C. Drinking Task
B. Fetch and Carry Task
The first experiment, considering the use of the real robot,
evaluates the complete system in fetch-and-carry tasks. The
goal was to transfer an object from one location to another,
e.g., from a shelf to the table, using the robot. To fulfill
such tasks the robot typically needs to execute four subtasks:
approach object location, grasp object, approach other location, drop object. The user was instructed to select a predefined goal using the EEG-controlled high-level planner.
Moreover, we selected a random initial placement for the
objects in each run, in order to cover different environment
states. The experiment was repeated ten times by the user.
Table II shows the averaged results for the experiment.
The second column indicates the overall number of desired
action calls, as scheduled by the high-level planner, as well
as the number of calls actually performed. The third to
fifth columns represent the success rate, mean and standard
deviation for the runtime of actions, respectively. Note,
that the number of scheduled and actually executed actions
The last experiment evaluates the direct interaction between user and robot. Therefore, we implemented an autonomous robotic drinking assistant. Our approach enables
the robot to fill a cup with a liquid, move the robot to the user
and finally provide the drink to the user by execution of the
corresponding drinking motion in front of the user’s mouth.
In addition to the techniques described above, successful
pouring and drinking using a robot requires the detection
of the liquid level in the cup and a reliable detection and
localization of the user’s mouth.
To detect the liquid level while pouring, we follow a
vision-based approach introduced by Do et al. [24]. Given
the camera’s viewing angle and the liquid’s index of refraction, the liquid height is determined from the depth measurement using a relationship based on Snell’s law (see [25] for
more details). Using this knowledge, we first detect the cup,
extract the depth values for the liquid and finally estimate the
real liquid height. The type of liquid and hence the index of
refraction is assumed to be given beforehand. The viewing
angle can be determined from the depth data. A Kalman
filter is then used to track the liquid level and compensate for
noise. Once it is detected that the liquid level has exceeded
a user defined value, a stop signal is sent to terminate the
pouring motion.
For detection and localizing of the user’s mouth, we adopt
a two step approach. In the first step, we segment the image
based on the output of a face detection algorithm in order
to extract the image region containing the user’s mouth and
eyes. Afterwards, we detect the position of the mouth of the
user, considering only the obtained image patch. Regarding
the mouth orientation, we additionally consider the position
of the eyes in order to obtain a more robust estimation of the
face orientation, hence compensating for slightly changing
angles of the head. The face, mouth and eye detectors are
implemented in OpenCV by applying an algorithm that uses
Haar cascades [26], [27].
Table III shows the averaged results for the experiment.
Here, only 3.75% of the 160 scheduled actions had to be
repeated in order to complete the task successfully. In one
run, plan recovery was not possible leading to abortion of
the task. Thus, our system achieved in total a success rate of
90% for the drinking task. Planning and execution required
on average 545.56±67.38 s. For the evaluation of the liquid
level detection approach, we specified a desired fill level and
executed 10 runs of the pour action. The resulting mean error
and standard deviation is 6.9±8.9 mm. In some instances the
bottle obstructed the camera view, resulting in poor liquid
level detection and a higher error.
VIII. C ONCLUSIONS
In this paper, we presented a thought-controlled mobile
robotic service assistant, capable of successfully performing
complex tasks, including close range interaction with the
user, in a continuously changing environment to increase
the independence of severely paralyzed patients. Through the
use of a high-level planner as an intermediate layer between
user and autonomous mobile robotic service assistant, we
overcome the curse of dimensionality typically encountered
in non-invasive BCI control schemes, thus opening up new
perspectives for human-robot interaction scenarios.
R EFERENCES
[1] D. Park, Y. K. Kim, Z. M. Erickson, and C. C. Kemp, “Towards assistive feeding with a general-purpose mobile manipulator,”
arXiv:1605.07996, 2016.
[2] C.-S. Chung, H. Wang, and R. A. Cooper, “Autonomous function of
wheelchair-mounted robotic manipulators to perform daily activities,”
in Rehabilitation Robotics (ICORR), 2013 IEEE International Conference on. IEEE, 2013, pp. 1–6.
[3] F. Achic, J. Montero, C. Penaloza, and F. Cuellar, “Hybrid bci system
to operate an electric wheelchair and a robotic arm for navigation
and manipulation tasks,” in Advanced Robotics and its Social Impacts
(ARSO), 2016 IEEE Workshop on. IEEE, 2016, pp. 249–254.
[4] C. Wang, B. Xia, J. Li, W. Yang, D. Xiao, A. C. Velez, and H. Yang,
“Motor imagery bci-based robot arm system,” in Natural Computation
(ICNC), 2011 Seventh International Conference on, vol. 1. IEEE,
2011, pp. 181–184.
[5] S. Schröer, I. Killmann, B. Frank, M. Völker, L. D. J. Fiederer, T. Ball,
and W. Burgard, “An autonomous robotic assistant for drinking,” in
Robotics and Automation (ICRA), 2015 IEEE International Conference
on. IEEE, 2015, pp. 6482–6487.
[6] K. Muelling, A. Venkatraman, J.-S. Valois, J. E. Downey, J. Weiss,
S. Javdani, M. Hebert, A. B. Schwartz, J. L. Collinger, and J. A.
Bagnell, “Autonomy infused teleoperation with application to brain
computer interface controlled manipulation,” Autonomous Robots, pp.
1–22, 2017.
[7] 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.
[8] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for
Image Recognition,” arXiv:1512.03385, 2015.
[9] T. N. Sainath, B. Kingsbury, G. Saon, H. Soltau, A.-r. Mohamed,
G. Dahl, and B. Ramabhadran, “Deep Convolutional Neural Networks
for Large-scale Speech Tasks,” Neural Networks, vol. 64, pp. 39–48,
2015.
[10] T. Sercu, C. Puhrsch, B. Kingsbury, and Y. LeCun, “Very deep
multilingual convolutional neural networks for LVCSR,” in IEEE
International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2016, pp. 4955–4959.
[11] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P.
Hung, and B. J. Lance, “Eegnet: A compact convolutional network
for eeg-based brain-computer interfaces,” CoRR, vol. abs/1611.08024,
2016.
[12] Y. R. Tabar and U. Halici, “A novel deep learning approach for
classification of EEG motor imagery signals,” Journal of Neural
Engineering, vol. 14, no. 1, p. 016003, 2017.
[13] R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger, M. Tangermann, F. Hutter, W. Burgard, and
T. Ball, “Deep learning with convolutional neural networks for brain
mapping and decoding of movement-related information from the
human eeg,” arXiv:1703.05051, 2017.
[14] K. K. Ang, Z. Y. Chin, H. Zhang, and C. Guan, “Filter Bank
Common Spatial Pattern (FBCSP) in Brain-Computer Interface,” in
IEEE International Joint Conference on Neural Networks (IJCNN),
2008, pp. 2390–2397.
[15] P. Bashivan, I. Rish, M. Yeasin, and N. Codella, “Learning Representations from EEG with Deep Recurrent-Convolutional Neural
Networks,” in arXiv: 1511.06448, 2016.
[16] S. Stober, “Learning Discriminative Features from Electroencephalography Recordings by Encoding Similarity Constraints,” in Bernstein
Conference 2016, 2016.
[17] D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and Accurate
Deep Network Learning by Exponential Linear Units (ELUs),” in
ArXiv e-prints, vol. 1511, 2016, p. arXiv:1511.07289.
[18] R. Dale and E. Reiter, “Computational interpretations of the gricean
maxims in the generation of referring expressions,” Cognitive science,
vol. 19, no. 2, pp. 233–263, 1995.
[19] M. Göbelbecker, “Assisting with Goal Formulation for Domain Independent Planning,” in KI 2015: Advances in Artificial Intelligence.
Springer, 2015, pp. 87–99.
[20] F. Burget, M. Bennewitz, and W. Burgard, “BI2 RRT*: An efficient
sampling-based path planning framework for task-constrained mobile
manipulation,” in IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), Daejeon, Korea, 2016.
[21] L. E. Kavraki and J.-C. Latombe, Probabilistic Roadmaps for Robot
Path Planning. John Wiley, 1998, pp. 33–53.
[22] M. Helmert, “The Fast Downward Planning System,” Journal of
Artificial Intelligence Research 26 (JAIR 2006), pp. 191–246, 2006.
[23] K. Pauwels and D. Kragic, “Simtrack: A simulation-based framework
for scalable real-time object pose detection and tracking,” in IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS).
IEEE, 2015, pp. 1300–1307.
[24] C. Do, T. Schubert, and W. Burgard, “A probabilistic approach to
liquid level detection in cups using an RGB-D camera,” in IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS),
Daejeon, Korea, 2016.
[25] Y. Hara, F. Honda, T. Tsubouchi, and A. Ohya, “Detection of Liquids
in Cups Based on the Refraction of Light with a Depth Camera Using
Triangulation,” in IROS, 2014.
[26] P. Viola and M. Jones, “Rapid object detection using a boosted cascade
of simple features,” in Computer Vision and Pattern Recognition,
2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society
Conference on, vol. 1. IEEE, 2001, pp. I–I.
[27] R. Lienhart and J. Maydt, “An extended set of haar-like features for
rapid object detection,” in Image Processing. 2002. Proceedings. 2002
International Conference on, vol. 1. IEEE, 2002, pp. I–I.
| 2cs.AI
|
arXiv:1707.01415v1 [cs.AI] 5 Jul 2017
Machine Learning, Deepest Learning:
Statistical Data Assimilation Problems
Henry Abarbanel,
Department of Physics
and
Marine Physical Laboratory (Scripps Institution of Oceanography)
Center for Engineered Natural Intelligence
[email protected]
Paul Rozdeba
and
Sasha Shirman
Department of Physics
University of California, San Diego
9500 Gilman Drive, Mail Code 0374
La Jolla, CA 92093-0374, USA
Draft of July 6, 2017
1
1
Abstract
We formulate a strong equivalence between machine learning, artificial intelligence methods and the formulation of statistical data assimilation as
used widely in physical and biological sciences. The correspondence is that
layer number in the artificial network setting is the analog of time in the
data assimilation setting. Within the discussion of this equivalence we show
that adding more layers (making the network deeper) is analogous to adding
temporal resolution in a data assimilation framework.
How one can find a candidate for the global minimum of the cost functions
in the machine learning context using a method from data assimilation is
discussed. Calculations on simple models from each side of the equivalence
are reported.
Also discussed is a framework in which the time or layer label is taken
to be continuous, providing a differential equation, the Euler-Lagrange equation, which shows that the problem being solved is a two point boundary
value problem familiar in the discussion of variational methods. The use of
continuous layers is denoted “deepest learning”. These problems respect a
symplectic symmetry in continuous time/layer phase space. Both Lagrangian
versions and Hamiltonian versions of these problems are presented. Their
well-studied implementation in a discrete time/layer, while respected the
symplectic structure, is addressed. The Hamiltonian version provides a direct rationale for back propagation as a solution method for the canonical
momentum.
2
Introduction
Through the use of enhanced computational capability two, seemingly unrelated, ‘inverse’ problems have flourished over the past decade. One is machine learning in the realm of artificial intelligence [1, 12, 21] with developments that often go under the name “deep learning”. The other is data
assimilation in the physical and life sciences. This describes the transfer of
information from observations to models of the processes producing those
observations [4, 9, 2].
This paper is directed towards demonstrating that these two areas of
investigation are the same at a fundamental level. Each is a statistical physics
problem where methods utilized in one may prove valuable for the other.
2
The main goal of the paper is to point out that many developments in data
assimilation can be utilized in the arena of machine learning. We also suggest
that innovative methods from machine learning may be valuable in data
assimilation.
Two areas of focus are attended to here: (1) a variational annealing (VA)
method for the action (cost function) for machine learning or statistical data
assimilation that permits the location of the apparent global minimum of
that cost function. (2) The notion of analyzing each problem in continuous
time or layer, which we call deepest learning, wherein it is clear that one
is addressing a two point boundary value problem [33, 11] with an underlying
symplectic structure. Methods abound for solving such two point boundary
value problems [27] and for assuring that symplectic structures are respected
when time (or layer) is discretized. These may be quite fruitful in machine
learning.
This paper primarily discusses multilayer perceptrons or feedforward networks [12] though it also makes it is clear that the discussion carries over to
recurrent networks as well [14, 8, 26].
3
3.1
Background
Machine Learning; Standard Feedforward Neural
Nets
We begin with a brief characterization of simple architectures in feedforward
neural networks [1, 12, 21]. The network we describe is composed of an input
layer l0 and output layer lF and hidden layers l1 , l2 , ..., lF − 1. Within each
layer we have N active units, called ‘neurons’, each of which has d degreesof-freedom. (Usually d is chosen to be 1.)
Each layer has N neurons with d degrees of freedom, so each layer has
D = N d degrees-of-freedom. For our purposes the ‘neurons’ in each layer
are the same structure. This can be generalized to different numbers and
different types of neurons in each layer at the cost of a notation explosion.
Data is available to layer l0 and layer lF in M pairs of L-dimensional
input, at l0 , and output, at lF . These are sets of vectors: {yr(k) (l0 ), yr(k) (lF )}
where k = 1, 2, ..., M labels the pairs, r is an index on the L-dimensional
data r = 1, 2, 3, ..., L ≤ D.
The activity of the units in each hidden layer l, x(k)
α (l); l0 < l < lF , is
3
determined by the activity in the previous layer. The index α combines the
neuron number j and the neuron degrees-of-freedom a into one label: α =
1, 2, ..., N d = D; j = 1, 2, ...N ; a = 1, 2, ...d. This connection is described by
the nonlinear function fα (•) via
x(k)
α (l) = fα (x(l − 1), l) = fα
NX
d=D
(k)
Wα β (l)xβ (l − 1) ,
(1)
β=1
(k)
(k)
(k)
where x(k) (l) = {x(k)
α (l)} = {x1 (l), x2 (l), ..., xN d (l)}. The summation over
weights Wα β (l) determines how the activities in layer l − 1 are combined
before allowing fα (•) to act, yielding the activities at layer l. There are
numerous choices for the manner in which the weight functions act as well
as numerous choices for the nonlinear functions, and we direct the reader to
the references for the discussion of the virtues of those choices [1, 12, 21].
At the input and the output layers l0 , lF the network activities are compared to the observations, and the network performance is assessed using an
error metric, often a least squares criterion, or cost function
CM
M
L
1 X
1 X
(k)
(k)
2
(k)
(k)
2
=
Rm (r) [xr (l0 ) − yr (l0 )] + [xr (lF ) − yr (lF )]
, (2)
M k=1 2 L r=1
where Rm (r) permits an inhomogeneous weighting in the comparison of the
network activities and the data. Minimization of this cost function over all
xα(k) (l) and weights Wα β (l), subject to the network model Eq. (1), is used
to determine the weights, the variables x(k)
α (l)in all layers, and any other
parameters appearing in the architecture of the network.
Before moving along to data assimilation, we note that one wishes to
find the global minimum of the cost function Eq. (2), which is a nonlinear
function of the neuron activities, the weights and any other parameters in
the functions at each layer. This is an NP-complete problem [25] and as
such suggests one cannot find a global minimum of the machine learning
problem, as set, unless there is a special circumstance. We will see just such
a circumstance in a data assimilation problem equivalent to machine learning
tasks.
The machine learning problem as described here assumes there is no error
in the model itself, so that the minimization of the cost function Eq. (2)
is subject to strong equality constraints through the model. This results
in the output at layer lF x(k) (lF ) being a very complicated function of the
parameters in the model and the activities at layers x(k) (l ≤ lF ). This is likely
4
connected with the many local minima associated with the NP-complete
nature of the search problem.
We introduce a variational annealing (VA) method in the next section
which regularizes this by moving the equality constraint into the cost function
via a penalty function. This introduces a hyperparameter allowing us to
systematically vary the complexity of the search process.
3.2
Standard Statistical Data Assimilation
Now we describe the formulation of a statistical data assimilation problem.
In data assimilation observations are made of a sparse set of dynamical
variables, associated with a model of the processes producing the observations. This will allow the estimation of parameters in the model and of the
unobserved state variables of the model. The goal is to estimate any unknown parameters in the model, and because not all of the dynamical state
variables in the model may be observed, to estimate those unmeasured state
variables as well. After a certain time window in which information is transferred to the model, we have an estimate of the full model including an initial
condition for all state variables, and predictions are made with the completed
model and compared to new observations. This validation by prediction is
essentially the same as the question of generalization as addressed in machine
learning [12].
In data assimilation, one has a window in time [t0 , tF ] during which observations are made at times t = {τ1 , τ2 , ..., τF } which lie in [t0 ≤ τs ≤ tF ]; s =
1, 2, ..., F . At each observation time L measurements yl (τs ); l = 1, 2, ..., L
are made, L ≤ D. Through knowledge of the measurement instruments the
observations are related to the state variables of the model via ‘measurement
functions’ hl (x) : yl (τk ) = hl (x(τk )); l = 1, 2, ..., L.
Using what knowledge we have of the processes producing the observations, we develop a dynamical model for the state variables. These satisfy a
set of D dynamical differential equations
dxa (t)
= Fa (x(t), t); a = 1, 2, ..., D.
dt
(3)
The time dependence of the vector field F(x, t) may come from external
forcing functions driving the dynamics.
This set of differential equations will necessarily be represented in discrete
time when solving them numerically, resulting in a map xa (tk ) → xa (tk+1 ) =
5
fa (x(tk ), tk ) in which the discrete time vector field fa (•) is related to Fa (x, t)
via the integration procedure one chooses for Eq. (3).
Starting from an initial condition at t0 , xa (t0 ) we use the discrete time
model
xa (tk+1 ) = fa (x(tk ), tk )
(4)
to move forward to the first observation time τ1 , then to τ2 , ... eventually
reaching the end of the observation window at tF . Altogether by making N
model integration steps in each of the intervals [τn , τn+1 ] we make (F + 1)N
time steps.: t0 → τ1 → τ2 ... → τF → tF .
The measurements are noisy, and the models have errors; so this is a statistical problem. Our goal is to construct the conditional probability distribution, P (X|Y), of the model states X(tF ) = {x(t0 ), x(t1 ), ..., x(tN ), ..., x(tF )},
conditioned on the LF measurements Y(τF ) = {y(τ1 ), y(τ2 ), ..., y(τk ), ..., y(τF )}.
Assuming the transition to the model state at time tk+1 depends only
on the model state at time tk (that is, the dynamics in Eq. (4) is Markov)
and using identities on conditional probabilities [2], one may write the action
A(X) = − log[P (X|Y)] (suppressing the dependence on the observations Y
in A(X)) as
A(X) = −
F
X
CM I[X(τn ), y(τn )|Y(τn−1 )]
n=1
N (F +1)−1
−
X
log[P (x(tn+1 )|x(tn ))] − log[P (x(t0 ))],
(5)
n=0
where
the conditional
mutual information is given as [10] CM I(a, b|c) =
log
P (a,b|c)
P (a|c) P (b|c)
. If the model is error free, P (x(tn+1 )|x(tn )) is a delta function:
P (x(tn+1 )|x(tn )) = δ D (x(tn+1 ) − f (x(tn ), tn )).
With a representation of P (X|Y) we may evaluate conditional expected
values of functions G(X) on the path X(N F ) of the model through the
observation window [t0 , tF ] as
R
E[G(X)|Y] = hG(X)i =
dX G(X) exp[−A(X)]
R
,
dX exp[−A(X)]
in which
A(X) = −
F
X
log[P (y(τn )|X(τn ), Y(τn−1 )]
n=1
6
(6)
N (F +1)−1
X
−
log P [(x(tn+1 )|x(tn ))] − log[P (x(t0 ))],
(7)
n=0
and terms depending only on the observations were canceled between numerator and denominator in the expected value.
If the observations at the times τk are independent, and if the measurement function is the identity yr (τk ) = hr (x(τk )) = xr (τk ), and if the noise
in the measurements is Gaussian, with a diagonal inverse covariance matrix
Rm (r, τk ), the first term in the action A(X), the measurement error term,
takes the form
2
F X
L
X
Rm (r, τn )
(8)
xr (τn ) − yr (τn ) .
2
n=1 r=1
If no measurement is made at τk , Rm (r, τk ) = 0.
If the error in the model Eq. (4) is taken as additive and Gaussian with
diagonal inverse covariance matrix Rf (a) the second term in A(X), the model
error term, becomes
N (F +1)−1 D
X
X
n=0
2
Rf (a)
xa (tn+1 ) − fa (x(tn ), tn )
2
a=1
.
(9)
In each term constants having to do with normalizations of the Gaussians
cancel in the expected value. If we keep these constants and take the limit
Rf (a) → ∞, we would restore the delta function in the dynamics.
Finally, if one accepts ignorance of the distribution of initial conditions
P (x(t0 )) and selects it as uniform over the dynamical range of the model
variables, hG(X)i is evaluated with
A0 (X) =
L
F X
X
Rm (r, τn )
2
n=1 r=1
N (F +1)−1 D
+
X
X Rf (a)
n=0
a=1
2
R
hG(X)i =
2
xr (τn ) − yr (τn )
xa (tn+1 ) − fa (x(tn ), tn )
dX G(X) exp[−A0 (X)]
R
,
dX exp[−A0 (X)]
2
.
(10)
(11)
This is the desired connection between the machine learning formulation
(with model error) and the statistical data assimilation formulation: identify
layer labels as time l ⇔ t.
7
This alone could be of passing interest. However, there is much in the
connection that may be of some utility. We will discuss these items in the
data assimilation language, but the translation should be easy at this point.
For statistical data assimilation we call this A0 (X) the standard model.
The critical suggestion here relative to standard practice in machine learning [1, 12, 21], is that by allowing Rf to be finite from the outset–so acknowledging model error–we may add an additional tool for exploration in machine
learning environments where typically no account for model error is introduced. Further, the hyper-parameter Rf serves as a regulating device for the
complexity of the surface in path space in which the estimation of states and
parameters occurs.
4
4.1
Data Assimilation Developments of Use in
Machine Learning
Finding the Global Minimum
The key to establishing estimates for unobserved state variables (L < D) and
for unknown parameters in the model is to perform, approximately of course,
the integral Eq. (11). One can do this using Monte Carlo methods [18] or
by the method of Laplace [19, 20]. In the Laplace method one seeks minima
of the action A0 (X), Eq. (10). The integral is not Gaussian. If it were, we
would just do it. As the functions fa (•) are nonlinear, we must perform a
numerical evaluation of Eq. (6).
The Laplace method approximates the integral with contributions from
the lowest minima of the action, if one can find them. Minima associated
with paths X having larger action give exponentially smaller contributions
to expected values, Eq. (11), than paths with smaller action. This allows
one to circumvent a search for the global minimum if the parameters, hyperparameters [12], and other aspects of the model and data yield a set of action
levels connected with minima of the action such that one path yields an action level much smaller than any other path. For numerical accuracy one
may use that smallest minimum path (comprised of parameters and ‘hidden’
(unobserved) state variables) neglecting larger minima of the action.
We have developed a variational annealing (VA) approach [32, 33], to
finding the path with the smallest value of the action. While we have no
proof that the global minimum is found, our numerical results indicate this
8
may be the case. The VA method produces a set of minima of the action
giving a numerical clue as to the roughness of the surface in path X space.
In data assimilation the surface depends, among other items, on the number of measurements L at each observation time τk , on the hyper-parameter
Rf , and on the number of model time steps between measurement times τn .
This translates directly to the analogous machine learning problem with time
→ layer. As the number of model time steps between measurement times
increases, the number of hidden layers increases and the model architecture
deepens.
VA proceeds by a kind of numerical continuation [3] in Rf of the requirement that varying over all X and all parameters in A0 (X) minimizes A0 (X).
The procedure begins by taking Rf → 0, namely the complete opposite of
the value found in usual machine learning where Rf → ∞ (deterministic,
error free layer to layer maps) from the outset. In the Rf = 0 limit, the
action is just a quadratic function of the model variables x(τk ) at the times
measurements are made, and the minimization is simple: xr (τk ) = yr (τk ) for
the r = 1, 2, ..., L ≤ D data presented at the input and output layers. The
minimum can be degenerate as we know only L ≤ D values for the state
variables.
At the first step of VA we choose as a solution to the optimization problem
xr (τk ) = yr (τk ) and select the other D−L states as drawn from a uniform distribution with ranges known from the dynamical range of the state variables.
One can learn that well enough by solving the underlying model forward for
various initial conditions. We make this draw K times, and now have K paths
X0 as candidates for the VA procedure.
Now we select a small value for Rf , call it Rf 0 . With the previous K
paths X0 as K initial choices in our minimization algorithm, we find K paths
X1 for the minimization problem with Rf = Rf 0 . This gives us K values of
the action A0 (X1 ) associated with the new paths X1 .
Next we increase the value of Rf to Rf = Rf 0 α where α > 1. (We have
found values of α in the range 1.1 to 2 to be good choices). For this new value
of Rf , we perform the minimization of the action starting with the K initial
paths X1 from the previous step to arrive at K new paths X2 . Evaluating the
action on these paths A0 (X2 ) now gives us an ordered set of actions that are
no longer as degenerate. Many of the paths X2 may give the same numerical
value of the action, however,
typically the ‘degeneracy’ lies within the noise
√
level of the data ≈ (1/ Rm ).
This procedure is continued until Rf is ‘large enough’ which is indicated
9
by at least one of the action levels becoming substantially independent of
Rf . As a check on the calculation, we observe that if the action A0 (X) is
independent of Rf , its expected value is that of the measurement error term.
As the measurement errors were taken to be Gaussian, this term in the action
is distributed as χ2 , and its expected value is readily evaluated. If the action
levels are at this expected value of χ2 for large Rf , the procedure is consistent
and no further increases in Rf are required.
Effectively VA starts with a problem (Rf = 0) where the global minimum is apparent and systematically tracks it and many other paths through
increases in Rf . In doing the ‘tracking’ of the global minimum, one must
check that the selected value of α is not too large lest one leave the global
minimum and land in another minimum. Checking the result using smaller
α is worthwhile.
It is important to note that simply starting with a large value of Rf places
one in the undesirable situation of the action A0 (X) having multiple local
minima into which any optimization procedure is quite likely to fall.
In the dynamical problems we have examined, one typically finds that as
the number of measurements L at each τk is increased, fewer and fewer minima of the action remain and when L is large enough there is one minimum.
This we attribute to the additional information from the augmented set of
measurements, and this will be manifest in the discussion below where the
additional information effectively controls unstable directions in the phase
space.
4.2
Smallest Minimum; Not Necessarily a Convex Action
As our goal is to provide accurate estimations of the conditional expected
value of functions G(X) where X, a path in model space, is distributed as
exp[−A(X)], we actually do not require convexity of A(X) as a function in
path space. From the point of view of accurately estimating expected values,
it is sufficient that the lowest action level be much smaller than the second
lowest action level. If the action value at the lowest level A(Xlowest ) is
much smaller than the action value at the next minimum A(Xsecond lowest ),
then by a factor exp[−{A(Xlowest ) − A(Xsecond lowest )}], the lowest path
Xlowest dominates the integral to be done and provides a sensible choice for
the path at which to evaluate the integral.
10
5
Examples from Feedforward Neural Networks and from Data Assimilation
In this section we examine one example from multi-layer perceptrons and one
example from statistical data assimilation. The latter utilizes a differential
equation model introduced by Lorenz in 1996 [23] which permits one to easily
increase the number of dimensions of the phase space, to easily select the
number of observations within a designated measurement window, and to
easily choose the number of model evaluations between measurement times.
The latter is analogous to increasing the number of layers in a multi-layer
perceptron.
In each case we perform a ‘twin experiment’. We use a model to generate
solutions from some initial conditions. These solutions, when Gaussian noise
is added to them, become our noisy data. Using the noisy data we use
VA to estimate the unobserved state variables (hidden layer variables) and
parameters/weights.
5.1
Data Assimilation for Lorenz96 Model
We begin by examining the dynamical equations introduced by [23]:
dxa (t)
= xa−1 (t)(xa+1 (t) − xa−2 (t)) − xa (t) + ν
dt
(12)
and a = 1, 2, ..., D; x−1 (t) = xD−1 (t); x0 (t) = xD (t); xD+1 (t) = x1 (t). ν is a
fixed parameter which we take to be 10.0 where the solutions to the dynamical
equations are chaotic [16]. The equations for the states xa (t); a = 1, 2, ..., D
are meant to describe ‘weather stations’ on a periodic spatial lattice. This
model is widely used in atmospheric science as a testbed for the exploration
of innovative data assimilation ideas.
Our example selects D = 11, and displays the action level plot for L =
2, 4, 5, and 6 observations at each measurement time within the window
[t0 , tF ]. We perform a ‘twin experiment’ wherein we generate D time series
{xa (t); a = 1, 2, ..., D} for Eq. (12) using a standard adaptive fourth order
Runge-Kutta algorithm with a time step ∆t = 0.025 and an initial condition
x(t0 ) drawn from a uniform distribution over the range of the variables x(t),
namely [-10, +10]. To these solutions of Eq. (12) we add Gaussian noise
with mean zero and variance σ 2 = 0.2 to each time series xa (t). These noisy
11
versions of our model time series constitute our ‘data’ {ya (t)}. L of these D
time series are presented to the model at times τn ; t0 ≤ τn ≤ tF .
The measurement window is from t0 = 0 to tF = 4.125. L = 2, 4, 5, 6
‘measurements’ are made at each time step; these are the y(τn ). The measurement error matrix Rm is taken to have diagonal elements at each measurement time τn and is zero at other times. Its magnitude is taken as
Rm = 1/σ 2 = 5.
The model error matrix Rf (a) is also taken as diagonal, with elements
along the diagonal Rf = Rf 0 2β , in performing the VA procedure, and we
take β = 0, 1, 2, . . . . Rf 0 was chosen 0.01.
The minimizations of nonlinear objective functions in the example using the Lorenz96 model was performed using the public domain software
IPOPT [30] with a front end script written in Python.
In Fig. (1) we display action level plots for L = 2, 3, 4, and 6 observations
at each measurement time. As we can see in the TopLeftPanel, where
L = 2, there are numerous local minima in the action A0 (X) for all values of
Rf ≥ Rf 0 , and these remain to large Rf . None of these minima is very far
separated from the paths with the smallest minimum, so that the evaluation
of the expected value integrals Eq. (11) would require contributions from
many maxima of the conditional probability distribution.
When L = 4, TopRightPanel we begin to see an isolated action level
whose contribution to the expected value integral is overwhelmingly larger
than the contribution from path giving rise to the next largest action level.
The value L = 4 is consistent with the observation in [16] that around 0.4D
the instabilities in the Lorenz96 state space appear to be controlled by the
data assimilation process.
At L = 5 or 6, BottomPanels, we see that the dominance of the lowest
action level is even clearer. The horizontal olive colored line is the expected
value of the measurement error term in the action. This is a sign of the
consistency of the data assimilation calculations.
In Fig. (2) we explore another aspect of the action level plots. We
still use D = 11, and we hold L = 6 fixed. The number of observations
within the window [t0 , tF ] has been reduced from 165 to 28 and we move
the model forward between observations 0, 2, 5 or 11 times. This is to
provide an analogy to how many layers are present in an equivalent machine
learning example. Our example here differs by having many entry points in
the measurement window while the machine learning example has only one.
We display in the LeftPanel the action level plots for the selected number
12
of model evaluation steps. As one can see for 0 and 2 intermediate steps we
have many persisting minima of the action. At 5 and 11 intermediate steps,
there is only a single minimum that is found, and for large Rf it comes to the
same action level as with 2 intermediate steps. All are consistent with the
expected value of the measurement error term. This calculation, performed
in a machine learning context provides information on how many hidden
layers are required to achieve a desired accuracy.
In the RightPanel we display the accuracy of the estimation of the single
parameter in the Lorenz96 model. It has been set at ν = 10.0 in producing
the data, and that value is clearly selected for 5 or 11 intermediate model
evaluations, while it is not so clearly selected for 2 intermediate steps and
with zero intermediate steps there is a consistent few percent error.
We see, in this collection of calculations, as noted earlier [32, 33], the ability to identify the dominant minimum of the action depends on the number
of measurements presented during the statistical data assimilation procedure
embodying transfer of information from the data to the model. In data assimilation this is associated with the number of positive conditional Lyapunov
exponents [16] of the model. In the machine learning instantiation it may
play the same role when the number of data presented at the output layer is
not sufficient to determine the parameters and hidden states in each layer.
We also see the analogue of deepening the network produces higher accuracy estimates of conditional expected values.
5.2
A Multi-Layer Perceptron; Feedforward Network
We constructed a feedforward network with lF layers: One input layer at
l0 and one output layer at lF . This network has lF − 2 hidden layers. We
analyzed lF = 20, 30, 50 and 100. There are N = 10 ‘neurons’ in each layer.
The activity in neuron j in layer l xj (l) is related to xj (l − 1) in the previous
layer as
z
(13)
xj (l) = g(W (l)x(l − 1)); g(z) = 0.5[1 + tanh( )].
2
We also investigated the “ReLU”-like function g(z) = log[1 + ez ], but we do
not report on those results. The activations at the input layer are drawn from
a Gaussian N (0, 1). The weights are selected from a uniform distribution
U [−0.1, 0.1]. Gaussian measurement noise was added to the ‘data’ generated
by the model; this has zero mean and variance 0.0025.
13
Figure 1: Action level plots for the Lorenz96 model Eq. (12) with D = 11
and ν = 10.0. The variational annealing procedure is performed with 100
different initial conditions for the minimization of the action at each value
of Rf = Rf 0 αβ ; β = 0, 1, ...; α = 2. Top Left Panel: L = 2 measurements
at each measurement time. At L = 2, there are many minima, but none so
much smaller than the others that it dominates the expected value integral
Eq. (11). Top Right Panel: L = 4 measurements at each measurement
time. At L = 4, the action in path space X has numerous local minima. The
lowest minimum has an action value much smaller than the action values
from the other minima, and this dominates the expected value integral Eq.
(11). Bottom Left Panel: L = 5 measurements at each measurement time.
At L = 5, the number of minima found is only two, and again the lowest
minimum dominates the expected value integral. Bottom Right Panel L
= 6 measurements at each measurement time. At L = 6, there is only one
minimum of the action. The solid green line is the expected value of the
measurement error term. This is distributed as χ2 . As the action becomes
independent of Rf , its expected value should equal this value.
14
Figure 2: Parameter estimation and action level results for the Lorenz96
model, D = 11, L = 6. The parameter value ν = 10.0 was used in the twin
experiments on the model. Observations were made every ∆tobs = 0.15, and L
= 6 measurements were made at each observation time. Left Panel: Action
Level Estimates These are the action levels when L = 6 observations are
made at each measurement time and with the choice of 0, 2, 5, and 11
model evaluation steps between measurement times. The horizontal olive
green line indicates the expected action level for large Rf . Right Panel
Parameter estimates Between the observations the model was evaluated
0, 2, 5 and 11 times leading to ∆tmodel = 1.0, 0.33, 0.16, and 0.08 ∗∆tobs .
The parameter estimates are quite accurate for 5 and for 11 model time
steps between observations. They are more distributed for 0 or 2 model step
between observations. One can associate the idea of increasing the number
of model steps between observations as equivalent to ‘deepening’ the hidden
(unobserved) layers. The horizontal olive green line indicates the parameter
value, 10.0, used in generating the data.
15
Our data are for the twin experiment, where we know all weights, all
inputs and generate all values of x(l) in each layer [l0 , lF ]. These are recorded,
and Gaussian noise with mean zero and variance 0.0025 is added to the data
at layers l0 = 1 and lF starting from known values x(l0 ). These data are
yi (1) for the input layer and yi (20) for the output layer.
M input/output pairs are presented to the model with L = 1, 5, and 10
(k)
(k)
inputs yi (l0 ) at layer l0 and L = 1, 5, 10 data outputs yi (lF ) at layer lF .
k = 1,2,...,M, and we investigated M = 1, 10, 100.
We minimize the action over all the weights and the states x(k)
a (l) at all
layers of the model:
M
L
Rm X
1 X
(k)
(k)
(k)
(k)
2
2
(xr (l0 ) − yr (l0 )) + (xi (lF ) − yi (lF ))
AM (X) =
M k=1 2L r=1
lF
−1 X
N
X
Rf
x(k) (l + 1) − g(W (l)x(k) (l))
N (lF − 1) l=1 a=1 a
2
+
,
(14)
where we have N = 10 neurons in each layer and L ≤ N data at the input l0
and at the output layers lF .
We use the variational annealing procedure described above to identify
the action levels for various paths through the network. The initial value of
Rf 0 /Rm is taken to be 10−8 and this is incremented via Rf /Rm = Rf 0 αβ
with α = 1.1 and β = 0, 1, ...,.
In the numerical optimizations for the machine learning example we used
L-BFGS-B [5, 34].
There are at least two ways to present more information to the model in
this setting:
• increase the number of training pairs available to the network at l0 and
lF ; this is our number M . M can be chosen as large as the user wishes.
• increase the number of components of the input/output pair vectors;
this is our number L. L ≤ N , the number of neurons in l0 or lF .
The resolution of the model in its ability to capture variations in activity
from layer to layer is improved by increasing the number of layers lF .
In Fig. (3) we explore the action levels as a function of Rf /Rm as we
increase the number of layers lF in the model: lF = 20, 30, 50, 100. We hold
fixed the number of neurons N = 10, the number of training pairs M = 100
and the number of inputs and outputs L = 10 at l0 and lF .
16
In each case there are many local minima, but only when lf = 50 does
the lowest action minimum significantly split from the other action minima
and qualify to dominate the expected value integral Eq. (11). When lF is
increased from 50 to 100, the lowest action minimum comes closer to the
second lowest minimum. This is seen as a result of the much larger number
of weights to be estimated at the latter value of lF while we are holding
fixed through the values of L and M the information available to make those
estimations.
In Fig. (4) we hold M fixed at 100, and lF fixed at 50 while we look at
L = 1, 5, 10 values of the dimension of input/output pairs.
In Fig. (5) we hold fixed the number of layers lF , the number of neurons
in each layer N = 10 and the dimension of the input/output vectors yr(k) (l0 )
and yr(k) (lF ); r = 1, 2, ..L = 10. We show the effect of increasing the number
of input/output pairs from M = 1 to M = 10 to M = 100. The emergence
of a lowest action minimum as M increases is displayed. This can serve as a
candidate for approximating Eq. (11).
In Fig. (6) we display the error in prediction after the mode, with lF = 50
layers, has been set by the estimation of the weights. This error is constructed
by selecting MP new input output pairs. Using each of the input elements
for L ≤ N components, we use the model with our estimated weights to
(k)
evaluate x(k)
r (lF ) and compare that with yr (lF ) from each of the MP pairs.
The square error averaged over L presented components and over MP pairs
MP X
L
1 X
(xr(k) (lF ) − yr(k) (lF ))2
LMP k=1 r=1
(15)
is displayed.
We see that increasing the information presented via increasing L or M
leads to decreased average prediction errors when choosing the path corresponding to the lowest action level, Top Layers, or choosing the path
associated with the second lowest action level Bottom Panel. The differences in quality of prediction (or generalization) in these examples among
the cases analyzed is not large, and this has been noted [12].
5.3
Recurrent Networks
In this network architecture one allows both interactions among neurons
from one layer to another layer as well as interactions among neurons within
17
Figure 3: Holding the number of neurons in each layer fixed at 10, the
number of input/output pairs fixed at M = 100, and the number of inputs
and outputs at l0 and lF fixed at L = 10, we vary the number of layers (the
deepening of the network) and examine the action level plots arising from the
variational annealing procedure. Upper Left Panel lF = 20 Top Right
Panel lF = 30 Bottom Left Panel lF = 50 and Bottom Right Panel
lF = 100.
18
Figure 4: Holding the number of neurons fixed N = 10, the number of layers
fixed lF = 50, and the number of input/output training pairs fixed M = 100,
we display the action levels as we vary the number of inputs at l0 and the
number of outputs at lF . Top Left Panel L = 1. Top Right Panel L = 5.
Bottom Panel L = 10.
19
Figure 5: Holding the number of neurons fixed N = 10, the number of layers
fixed lF = 50, and the number of inputs L = 10 at l0 and the number of
outputs at lF , we display the action levels as we vary the number of training
pairs M . Top Left Panel M = 1. Top Right Panel M = 10. Bottom
Panel M = 100.
20
Figure 6:
Prediction Errors for the AI/Machine Learning Network
(k)
(k)
2
r=1 (xr (lF ) − yr (lF )) averaged over MP new input/output
k=1
pairs. A noisy input yr(k) (l0 ) produces the output x(k)
r (lF ) using the model estimated using the M training sets. This is compared with the output yr(k) (lF )
produced with the original model used to produce the ‘data’ for the twin
experiment. In each case the number of neurons is N = 10 and lF = 50. Top
Left Panel Using the model associated with lowest Action Level: L = 10
and M = 1 and M = 10. Top Right Panel Using the model associated with
lowest Action Level: L = 5 and L = 10; M = 100. Bottom Panel Using
the model associated with the second lowest Action Level: L = 5 and 10; M
= 100.
1
LMP
PMP PL
21
a single layer [14, 8]. The activity xj (l) of neuron j, j = 1,2,...,N in layer l
P
{l0 , l1 , ..., lF } is given by xj (l) = f [ i wji (l)xi (l − 1)] in a feedforward, layer
goes to the next layer, network.
We can add interactions with a layer in the same fashion, and to give
some ‘dynamics’ to this within-layer activity we introduce a sequential label
σ to the activity of neuron j in layer l: xj (l, σ). The mapping from layer to
layer and within a layer can be summarized by
xj (l, σ) = f [
X
wji (l)xi (l − 1, σ) +
i
X
Wji (l)xj (l, σ − 1)],
(16)
i
Another version of this allows the nonlinear function to be different for layerto-layer connections and within-layer connections, so
xj (l, σ) = f [
X
X
wji (l)xi (l − 1, σ)] + g[
i
Wji (l)xj (l, σ − 1)],
(17)
i
where f (x) and g(x) can be different nonlinear functions.
We can translate these expressions into the DA structure by recognizing
that xj (l) is the ‘model variable’ in the layer-to-layer function while in the
recurrent network, the state variables become xj (l, σ). It seems natural that
as dimensions of connectivity are added–here going from solely feedforward
to that plus within-layer connections–that additional independent variables
would be aspects of the ‘neuron’ state variables’ representation.
In adding connections among the neurons within a layer we have another
independent variable, we called it σ, and the ‘point’ neurons depending on
layer alone become fields xj (l, σ). In the machine learning/AI networks
we have no restrictions on the number of independent variables. This may
lead to the investigation of ‘neural’ fields φj (v) where v is a collection of
independent variables indicating which layers are involved in the progression
of the field from an input to an output layer.
However many independent variables and however many ‘neurons’ we
utilize in the architecture of our model network, the overall goal of identifying
the conditional probability distribution P (X|Y) and estimating the moments
or expected values of interest still comes down to one form or another in the
approximation of integrals such as Eq. (6).
22
5.4
Making Time Continuous; Continuous Layers: Deepest Learning
There is much to learn about the data assimilation or machine learning problem as the number of layers or equivalently the number of time points within
an epoch becomes very large. The limit of the action where the number
of layers in an epoch becomes a continuous variable is, in data assimilation
notation [15],
Z tF
A0 (x(t), ẋ(t)) =
dt L(x(t), ẋ(t), t)
t0
L
X
2
Rm (l, t)
L(x(t), ẋ(t), t) =
xr (t) − yr (t)
2
r=1
+
D
X
Rf (a)
a=1
2
2
x˙a (t) − Fa (x(t))
.
In this formulation the quantity Rm (l, t) is non-zero only near the times
t ≈ τk . It can be taken as proportional to δ(t − τk ).
Within the machine learning context we call this ‘deepest learning’ as the
number of layers goes to infinity in a useful manner.
The minimization of the action now requires that the paths
x(t) in {x(t), ẋ(t)}
d ∂L(x(t),ẋ(t),t)
= ∂L(x(t),ẋ(t),t) ,
space satisfy the Euler-Lagrange equation
dt
∂ ẋa (t)
∂xa (t)
along with the boundary conditions δxa (t0 )pa (t0 ) = 0; δxa (tF )pa (tF ) = 0
where pa (t) = ∂L(x(t), ẋ(t), t)/∂ ẋa (t) is the canonical momentum.
For the standard model, the Euler-Lagrange equations take the form
Rf [
dxb (t)
d
δab + DFab (x(t))][
− Fb (x(t))] = Rm (l, t)δar (xr (t) − yr (t)), (18)
dt
dt
where we have DF(x) = ∂F(x)/∂x.
The E-L equations are the necessary condition, along with the accompanying boundary conditions, that show how errors represented on the right
hand side of the E-L equation drive the model variables at all layers to produce x(l) → y(l) where data is available.
In the integral for < G(X) >, the coordinates x(t0 ) and x(tF ) are not
restricted, so we have the ‘natural’ boundary conditions [11, 17, 22] pa (t0 ) = 0
and pa (tF ) = 0.
This shows quite clearly that the minimization problem requires a solution
of a two point boundary value problem in {x(t), v(t) = ẋ(t)} space. One way
to address two point boundary value problems is to start at one end, t0 with
a value of x(t0 ) and proceed from tF with a value of x(tF ) and integrate both
23
ways requiring a match [27]. Furthermore, the residual of the measurement
error term on the right hand side of Eq. (18) nudges the solution in x(t) to
the desired output.
If one were to specify x(t0 ), but not x(tF ), then the boundary conditions
for the Euler-Lagrange equation are the given x(t0 ) (δx(t0 ) = 0) and require
the canonical momentum pa (tF ) = 0. Examining the Hamiltonian dynamics
for this problem then suggest integrating the x(t) equation forward from
t0 and the canonical momentum equation backward from tF . This is back
propagation.
5.4.1
Hamiltonian Dynamics Realization
If one moves from the Lagrangian realization of the variational problem to a
Hamiltonian version by trading in the phase space from {x(t), v(t)} to canonical coordinates {x(t), p(t)}, then the Hamiltonian H(x, p) for the standard
model reads
H(x, p, t) =
D
X
pa (t)pa (t)
a=1
2Rf (a)
+ pa (t)Fa (x(t)) −
L
X
Rm (r, t)
r=1
2
(xr (t) − yr (t))2 . (19)
In these coordinates the equations of motion are then given by Hamilton’s
equations
dpa (t)
∂Fb (x(t))
= −pb (t)
+ δar Rm (r, t)(xr (t) − yr (t))
dt
∂xa (t)
dxa (t)
pa (t)
= Fa (x(t)) +
.
dt
Rf (a)
(20)
Returning from this to discrete time (or layers) we see that if the variational principle is carried out in {x, p} space, the boundary conditions
pa (t0 ) = pa (tF ) = 0 are quite easy to impose while the other variables,
all the xa (tk ) and the pa (tk ); k 6= 0, F , are varied. Going forward in x and
backward in p is neither required nor suggested by this formulation. It is
worth noting that in either {x, v} space or {x, p} space, the continuous time
(layer) formulation has a symplectic symmetry [11, 15]. This not automatically maintained when the discrete time (layer) problem is reinstated [24, 31];
however, many choices of integration procedure in which time/layer becomes
discrete and the symplectic symmetry is maintained are known [24, 31, 13]
In a detailed analysis [15, 33] of the variational problem in Lagrangian
and Hamiltonian formulations, it appears that the direct Lagrangian version
24
in which the state variables x(tn ) or x(ln ) are varied, the symplectic structure
can be maintained and the boundary conditions on the canonical momentum
respected [24, 31].
In practice, this means that the direct variational methods suggested for
the machine learning problems taking into account model error (Rf 6= ∞)
may skirt issues associated with back propagation. This issue may be seen a
bit more directly by comparing how one moves in {x(t), ẋ(t)} space organized
by Eq. (18) with the motion in {x(t), p(t)} space guided by Eq. (20). These
are equivalent motions of the model in time/layer, connected by a Legendre
transformation from {x(t), ẋ(t)} → {x(t), p(t)}.
In the Hamiltonian form, where Rf → ∞ is the limit where one usually
works, moving in regions where DF(x) may have saddle points may ‘slow
down’ the progression in the canonical momentum p(t). This may occur at a
maximum, at a minimum, or at a saddle point of DF(x). At any of these the
observation in [21]: “The analysis seems to show that saddle points with only
a few downward curving directions are present in very large numbers, but
almost all of them have very similar values of the objective function. Hence,
it does not much matter which of these saddle points the algorithm gets
stuck at.” may apply. In the Lagrangian formulation Eq. (18) the manner
in which DF(x) enters the motion is quite different and may well avoid this
confounding property. We have pointed out that Eq. (20) is backprop. The
use of the Lagrangian variational principle [24, 31] solves the same problem,
so may have an unexpected virtue.
6
Summary and Discussion
This paper has been directed to drawing a direct analogy between the formulation of a much utilized class of machine learning problems and a set of
equivalent problems in data assimilation as encountered in many physical,
biological and geoscience problems as well as in many engineering analyses
where data driven model development and testing is a goal. The fundamental
equivalence of the two inquiries is the core of this paper.
The analogy allows us to identify methods developed in data assimilation
as potentially quite useful in machine learning contexts. In particular the
possibility of using variational annealing to produce the global minimum of
the action (cost function) of the standard model of data assimilation with
both observation error and model error appears potentially of value.
25
The idea of making time continuous for purposes of exploring properties of
data assimilation suggests a similar tactic in machine learning. The machine
learning step of making layers continuous we have called “deepest learning”
as deep learning appears to result from increasing the number of layers. In
the continuous layer (time) formulation, we see clearly that the problem to
be solved is a two point boundary value problem. This may lead to the
construction and solution of tractable models that may helpfully illuminate
how deep learning networks operate successfully and expand the possibilities
of utilizing them employing additional methods for numerical calculations
and for interpretation.
In the formulation of the statistical data assimilation problem at the general level expressed in Eq. (5) we see that the measurement error term which
is where information from data is passed to the model, it is explicitly information through the conditional mutual information that is being passed from
observations to the model. This suggests that the idea that deep learning
works because of major increases in computing power as well as in having
large data sets, however, the attribute of the data sets is not so much as they
are large but that they possess information, in a precise manner, that can be
utilized to learn about models. The conjunction of information transfer and
state and parameter estimation is embodied in the work of Rissanen [28, 29]
where he identifies the cost of estimating a parameter or a state variable at
some time. The arguments we have presented suggest evaluating how much
information in a data set is available to inform a model is of greater utility
than just the size of the data set itself.
One point not made explicit in the main text, but worth noting, is that
once we have formulated the data assimilation or machine learning problems
as accurately performing high dimensional integrals such as Eq. (6), the
Laplace approximation method, namely the usual variational principle, permits the investigation of corrections through further terms in the expansion
of the action about the path leading to the global minimum. In [33] it is
shown that corrections to this first approximation are small as Rf becomes
large when analyzing the standard model. This need not be the case for
other choices of noise distributions in the measurement error or model error
terms in the action.
Another item of interest is the argument noted in [21] that as the dimension of a model increases, one may find fewer and fewer local minima
confounding the search in path space for a global minimum, and in that
situation many more unstable saddle points in path space will arise [7, 6].
26
In the case of a chaotic system such as the Lorenz96 model, the evidence
is that however large the dimension of the model itself and the paths over
which one may search, there are multiple local minima until the number of
measurements at any observation time is large enough and the information
transferred to the model is sufficient. The role of the number of model
evaluations between observations, suggested in some of the arguments here,
also play a significant part in establishing whether the action surface has
many local minima.
The view of a deep network as moving from a few hidden layers to many
may also be illuminated by our arguments. One idea is that by increasing
the number of hidden layers one is increasing the resolution in the analog
of ‘time’ in data assimilation. When one does that in data assimilation, we
see it as probing the variation of the underlying model as it evolves from
an initial condition through ‘layer’ = ‘time.’ Missing the higher frequency
variations in time by employing a coarse grid in discrete time should have its
counterpart role in the feedforward networks discussed here.
It is recognized that the ‘neuron’ models widely utilized in machine learning applications have little in common with properties of biological neurons,
the construction and implementation of large networks that have successful
function within machine learning may prove a useful guide for the construction and implementation of functional natural neural networks.
Finally, it is important to comment that while the analogy drawn and
utilized here may improve the testing and validation of models supported by
observational data, it does not assist in the selection of the models and their
formulation. That is still a task to be addressed by the user.
Acknowledgments
We express our appreciation to our colleagues Dan Breen and Jeff Elman
for important discussions. Also the CENI team made finding the analogs
in this paper possible. Many thanks to Tim Gentner, Gert Cauwenberghs,
and especially Gabe Silva improved the manuscript. Jon Shlens provided
a detailed view from the active world of machine learning. Partial support from the MURI Program (N00014-13-1-0205) sponsored by the Office
of Naval Research is acknowledged as is support for A. Shirman from the
ARCS Foundation.
27
References
[1] Perspectives on Research in Artificial Intelligence and Artificial
General Intelligence Relevant to DoD, 2017. https://www.google.
com/?gws_rd=ssl#q=perspectives+on+research+in+artificial+
intelligence+and+artificial+general+intelligence.
[2] Henry D. I. Abarbanel. Predicting the Future: Completing Models of
Observed Complex Systems. Springer, 2013.
[3] E. L. Allgower and K. Georg. Numerical Continuation Methods: An
Introduction. Springer-Verlag, 1990.
[4] A. F. Bennett. Inverse Methods in Physical Oceanography. Cambridge
University Press, 1992.
[5] R. H. Byrd, P. Lu, and J. Nocedal. A limited memory algorithm for
bound constrained optimization. SIAM Journal on Scientific and Statistical Computing, 16:1190–1208, 1995.
[6] Anna Choromanska, Mikael Henaff, Michael Mathieu, Ge’rard Ben
Arous, and Yann LeCun. The loss surfaces of multilayer networks. In
18th International Conference on Artificial Intelligence and Statistics
(AISTATS), 2015.
[7] Yann N. Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho,
Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In Proceedings of the 27th International Conference on Neural Information
Processing Systems, NIPS’14, pages 2933–2941, Cambridge, MA, USA,
2014. MIT Press.
[8] J. L. Elman. Finding structure in time. Cognitive Science, 14:179–211,
1990.
[9] Geir Evensen.
Springer, 2009.
Data Assimilation:
The Ensemble Kalman Filter.
[10] Robert M. Fano. Transmission of Information; A Statistical Theory of
Communication. MIT Press, 1961.
28
[11] I. M. Gelfand and S. V. Fomin. Calculus of Variations. Dover Publications, Inc., 1963.
[12] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, Cambridge, MA; London, UK, 2016. http://www.
deeplearningbook.org.
[13] Ernst Hairer, Gerhard Wanner, and Christian Lubich. Geometric Numerical Integration: Structure-Preserving Algorithms for Ordinary Differential Equations. Springer Series in Computational Mathematics; Volume 31, 2nd edition, 2006.
[14] M. Jordan. Attractor dynamics and parallelism in a connectionist sequential machine. In Proceedings of the Eighth Conference of the Cognitive Science Society, pages 531–546. Cognitive Science Society, 1986.
[15] N. Kadakia, D. Rey, J. Ye, and H. D. I. Abarbanel. Symplectic structure
of statistical variational data assimilation. Quarterly Journal of the
Royal Meteorological Society, 143(703):756–771, 2017.
[16] Mark Kostuk. Synchronization and statistical methods for the data
assimilation of hvc neuron models. PhD Dissertation in Physics, University of California, San Diego, 2012.
[17] Mark Kot. A First Course In the Calculus of Variations. American
Mathematical Society; Providence, Rhode Island., 2014.
[18] Rubin H. Landau, Manuel Jose’ Paez, and Cristian C. Bordeianu. A
Survey of Computational Physics: Introductory Computational Science.
Princeton University Press, 2010.
[19] Pierre Simon Laplace. Memoir on the probability of causes of events.
Mémoires de Mathématique et de Physique, Tome Sixième, pages 621–
656, 1774.
[20] Pierre Simon Laplace. Memoir on the probability of the causes of events.
Statistical Science, 1(3):364–378, 1986. Translation to English by S. M.
Stigler.
[21] Yann LeCunn, Yoshua Bengio, and Geoffrey Hinton. Deep learning.
Nature, 521:436–444, 2015.
29
[22] Daniel Liberzon. Calculus of Variations and Optimal Control Theory.
Princeton University Press, 2012.
[23] Edward N Lorenz. Predictability: A problem partly solved. In Tim
Palmer and Renate Hagedorn, editors, Predictability of weather and climate. Cambridge, 2006.
[24] J. E. Marsden and M. West. Discrete mechanics and variational integrators. Acta Numerica, pages 357–514, 2001.
[25] K. G. Murty and S. N. Kabadi. Some np-complete problems in quadratic
and nonlinear programming. Mathematical Programming, 39:117–129,
1987.
[26] Alexander G. Parlos, Kil T. Chong, and Amir F. Atiya. Application
of the recurrent multilayer perceptron in modeling complex process dynamics. IEEE Transactions on Neural Networks, 5:255–266, 1994.
[27] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery.
Numerical Recipes: The Art of Scientific Computing, Third Edition.
Cambridge University Press, 2007.
[28] Jorma Rissanen. Stochastic Complexity in Statistical Inquiry Theory.
World Scientific Publishing Co., Inc., River Edge, NJ, USA, 1989.
[29] Jorma Rissanen. Information and Complexity in Statistical Modeling.
Springer, 2007.
[30] A. R. Wachter and L. T. Biegler. On the implementation of an interiorpoint filter line-search algorithm for large-scale nonlinear programming.
Mathematical programming, 106:25–57, 2006.
[31] J. M. Wendlandt and J. E. Marsden. Mechanical integrators derived
from a discrete variational principle. Physica D: Nonlinear Phenomena,
106:223–246, 1997.
[32] J. Ye, N. Kadakia, P. J. Rozdeba, H. D. I. Abarbanel, and J. C. Quinn.
Improved variational methods in statistical data assimilation. Nonlinear
Processes in Geophysics, 22(2):205–213, 2015.
30
[33] J. Ye, Daniel Rey, Nirag Kadakia, Michael Eldridge, Uri Morone, Paul
Rozdeba, Henry D. I. Abarbanel, and John C. Quinn. A systematic variational method for statistical nonlinear state and parameter estimation.
Physical Review E, 2015.
[34] C. Zhu, R. H. Byrd, and J. Nocedal. L-bfgs-b: Algorithm 778: L-bfgs-b,
fortran routines for large scale bound constrained optimization. ACM
Transactions on Mathematical Software, 23:550–560, 1997.
31
| 2cs.AI
|
Strongly connected components
-Algorithm for finding the strongly connected components of a graphAuthor: Vlad-Andrei Munteanu
1. Abstraction
A directed graph G (V, E) is strongly connected if and only if, for any pair of vertices X and Y from
V, there exists a path from X to Y and a path from Y to X. In Computer Science, the partition of a
graph in strongly connected components is represented by the partition of all vertices from the
graph, so that for any two vertices, X and Y, from the same partition, there exists a path from X
to Y and a path from Y to X and for any two vertices, U and V, from different partition, the
property is not met. The algorithm presented below is meant to find the partition of a given graph
in strongly connected components in O (numberOfNodes + numberOfEdges * log*
(numberOfNodes)), where log* function stands for iterated logarithm.
2. Algorithm
For a better understanding of the algorithm it is recommended a revision of Disjoint-set Data
Structure (section 6) and DFS oriented graph traversing (section 7). We will consider a strongly
connected component as a set. With no doubt, two strongly connected components represent
two disjoint sets because, if there exists a vertex X which belongs to the first strongly connected
component as well as to the second strongly connected component, it means that for any vertex
U from the first strongly connected component and any vertex V from the second strongly
connected component, we can find both, a path from V and a path from U to V through X. This
holds because X belongs to the same strongly connected component as U and V. From here, it
results that the two strongly connected components we have selected can be joined in a bigger
one. Using induction, we can prove that, if two or more strongly connected components are
linked by one or more vertices, that means that they can be jointed. Therefore, the strongly
connected components of a graph are pairwise disjointed. Initially, we will consider that every
vertex is in its own strongly connected component (by definition, a vertex is a strongly connected
component) and, using a DFS traversal, we will link those vertices that are in the same strongly
connected component). The Algorithm is recursive type, and in the following lines, I will present
the steps that must be followed:
1. the X vertex is added in the DFS stack and it is marked as being visited (either as DFS
source, either as adjacent vertex to the top vertex of the stack);
2. iterate through all adjacent vertices of X in any order and for every neighbor, Y, it will be:
-
the case in which Y had already been visited, so the algorithm just verifies if any vertex
from the stack is in the same strongly connected component as Y, and, if case positive, it
joins the strongly connected components of X and Y;
-
the case in which the neighbor has not been visited yet, so it must be introduced in the
stack. The algorithm is executed for the vertex Y, starting from the first step. After the
execution of the algorithm for Y, it is popped out from the stack and, now being visited,
the checking mentioned above between the nodes in the stack and Y needs to be done.
If case positive, the strongly connected components of X and Y must be joined as above.
3. the vertex X is popped out from the stack;
A crucial point of the algorithm is how we may know that a vertex from the stack is in the same
connected components as the neighbor. In order to solve this problem, we must hold the
minimum level from the DFS stack of each strongly connected component, which will be
initialized with infinity. If the level of the strongly connected component of Y is infinity, there is
no vertex in the stack which belongs to it. We have to reset to infinity the level of a strongly
connected component when the vertex with the lowest level which belongs to that component
is to be popped out from the DFS stack. When two strongly connected components are joined,
the dominant one will hold the minimum level between its level and the level of the other one.
3. Proof of correctness
The last paragraphs focused on the algorithm itself and on the most important observation which
allows us to use Disjoint-set Data Structure. The following paragraphs will focus on proving the
correctness of the algorithm. We will take into account all the situations which could appear with
the adding of a new vertex in the stack (let`s assume that the new added vertex is X):
-Case 1: The vertex X is the first pushed to the stack and the algorithm begins.
-Case 2: The vertex X has a father, vertex Y, and they belong to the same strongly connected
component.
-Case 3: The vertex X has a father, vertex Y, and they belong to different strongly connected
components.
Case 1: If X is the first pushed vertex, it does not have a father, so this will not be an attempt to
change the strongly connected component (joining attempt is done only for the pairs {father,
vertex}.
Case 2: If the X vertex has a father, node Y and they are found in the same strongly connected
component, it means that there exists a path from X to Y (the path from Y to X being a direct
edge). We will consider the path as X->Vertex1->Vertex2->Vertex3->…->Vertexk->Y (in the case in
which there are more valid paths, consider the first which will be taken into account by the DFS
traversal). Due to the properties of the DFS traversal which guarantees that all the paths from
the current node which have not been visited until then will be taken into consideration, we could
jump to the conclusion that if the vertex Vertexi is not visited, then all vertices Vertexi,
Vertex(i+1),…, Vertexk and Y are visited and they are in the DFS stack (if at least one from these is
not visited it means that the chosen path is not the first one selected by the DFS traversal, and if
they would not be in the stack and but visited, it would mean that a complete DFS traversal has
been executed, so X, being previously visited too, cannot be added now) and, moreover, any
vertex Z, which belongs to a path from X to Y, but does not belong to the chosen path, it is not
visited (in the case in which it was previously visited, either a complete execution of DFS traversal
has been executed, which leads to the contradiction discussed before, or the chosen path is not
the first one -the first path being the one which goes through Z-, which leads to a contradiction
of the hypothesis). Therefore, there exists a vertex, Vertexj, with the property that all vertices
Vertexj, Vertex(j+1), …, Vertexk, Y are visited and in the stack and all vertices X, Vertex1, Vertex2,…,
Vertex(j-1) are not visited. From the same property of the DFS traversal, it is guaranteed that it
must lead to the traversal of X, Vertex1, Vertex2, …, Vertex(j-1) at some point. Arrived in Vertex(j-1),
we will find Vertexj which is visited and in the stack, so the strongly connected components of
Vertex(j-1) and Vertexj are joined. Back from recursion to Vertex(j-2), it is confirmed the fact that
Vertex(j-1) it is in the stack through the link with Vertexj. Inductively, the link would be made
between Vertex(j-2) and Vertex(j-3), between Vertex(j-3) and Vertex(j-4), …, between Vertex1 and X,
and, finally, between X and Y, being in the stack through Vertexj. In conclusion, if two vertices X
and Y which are connected by an edge are in the same component, the algorithm will join them.
Case 3: If the vertex X has a father, vertex Y, and they are in different strongly connected
components, it means that there exists no path from X to Y (the path from Y to X exists through
direct edge). As there exists no path from X to Y, it means that for any vertex Y` with the property
that there exists a path from Y` to Y and vertex X` with the property that there exists a path from
X to X`, there exists no path from X` to Y`. Therefore, the DFS traversal, would not find any path
from any “son" of Y to any “father" of X, and so, no vertex which is in the same set as Y ( Y
included) will be connected with no vertex which is in the same set as X( X included).
In conclusion, the behavior of the algorithm is as desired in all possible cases, so the correctness
is demonstrated.
4. Complexity
The complexity is given by the DFS traversal and by a possible connection between of any two
vertices which are connected through and edge, so we have O(numberOfNodes+numberOfEdges
+ numberOfEdges * log* (numberOfNodes)), which is O(numberOfNodes + numberOfEdges *
log* (numberOfNodes)), where log* stands for the iterated logarithm.
5. Implementation
A simple and clean implementation of the algorithm can be done through five methods (you can
see the C++ code below):
1. The method which generate the set of a specific element (vertex)
int getComponent (int currentElement) {
int currentKing = currentElement;
while (currentSCC[currentKing] != currentKing) {
currentKing = currentSCC[currentKing];
}
while (currentElement != currentKing) {
int copyOfElement = currentElement;
currentElement = currentSCC[copyOfElement];
currentSCC[copyOfElement] = currentKing;
}
return currentKing;
} //getComponent
2. The method which joins the sets of two given elements (vertices)
void uniteComponents (int firstElement, int secondElement) {
int firstComponent = getComponent(firstElement);
int secondComponent = getComponent(secondElement);
if (firstComponent == secondComponent) {
return;
}
if (heightOfSCC[firstComponent] < heightOfSCC[secondComponent]) {
swap (firstComponent, secondComponent);
}
if (heightOfSCC[firstComponent] == heightOfSCC[secondComponent]) {
++heightOfSCC[firstComponent];
}
currentSCC[secondComponent] = firstComponent;
minLevelInDFS[firstComponent] = min (minLevelInDFS[firstComponent],
minLevelInDFS[secondComponent]);
} //uniteComponents
3. DFS method
void solveIt (int node, int currentLevel) {
int componentOfNode = getComponent(node);
minLevelInDFS[componentOfNode] = min (minLevelInDFS[componentOfNode],
currentLevel);
visited[node] = true;
for (auto currentNeighbour : graph[node]) {
if (visited[currentNeighbour] == false) {
solveIt(currentNeighbour, currentLevel + 1);
}
int componentOfNeighbour = getComponent(currentNeighbour);
if (minLevelInDFS[componentOfNeighbour] < currentLevel) {
uniteComponents(node, currentNeighbour);
}
}
componentOfNode = getComponent(node);
if (minLevelInDFS[componentOfNode] == currentLevel) {
minLevelInDFS[componentOfNode] = MAXLevel;
}
} //solveIt
4. Main method
int main() {
ofstream fout ("output");
int numberOfNodes, numberOfEdges;
readIt(numberOfNodes, numberOfEdges);
prepareIt(numberOfNodes);
for (int currentNode = 1; currentNode <= numberOfNodes; ++currentNode) {
if (visited[currentNode] == false) {
solveIt(currentNode, 1);
}
}
int howmanySCC = 0;
for (int currentNode = 1; currentNode <= numberOfNodes; ++currentNode) {
if (currentNode == currentSCC[currentNode]) {
++howmanySCC;
}
}
fout << howmanySCC << '\n';
int currentIndexSCC = 0;
for (int currentNode = 1; currentNode <= numberOfNodes; ++currentNode) {
int currentSCC = getComponent(currentNode);
if (buildFinalSCC[currentSCC]) {
SCC[buildFinalSCC[currentSCC]].push_back(currentNode);
}
else {
++currentIndexSCC;
buildFinalSCC[currentSCC] = currentIndexSCC;
SCC[buildFinalSCC[currentSCC]].push_back(currentNode);
}
}
for (int currentSCC = 1; currentSCC <= howmanySCC; ++currentSCC) {
for (auto currentNode : SCC[currentSCC]) {
fout << currentNode << ' ';
}
fout << '\n';
}
return 0;
} //main
5. Declarations
//Declarations - begin
const int MAXN = 1e5 + 5;
const int MAXLevel = 1e8;
vector <int> graph[MAXN];
vector <int> SCC[MAXN];
int currentSCC[MAXN];
int heightOfSCC[MAXN];
int minLevelInDFS[MAXN];
int buildFinalSCC[MAXN];
bool visited[MAXN];
//Declarations - end
6. Disjoint-set Data Structure
Disjoint-set is a data structure used for keeping disjoint sets and it was first described by Bernard
A. Galler and Michael J. Fischer in 1964. It allows us to find the set which a specific element
belongs to or to join the sets of two specific elements in complexity O (log*
(numberOfElementsInTheSets)). The main idea is to represent each set as a tree, where the root
is the father of the whole set. When two sets are to be joined, the set with the higher tree is
considered the dominant one and its father becomes the father for the both sets, and when we
need to find the father of a specific element, we will change the fathers of all elements on the
path between it and the root, linking each of those to the root. These two optimizations generate
the complexity mentioned above. More information about it may be found following the links in
bibliography.
7. DFS Traversal
Depth-First Search is an algorithm for traversing trees or graphs (directed and undirected). A
version of it was developed by French mathematician Charles Pierre Tremaux in the 19th century.
It makes use of a stack, the top of it being considered the current node. The idea is to iterate
through the neighbors of the current node and, if one of those has not been visited yet, it will be
pushed in the stack and it will be set as the current node. More information about it may be
found following the links in bibliography.
8. Bibliography
https://en.wikipedia.org/wiki/Depth-first_search
https://en.wikipedia.org/wiki/Disjoint-set_data_structure
| 8cs.DS
|
arXiv:1701.03322v2 [cs.AI] 28 Apr 2017
From First-Order Logic to Assertional Logic
Yi Zhou
Abstract
First-Order Logic (FOL) is widely regarded as one of the most important foundations for knowledge representation. Nevertheless, in this paper, we argue that
FOL has several critical issues for this purpose. Instead, we propose an alternative called assertional logic, in which all syntactic objects are categorized as set
theoretic constructs including individuals, concepts and operators, and all kinds of
knowledge are formalized by equality assertions. We first present a primitive form
of assertional logic that uses minimal assumed knowledge and constructs. Then,
we show how to extend it by definitions, which are special kinds of knowledge,
i.e., assertions. We argue that assertional logic, although simpler, is more expressive and extensible than FOL. As a case study, we show how assertional logic can
be used to unify logic and probability, and more building blocks in AI.
1 Introduction
Classical First-Order Logic (FOL) is widely regarded as one of the most important
foundations of symbolic AI. FOL plays a central role in the field of Knowledge Representation and Reasoning (KR). Many of its fragments (such as propositional logic,
modal and epistemic logic, description logics), extensions (such as second-order logic,
situation calculus and first-order probabilistic logic) and variants (such as Datalog and
first-order answer set programming) have been extensively studied in the literature
[6, 22].
Nevertheless, AI researchers have pointed out several issues regarding using FOL
for knowledge representation and reasoning, mostly from the reasoning point of view.
First, FOL is computationally very difficult. Reasoning about FOL is a well-known
undecidable problem. Also, FOL is monotonic in the sense that adding new knowledge into a first-order knowledge base always results in more consequences. However,
human reasoning is sometimes nonmonotonic.
In this paper, we argue that FOL also has some critical issues from the knowledge
representation point of view. First of all, although FOL is considered natural for welltrained logicians, it is not simple and flexible enough for knowledge engineers with
less training. One reason is the distinction and hierarchy between term level, predicate
level and formula level. From my own experience as a teacher in this subject, although
strongly emphasized in the classes, many students failed to understand why a predicate
or a formula cannot be in the scope of a function. Another reason is the notion of free
occurrences of variables. For instance, it is not easily understandable for many students
why the GEN inference rule has to enforce the variable occurrence restrictions. Last
1
but not least, arbitrary nesting also raises issues. Although natural from a mathematical
point of view, a nested formula, e.g., (x∨¬(y ∧z))∧(¬y ∨¬x) is hard to be understood
and used.
Secondly, FOL has limitations in terms of expressive power. FOL cannot quantify
over predicates/functions. This can be addressed by extending FOL into high-order
logic. Nevertheless, high-order logic still cannot quantify over formulas. As a consequence, FOL and high-order logic are not able to represent an axiom or an inference
rule in logic, such as Modus Ponens. Flexible quantification beyond the term level is
needed in applications. As an example, in automated solving mathematical problems,
we often use proof by induction. To represent this, we need to state that for some statement P with a number parameter, if that P holds for all numbers less than k implies
that P holds for the number k as well, then P holds for all natural numbers. Here, P
is a statement at a formula level, possibly with complex sub-statements within itself.
Hence, in order to represent and use proof by induction, we need to quantify over P
that is at a formula level.
Thirdly, FOL is hard to be extended with new building blocks. FOL itself cannot
formalize some important AI notions including probability, actions, time etc., which
are needed in a wide range of applications. For this purpose, AI researchers have
made significant progresses on extending FOL with these notions separately, such as
first-order probabilistic logic [4, 12], situation calculus [15, 16], CTL [7] etc. Each is
a challenging task in the sense that it has to completely re-define the syntax as well
as the semantics. However, combing these notions together, even several of them,
seems an extremely difficult task. Moreover, there are many more building blocks to
be incorporated in applications. For instance, consider task planning for home service
robots [13]. It is necessary to represent actions, probability, time and more building
blocks such as preferences altogether at the same time.
To address these issues, we propose assertional logic, in which all syntactic objects
are categorized as set theoretic constructs including individuals, concepts and operators, and all kinds of knowledge are uniformly formalized by equality assertions of
the form a = b, where a and b are either atomic individuals or compound individuals.
Semantically, individuals, concepts and operators are interpreted as elements, sets and
functions respectively in set theory, and knowledge of the form a = b means that the
two individuals a and b are referring to the same element.
We first present the primitive form of assertional logic that uses minimal assumed
knowledge and primitive constructs. Then, we show how to extend it with more building blocks by definitions, which are special kinds of knowledge, i.e., assertions used
to define new individuals, concepts and operators. Once these new syntactic objects
are defined, they can be used as a basis to define more. As an example, we show how
to define multi-assertions by using Cartesian product, and nested assertions by using
multi-assertions.
We show that assertional logic, although simpler, is more expressive and extensible
than FOL. As a case study, we show how to extend assertional logic for unifying logic
and probability, and more important AI building blocks including time. Note that our
intention is not to reinvent the wheel of these building blocks but to borrow existing
excellent work on formalizing these building blocks separately and to assemble them
within one framework (i.e., assertional logic) so that they can live happily ever after.
2
2 Meta Language and Prior Knowledge
One cannot build something from nothing. Hence, in order to establish assertional
logic, we need some basic knowledge. Of course, for the purpose of explanation, we
need an informal meta language whose syntax and semantics are pre-assumed. As
usual, we use a natural language such as English. Nevertheless, this meta language is
used merely for explanation and it should not affect the syntax as well as the semantics
of anything defined formally.
Only a meta level explanation language is not enough. Other than this, we also need
some core objects and knowledge, whose syntax and semantics are pre-assumed as
well. These are called prior objects and prior knowledge. For instance, when defining
real numbers, we need some prior knowledge about natural numbers; when defining
probability, we need some prior knowledge about real numbers.
In assertional logic, we always treat the equality symbol “=” as a prior object.
There are some prior knowledge associated with the equality symbol. For instance, “=”
is an equivalence relation satisfying reflexivity, symmetricity, and transitivity. Also,
“=” satisfies the general substitution property, that is, if a = b, then a can be used
to replace b anywhere. Other than the equality symbol, we also assume some prior
objects and their associated prior knowledge in set theory [11], including set operators
such as set union and Cartesian product, Boolean values, set builder notations and
natural numbers.
3 Assertional Logic: the Primitive Form
In this section, we present the primitive form of assertional logic. As the goal of assertional logic is to syntactically represent knowledge in application domains, there are
two essential tasks, i.e., how to capture the syntax of the domain and how to represent
knowledge in it.
3.1 Capturing the syntax
Given an application domain, a syntactic structure (structure for short if clear from
the context) of the domain is a triple hI, C, Oi, where I is a collection of individuals,
representing objects in the domain, C a collection of concepts, representing groups
of objects sharing something in common and O a collection of operators, representing
relationships and connections among individuals and concepts. Concepts and operators
can be nested and considered as individuals as well. If needed, we can have concepts
of concepts, concepts of operators, concepts of concepts of operators and so on.
An operator could be multi-ary, that is, it maps a tuple of individuals into a single
one. Each multi-ary operator O is associated with a domain of the form (C1 , . . . , Cn ),
representing all possible values that the operator O can operate on, where Ci , 1 ≤ i ≤
n, is a concept. We call n the arity of O. For a tuple (a1 , . . . , an ) matching the domain
of an operator O, i.e., ai ∈ Ci , 1 ≤ i ≤ n, O maps (a1 , . . . , an ) into an individual,
denoted by O(a1 , . . . , an ).
3
Operators are similar to functions in FOL but differs in two essential ways. First,
operators are many-sorted as C1 , . . . , Cn could be different concepts. More importantly, C1 , . . . , Cn could be high-order constructs, e.g., concepts of concepts, concepts
of operators.
3.2 Representing knowledge
Let hI, C, Oi be a syntactic structure. A term is an individual, either an atomic individual a ∈ I or the result O(a1 , . . . , an ) of an operator O operating on some individuals
a1 , . . . , an . We also call the latter compound individuals.
An assertion is of the form
a = b,
(1)
where a and b are two terms. Intuitively, an assertion of the form (1) is a piece of
knowledge in the application domain, claiming that the left and the right side refer to
the same object. A knowledge base is a set of assertions. Terms and assertions can be
considered as individuals as well.
Similar to concepts that group individuals, we use schemas to group terms and assertions. A schema term is either an atomic concept C ∈ C or of the form O(C1 , . . . , Cn ),
where Ci , 1 ≤ i ≤ n are concepts. Essentially, a schema term represents a set of
terms, in which every concept is grounded by a corresponding individual. That is,
O(C1 , . . . , Cn ) is the collection {O(a1 , . . . , an )}, where ai ∈ Ci , 1 ≤ i ≤ n are individuals. Then, a schema assertion is of the same form as form (1) except that terms
can be replaced by schema terms. Similarly, a schema assertion represents a set of
assertions.
We say that a schema term/assertion mentions a set {C1 , . . . , Cn } of concepts if
C1 , . . . , Cn occur in it, and only mentions if {C1 , . . . , Cn } contains all concepts mentioned in it. Note that it could be the case that two or more different individuals are
referring to the same concept C in schema terms and assertions. In this case, we need
to use different copies of C, denoted by C 1 , C 2 , . . . , to distinguish them. For instance,
all assertions x = y, where x and y are human, are captured by the schema assertion
Human1 = Human2 . On the other side, in a schema, the same copy of a concept C
can only refer to the same individual. For instance, Human = Human is the set of
all assertions of the form x = x, where x ∈ Human.
3.3 The semantics
We propose a set theoretic semantics for assertional logic. Since we assume set theory
as the prior knowledge, in the semantics, we freely use those individuals (e.g., the
empty set), concepts (e.g., the set of all natural numbers) and operators (e.g., the set
union operator) without explanation.
An interpretation (also called a possible world) is a pair h∆, .I i, where ∆ is a
domain of elements, and .I is a mapping function that admits all prior knowledge, and
maps each individual into a domain element in ∆, each concept into a set in ∆ and each
n-ary operator into an n-ary function in ∆. The mapping function .I is generalized for
4
terms by mapping O(a1 , . . . , an ) to OI (aI1 , . . . , aIn ). Similar to terms and assertions,
interpretations can also be considered as individuals to be studied.
It is important to emphasize that an interpretation has to admit all prior knowledge.
For instance, since we assume set theory, suppose that an interpretation maps two individuals x and y as the same element a in the domain, then the concepts {x} and {y}
must be interpreted as {a}, and x = y must be interpreted as a = a.
Let I be an interpretation and a = b an assertion. We say that I is a model of
a = b, denoted by I |= a = b iff .I (a) = .I (b), also written aI = bI . Let KB be a
knowledge base. We say that I is a model of KB, denoted by I |= KB, iff I is a model
of all assertions in KB. We say that an assertion A is a property of KB, denoted by
KB |= A, iff all models of KB are also models of A. In particular, we say that an
assertion A is a tautology iff it is modeled by all interpretations.
Since we assume set theory as our prior knowledge, we directly borrow some set
theoretic constructs. For instance, we can use ∪(C1 , C2 ) (also written as C1 ∪ C2 ) to
denote a new concept that unions two concepts C1 and C2 . Applying this to assertions,
we can see that assertions of the primitive form (1) can indeed represent many important features in knowledge representation. For instance, the membership assertion,
stating that an individual a is an instance of a concept C, is the following assertion
∈ (a, C) = ⊤ (also written as a ∈ C). The containment assertion, stating that a concept C1 is contained by another concept C2 , is the following assertion ⊆ (C1 , C2 ) = ⊤
(also written as C1 ⊆ C2 ). The range declaration, stating that the range of an operator O operating on some concepts C1 , . . . , Cn equals to another concept C, is the
following assertion O(C1 , . . . , Cn ) = C.
4 Extensibility via Definitions
As argued in the introduction section, extensibility is a critical issue for knowledge
representation. In assertional logic, we use definitions for this purpose. Definitions
are (schema) assertions used to define new syntactic objects (including individuals,
concepts and operators) based on existing ones. Once these new syntactic objects are
defined, they can be used to define more. Note that definitions are nothing extra but
special kinds of knowledge (i.e. assertions).
We start with defining new individuals. An individual definition is an assertion of
the form
a = t,
(2)
where a is an atomic individual and t is a term. Here, a is the individual to be defined.
This assertion claims that the left side a is defined as the right side t. For instance,
0 = ∅ means that the individual 0 is defined as the empty set.
Defining new operators is similar to defining new individuals except that we use
schema assertions instead. Let O be an operator to be defined and (C1 , . . . , Cn ) its
domain. An operator definition is a schema assertion of the form
O(C1 , . . . , Cn ) = T,
where T is a schema term that mentions concepts only from C1 , . . . , Cn .
5
(3)
Since a schema assertion represents a set of assertions, essentially, an operator
definition of the form (3) defines the operator O by defining the value of O(a1 , . . . , an )
one-by-one, where ai ∈ Ci , 1 ≤ i ≤ n. For instance, for defining the successor
operator Succ, we can use the schema assertion Succ(N) = {N, {N}}, meaning that,
for every natural number n, the successor of n, is defined as {n, {n}}.
Defining new concepts is somewhat different. As concepts are essentially sets, we
directly borrow set theory notations for this purpose. There are four ways to define a
new concept.
Enumeration Let a1 , . . . , an be n individuals. Then, the collection {a1 , . . . , an } is a
concept, written as
C = {a1 , . . . , an }.
(4)
For instance, we can define the concept Digits by Digits = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}.
Operation Let C1 and C2 be two concepts. Then, C1 ∪ C2 (the union of C1 and
C2 ), C1 ∩ C2 (the intersection of C1 and C2 ), C1 \ C2 (the difference of C1 and C2 ),
C1 × C2 (the Cartesian product of C1 and C2 ), 2C1 (the power set of C1 ) are concepts.
Operation can be written by assertions as well. For instance, the following assertion
C = C1 ∪ C2
(5)
states that the concept C is defined as the union of C1 and C2 . As an example, one can
define the concept M an by M an = Human ∩ M ale.
Restricted Comprehension Let C be a concept and A(C) a schema assertion that only
mentions concept C. Then, individuals in C satisfying A, denoted by {x ∈ C|A(x)}
(or simply C|A(C)), form a concept, written as
C ′ = C|A(C).
(6)
For instance, we can define the concept M ale by M ale = {Animal |Sex(Animal) =
male}, meaning that M ale consists of all animals whose sexes are male.
Replacement Let O be an operator and C a concept on which O is well defined. Then,
the individuals mapped from C by O, denoted by {O(x) | x ∈ C} (or simply O(C)),
form a concept, written as
C ′ = O(C).
(7)
For instance, we can define the concept P arents by P arents = P arentOf (Human),
meaning that it consists of all individuals who is a P arentOf some human.
Definitions can be incremental. We may define some syntactic objects first. Once
defined, they can be used to define more. One can always continue with this incremental process. For instance, in arithmetic, we define the successor operator first. Once
defined, it can be used to define the add operator, which is further served as a basis to
define more.
Since terms and assertions can be considered as individuals, we can define new type
of terms and assertions by definitions. As an example, we extend assertions of the form
(1) into multi-assertions by using Cartesian product. We first define multi-assertions of
a fixed number of assertions. Given a number n, we define a new operator Mn with
arity n by the following schema assertion:
Mn (C1 = D1 , . . . , Cn = Dn ) = (C1 , . . . , Cn ) = (D1 , . . . , Dn ),
6
where Ci , Di , 1 ≤ i ≤ n, are concepts of terms. Notice that, (C1 , . . . , Cn ) =
(D1 , . . . , Dn ) is a single assertion of the form (1). In this sense, an n-ary multiassertion is just a syntax sugar. Then, we define the concept of multi-assertions:
[
M ulti − Assertion =
Mi (A1 , . . . , Ai ),
1≤i≤∞
where A1 , . . . , Ai are i copies of standard assertions. For convenience, we use Assertion1 , . . . , Assertionn
to denote an n-ary multi-assertion. Once multi-assertion is defined, it can be used to
define more syntactic objects.
As an example, we use multi-assertion to define nested assertions. We first define
nested terms as follows:
N ested − T erm =
N − T erm =
T erm ∪ N − T erm
Op(N ested − T erm).
Then, nested assertions can be defined as
N ested − Assertion = N ested − T erm = N ested − T erm.
Again, once nested assertion is defined, it can be used as basis to define more, so
on and so forth. Using nested assertions can simplify the representation task. However,
one cannot overuse it since, essentially, every use of a nested term introduces a new
individual.
5 Embedding FOL into Assertional Logic
In the previous section, we show how to extend assertions of the primitive form (1)
into multi-assertions and nested assertions. In this section, we continue with this task to
show how to define more complex forms of assertions with logic connectives, including
propositional connectives and quantifiers.
We start with the propositional case. Let A be the concept of nested assertions.
We introduce a number of operators over A in assertional logic, including ¬(A) (for
negation), ∧(A1 , A2 ) (for conjunction), ∨(A1 , A2 ) (for disjunction) and → (A1 , A2 )
(for implication).
There could be different ways to define these operators in assertional logic. Let
a = a′ and b = b′ be two (nested) assertions. The propositional connectives are
defined as follows:
¬(a = a′ ) =
∧(a = a′ , b = b′ ) =
′
′
∨(a = a , b = b ) =
→ (a = a′ , b = b′ ) =
{a} ∩ {a′ } = ∅
({a} ∩ {a′ }) ∪ ({b} ∩ {b′ }) = {a, a′ , b, b′ }
({a} ∩ {a′ }) ∪ ({b} ∩ {b′ }) 6= ∅
({a, a′ } \ {a} ∩ {a′ }) ∪ ({b} ∩ {b′ }) 6= ∅,
where a 6= a′ is used to also denote ¬(a = a′ ). One can observe that the ranges of
all logic operators are nested assertions. Hence, similar to multi-assertion and nested
assertion, propositional logic operators are syntactic sugar as well.
7
Now we consider to define operators for quantifiers, including ∀ (for the universal
quantifier) and ∃ (for the existential quantifier). The domain of quantifiers is a pair
(C, A(C)), where C is a concept and A(C) is a schema assertion that only mentions
C.
The quantifiers are defines as follows:
∀(C, A(C))
= C|A(C) = C
(8)
∃(C, A(C))
= C|A(C) 6= ∅
(9)
Intuitively, ∀(C, A(C)) is true iff those individuals x in C such that A(x) holds equals
to the concept C itself, that is, for all individuals x in C, A(x) holds; ∃(C, A(C)) is
true iff those individuals x in C such that A(x) holds does not equal to the empty set,
that is, there exists at least one individual x in C such that A(x) holds. We can see that
the ranges of quantifiers are nested assertions as well. In this sense, quantifiers are also
syntactic sugar of the primitive form.
Note that quantifiers defined here are ranging from an arbitrary concept C. If C is
a concept of all atomic individuals and all quantifiers range from the same concept C,
then these quantifiers are first-order. Nevertheless, the concepts could be different. In
this case, we have many-sorted FOL. Moreover, C could be complex concepts, e.g., a
concept of all possible concepts. In this case, we have monadic second-order logic. Yet
C could be many more, e.g., a concept of assertions, a concept of concepts of terms
etc. In this sense, the quantifiers become high-order. Finally, the biggest difference
is that C can even be a concept of assertions so that quantifiers in assertional logic
can quantify over assertions (corresponding to formulas in classical logics), while this
cannot be done in classical logics including high-order logic.
It can be verified that all tautologies in propositional logic and FOL (e.g., DeMorgan’s laws) are also tautologies in assertional logic. For space reasons, we leave
the theorems and their proofs to a full version of this paper.
6 Incorporating Probability and More
Probability is another important building block for knowledge representation. In the
last several decades, with the development of uncertainty in artificial intelligence, a
number of influential approaches [4, 8, 10, 17, 18, 20] have been developed, and important applications have been found in machine learning, natural language processing
etc.
Normally, to incorporate probability in logic, one has to complete redefine the
whole semantics since the integrations between probability and logic connectives and
quantifiers are complicated. In this section, we show how this can be easily done in
assertional logic. The key point is, although the interactions between logic and probability are complicated, their interactions with assertions of the basic form (1) is relatively simple. As shown in the previous section, the interactions between logic and
assertions can be defined by a few lines. In this section, following Gaifman’s idea [8],
we show that this is indeed the case for integrating assertions with probability as well.
Then, the interactions between logic and probability will be automatically established
via assertions.
8
6.1 Integrating assertions with probability
Since operations over real numbers are involved in defining probability, we need to
assume a theory of real number as our prior knowledge.
Gaifman [8] proposed to define the probability of a logic sentence by the sum of
the probabilities of the possible worlds satisfying it. Following this idea, in assertional
logic, we introduce an operator P r (for probability) over the concept A of assertions.
The range of P r is the concept of real numbers. For each possible world w, we assign
an associated weight Ww , which is a positive real number. Then, for an assertion A,
the probability of A, denoted by P r(A), is define by the following schema assertion:
P r(A) =
Σw,w|=A Ww
.
Σw Ww
(10)
This definition defines the interactions between probability and assertions. In case that
there are a number of infinite worlds, we need to use measure theory. Nevertheless,
this is beyond the scope of our paper.
Once we have defined the probability P r(A) of an assertion A as a real number, we
can directly use it inside other assertions. In this sense, P r(A) = 0.5, P r(A) ≥ 0.3,
P r(A) ≥ P r(∀(C, B(C))) − 0.3, P r(A) × 0.6 ≥ 0.4 and P r(P r(A) ≥ 0.3) ≥ 0.3
are all valid assertions. We are able to vefiry some properties about probability, for
instance, Kolmogorov’s first and second probability axioms.
We also extend this definition for conditional probability. We again introduce a
new operator P r over pairs of two assertions. Following a similar idea, the conditional probability P r(A1 , A2 ) of an assertion A1 providing another assertion A2 , also
denoted by P r(A1 |A2 ), is defined by the following schema assertion:
P r(A1 |A2 ) =
Σw,w|=A1,w|=A2 Ww
.
Σw,w|=A2 Ww
(11)
Again, once conditional probability is defined as a real number, we can use it arbitrarily inside other assertions. Similarly, we can verify some properties about conditional probabilities, including the famous Bayes’ theorem, i.e.,
P r(A1 ) × P r(A2 |A1 ) = P r(A2 ) × P r(A1 |A2 ).
for all assertions A1 and A2 .
6.2 Interactions between logic and probability via assertions
Although we only define probabilities for assertions of the basic form, the interactions
between probability and other building blocks, e.g., logic, are automatically established
since assertions connected by logic operators can be reduced into the primitive form.
In this sense, we can investigate some properties about the interactions between logic
and probability. For instance, it can be observed that Kolmogorov’s third probability
axiom is a tautology in assertion logic. That is, let A1 , . . . , An be n assertions that are
pairwise disjoint. Then, P r(A1 ∨ · · · ∨ An ) = P r(A1 ) + · · · + P r(An ).
9
It can be verified that many axioms and properties regarding the interactions between logic and probability are tautologies in assertional logic, for instance, the additivity axiom: P r(φ) = P r(φ ∧ ψ) + P r(φ ∧ ¬ψ) and the distributivity axiom: φ ≡ ψ
implies that P r(φ) = P r(ψ), for any two assertions φ and ψ. In this sense, assertional
logic can also be used to validate existing properties about the interactions of logic and
probability. In addition, it may foster new discoveries, e.g., the interactions between
higher-order logic and probability and some properties about nested probabilities.
Note that we do not intend to reinvent the wheel of defining probability nor its
interactions with logic. All definitions about (conditional) probability are borrowed
from the literature. Instead, we take probability as a case study to show how one
building block (e.g., logic) and another (e.g., probability) can be interacted through
assertions without going deeper into the interactions between themselves.
6.3 More building blocks
More critically, there are many more important building blocks to be incorporated. It
is barely possible to clarify the interactions among them all. Nevertheless, it becomes
possible to unify them altogether in assertional logic as one only needs to consider the
interactions between these building blocks and the basic form of assertions separately.
Consequently, the interactions among these building blocks themselves will be automatically established via assertions, as what we did for unifying logic and probability.
As another case study, we consider how to formalize time in assertional logic. Time
itself can be understood in different ways such as time points, time interval, LTL and
CTL [1, 7, 19]. Following the same idea of incorporating probability, we only need
to consider the interactions between time and assertions. In this paper, we only report
the simple case of integrating assertions with time points. Let T p be a concept of time
points. We introduce a new operator t whose domain is a pair (I, T p). Intuitively,
t(i, tp), i ∈ T , tp ∈ T p, is the value of individual i at time point tp. Then, we introduce temporal formulas, a new Boolean operator T whose domain is a pair (A, T p),
by the following schema assertion:
T (a = b, tp) = t(a, tp) = t(b, tp).
(12)
Then, the interactions between time points and logic connectives and probability can
be automatically established. We are able to investigate some properties. For instance,
for all assertions A and B, T (A, tp) = ⊤ iff T (¬A, tp) = ⊥; T (A ∧ B, tp) = ⊤ |=
T (A, tp) = ⊤ etc. Hence, we have an integrated formalism combing logic, probability
and time points in assertional logic.
7 Discussion, Related Work and Conclusion
In this paper, we argue that, for the purpose of knowledge representation, classical firstorder logic has some critical issues, including simplicity, flexibility, expressivity and
extensibility. To address these issues, we propose assertional logic instead, in which the
syntax of an application domain is captured by individuals (i.e., objects in the domain),
concepts (i.e., groups of objects sharing something in common) and operators (i.e.,
10
connections and relationships among objects), and knowledge in the domain is simply
captured by equality assertions of the form a = b, where a and b are terms.
In assertional logic, without redefining the semantics, one can extend a current system with new syntactic objects by definitions, which are special kinds of knowledge
(i.e., assertions). Once defined, these syntactic objects can be used to define more.
This can be done for assertional logic itself. We extend the primitive form of assertional logic with multi-assertions and nested assertions as well as logic connectives
and quantifiers. We further consider how to extend assertional logic with other important AI building blocks. The key point is that, when one wants to integrate a new
building block in assertional logic, she only needs to formalize it as syntactic objects
(including individuals, concepts and operators) and defines its interactions with the basic form of assertions (i.e., a = b). Then, the interactions between this building block
and others will be automatically established since all complicated assertions can essentially be reduced into the basic form. As a case study, we briefly discuss how to
incorporate probability and time points in this paper.
Of course, assertional logic is deeply rooted in first-order logic. Individuals, concepts and operators are analogous to constants, unary predicates and functions respectively, and assertions are originated from equality atoms. Nevertheless, they differ from
many essential aspects. Firstly, individuals can be high-order objects, e.g., concepts
and assertions, so are concepts and operators. Secondly, assertional logic is naturally
many-sorted, that is, the domain of an operator can be a tuple of many different concepts including high-order ones. Thirdly, concepts play a central role in assertional
logic, which is natural for human knowledge representation. While concepts can be
formalized as unary predicates in FOL, they are not specifically emphasized. Fourthly,
in assertional logic, all kinds of knowledge are uniformly formalized in the same form
of equality assertions. As shown in Section 5, complicated logic sentences are defined
as equality assertions as well by embedding connectives and quantifiers as operators
over assertions. Fifthly, following the above, although connectives, quantifiers and
nesting can be defined in assertional logic, they are not considered as primitive constructs. In this sense, they will only be used on demand when necessary. We argue that
this is an important reason that makes assertional logic simpler than FOL. Sixthly, in
assertional logic, the simple form of a = b is expressive as a and b can be high-order
constructs and can be inherently related within the rich syntactic structure. In contrast,
equality atoms in FOL do not have this power. Last but not least, assertional logic
directly embraces extensibility within its own framework by syntactic definitions. For
instance, to define quantifiers, assertional logic only needs two lines (see Equations 8
and 9) without redefining a whole new semantics, which is much simpler than FOL.
Assertional logic is also inspired by many approaches in symbolic AI. Traditionally in FOL, there is a strict hierarchy from the term level to the formula level. To
some extent, this is broken in situation calculus [15, 16] and game description language [21] that have to quantify over situations, actions and fluents and to directly
talk about whether a fluent holds in a particular situation by a meta-predicate Hold.
Assertional logic goes much further by completely demolishing the distinction and hierarchy between term level, predicate level and formula level. In assertional logic, one
can flexibly use, e.g., atoms and predicates in the scope of a function as long as they
match its domain definition. Also, one can quantify over any well-defined concepts,
11
including a concept of assertions. This makes assertional logic even more expressive
than high-order logic that cannot quantifier over formulas.
Another inspiration comes from the family of description logics [2], where the terminologies individuals and concepts are borrowed from. The family of description
logics allows a certain level of extensibility. By interpreting individuals, concepts and
roles as domain elements, unary predicates and binary predicates respectively, one can
extend the basic description logics with more building blocks (e.g., nominal, number
restrictions, inverse and transitive roles etc.) based on the same foundational semantics. Also, one can freely assemble those building blocks into different fragments of
description logics such as ALC, SHIQ, SHION and so on. However, many important
AI building blocks, e.g., actions, probability, time, rules, etc. are still difficult to be
incorporated by this method. Some interesting pioneering work have been done to
consider more extensibility in description logics [3, 5, 9, 14]. Nevertheless, they differ from assertional logic that embraces extensibility at a syntactic level instead of a
semantic one.
This paper is only concerned with the representation task and the definition task.
We leave the reasoning task to our future work. Certainly, complete reasoning in assertional logic is undecidable as FOL can be embedded in it. Traditionally, the way
to address this issue is to find decidable fragments. Nevertheless, we shall propose an
alternative approach that focuses on efficient but not necessarily complete reasoning.
We have developed such an approach, in which the flexibility and extensibility of assertional logic play a critical role. We shall present this in another paper. Nevertheless, we
argue that representation and definition are worth study on their own merits. Such successful stories include entity-relationship diagram, semantic network and many more.
Besides, extending assertional logic with some important AI building blocks, e.g., actions and their effects, is indeed challenging and worth pursuing.
Acknowledgement
The author gratefully acknowledges Prof. Fangzhen Lin’s comments on a first draft of
this paper.
References
[1] James F. Allen. Maintaining knowledge about temporal intervals. Commun. ACM,
26(11):832–843, November 1983.
[2] Franz Baader, Diego Calvanese, Deborah L. McGuinness, Daniele Nardi, and
Peter F. Patel-Schneider, editors. The Description Logic Handbook: Theory, Implementation, and Applications. Cambridge University Press, New York, NY,
USA, 2003.
[3] Franz Baader and Philipp Hanschke. A scheme for integrating concrete domains
into concept languages. In Proceedings of the 12th International Joint Conference
12
on Artificial Intelligence - Volume 1, IJCAI’91, pages 452–457, San Francisco,
CA, USA, 1991. Morgan Kaufmann Publishers Inc.
[4] Fahiem Bacchus. Representing and Reasoning with Probabilistic Knowledge: A
Logical Approach to Probabilities. MIT Press, Cambridge, MA, USA, 1990.
[5] Alexander Borgida. Extensible knowledge representation: the case of description
reasoners. J. Artif. Intell. Res. (JAIR), 10:399–434, 1999.
[6] Ronald J. Brachman and Hector J. Levesque. Knowledge Representation and
Reasoning. Elsevier, 2004.
[7] Edmund M. Clarke and E. Allen Emerson. Design and synthesis of synchronization skeletons using branching-time temporal logic. In Logic of Programs,
Workshop, pages 52–71, London, UK, UK, 1982. Springer-Verlag.
[8] Haim Gaifman. Concerning measures in first order calculi. Israel J. Math., 2:1–
18, 1964.
[9] Giuseppe De Giacomo, Maurizio Lenzerini, and Riccardo Rosati. Higher-order
description logics for domain metamodeling. In Proceedings of the Twenty-Fifth
AAAI Conference on Artificial Intelligence, AAAI’11, pages 183–188. AAAI
Press, 2011.
[10] Theodore Hailperin. Probability logic. Notre Dame J. Formal Logic, 25:198–212,
1984.
[11] Paul Halmos. Naive Set Theory. Van Nostrand, 1960. Reprinted by SpringerVerlag, Undergraduate Texts in Mathematics, 1974.
[12] Joseph Y. Halpern. An analysis of first-order logics of probability. Artif. Intell.,
46(3):311–350, 1990.
[13] Thomas Keller, Patrick Eyerich, and Bernhard Nebel. Task planning for an autonomous service robot. In Proceedings of the 33rd Annual German Conference
on Advances in Artificial Intelligence, KI’10, pages 358–365, Berlin, Heidelberg,
2010. Springer-Verlag.
[14] O. Kutz, C. Lutz, F. Wolter, and M. Zakharyaschev. E-connections of abstract
description systems. Artificial Intelligence, 156(1):1–73, 2004.
[15] Hector Levesque, Fiora Pirri, and Ray Reiter. Foundations for the situation calculus. Electronic Transactions on Artificial Intelligence, Vol. 2(1998), Issue 3-4:pp.
159–178, 1991.
[16] Fangzhen Lin. Situation calculus. In Handbook of Knowledge Representation,
pages 649–669. 2008.
[17] Brian Christopher Milch. Probabilistic Models with Unknown Objects. PhD
thesis, Berkeley, CA, USA, 2006. AAI3253991.
13
[18] Judea Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA,
1988.
[19] Amir Pnueli. The temporal logic of programs. In Proceedings of the 18th Annual Symposium on Foundations of Computer Science, SFCS ’77, pages 46–57,
Washington, DC, USA, 1977. IEEE Computer Society.
[20] Matthew Richardson and Pedro Domingos. Markov logic networks. Machine
learning, 62(1-2):107–136, 2006.
[21] Michael Thielscher. GDL-III: A proposal to extend the game description language to general epistemic games. In ECAI 2016 - 22nd European Conference on
Artificial Intelligence, 29 August-2 September 2016, The Hague, The Netherlands
- Including Prestigious Applications of Artificial Intelligence (PAIS 2016), pages
1630–1631, 2016.
[22] Frank van Harmelen, Vladimir Lifschitz, and Bruce W. Porter, editors. Handbook
of Knowledge Representation, volume 3 of Foundations of Artificial Intelligence.
Elsevier, 2008.
14
| 2cs.AI
|
Two Distinct Seasonally Fractionally
Differenced Periodic Processes
arXiv:1803.03154v1 [math.ST] 8 Mar 2018
1
Ahmed BENSALMA 1,2
ENSSEA Pôle universitaire Koléa, Tipaza (ex: INPS 11 Doudou mokhtar
Benaknoun) Algeria)
2
Faculty of Mathematics, University of Science and Technology Houari
Boumediene, Algeria
Abstract
This article is devoted to study the effects of the S-periodical fractional differencing filter (1 − LS )Dt . To put this effect in evidence, we
have derived the periodic auto-covariance functions of two distinct univariate seasonally fractionally differenced periodic models. A multivariate
representation of periodically correlated process is exploited to provide
the exact and approximated expression auto-covariance of each models.
The distinction between the models is clearly obvious through the expression of periodic auto-covariance function. Besides producing different
auto-covariance functions, the two models differ in their implications.
In the first model, the seasons of the multivariate series are separately
fractionally integrated. In the second model, however, the seasons for the
univariate series are fractionally co-integrated. On the simulated sample,
for each models, with the same parameters, the empirical periodic autocovariance are calculated and graphically represented for illustrating the
results and support the comparison between the two models.
1
Introduction
Since their introduction by Gladyshev (1961, 1963) much attention has
been given to periodically correlated processes. The interest, for such
processes is due to their potential use in modeling of cyclical phenomena
appearing in hydrology, climatology and in econometrics. Following pioneer work of Gladyshev (1963), an important part of the literature has
been devoted to the periodically correlated discrete time processes. A discrete time process is periodically correlated, if there is a non zero integer
S such that
E (Xt+S ) = E (Xt ) and Cov(Xt1 +S , Xt2 +S ) = Cov(Xt1 , Xt2 ).
A review of the periodically correlated discrete time processes is proposed in Lund and Basawa (1999), Bentarzi and Hallin (1994) give invertibility conditions for periodic moving average.. A large part of the
literature on the subject is devoted to the periodic ARM A (P ARM A)
models, which have the following representation:
Xt −
pt
X
i=0
φi,t Xt−i =
qt
X
θj,t ut−j , t = 0, ±1, ±2, ....,
j=0
1
where ut is a zero-mean white noise with variance σ 2t . Among searchers
who were interested with the periodic autoregressifs processes not periodically stationary, we cite Boswijk and Franses (1995) which studied the
problem of the presence of a unit root in a periodic autoregression model
of order p (P AR(p)) and Boswijk, Franses and Haldrup (1997) which studied the presence of multiple unit roots in a periodic autoregression model
of order p. All work cited above were made under the assumption that
the processes are periodically integrated of order zero (P I(0)), integrated
of order one (I(1)) or periodically integrated of order one ((P I(1)). However currently, it well-known that in the scientific fields mentioned above
(hydrology, meteorology, econometrics) much of sets of data that have
a certain periodicity; have also a long range dependence (or long memory). Such phenomena can be modeled by stationary processes. The
stationary processes with seasonal long memory are well know (see for example Gray, Zhang and Woodward (1989): Garma models; Purter-Hudak
(1990):Seasonal ARF IM A; Oppenheim, G. and al (2000); ould Haye and
al (2003) for references, properties and simulations). Another alternative, to take account of certain periodic phenomena with long memory
is to consider nonstationary models (but periodically stationary) such as
the periodically correlated processes with long memory. The periodically
correlated processes, within the meaning of Gladyshev (1963), with long
memory did not receive much attention on behalf of the statisticians and
the probabilists. Among works associating periodicity within the meaning
of Gladyshev (1963) and the presence of long memory we cite, Hui and
Li (1995), Franses and Ooms (1997), Ooms and Franses (2001).
For modelling of the Hong Kong United Christian Hospital attendance
series, Hui and Li (1995) propose a 2-periodic correlated process,
(1 − L)dt Yt = ut ,
(1.1)
where {ut , t ∈ Z} is a zero mean white noise with variance σ2t , and dt the
2-periodic fractional parameter. The empirical series yt concerns seventy
five (approximately one and half years) data on the average number of
people entering the emergency unit on weekday and weekend.
On the other hand, in order to analyzes the long-memory properties in
the conditional mean of the quarterly inflation rate in the United Kingdom
Franses and Ooms (1997) propose a 4-periodic correlated process,
Yt = (1 − L)−dt ut ,
(1.2)
where {ut , t ∈ Z} is defined as above and dt is the 4-periodic fractional
parameter.
Finally, for the monthly empirical data, which concern the log transformed data of the monthly mean river flow in cubic feet per second,
Ooms and Franses (2001) propose to use the seasonal periodic fractional
operator defined, in simple framework as follows,
Yt = (1 − LS )−Dt ut , S = 12
(1.3)
where {ut , t ∈ Z} is defined as above and S = 12.
The main difference between, the one hand, the models (1.1) and (1.2)
and the other hand, the model (1.3), is in the unit of lag to which the
2
fractional difference operator is applied. In the models (1.1) and (1.2)
the fractional difference operator was applied to weekly and quarterly
lags, respectively, corresponding to the basic time interval of the time
series analyzed. In the model (1.3) the fractional difference operator was
applied to yearly, which is the seasonal lag of the time series analyzed.
Indeed, by using a binomial expansion for the difference operator (1−L)dt ,
(1−L)−dt , (1−LS )−dt we can rewrite, respectively, the models (1.1), (1.2)
and (1.3) as the following,
∞
X
j=0
Γ (j − dt )
Yt−j = ut ,
Γ (j + 1) Γ (−dt )
Yt =
∞
X
Γ (j + dt )
ut−j ,
Γ (j + 1) Γ (dt )
(1.5)
Γ (j + Dt )
ut−Sj .
Γ (j + 1) Γ (Dt )
(1.6)
j=0
Yt =
∞
X
j=0
where
Γ (z) =
(1.4)
R +∞
0
sz−1 e−z ds,
∞
if z > 0
if z = 0,
if z < 0, Γ (z) is defined in terms of the above expressions and the recurrence formula zΓ (z) = Γ (z + 1).
While, the invertibility and stationarity conditions of the model (1.3)
are known (see Ooms and Franses 2001), apart when dt = d is a constant,
nothing is clear about the models (1.1) and (1.2). More precisely, no thing
is clear about the stationarity conditions for the model (1.1), because his
infinite moving average representation is unknown and no thing is clear
about the invertibility conditions for the model (1.2), because his infinite
autoregressive representation is unknown. The model (1.3) is invertible
and stationary if −0.5 < Dt < 0.5 and it is easy to show in this case that
the infinite autoregressive representation of the process yt is given by
∞
X
j=0
Γ (j − Dt )
Yt−Sj = ut .
Γ (j + 1) Γ (−Dt )
For the model (1.4), at any case, in general, we have,
Yt 6=
∞
X
j=0
Γ (j + dt )
ut−j ,
Γ (j + 1) Γ (dt )
and for the model (1.5), at any case, in general, we have,
∞
X
j=0
Γ (j − dt )
Yt−j 6= ut .
Γ (j + 1) Γ (−dt )
For the particular periodic ARF IM A(0, dt , 0), namely (1 − L)dt yt =
ut , ut ∼ i.i.d(0, σ 2t ), the infinite moving average representation is unknown. In this paper, we give the closed form of this representation. It
3
is important to known such representation in order to deduce the stationarity condition of this type of model. Unfortunately, the closed form
obtained is not easy to handle due to her parametric complexity (see Appendix).
Since the P ARF IM A(p, dt , q) is not easy to handle. The work that
we present in this article is concerned only on the Seasonal periodical fractional operator, namely (1−LS )Dt . More Precisely, in this work we are interested in certain theoretical properties of the SP ARF IM A(p, 0, 0)(0, Dt , 0)S
(Seasonal periodic ARF IM A). The study of the theoretical properties of this class of models remains to be made; because among works
which evoke this class, only one exists; that is of Franses and Ooms
(2001). The work of Franses and Ooms has to consist in adjusting a
SP ARF IM A(p, 0, 0)(0, Dt , 0)S to a set of real data. Precisely the model
considered by Ooms and Franses is defined as follows:
φt (L) (Xt − µt ) = η t , t ∈ N∗ , with η t = (1 − LS )−Dt ut ,
where µt is S-periodical constant such as µt = µt+S , φt (L) = 1 − φt,1 L −
φt,2 L2 − ... − φt,p Lp . The parameters φt,i i = 1, ..., p are periodic functions
in t, and η t a white noise seasonally fractionally integrated of order Dt ,
where Dt is S-periodical fractional parameter. The model above, if 0 ≤
Dt < 0.5, ∀t can be written as follows:
(1 − LS )Dt φt (L) (Xt − µt ) = ut , t ∈ Z.
(M odel(I))
There is another class of models SP ARF IM A(p, 0, 0)(0, Dt , 0)S distinct from that used by Franses and Ooms (2001); this class is defined as
follows:
φt (L)(1 − LS )Dt (Xt − µt ) = ut , t ∈ Z,
(M odel(II))
where µt , φt (L), Dt are defined like above. These two classes coincide, only if Dt = D, ∀t, since, generally, the composition of φt (L) and
(1 − LS )Dt is not necessarily commutative. To convince, it is sufficient
to notice that the S-variate representation of the model (I) is a V ARF I
model (vector autoregressive model, driven by fractionally integrated innovation) whereas the multivariate writing of the model (II) is a F IV AR
model (fractionally integrated vector autoregression) (see Rebecca Sela
and Clifford Mr. Hurvich (2008)). These two distinct classes, generalize
the univariate model ARF IM A, the first is closely related to the cointegrated processes, whereas the second is closely related to the integrated
processes. Consequently, in our case, the model (I) is closely related to the
cointegrated season and the model (II) is closely related to the integrated
season.
In order to distinguish between the model (I) and (II), we note them,
respectively as the following: P AR(p) − P SF I(Dt ) and P SF I(Dt ) −
P AR(p). The rest of this paper is organized as follows: section 2 is
devoted to defined two class of processes; the periodic autoregressive
of order p process with periodic seasonal fractional integrated of order
Dt innovation, namely P AR(p) − P SF I(Dt ) and the periodic seasonal
fractional integrated process, periodic autoregressive of order p, namely
P SF I(Dt ) − P AR(p). In section 3, for each model defined in section 2,
4
we provide the exact and approximated expression of the periodic autocovariances function. In the section 4, on the simulated samples for each
model, with the same parameters for the model (I) and (II), the empirical periodic autocovariances are calculated and graphically represented
for illustrating the theoretical results and comparison between the two
models.
Without restricting the generality, we suppose that all processes defined below have zero mean.
2
Representation and notation
2.1 S-periodical seasonally fractionally integrated,
periodic autoregressive process (P SF I(Dt)−P AR(p))
A periodically correlated process {Yt , t ∈ Z} is said S-periodical seasonally
fractionally integrated of order Dt , periodic autoregressive of order p; if
it has the following representation:
Φt (L)(1 − LS )Dt Yt = ut ,
t ∈ Z,
(2.1)
where {ut , t ∈ Z} is a zero mean white noise with variance σ2t and (1 −
LS )Dt are defined like above. φt (L) = 1 − φt,1 L − φt,2 L2 − ... − φt,p Lp
where φt,1 , ..., φt,p are S-periodical parameters.
Letting Yτ = (Y1,τ , ..., Ys,τ , ..., YS,τ )′ and uτ = (u1,τ , ..., us,τ , ..., uS,τ )′
with Ys,τ = Ys+Sτ and us,τ = us+Sτ then the process (2.1) can be rewritten in the S variate form
D
Φ0 ∇S (L)Y τ −
P
X
D
Φi ∇S (L)Y τ −i = uτ ,
(2.2)
i=1
+ 1, with [x] denotes the smallest integer than or equal
where P = p+1
S
D
to x, ∇S (L) is defined like above. The autoregressive coefficient matrices
are given by
1
s = j,
0
s < j,
(Φ0 )s,j =
−φs−j,j
s > j,
and
(Φi )s,j = φiS+s−j,s , s, j = 1, ..., S and 1 ≤ i ≤ P.
The periodic stationarity condition of the model (2.2) is the same as
the stationarity condition of it equivalent fractional integrated vector autoregression, namely F IV AR, (Rebecca Sela and Clifford Hurvich (2008))
representation (2.2), which means that the roots of the determinantal
equation
!
P
X
P
−1
P−i
det IS z −
Φ0 Φi z
= 0,
i=1
are less than 1 in absolute value (Hannan (1970), Fuller (1976)) and
0 ≤ Ds < 0.5, for all s = 1, ..., S,
5
(Hosking (1981)). If the process (2.2) is stationary, then it has an infinite
moving average representation given by
D
Y τ = ∇S (L)−1 Φ (L)−1 uτ ,
!
!
∞
∞
P
P
j
j
Ψj L
Πj L uτ ,
=
j=0
=
∞
P
j=0
=
∞
P
j=0
j
P
Ψk Πj−k uτ −j ,
k=0
(2.3)
Cj uτ −j ,
j=0
D
where ∇S (L) = diag (1 − L)D1 , · · · , (1 − L)Ds , · · · , (1 − L)DS , Φ (L) =
∞
P
Φ0 − Φ1 L − ... − ΦP LP and [Φ (L)]−1 = Π(L) =
Πj Lj , with (Πj )j∈N
j=0
is sequence of absolutely summable matrix i.e.
∞
P
|Πj (l, k)| < ∞, ∀l ∈
j=0
{1, ..., S} and ∀k ∈ {1, ..., S}. Cj =
j
P
Ψk Πj−k with Ψj defined like
k=0
above. The ith element of Y τ , Yi,τ is written as follows
Yi,τ = (1 − LS )Di Φ (L)−1 i uτ
(2.4)
where Φ (L)−1 i is the ith rows of Φ (L)−1 . From (2.4) we see clearly
that,
Yi,τ is integrated of order Di , i = 1, ..., S.
2.2 Periodic autoregressive, S-periodical seasonally fractionally integrated process (P AR(p)−P SF I(Dt))
A periodically correlated process {Zt , t ∈ Z} is said, periodic autoregressive of order p; S-periodical seasonally fractionally integrated of order Dt
if it has the following representation:
Φt (L)Zt = (1 − LS )−Dt ut ,
t ∈ Z,
(2.5)
LS )Dt
where {ut , t ∈ Z}, (1 −
and φt (L) are defined like above. Letting Zτ = (Z1,τ , ..., Zs,τ , ..., ZS,τ )′ and uτ = (u1,τ , ..., us,τ , ..., uS,τ )′ with
Zs,τ = Zs+Sτ and us,τ = us+Sτ then the process (2.5) can be rewritten
in the S variate form
Φ0 Zτ −
P
X
−D
Φi Zτ −i = ∇S (L)uτ ,
(2.6)
i=1
+ 1, with [x] denotes the smallest integer than or equal
where P = p+1
S
(L),
Φ
to x, ∇−D
0 and Φi , i = 1, ..., P are defined like above. The
S
model (2.6) is vector autoregression with fractional integrated innovation, namely V ARF I (Rebecca Sela and Clifford Hurvich (2008)). The
6
periodic stationarity condition of the model (2.6) is the same than the
model (2.2). The ith relation of (2.6) is written
(Φ(L))i Zτ = (1 − LS )−Di ui,τ
where Φ (L) = Φ0 −Φ1 L−...−ΦP LP and (Φ(L))i is the ith rows of Φ (L),
this means that the ith relation of (2.6) is integrated of order Di . Among
the S relations of (2.6), those which are integrated of order lower than
max Di are relations of cointegration. If all the values Di are different,
1≤i≤S
then we can say that there are (S − 1) relations of cointegrations. If
D1 = .... = DS it does not exist any relation of cointegration. Generally,
when we have D1 < D2 < ...DS−R−1 < (DS−R = DS−R+1 = ... = DS )
that means that there are (S − R − 1) relations of cointegrations between
the S seasons. If the model (2.6) is stationary, it has an infinite moving
average representation given by
−D
Zτ = Φ (L)−1 ∇S (L)uτ
!
!
∞
∞
P
P
j
j
Πj L
Ψj L uτ
=
j=0
=
∞
P
j=0
=
∞
P
j=0
j
P
Πk Ψj−k
k=0
uτ
(2.7)
Hj uτ −j ,
j=0
where Hj =
j
P
Πk Ψj−k . The ith element of Zτ , Zi,τ is written as follows
k=0
Zi,τ =
S
X
s=1
−1
Φ(L)−1
i,s
(1 − L)−DS us,τ ,
where Φ(L)
is (i, s) th element of the matrix Φ(L)−1 . Zi,τ is
i,s
written like linear combination of S independent processes, respectively,
integrated of order D1 , ..., Ds , ..., DS ; consequently Zi,τ is integrated of
order max Di (Granger 1986).
1≤i≤S
3
Periodic autocovariances
This section deals with the determination of theoretical periodic autocovariances of periodically correlated processes defined in precedent section.
3.1
P SF I(Dt) − P AR(p) periodic autocovariances
Theorem 1 Given the stationary S-variate process Yτ defined by (2.2),
we have
h
i
h
i
(3.1)
ΓYτ (h) ∼ ∆ hD−0.5 A∆ hD−0.5 , as h → ∞
where the (i, k) th element of S × S matrix A is:
Γ (1 − Di − Dk ) ′
Π ΩΠk
Γ (Dk ) Γ (1 − Dk ) i
7
with Π′i is the ith rows of the matrix Π and Ω = diag(σ21 , ..., σ 2s , ..., σ2S ).
Proof. See Ching-Fan Chung (2002).
Corollary 2 Given the process Yt defined in (2.1), we have:
Γ(1 − Ds − Ds+ν−S∗δ )
γ (s) (j) ∼ (h+δ)Ds +Ds+ν−S∗δ −1
Π′s ΩΠs+ν−S∗δ ,
Γ(Ds+ν−S∗δ )Γ(1 − Ds+ν−S∗δ )
(3.2)
where h and ν are integers such as j = h × S + ν, and j > 0, i.e. j ≡ ν [h]
with 0 ≤ ν < S − 1 and δ is defined as follows:
δ = 0,
if 1 ≤ s + ν < S,
δ = 1,
if S + 1 ≤ s + ν ≤ 2S − 1,
Π′s is the sth rows of the matrix Π and Ω = diag(σ21 , ..., σ 2s , ..., σ2S ).
Proof. The proof of the corollary, rises directly from theorem 1. From
theorem 1, we have :
(i,k)
ΓY
τ
(h) ∼ h(Di +Dk −1)
Γ(1 − Di − Dk ) ′
Πi ΩΠk ,
Γ(Dk )Γ(1 − Dk )
(3.3)
(i,k)
where ΓY τ (h) = Cov(Yi,τ , Yk,τ +h ) are the (i, k) th element of the
covariance matrix of ΓYτ (h). Moreover, it is known that
γ (s) (j) = Cov(YSτ +s , YSτ +s+j )
= Cov(Ys,τ , Ys+j,τ ).
(3.4)
Putting j = Sh + ν with 0 ≤ ν < S − 1, by replacing j by Sh + ν in (3.4),
we have
γ (s) (j) = Cov(Ys,τ , Ys+ν,τ +h ).
(3.5)
According to the value of (s + ν), the equality (3.5), becomes
( (s,s+ν)
ΓY τ
(h),
if 1 ≤ s + ν < S,
(s)
γ (j) =
(s,s+ν−S)
(h + 1), if (S + 1) ≤ s + ν ≤ 2S − 1.
ΓY τ
By using the approximation (3.3), we have:
Γ(1−Ds −Ds+ν )
Π′s ΩΠs+ν ,
h(Ds +Ds+ν −1) Γ(Ds+ν )Γ(1−D
s+ν )
if 1 ≤ s + ν < S,
γ (s) (j) ∼
Γ(1−Ds −Ds+ν−S∗ν )
(h + 1)(Ds +Ds+ν−S∗δ −1) Γ(Ds+ν )Γ(1−D
Π′s ΩΠs+ν−S∗δ ,
s+ν−S∗δ )
if (S + 1) ≤ s + ν ≤ 2S − 1,
where Π′s is the sth rows of the matrix Π and Ω = diag(σ21 , ..., σ2s , ..., σ 2S ).
From corollary 1; emerges several remarks, the most important are
Remark 3 : The periodic autocovariances γ (s) (j) s = 1, ..., S taper off at
different hyperbolic rates. If we suppose that min Di = D1 and max Di =
1≤i≤S
1≤i≤S
DS (this does not restrict the generality) than γ (1) (j), with j ≡ 0[S] has
the more speedy taper off hyperbolic rate (∝ h2D1 −1 ) and γ (S) (j), with
j ≡ 0[S] has the lowest taper off hyperbolic rate (∝ h2DS −1 ).
8
This remark will be largely clarified graphically (see section 4, couples
of figures (1a, 1b) to (5a, 5b). The advantage which offer by the periodic
process is the possibility of representing the graph of the autocovariances
in various manners. The autocovariances functions γ (s) (j), s = 1, ..., S
can be represented in the same plot (hui ad Li 1995), or separately. For
j = Sh + ν, with 0 ≤ ν < S − 1 we can also represented γ (s) (Sh + ν),
ν = 0, ..., S − 1 in the same plot. These are the three kinds of graphs
which we will use in the next section.
3.2
P AR(p) − P SF I(Dt) periodic autocovariances
Before stating the main result of this section, we need some further notation. Let Dmax = max Ds and define Before stating the main result of
1≤s≤S
this section, we need some further notation. Let Dmax = max Ds and de1≤s≤S
fine F = {1, ..., s, ..., S}, F1 = {i, i ∈ F / Di = Dmax }, with |F1 | = R and
F2 = {i, i ∈ F / Di < Dmax }, with |F2 | = S − R. We have F1 ∩ F2 = ∅
and F1 ∪ F2 = F .
Theorem 4 Given the stationary S-variate process Zτ defined by (2.6),
we have
ΓZτ (j) ∼ j 2Dmax −1 A, as j → ∞
(3.6)
where the (i, k) th element of S × S matrix A, is:
A(l, m) =
X
Γ (1 − 2Dmax )
Π(l, i)Π(m, i)σ2i
Γ (Dmax ) Γ (1 − Dmax ) i∈F
1
where Π = [Φ (1)]−1 =
∞
P
Πj and Π(l, i) is (l, i) th element of Π.
j=0
Proof. See Ching-Fan Chung (2002).
The corollary below, gives the approximated expression, as j → ∞, of
the periodic autocovariances function, γ (s) (j) = cov(ZSτ +s , ZSτ +s+j ) of
the process Zt , defined in (2.5).
Corollary 5 Given the process Zt defined in (2.5), we have,
X
Γ(1 − 2Dmax )
γ (s) (j) ∼ (h+δ)2Dmax −1
Π(s, i)Π(s+ν−Sδ, i)σ2i
Γ(Dmax )Γ(1 − Dmax ) i∈F
1
(3.7)
where h and ν are integers such as j = hS + ν, and j > 0, i.e. j ≡ ν [h]
with 0 ≤ ν < S − 1, and δ is defined as follows:
δ = 0,
if 1 ≤ s + ν < S
δ = 1,
if S + 1 ≤ s + ν ≤ 2S − 1
and Π (i, s) is the (i, s) th element of the matrix Π = [Φ (1)]−1 =
∞
P
Πj .
j=0
Proof. The proof of the corollary, rises directly from theorem 3. From
theorem 3, we have :
X
Γ (1 − 2Dmax )
(l,k)
ΓZ (j) ∼ j 2Dmax −1
Π(l, i)Π(k, i)σ 2i (3.8)
τ
Γ (Dmax ) Γ (1 − Dmax ) i∈F
1
9
(l,k)
where ΓZτ (j) = Cov(Zl,τ , Zk,τ +j ) is the (i, k) th element of the covariance matrix ΓZτ (j). Moreover, it is known that
γ (s) (j) = Cov(ZSτ +s , ZSτ +s+j )
= Cov(Zs,τ , Zs+j,τ )
(3.9)
Putting j = Sh + ν with 0 ≤ ν < S − 1, by replacing j by Sh + ν in (3.9),
we have
γ (s) (j) = Cov(Zs,τ , Zs+ν,τ +h )
(3.10)
According to the value of (s + ν), the equality (3.10), becomes
( (s,s+ν)
Γ Zτ
(h),
if 1 ≤ s + ν < S
(s)
γ (j) =
(s,s+ν−S)
(h + 1), if (S + 1) ≤ s + ν ≤ 2S − 1
Γ Zτ
By using the approximation (3.8), we have,
P
2
max )
h(2Dmax −1) Γ(DΓ(1−2D
i∈F1 Π(s, i)Π(s + ν, i)σ i ,
max )Γ(1−Dmax )
if 1 ≤ s + ν < S
(s)
γ (j) ∼
P
Γ(1−2Dmax )
2
(h + 1)(2Dmax −1)
i∈F1 Π(s, i)Π(s + ν − S, i)σ i ,
Γ(Dmax )Γ(1−Dmax )
if (S + 1) ≤ s + ν ≤ 2S − 1
Remark 6 If D1 = D2 = ... = DS the periodic autocovariances γ (s) (j)
of the model (2.2) coincide with those of model (2.8)
Remark 7 From corollary 4; we see that the periodic autocovariances
γ (s) (j) s = 1, ..., S taper off at the same hyperbolic rates.
4
Simulation
In this section we compare the finite sample of the periodic autocovariances γ (s) (j) s = 1, ..., 4 of the models (1.3), (2.1) and (2.5) for different value of D = (D1 , D2 , D3 , D4 ). The sample size for each model is
T = 1000.
The model we consider for the simulation study are
• Model A
(1 − L4 )Dt Xt = εt
which has the following S-variate representation
(1 − L)D1
0
0
0
D2
0
(1
−
L)
0
0
0
0
(1 − L)D3
0
0
0
0
(1 − L)D4
10
X = u
τ
τ
• Model B
Φt (L)(1 − L4 )Dt Yt = ut
which has the following S-variate representation
(1 − L)D1
0
1
0
0
−0.7
−0.8
0
(1
−
L)D2
1
0
0
0
−0.6
1
0
0
0
0
0
−0.4
1
0
0
• Model C
0
0
(1 − L)D3
0
(1 − L4 )Dt Φt (L)Zt = ut
which has the following S-variate representation
(1 − L)D1
0
0
0
D2
0
(1
−
L)
0
0
0
0
(1 − L)D3
0
0
0
0
(1 − L)D4
1
−0.8
0
0
0
1
−0.6
0
where uτ are i.i.d N (0, Ω) with Ω = diag(1, 1, 1, 1).
4.1
0
0
Y = u
τ
τ
0
D4
(1 − L)
Simulated Autocovariances of model A
In figures 1 to 4, we represent the empirical autocovariances function
γ (s) (j) s = 1, ..., 4 in the same plot, for the model A for different value of
11
0
0
1
−0.4
−0.7
0
Z = u
τ
0 τ
1
D = (D1 , D2 , D3 , D4 ).
Figure 1: The periodic autocovariances γ
b(s) (j), s = 1, ..., 4, for lag j = 1 to 25
with D = (0.1, 0.2, 0.3, 0.4) for model A
Figure 2: The periodic autocovariances γ
b(s) (j), s = 1, ..., 4, for lag j = 1 to 25
with D = (0.1, 0.2, 0.4, 0.4) for model A
Figure 3: The periodic autocovariances γ
b(s) (j), s = 1, ..., 4, for lag j = 1 to 25
with D = (0.1, 0.4, 0.4, 0.4) for model A
12
Figure 4: The periodic autocovariances γ
b(s) (j), s = 1, ..., 4, for lag j = 1 to 25
with D = (0.4, 0.4, 0.4, 0.4) for model A
The figures 1, illustrate well the theoretical result of theorem 1 and also
states that the periodicity is caused by the fractional parameters D =
(0.1, 0.2, 0.3, 0.4) (the auto-covariances γ (s) (j) s = 1, ..., 4 for lag j ≡ 0[4]
taper off, respectively, at hyperbolic rates, according the value of D.
4.2
Simulated autocovariances of model B
For D = (0.1, 0.2, 0.3, 0.4), the figures (1a) and (1b) represents the empirical autocovariances b
γ (s) (j) s = 1, ..., 4, respectively, in spike graph and in
line graph of the model B. The couples of figures (2a, 2b) to (5a, 5b) represents the empirical autocovariances γ
b(1) (4h+ν), ν = 0, ..., 3 to b
γ (4) (4h+ν),
ν = 0, ..., 3, for h = 1 to 25, respectively, in spike graph and line graph,
for the model B.
Figure 1a
Figure 1b
(s)
The periodic autocovariances b
γ (j), s = 1, ..., 4, for lag j = 1 to 100,
of model B, with D = (0.1, 0.2, 0.3, 0.4), taper off at different hyperbolic rates
figure 2a
figure 2b
The figure 2a and 2b represents, respectively, the speedy and the lowest
taper off hyperbolic rate of autocovariances of model (B)
13
Figure 2a
Figure 2b
(1)
The periodic autocovariances γ (4h + ν), ν = 0, ..., 3, for fixed h (h = 1 to 25)
have tendency to increase according with the value of D1 + D1+ν−4δ
Figure 3a
Figure 3b
(2)
The periodic autocovariances γ
b (4h + ν), ν = 0, ..., 3, for fixed h (h = 1 to 25)
have tendency to increase according with the value of D2 + D2+ν−4δ
Figure 4a
Figure 4b
The periodic autocovariances γ (3) (4h + ν), ν = 0, ..., 3, for fixed h (h = 1 to 24)
have tendency to increase according with the value of D3 + D3+ν−4δ
14
figure 5a
figure 5b
(4)
The periodic autocovariances γ
b (4h + ν), ν = 0, ..., 3, for fixed h (h = 1 to 25)
have tendency to increase according with the value of D4 + D4+ν−4δ
4.3
Simulated autocovariances of model C
The figures (1c) and (2c) represents the empirical autocovariances γ
b(s) (j)
s = 1, ..., 4, respectively, in spike graph and in line graph of the model C.
The difference between the periodic autocovariances γ (s) (j), s = 1, ..., 4,
for lag j = 1 to 100, decreases at the same manner, mainly because they
taper off at the same hyperbolic rates
Figure 1c
Figure 2c
The periodic autocovariances γ
b(s) (j), s = 1, ..., 4, for lag j = 1 to 100 and
D = (0.1, 0.2, 0.4, 0.4), taper off at the same hyperbolic rates
4.4 Simulated comparison between autocovariances
of model B and C
In order to compare, both autocovariances γ
b(s) (j), s = 1, ..., 4 for model
(B) and model (C) we represent them graphically in the same scale for
different value of D = (D1 , D2 , D3 , D4 ) (see below).
15
Figure 5: The periodic autocovariances b
γ (s) (j), s = 1, ..., 4, for lag j = 1 to 100,
D = (0.1, 0.2, 0.3, 0.4) for respectively, model (B) and model (C)
Figure 6: The periodic autocovariances b
γ (s) (j), s = 1, ..., 4, for lag j = 1 to 100,
D = (0.1, 0.2, 0.4, 0.4) for respectively, model (B) and model (C)
Figure 7: The periodic autocovariances b
γ (s) (j), s = 1, ..., 4, for lag j = 1 to 100,
D = (0.1, 0.4, 0.4, 0.4) for, respectively, model (B) and model (C)
16
Figure 8: The periodic autocovariances b
γ (s) (j), s = 1, ..., 4, for lag j = 1 to 100,
D = (0.4, 0.4, 0.4, 0.4) for, respectively, model (B) and model (C)
In figures 5 to 8 we plot the autocovariance sequences b
γ (s) (j), s =
1, ..., 4 of model B and model C in the same scale and with identical
parameters (Φ(L), Ω, and D). The autocovariances sequences differ dramatically. Rebecca Sela and Clifford Hurvich (2008) presents a similar
conclusion for cross-covariance sequences of bivariate F IV AR(1, D) and
V ARF I(1, D) processes with the same parameters. They point out that
the first model have the series integrated separately (in our case the seasons are integrated separately) and in the second there is cointegration
relation between the two series (in our case there are 3 cointegrations
relations between the four seasons). This fact, does not explain clearly
why there is such difference between the autocovariances of model B and
model C. Further more, the taper off hyperbolic rates of the autocovariances of model (C) is equal than the lowest tapper off hyperbolic rate of
the autocovariances of model (B), so why the autocovariance sequences
differ dramatically? The explanation is in explicit results of corollary 3.1
and corollary 3.2. Generally, in the literature of long memory models,
attention is focused on the fractional parameters (which associate with
hyperbolic tapper off of autocovariance) rather than on autoregressive or
moving average parameters and V (εt ) included in expression of autocovariance. In the expression (3.4), the autoregressive parameters and V (εt )
appears in the following form: π′s Ωπ s+ν−S∗δ and in expression (3.6) it appears in the following form: Π(s, S)Π(s + ν − S ∗ δ, S)σ2S . From model
(B) and model (C), the set, of possible values, of these two quantities are
respectively:
2.131 1.8989 1.4628 1.3938
1.8989 2.6744 1.8634 1.3923
(4.1)
1.4628 1.8634 2.2734 1.2975
1.3938 1.3923 1.2975 1.6743
and
0.65398
0.52318
0.31391
0.93428
0.52318
0.41854
0.25113
0.74742
17
0.31391
0.25113
0.15068
0.4445
0.93428
0.74742
0.44845
1.3347
(4.2)
It seen that all, possible values, of π ′s Ωπ s+ν−S∗δ are greater than 1 (some
are greater than 2, see the diagonal of matrix (4.1)). On the other hand,
all values of Π(s, S)Π(s + ν − S ∗ δ, S)σ 2S are lower than 1 (except the last
value in the diagonal of matrix (4.2).
5
Conclusion
For Seasonal-Periodic-ARF IM A(p, 0, 0)(0, D, 0) model, allowing the seasonal fractional parameter D to be S-periodic rather than constant we have
highlighted the existence of two distinct models (see section 1, model(I)
and model (II)). For these two distinct models we have established the exact and approximated expression of the periodic autocovariance. On the
simulated sample, for each model, the empirical periodic autocovariance
are calculated and graphically represented.
It is clear, through, theoretical and simulated results that it is not
easy to distinguish between these two models (the shape of the autocovariance for each model is not sufficient). If we consider the general model,
namely, Seasonal-Periodic-ARF IM A(p, dt , q)(P, Dt , Q) the situation becomes more complex to handle, because the number of different models we
can distinguish is more than two models. Furthermore, the non seasonal
part of the general model (i.e. P ARF IM A(p, dt , q)) did not receive much
attention on behalf of the statisticians and the probabilists.
References
[1] Peter Boswijk, H. & Franses, Philip Hans, 1995. Testing for periodic
integration, Economics Letters, Elsevier, vol. 48(3-4), pages 241-248.
[2] Boswijk, H. Peter & Franses, Philip Hans & Haldrup, Niels, 1997.
Multiple unit roots in periodic autoregression, Journal of Econometrics, Elsevier, vol. 80(1), pages 167-193.
[3] Chung, Ching-Fan. (2002). Sample Means, Sample Autocovariances,
and Linear Regression of Stationary Multivariate Long Memory Processes,” Econometric Theory, 18(1), 51-78.
[4] Franses, P.H. and Ooms, M. (1997). A periodic long memory model
for quartely UK inflation. International Journal of Forecasting 13,
117-126.
[5] Gladyshev, E. G. (1961). Periodically correlated random sequences.
Soviet. Mathematics, 2.
[6] Gladyshev, E. G. (1963). Periodically and almost PC random processes with continuous time parameter. Theory Probability and its
Applications, 8.
[7] Gray, H. L., Zhang, N.-F., and Woodward, W. (1989). On generalized
fractional processes. J. Time Ser. Anal., 10:233–258.
[8] Granger, C.W. J. (1980). Long Memory Relationships and the Aggregation of Dynamic Models. Journal of Econometrics, 14 227–238.
18
[9] Granger, C.W. J. (1986). Developements in the study of Cointegrated
Economic Variables, Oxford Bulletin of Economics and Statistics, 48
(3), 213-228.
[10] Hosking, J. R. M. (1981) Fractional differencing, Biometrica, 68,16576.
[11] Hui, Y. and Li, W. (1995). On fractionally differenced periodic processes. Sankhya,Ser. B, 57(1):19–31.
[12] Lund, R. B. and Basawa, I. V. (1999). Modeling and inference for
periodically correlated time series. In Asymptotics, nonparametrics,
and time series, volume 158 of Statist. Textbooks Monogr., pages
37–62. Dekker, New York.
[13] Rebecca J. Sela & Clifford M. Hurvich, (2009). Computationally efficient methods for two multivariate fractionally integrated models,
Journal of Time Series Analysis, Wiley Blackwell, vol. 30(6), pages
631-651.
[14] Ooms, M. and Franses, P.H. (2001). A seasonal periodic long memory
model for monthly river flows. Envoronmental Modelling and Software. Volume 16, Issue 6, pages 495-581.
[15] Oppenheim, G., Ould Haye, M., and Viano, M.-C. (2000). Long memory with seasonal effects. Statist. Inf. for Stoch. Proc., 3:53–68.
[16] Ould Haye, M. and Viano, M.-C. (2003). Limit theorems under seasonal longmemory. In Doukhan, Paul (ed.) et al., Theory and applications of long-range dependence. Boston, MA: Birkhuser. 101-110.
[17] Porter Hudak, S. (1990). An aplication of the seasonal fractionally di
erenced model to the monetary aggegrates. Journal of the American
Statistical Association, 85, 338-344.
6
Appendix A
Proposition: The infinite moving average representation of the process,
{yt , t ∈ Z} , defined by (1.1), is given by
yt = ut +
∞
P
j=1
j
P
(−1)
k
k=1
P
π i1 (−dt )πi2 (−dt−i1 )πi3 (−dt−i1 −i2 ) · · · π ik
i1 +i2 +···+ik =j
t +i1 −1)
. The numwhere π 0 (−dt ) = 1 and πi1 (−dt ) = −dt (−dt +1)···(−d
(i1 )!
P
j−1
ber terms in the sum ” i1 +i2 +···+ik =j ” is equal 2 . The number 2j−1
represent the cardinal sets of k positive integers, namely, (i1 , i2 , · · · , ik ),
which when summed together give j.
Proof. Putting
Π0 (−dt ) = 1 and
Γ (j − dt )
= Πj (−dt ),
Γ (j + 1) Γ (−dt )
we can rewrite (1.4) as
yt +
∞
X
Πj (−dt )yt−j = ut .
j=1
19
(A1)
−d Pk−1
t−
l=1
il
!
ut−j ,
More generally, we have
yt−j +
∞
X
Πj (−dt−j )yt−j−k = ut−j .
k=1
Suppose that the infinite moving average representation of (A1) is
given by
∞
X
Ψj (t)ut−j , with Ψ0 = 1,
(A2)
yt = Ψ0 (t)ut +
j=1
we have for the lagged variable yt−j ,
yt−j = ut−j +
∞
X
Ψk (t − j)ut−j−k .
(A3)
k=1
By replacing yt−j by ut−j +
Yt +
∞
X
Πj (−dt )ut−j +
j=1
P∞
k=1
∞
∞ X
X
Ψk (t − j)ut−j−k in (A1), we obtain
Πj (−dt )Ψk (t − j)ut−j−k = ut .
(A4)
j=1 k=1
Putting k′ = k + j, (A4) becomes
Yt +
∞
X
j=1
Πj (−dt )ut−j +
∞
∞
X
X
Πj (−dt )Ψk′ −j (t − j)ut−k′ = ut . (A5)
j=1 k′ =j+1
Let Φj,k′ (t − j) = Πj (−dt )Ψk′ −j (t − j), then we can rewrite (A5) as,
20
Yt +Π1 (−dt )ut−1
+
∞
P
k′ =j+1
∞
P
j=2
Πj (−dt )ut−j +
∞
P
k′ =2
Φ1,k′ (t − 1)ut−k′ +
Φj,k′ (t − j)ut−k′ + · · · = ut , (A6)
∞
P
k′ =3
Φ2,k′ (t − 2)ut−k′ + · · ·
We can rewrite (A6) as,
Yt + Π1 (−dt )ut−1 + (Π2 (−dt ) + Φ1,2 (t − 1)) ut−2
(A7)
+ (Π3 (−dt ) + Φ1,3 (t − 1) + Φ2,3 (t − 2)) ut−3
+ (Π4 (−dt ) + Φ1,4 (t − 1) + Φ2,4 (t − 2) + Φ3,4 (t − 3)) ut−4
..
.
+ (Πh (−dt ) + Φ1,h (t − 1) + Φ2,h (t − 2) + · · · + Φh−1,h (t − h + 1)) ut−h
.
+ ..
Let Πh (−dt ) + Φ1,h (t − 1) + Φ2,h (t − 2) + · · · + Φh−1,h (t − h + 1) =
β h (t − h + 1), for h ≥ 1, then we can rewrite (A7) as
Yt + β 1 (t)ut−1 + β 2 (t − 1)ut−2 + · · · + β h (t − h + 1)ut−h + · · · = ut . (A8)
From (A8), the infinite moving average representation of the process
yt is
∞
X
−β j (t) ut−j .
(A9)
Yt = ut +
j=1
By identification between (A2) and (A9), we obtain
Ψj (t) = −β j (t − j + 1).
(A10)
From (A10), the first three coefficients, (Ψ1 (t), Ψ2 (t), Ψ3 (t)) are:
• Ψ1 (t) = −β 1 (t),
= −Π1 (−dt ).
• Ψ2 (t) = −β 2 (t − 1)
= − (Π2 (−dt ) + Φ1,2 (t − 1))
= −Π2 (−dt ) − Π1 (−dt )Ψ1 (t − 1)
= −Π2 (−dt ) + Π1 (−dt )Π1 (−dt−1 )
=
2
P
P
π i1 (−dt ) +
(−1)k
k=1
i1 =2
i1 6=0
with il 6= 0, l ∈ 1, 2
P
i1 +i2 =2
i1 6=0; i2 6=0
πi1 (−dt )π i2 (−dt−i1 )
• Ψ3 (t) = −β 3 (t − 2)
= − (Π3 (−dt ) + Φ1,3 (t − 1) + Φ2,3 (t − 2))
= − (Π3 (−dt ) + Π1 (−dt )Ψ2 (t − 1) + Π2 (−dt )Ψ1 (t − 2))
= −Π3 (−dt )+Π1 (−dt )Π2 (−dt−1 )−Π1 (−dt )Π1 (−dt−1 )Π1 (−dt−2 )+
Π2 (−dt )Π1 (−dt−2 )
3
P
P
P
π i1 (−dt )πi2 (−dt−i1 ) +
(−1)k
π i1 (−dt ) +
=
i1 +i2 =3
i1 =3
k=1
P
πi1 (−dt )π i2 (−dt−i1 )πi3 (−dt−i1 −i2 ),
i1 +i2 +i3 =3
21
with il 6= 0, l ∈ 1, 3.More generally, we have,
Ψj (t) = −β j (t − j + 1)
j
X
X
= (−1)k
k=1
i1 +i2 +···+ik =j
π i1 (−dt )πi2 (−dt−i1 )πi3 (−dt−i1 −i2 ) · · · π ik (−dt−j+ik ) ,
(A11)
with il 6= 0, l ∈ 1, k
When dt = d = constant, we have (Ψ1 (t), Ψ2 (t), Ψ3 (t)) = d,
22
d(d+1)
, d(d+1)(d+2)
1
2
.
| 10math.ST
|
Analyzing Attacks on Cooperative Adaptive Cruise
Control (CACC)
Rens van der Heijden, Thomas Lukaseder, Frank Kargl
arXiv:1710.05789v2 [cs.CR] 8 Dec 2017
Institute of Distributed Systems
Ulm University, Germany
[email protected], [email protected], [email protected]
Abstract—Cooperative Adaptive Cruise Control (CACC) is
one of the driving applications of vehicular ad-hoc networks
(VANETs) and promises to bring more efficient and faster
transportation through cooperative behavior between vehicles.
In CACC, vehicles exchange information, which is relied on to
partially automate driving; however, this reliance on cooperation
requires resilience against attacks and other forms of misbehavior. In this paper, we propose a rigorous attacker model
and an evaluation framework for this resilience by quantifying
the attack impact, providing the necessary tools to compare
controller resilience and attack effectiveness simultaneously. Although there are significant differences between the resilience of
the three analyzed controllers, we show that each can be attacked
effectively and easily through either jamming or data injection.
Our results suggest a combination of misbehavior detection
and resilient control algorithms with graceful degradation are
necessary ingredients for secure and safe platoons.
I. I NTRODUCTION
In this paper, we study cooperative adaptive cruise control (CACC), an application of Vehicular Ad-Hoc Networks
(VANETs) aimed at increasing road efficiency by partially
automating driving. CACC is essentially an extension of
existing cruise control (CC) and adaptive cruise control (ACC)
technologies, both of which are designed to allow the driver
to maintain constant speed while driving. In the case of ACC,
vehicular sensors, such as RADAR, LIDAR and cameras,
are used to measure the distance to the preceding vehicle,
in order to automatically respond to changes in its’ driving
behavior. CACC extends this concept by allowing vehicles
to communicate and create a platoon, consisting of a leader
vehicle and multiple followers, as shown in Figure 1. There
are many different controller implementations in the literature
that enable this behavior [1] since the proposal of CACC
proposal in the 90s. It has been shown that CACC can be
more efficient than ACC, and that even a constant spacing
between vehicles (i.e., independent of platoon speed) can be
theoretically achieved, whereas it has been shown that ACC
alone cannot achieve this goal [2].
Since CACC was proposed, VANETs have made significant
leaps in development, including the DSRC standard and a
variety of ETSI, IEEE and SAE standards that define the
various network layers. In parallel to these developments,
security standards were designed, in order to protect against
attacks common in other fields. In relatively early research,
©2017 IEEE
accepted for VNC 2017 (no DOI available yet)
aA = -30 m/s2
di
…
i+1
victims
i
A
attacker
…
0
v0, a0
leader
Fig. 1. This figure shows a platoon with a leader and several followers: in
our work, the red vehicle is the attacker, and the platoon length is 8. The
attacker transmits a malicious beacon with a false acceleration of −30m/s2 .
it became clear that integrity is the most significant security
goal for VANETs in general [3], although availability is also
significant. The IEEE 1609.2 standard defines how to protect
against common external attacks, by providing message integrity and message authenticity (i.e., only legitimate vehicles
can transmit messages). Both the C2C-CC in Europe and the
CAMP in the US have since started deployment of security
credential management systems (SCMSs) to manage what
constitutes legitimate vehicles. In parallel, one of the major
advances in the security community was the introduction of
misbehavior detection, as a mean to detect and revoke vehicles
that transmit malicious data. This type of behavior is at the
focus of this work, particularly because the desired increased
road usage with CACC requires vehicles to move in a way
that would be unstable if only ACC was deployed.
The research in misbehavior detection for VANETs concentrates mostly on being able to detect incorrect information in beacon messages. These messages, defined by the
previously mentioned standards, are periodically sent to all
neighboring vehicles to inform them of the state of the sender.
One of the main application areas studied to evaluate such
detection schemes is position-based routing, where multi-hop
communication is enabled by finding routes based on vehicle
positions. Because vehicle position is also used in order to
avoid potential accidents in normal traffic (e.g., by warning
the driver with an alarm tone), many detection algorithms
operate on position information specifically. However, CACC
typically uses acceleration and speed information to compute
the behavior of the local vehicle based on that of the leader and
predecessor vehicles (and in some cases, the entire platoon). If
this information becomes unavailable or is incorrect, accidents
may occur. The data for CACC controllers is essential to
protect, but the security of these controllers when considering
malicious insiders has only been studied by few authors [4],
[5], [6], [7].
Existing work shows, that attacks can impact safety (i.e.,
potentially cause accidents), but lack a rigorous analysis and
a framework within which such attacks and the resilience of
the respective controllers can be analyzed. In this paper, we lay
the groundwork for such a framework: we implement several
attacks and provide a concrete quantification to measure the
impact of an attack. Going beyond simply finding whether an
accident occurs or not, we analyze the significance of such
an attack (in terms of the ∆v of the accident) and quantify
how the stability of an individual platoon is affected even
if no accident occurs. We present results for three existing
controllers implemented in Plexe [8]. Because we use this extensible platooning simulator as a basis for our evaluation, our
work can easily be extended with new controllers. Similarly,
our attack implementation can easily be extended to study
the impact of more attacks. We publish the source code, data
analysis scripts, and recorded data in order to stimulate further
research in this area.
The remainder of this paper is organized as follows. Section
II discusses the existing analyses of attacks on CACC and
platooning, as well as some basics regarding the controllers
in Plexe. In Section III we divide existing attacks into three
categories, of which two are analyzed extensively in this paper.
We also discuss which proposed attacks are infeasible under
the assumption that current security standards are implemented
correctly. Section IV describes our simulation setup and the
corresponding evaluation methodology.We discuss these results and their implications for CACC in Section V, and close
with a conclusion in Section VI.
II. R ELATED W ORK
We distinguish two areas of related work: controllers, which
implement the functionality that CACC and ACC offer, and
existing studies on misbehavior in platooning and similar
settings.
A. Controllers
Platoons are vehicles driving in a group, attempting to keep
a minimal but still safe distance between the vehicles within
the group. This can be achieved using ACC and CACC; the
main reason for using CACC is that the safe distance can be
much smaller. In both settings, a platoon normally consists of
a single leader and a set of followers, together referred to as
the platoon (or sometimes, string). The set of followers can be
dynamic, although this usually requires some kind of join and
leave procedures. Each vehicle implements a controller that
determines the desired acceleration, which is used to configure
the vehicles’ engine and brake controller. In this work we focus
on the upper controller (i.e., setting the desired acceleration)
exclusively, since this is the controller that an attacker can
easily reach (either through communication or through attacks
on sensors).
There is a broad scope of literature from the control
theory and engineering side covering a variety of different
controllers [1]. This literature typically introduces two stability
metrics: local stability and string stability. Here, local stability
refers to a vehicles’ distance to the next vehicle, while string
stability is a property of the platoon that determines how
spacing errors propagate through the platoon. A platoon is
considered string stable when the ∞-norm of the transfer function is less than 1 (||Ĥ(s)||∞ ≤ 1) [2], which basically means
that the platoon will eventually converge to a stable state
if the leader keeps a specific velocity perfectly. Maintaining
the distance to the next vehicle can done by keeping a fixed
distance (i.e., meters) or by maintaining a speed-dependent
headway time (i.e., seconds), referred to as constant spacing
and constant-time spacing. Earlier work has shown that noncooperative methods (i.e., ACC) cannot achieve string stability
in a constant spacing policy [2].
In this work we focus on controllers implemented by
Segata et al. [8] in Plexe, which includes a constant spacing controller, a controller with graceful degradation and a
consensus controller. The constant spacing controller is taken
from Rajamani [2] and keeps a fixed distance based on the
measured distance and the received speed and acceleration of
the preceding vehicle and the leader vehicle. The controller by
Ploeg et al. [9] is designed to enable graceful degradation on
network errors: their degraded CACC controller predicts the
acceleration of the preceding vehicle if this information is not
received. The idea is that this degraded controller outperforms
ACC, and can thus be used to bridge communication gaps;
this is particularly interesting for jamming attacks. The last
controller is a consensus algorithm by di Bernardo et al. [10],
which uses the position, speed, and acceleration information of
all vehicles in the platoon. This controller has a larger spacing,
but aims to maintain stability bidirectionally (i.e., vehicles also
adapt their behavior to that of vehicles behind them); this is
particularly interesting with respect to data injection attacks.
B. Existing Attacks
Attacks on CACC have typically been studied from two
different perspectives, mostly working in parallel. Security
research in vehicular networks has already suggested new
types of attacks for the overall network in the previous
decade [3]. More recent work has concentrated on misbehavior detection [11] and data injection attacks, which are
uniquely suitable for vehicular networks and other cyberphysical systems. Platooning is one of the motivating examples
for these researchers. On the other hand, authors from the
control theory side have gained an understanding of potential
security concerns, and have started developing attack-resilient
control algorithms. In terms of attacks, some authors have
also investigated how to exploit knowledge of the controller
to show that attacks causing instability are possible.
Alipour-Fanid et al. [7] discussed platoon behavior under
jamming. In their attack scenario, the jammer is a drone
flying above the platoon, aiming to use its’ limited power
to disrupt the platoon from above. Jamming is thus done
reactively (i.e., with success rate < 100%) to conserve power,
although power consumption is not considered further. Their
work shows that their ACC controller is string stable at 2.2
seconds headway distance, while CACC under the influence of
their reactive jammer is string stable at 1.7 seconds headway
distance. Additionally, they show the attack works best when
the attacker is located near the first vehicle in the platoon.
Amoozadeh et al. [6] discuss a two-level controller model
and describe a variety of attacks, including application and
network layer attacks, as well as other issues such as sensor
tampering and privacy issues. They present a result of a
message falsification attack, which in their setting is a vehicle
external to the platoon that falsifies all beacons; the result
of the attack is platoon instability (i.e., the platoon does not
converge back to a stable state). They also explain that such
an attack is most effective when acceleration changes occur.
Similar results are presented for a jamming attack – here, the
authors’ controller downgrades to ACC automatically. In our
work, we aim to show that such a downgrade is not necessarily
sufficient, depending on how system parameters are chosen.
DeBruhl et al. [5] also use a two component controller,
consisting of a PD feedback controller (radar-based) to keep
distance and a feedforward controller (communication-based),
whose outputs are added together to form the desired acceleration that is provided to the vehicle drivetrain. This
controller has been shown to be string-stable in real networked
platoons, and DeBruhl et al. examine its’ behavior under
various attack strategies, including a collision induction attack.
The authors continue to develop an error calculation and
detection algorithm, which essentially estimates the expected
behavior of the vehicle in front and switches to ACC if this
vehicle appears to behave differently than it claims to.
Outside of the context of communication, but also relevant to secure and safe platooning, is the work from the
research group of Gerdes et al. [12] and Dadras et al. [4].
These authors looked at ACC-based platoons, and examined
adversarial behavior within such a platoon, where the attacker
systematically disrupts the distributed control algorithm using
another algorithm that manipulates control input of their own
vehicle to achieve their goal. In the first work, the authors show
that for a given PID controller, they can cause significantly
increased energy consumption in neighboring vehicles; as
reduced energy consumption is one of the main drivers of
CACC research, this impact is significant. The second work
shows that for a PD controller, the attacker can cause the
platoon to be asymptotically unstable, i.e., destabilize and
eventually dissolve the platoon.
III. ATTACKER M ODEL
In this section, we describe our attacker model, which covers
attacks proposed in the literature above, as well as elements
from the area of misbehavior detection [11] and attacks on
sensors [13]. In particular, we are interested in how realistic
the attacks proposed in the literature are, when considering
currently standardized security mechanisms. Standardized security mechanisms include the use of pseudonymous certificates to protect driver privacy, the corresponding private keys
of which are stored in a hardware security module (HSM),
which is somewhat protected against tampering. The core of
our attacker model is that an attacker may gain control over a
vehicle, and thus over legitimate key material. This can happen
either through software compromise (e.g., malware spread
through applications running on the entertainment system), by
physical manipulation of the vehicle, or through key extraction from an old HSM. However, unlike traditional network
attacker models, the attacker additionally must obey the laws
of physics that are observable by any network participant (e.g.,
an attacker is not omnipresent throughout the network).
For CACC specifically, we assume the software and protocol
implementations are suitable, i.e., the attacker cannot find an
exploit here, and must obey the protocol. In other words, we
restrict the target of the attack to the platooning application,
and assume other mechanisms are in place to filter out invalid
or improperly formatted messages. In our model, the attackers’
primary goal is to cause a crash within the platoon, with the
secondary goal of destabilizing it. We assume the attacker is
already part of the platoon, a position from which the attacks
can do the most damage. However, as the literature points
out [5], if the attacker controls the leader vehicle, they have
complete control over the platoon, and therefore attacks should
be performed by a non-leader vehicle. We follow this argument
and also assume attacks are executed by a follower in order
to study the potential impact of attacks. Another argument
for this is that leader vehicles can be placed under additional
scrutiny and replaced if necessary, while followers are more
difficult to replace if they are non-cooperative.
A. Jamming
Denial of service (DoS) attacks includes a variety of jamming attacks (regular and reactive), which are the most popular
variant of this attack within vehicular networks. This also includes various attacks on higher layers that influence reception
(e.g., violating the MAC protocol, causing others to wait).
Some authors [6] also suggest a DoS attack may be launched
on the network layer, in order to disrupt communication by
overloading the hardware security module (HSM). This leads
to a form of cryptographic packet loss, i.e., packet loss caused
because messages cannot be verified, or cannot be verified in
time [14]. In this work, we consider all of these attacks to have
the same effect, i.e., that messages will no longer arrive, and
we refer to it as jamming. This generalizes the attack compared
to the related work [7], where the authors considered reactive
jamming.
B. V2X Data Injection
The second class of attacks we identify is V2X Data Injection, which generalizes a variety of different attacks discussed
in related work [6], [5]. Any attack that requires the attacker
(or a vehicle compromised by the attacker) to send modified
packets falls into this category; it is particularly useful because
these types of attacks may be feasible to mitigate through
misbehavior detection. The specific implementation of this
attack is usually dependent on the type of data exchanged by
the controller, and thus the corresponding impact is different
for every controller. Similarly, the goal of an attacker is
significant for the implementation, and sometimes used as
a distinguishing factor in the literature (this includes the
reduced headway attack, the collision induction attack and the
misreporting attack, among others). However, these goals can
often be achieved through one of the other classes of attacks
(e.g., collision induction is possible through jamming), and
thus we avoid this terminology.
Spoofing and replay attacks can also be considered examples of this class, but these are considered mostly solved in
the vehicular networking security community. In vehicular
networks, where GPS is often assumed to be practically
ubiquitous, replay attacks can be protected against using (close
to) synchronized clocks. Similarly, because cryptographic keys
are authenticated through pseudonymous certificates, and messages are protected by digital signatures, spoofing attacks that
cause the receiver to misinterpret the origin of the message are
not possible. This also applies to message falsification attacks
(i.e., attacks that manipulate messages of other vehicles),
for which digital signatures are an adequate countermeasure.
However, it is possible to use multiple certificates to simulate
more than one vehicle – this is referred to as a Sybil attack (as
originally introduced by Douceur [15]). Whether this attack is
feasible depends on how easy it is for the attacker to obtain
pseudonymous certificates that can be used simultaneously.
C. Sensor Manipulation
The third and final class of attacks is sensor manipulation,
which constitutes any attack that causes the internal network
of the vehicle to report incorrect information to the controller.
An attacker could either have access to the internal network of
the vehicle and transmit false data here, causing the vehicle
to react incorrectly (or even destabilize the platoon through
the controller attack [4]). Although this could also be done
in software, i.e., manipulating memory before the controller
processes the inputs, attacks on the internal network are considered to be more interesting, because misbehavior detection
is a potential countermeasure. Software attacks, on the other
hand, are difficult to prevent through misbehavior detection, as
the detection system will likely also be compromised in such
a setting.
Another variant of sensor manipulation positions the attacker outside of the target vehicle: instead of attacking the
network, the attacker uses their sensors and other tools to blind
the sensors of the target vehicle, or cause them to behave in
a certain way. Researchers have shown that both LIDAR and
cameras are vulnerable to such attacks [13], and the effect
of such an attack is very similar to the other type of sensor
manipulation attack (i.e., the controller sees invalid inputs).
The only distinction is where the attacker is located, which
influences the cost of such an attack, but the expected results
for the platoon behavior are similar, and thus for the study of
attacks we place these attacks in the same category.
TABLE I
E XCERPT OF SIMULATION PARAMETERS : FULL PARAMETERS AVAILABLE
IN THE DATA AND SOURCE CODE REPOSITORIES .
transmit power
sensitivity
platoon length
attacker ID
controllers
CACC spacing
target speed
attack: speed value
attack: accel value
attack: position shift
100mW
-94dBm
8
3
CACC, Ploeg, Consensus
5, 7, 9, 11, 13, 20 m
50, 80, 100, 120, 150 km/h
-50, 0, 50, 100, 150 m/s
-30, -10, 0, 10, 30 m/s2
3, 5, 7, 9, 11 m/s
IV. ATTACK A NALYSIS
In this paper we focus on communication only: we want to
answer the question how CACC behaves under misbehavior,
i.e., what are the potential impacts of such attacks. The aim of
this work is to provide a general framework for attack analysis,
which is helpful for the development of resilient control
algorithms, as well as the misbehavior detection community.
A. Simulation Setup
We use Plexe (version 2.0) [8], which is a simulation
toolkit based on VEINS, as basis for our analysis due to its’
flexibility in implementing control algorithms and accurate
simulation of network behavior. Plexe also extends SUMO,
which provides realistic microsimulations of physical vehicles;
it extends this toolkit with various controllers: a constant
spacing controller [2], a loss-tolerant controller [9] and a
consensus controller [10]. For our implementation, we leave
Plexe mostly unchanged, and only extend it to include our
attacks. We implement four attacks: one jamming attack and
three data injection attacks.
Our jamming attack is designed to analyze the maximum
possible impact of an attack, and is therefore implemented
by locally dropping received messages after the attack starts.
This simulates any type of denial of service, including cryptographic packet loss and various jamming strategies, and
provides an upper bound on the attack success, although it
may overestimate it in some cases. The alternative would
be to implement a specific attack (e.g., as done by AlipourFanid et al. [7]), but this approach is not as generalizable as
ours.
Our data injection attacks are all implemented by exchanging some data with false data in the message an attacker
would send according to the normal control algorithm. We
analyze three different types: injecting false positions, false
speeds or false acceleration values. There are multiple ways
to implement these attacks; in our case, we set the attackers’
value to a constant value for speed and acceleration, which
greatly simplifies the interpretation of the results. In addition,
we again aim to provide an upper bound on the attack
effectiveness. For position falsification, we provide a position
error that increases linearly over time, and is added to the
vehicles’ current position for each beacon.
The simulation scenario we analyze is a standard scenario
from Plexe, which was used by the authors to analyze platoon
behavior: the sinusoidal scenario. This scenario initializes
the platoon and then provides the leader with a specific
acceleration profile that leads to a sinusoidal speed graph.
By performing attacks at different points in time, we can use
this scenario to approximate the impact of attacks on real
behavior. The significant simulation parameters can be found
in I; the complete configuration file and code is available on
Github1 . Because of the way Plexe is implemented, simulating
chain collisions reliably was not possible, so we collected
information about the first collision only.
50
80
30.0
100
120
150
50
80
30.5
100
B. Metrics
120
In this work, we aim to analyze the effectiveness of various
attacks on different CACC controllers. The goal of our attacker
is to either cause a crash or maximize instability of the platoon.
As such, we distinguish between two cases: a crash or some
instability. If a crash occurs, impact velocity ∆v is used to
quantify the impact of the attack (medical research suggests
that this is the best predictive factor for injury [16]). If there
is no crash, we quantify the instability in the platoon for
the remainder of the simulation (approximately 30 seconds).
Since all the studied controllers have been shown to be string
stable, we can assume that the platoon eventually stabilizes.
The option we chose is maxs,i,t (e), where s is the simulation
run, i is the ith vehicle in the platoon, t is time and e
is spacing error, which most accurately represent the worst
case. This was inspired by earlier work [7], where a similar
metric is used. The spacing error is defined as the difference
between the desired distance of the controller and the actual
distance between two vehicles. The alternative metrics were
avgs (avgi (maxt (e))), the average maximum spacing error,
and avgs (avgi (maxt (a))), the average maximum acceleration, both of which can be considered metrics for passenger
comfort. Usability studies have shown [17] that spacing is
important for user acceptance; we included the acceleration as
an additional metric, which is also common in transportation
research. Averaging over simulations and vehicles allow us to
estimate the average comfort of all vehicles in a platoon; however, the maximum spacing error more accurately represents
the potential inefficiency and risk due to an attack, which is
why we decided on this metric2 .
150
50
80
31.0
120
150
50
80
31.5
1 https://github.com/vs-uulm/vnc2017-CACC-code
2 The
data repository for this paper includes all metrics.
3 https://github.com/vs-uulm/vnc2017-CACC-data
100
120
150
50
80
32.0
100
120
150
50
80
32.5
100
120
C. Results
sim
ul
at
io
ge
n
t
tim
sp
ee
e
d
(k
ph
)
C
A
C
C
-5
C
A
C
C
-7
C
A
C
C
-9
C
A
C
C
-1
C
1
A
C
C
-1
C
3
A
C
C
-2
0
Pl
oe
C
g
on
se
ns
us
150
ta
r
This section presents the results of our attack experiments,
and shows how we used our metrics to gain insight into the
effectiveness of different attacks, as well as the resilience of
the existing CACC algorithms. The data on which these results
are based will also be published on Github3 , although they can
also be re-generated using the source code and configuration
mentioned above.
100
Fig. 2. Heat map showing attack success and impact for the constant jamming
attack at the given point in simulation time, for different target platoon
speeds and different controllers. Red/italics: attack leads to accident; number
is impact velocity in m/s. Blue/bold: no accident; number is maximum spacing
error in m.
leader target speed (in km/h)
leader target speed (in km/h)
150
50
140
120
100
80
120
100
80
Δ
60
20
C
A
C
C C5
A
C
C
-2
0
P
C loe
on
g
se
ns
us
40
20
A
C
C
A -5
C
C
-2
0
Pl
o
C
on eg
se
ns
us
C
140
40
C
A
C
C
A -5
C
C
-2
0
P
C loe
on
g
se
ns
us
C
C
C
A
C
C C5
A
C
C
-2
0
P
C loe
on
g
se
ns
us
Δ
60
150
C
A
C
C C5
A
C
C
-2
0
P
C loe
on
g
se
ns
us
3m
5m
7m
9m
11 m
100
-50 km/h
0 km/h
50 km/h
100 km/h
150 km/h
C
A
C
C
A -5
C
C
-2
0
P
C lo e
on
g
se
ns
us
100
C
50
Fig. 3. The effect of position falsification: blue areas represent points not
related to a crash.
Fig. 4. Attacking with false speeds: blue areas represent points not related
to a crash.
1) Jamming: We first discuss the results of our jamming
attack, which are shown in Figure 2. This figure includes
both successful attacks (i.e., crashes, marked in red) and
stability impact (i.e., no crash, marked in blue), as well as the
corresponding impact (see Section IV-B). The figure shows
aggregated values (using the previously discussed aggregation
approach) over five runs. In all cases, the impact of the attack
remained similar (i.e. all five runs resulted in collisions or
all five runs did not result in collisions). We simulated this
for different target speeds and different jamming times: the
target speed is the intended average speed of the platoon (i.e.,
the average speed of the leader vehicle), while the jamming
time is the time at which jamming starts. The jamming time
directly corresponds to the range in which acceleration of the
leader vehicle is positive; since the leader data is disseminated
throughout the platoon, the attacker can use this acceleration
profile to choose the time of attack. Based on the speed profile
generated from Plexe’s standard configuration, we determined
that the positive acceleration lies between 30 and 32.5 seconds
of simulation time (since the behavior is sinusoidal, this pattern
repeats every 5 seconds).
As can be seen in the figure, the target speed has no impact
on the effectiveness of the attack on CACC. However, the
attacks are more successful with higher velocities when Ploeg
or Consensus are used as the spacing of these algorithms depends on the platoon target speed. We can therefore conclude,
that spacing has a strong influence on attack effectiveness. On
the other hand, attack effectiveness is not directly influenced
by the target velocity. In general, with more spacing, attack
impact decreases. The figure shows that this is not always true
(e.g. CACC-13 vs. CACC-20 at 32s simulation time). This is
due to the fact that with CACC-13 and lower, the attack occurs
between the attacker (car 3) and the subsequent car 4. With
CACC-20, car 4 can avoid colliding with the attacker, however,
car 5 does collide with car 4.
2) Position Injection: The next step in our analysis covers position falsification attacks, often cited as one of the
most significant types of data injection attacks on vehicular
networks. In our model, the false position is a shift of the
attackers’ real position that increases over time. It turns out
that this attack is completely ineffective against the Ploeg and
constant spacing controllers – it only causes crashes when the
consensus controller is used. However, we can furthermore
see that the spacing error of the Ploeg controller increases
significantly with the leader target speed. This is because the
consensus controller is the only one that uses received position
information directly. Somewhat surprisingly, the impact of this
attack on the consensus controller depends only on the value
transmitted by the attacker, and not significantly on the target
speed with which the leader (and thus the platoon) is moving.
The full results are shown in Figure 3: for each shift parameter
(mark and color of the scatter plot), with different leader target
speeds (top, left to right), the graph includes the results for
four different controller configurations (bottom, left to right).
The impact speed of successful attacks (white background)
varies from 2 to 9 m/s, while the maximum spacing error
(gray background) is unaffected by the position falsification.
This means that all attacks that have any effect lead to an
accident.
3) Speed Injection: Our next attack was manipulating the
speed transmitted by the attacker. Here, we continuously output the same value regardless of platoon or controller behavior,
which causes the error between the attacker’s value and actual
value to vary from 10 to −10 km/h if it falsifies the platoon
target speed. The result of our analysis, displayed in Figure 4,
shows that speed injection is successful against constant spacing CACC only, where the most significant crashes happen
at lower speeds; the attacker claims to accelerate, but actually
does not. When the attacker claims negative speeds (e.g., −50)
the result is a high position error. The Ploeg controller shows
the same behavior as before (i.e., it is not affected by the
attack), and here the consensus controller is also unaffected,
although the spacing error is much higher. Due to a constant
falsified speed, the attacker does not achieve success in the
constant value matching the leaders’ speed; we expect this
will change if the leader behavior is less predictable.
4) Acceleration Injection: The final attack we performed
is manipulation of the transmitted acceleration. Here, we also
chose a relatively extreme scope, in order to discover the maximum potential of these attacks (−30, −10, 0, 10, 30 m/s2 ). The
resulting statistics are shown in Figure 5. As expected, this
attack affects all controllers, because acceleration is used in all
leader target speed (in km/h)
50
100
-30
-10
0
10
30
150
140
m/s2
m/s2
m/s2
m/s2
m/s2
120
100
80
Δ
60
40
20
A
C
C
A -5
C
C
-2
0
Pl
o
C
on eg
se
ns
us
C
C
C
A
C
C C5
A
C
C
-2
0
P
C loe
on
g
se
ns
us
C
A
C
C C5
A
C
C
-2
0
P
C loe
on
g
se
ns
us
0
Fig. 5. Behavior of controllers with falsified acceleration information: blue
areas represent points not related to a crash.
the control algorithms. However, the effect is widely different
per controller: the consensus controller can tolerate these large
errors in acceleration information (although the maximum
spacing error increases slightly), while the others fail almost
completely. This is likely due to the fact that the consensus
controller is the only bi-directionally coupled controller (or
more precisely, each vehicle considers all others in its’ control
algorithm). Similar to the speed injection attack, injecting
false acceleration data causes attack impact dependent on the
injected value; the attack impact is significant for positive
acceleration values. However, here the impact of false negative
acceleration is also significant — however, at some point, the
distance sensor notices that the distance is too large, and it
resets to ACC behavior.
V. D ISCUSSION
Our results show that most CACC controllers are unreliable
when subject to effective denial of service attacks, and that
the impact of injection is also significant for any data that
is used in the controller directly. This confirms earlier results
on individual jamming and injection experiments, and shows
that these results also hold when simulation parameters are
changed. Beyond this, our results show that the window of
opportunity for the attacker to jam or transmit false information is broad, i.e., the attacker does not need to hit precisely
the correct timing for the attack to work. Our results can
also be used as a guide for the attacker to further hone
the effectiveness of an attack against more resilient control
algorithms.
Overall, we can conclude that the Ploeg controller deals
well with jamming even considering that this controller was
only intended for intermittent, unreliable channels, and not
for complete failure. However, we still managed to cause some
accidents even with this algorithm, suggesting that an eventual
complete fallback to ACC may be advisable. The consensus
controller behaves differently, and only fails with a later starting point when jamming — future work could more thoroughly
investigate the cause of this behavior. Finally, we conclude
that the constant spacing controller is completely unsuitable
for unreliable communication; the window for causing a crash
is at least 2 seconds.
For data injection attacks, the actual value transmitted by
the attacker had an important influence over the attack impact,
but the target speed of the leader did not matter much in
comparison. We expected our attacks to increase in impact as
the overall platoon speed increases, because we expected the
controllers to react more quickly in those scenarios. Although
this is the case for some position errors, the crash impact
remains relatively consistent for the CACC controller. For
the Ploeg controller, the impact is dependent on the speed,
but this dependence is mainly due to the speed-dependent
gap that should occur between vehicles. The one exception
to all these observations is the consensus controller, which is
surprisingly resilient against attacks, showing no impact from
speed and acceleration falsification. However, the impact of
position shift on this controller is significant. Nevertheless,
this suggests that the consensus controller might be a good
starting point for building a resilient controller, despite its’
relative ineffectiveness, illustrated by a higher overall position
error.
Our results are likely to extend to similar scenarios; the
CACC and Ploeg controllers both rely only on the leader
and the preceding vehicle, so any platoon long enough to
contain the crashing vehicle is affected by potential crashes.
It is unclear how well the results transfer for the consensus
controller, since its’ intended behavior is quite different from
the other two controllers, considering all the vehicles in
the platoon for the control of each vehicle. However, the
overall methodology we applied here is not constrained to the
scenario — the tools we provide can easily be used to further
tune the parameters of the controllers for increased safety, or
the attackers’ parameters to improve attack effectiveness.
VI. C ONCLUSION & F UTURE W ORK
In this paper, we have discussed a more holistic attacker
model, implemented several attacks and tested their impact
on existing controllers using our proposed analysis methodology. Overall, the results show not only that controllers are
vulnerable to attack, but that there are significant differences
between controllers in terms of how they are affected by
attacks. The research community should critically discuss the
limitations of CACC under the influence of these attacks. Our
future work focuses on improving the attacks and studying
their detection; however, detection alone is likely insufficient,
because of the volume of crashes we encountered. Work
should therefore focus also on prevention, through the design
of resilient controllers on the one hand, and effective detection
and revocation mechanisms on the other.
Future work in false data injection could look at how
other types of controllers behave, or study the effectiveness
of simultaneously falsifying multiple fields in the beacon.
The latter is particularly relevant, because it circumvents
many typical misbehavior detection mechanisms (which could
detect most of the attacks we have described with reasonable
accuracy, in part due to the values we chose). An alternative
route that has not yet been studied very well is the injection
of false data into external sensors. Some research has shown
that camera and LIDAR sensors can be tricked easily in a lab
environment [13]. Once Plexe has been extended to include
some sensor error models, experimenting with these attacks
and combining them with false data injection will likely be
an interesting path. Even though attacks on sensors would
rely on line-of-sight and additional equipment (i.e., it cannot
be done by manipulating a vehicle with malware), we have
also found that some injection attacks cause crashes between
vehicles further behind the attacker.
For the development of future control algorithms, we suggest that authors consider a fallback mechanism similar to
the Ploeg controller. However, we suggest that this eventually
downgrades from degraded CACC to pure ACC, because
the error in the estimated acceleration will increase over
time if a jamming attack is maintained long enough. To
improve reliability, using all available inputs may be a feasible
approach to reduce the impact of attacks (e.g., as partially
demonstrated by the consensus controller), especially because
common controllers typically only use leader and preceding
vehicle information. This may not increase overall security, but
will make it more difficult to successfully fine-tune attacks and
achieve a desired impact. Should the information show inconsistencies, a fallback mechanism could again be considered;
here the challenge is ensuring that the entire platoon mutually
decides on performing a fallback, to avoid a heterogeneous
controller environment.
The attacks discussed in this work may be detected with
relative ease using plausibility checks; however, the main
aim of this work was to show how attack impact may be
maximized. Future work may look at the maximum tolerance
of the attacks we designed; however, our currently proposed
metrics are bound to the behavior of the leader. The next
step is then to decide how to use this information, apart from
invalidating the vehicle in the SCMS [18], preventing damage
from ongoing attacks is also a goal. If using a controller
with graceful degradation (similar to Ploeg), the detection
may trigger this functionality: when combined with effective
revocation, this will ensure that temporarily degraded service
is the best possible attack. An interesting side aspect is that
not all detection algorithms are compatible with platooning,
because the corresponding behavior is different: it may thus
be a feasible path to design specialized misbehavior detection
mechanisms for the CACC scenario specifically.
ACKNOWLEDGMENT
The authors would like to thank Johannes Diebold for
the thesis upon which this work expands, and Henning
Kopp for proof reading the paper. Experiments for this work
were performed on the computational resource bwUniCluster funded by the Ministry of Science, Research and the
Arts Baden-Württemberg and the Universities of the State of
Baden-Württemberg, Germany, within the framework program
bwHPC. This work was supported in part by the BadenWürttemberg Stiftung gGmbH Stuttgart as part of the project
IKT-05 AutoDetect of its IT security research programme.
R EFERENCES
[1] D. Jia, K. Lu, J. Wang, X. Zhang, and X. Shen, “A survey on platoonbased vehicular cyber-physical systems,” IEEE Communications Surveys
Tutorials, vol. 18, no. 1, pp. 263–284, Firstquarter 2016.
[2] R. Rajamani, Vehicle Dynamics and Control. Springer US, 2012.
[3] M. Raya and J.-P. Hubaux, “Securing vehicular ad hoc networks,”
Journal of Computer Security, vol. 15, pp. 39–68, 2007.
[4] S. Dadras, R. M. Gerdes, and R. Sharma, “Vehicular platooning in an
adversarial environment,” in Proceedings of the 10th ACM Symposium
on Information, Computer and Communications Security, ser. ASIA
CCS ’15. New York, NY, USA: ACM, 2015, pp. 167–178.
[5] B. DeBruhl, S. Weerakkody, B. Sinopoli, and P. Tague, “Is your
commute driving you crazy?: A study of misbehavior in vehicular
platoons,” in Proceedings of the 8th ACM Conference on Security
& Privacy in Wireless and Mobile Networks, ser. WiSec ’15. New
York, NY, USA: ACM, 2015, pp. 22:1–22:11. [Online]. Available:
http://doi.acm.org/10.1145/2766498.2766505
[6] M. Amoozadeh, A. Raghuramu, C. n. Chuah, D. Ghosal, H. M. Zhang,
J. Rowe, and K. Levitt, “Security vulnerabilities of connected vehicle
streams and their impact on cooperative driving,” IEEE Communications
Magazine, vol. 53, no. 6, pp. 126–132, June 2015.
[7] A. Alipour-Fanid, M. Dabaghchian, H. Zhang, and K. Zeng, “String
stability analysis of cooperative adaptive cruise control under jamming
attacks,” in 2017 IEEE 18th International Symposium on High Assurance
Systems Engineering (HASE), Jan 2017, pp. 157–162.
[8] M. Segata, S. Joerer, B. Bloessl, C. Sommer, F. Dressler, and
R. Lo Cigno, “PLEXE: A Platooning Extension for Veins,” in 6th IEEE
Vehicular Networking Conference (VNC 2014). Paderborn, Germany:
IEEE, December 2014, pp. 53–60.
[9] J. Ploeg, E. Semsar-Kazerooni, G. Lijster, N. van de Wouw, and
H. Nijmeijer, “Graceful degradation of cacc performance subject to unreliable wireless communication,” in Intl. IEEE Conference on Intelligent
Transportation Systems (ITSC 2013), Oct 2013, pp. 1210–1216.
[10] M. di Bernardo, A. Salvi, and S. Santini, “Distributed consensus strategy
for platooning of vehicles in the presence of time-varying heterogeneous
communication delays,” IEEE Transactions on Intelligent Transportation
Systems, vol. 16, no. 1, pp. 102–112, Feb 2015.
[11] R. W. van der Heijden, F. Kargl, O. M. F. Abu-Sharkh, and A. AlMomani, “Enhanced position verification for vanets using subjective
logic,” in Vehicular Technology Conference Fall. IEEE, 2016.
[12] R. M. Gerdes, C. Winstead, and K. Heaslip, “CPS: An efficiencymotivated attack against autonomous vehicular transportation,” in Proc.
Annual Computer Security Applications Conference, ser. ACSAC ’13.
New York, NY, USA: ACM, 2013, pp. 99–108.
[13] J. Petit and S. E. Shladover, “Potential Cyberattacks on Automated
Vehicles,” IEEE Transactions on Intelligent Transportation Systems,
vol. 16, no. 2, pp. 546–556, Apr. 2015.
[14] M. Feiri, J. Petit, and F. Kargl, “Congestion-based certificate omission in
vanets,” in Intl. Workshop on Vehicular Inter-networking, Systems, and
Applications, ser. VANET ’12. NY, USA: ACM, 2012, pp. 135–138.
[15] J. R. Douceur, “The Sybil Attack,” in Peer-to-Peer Systems, ser. Lecture
Notes in Computer Science, P. Druschel, F. Kaashoek, and A. Rowstron,
Eds., vol. 2429. Springer Berlin Heidelberg, Mar. 2002, pp. 251–260.
[16] H. C. Joksch, “Velocity change and fatality risk in a crash – a rule of
thumb,” Accident Analysis & Prevention, vol. 25, no. 1, pp. 103–104,
1993.
[17] M. Larburu, J. Sanchez, and D. J. Rodriguez, “Safe road trains for
environment: Human factors’ aspects in dual mode transport systems,”
in ITS World Congress, Busan, Korea, 2010.
[18] W. Whyte, A. Weimerskirch, V. Kumar, and T. Hehn, “A security
credential management system for v2v communications,” in 2013 IEEE
Vehicular Networking Conference, Dec 2013, pp. 1–8.
| 3cs.SY
|
arXiv:1410.8512v2 [math.GR] 16 Sep 2016
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED
ARTIN GROUPS I: THE FINITE OUT CASE
JINGYIN HUANG
Abstract. Let G and G0 be two right-angled Artin groups. We show they
are quasi-isometric if and only if they are isomorphic, under the assumption
that the outer automorphism groups Out(G) and Out(G0 ) are finite. If we
only assume Out(G) is finite, then G0 is quasi-isometric G if and only if G0
is isomorphic to a subgroup of finite index in G. In this case, we give an
algorithm to determine whether G and G0 are quasi-isometric by looking at
their defining graphs.
Contents
1. Introduction
1.1. Backgrounds and Summary of Results
1.2. Comments on the Proof
1.3. Organization of the Paper
1.4. Acknowledgement
2. Preliminaries
2.1. Notation and Conventions
2.2. CAT (0) space and CAT (0) cube complex
2.3. Coarse intersections of convex subcomplexes
2.4. Right-angled Artin groups
3. Stable subgraph
3.1. Coarse intersection of standard subcomplexes and flats
3.2. Standard flats in transvection free RAAG’s
3.3. Standard flats in general RAAG’s
4. From quasi-isometries to isomorphisms
4.1. The extension complexes
4.2. Reconstruction of quasi-isometries
4.3. The automorphism groups of extension complexes
5. Quasi-isometries and special subgroups
5.1. Preservation of extension complex
5.2. Coherent ordering and coherent labeling
6. The geometry of finite index RAAG subgroups
6.1. Constructing finite index RAAG subgroups
6.2. Rigidity of RAAG subgroups
6.3. Generalized star extension
References
1
2
2
5
7
7
7
7
8
10
11
13
13
17
20
25
25
28
31
34
34
36
41
41
44
50
53
2
JINGYIN HUANG
1. Introduction
1.1. Backgrounds and Summary of Results. Given a finite simplicial graph Γ
with vertex set {vi }i∈I , the right-angled Artin group (RAAG) with defining graph
Γ, denoted by G(Γ), is given by the following presentation:
{vi , for i ∈ I | [vi , vj ] = 1 if vi and vj are joined by an edge}
{vi }i∈I is called a standard generating set for G(Γ) (cf. Section 2.4).
The class of RAAG’s enjoys a balance between simplicity and complexity. On
one hand, RAAG’s have many nice geometric, combinatorial and group theoretic
properties (see [Cha07] for a summary); on the other hand, this class inherits the
full complexity of the collection of finite simplicial graphs, and even a single RAAG
could have very complicated subgroups (see, for example [BB97]).
In recent years, RAAG’s have become important models to understand other
unknown groups, either by (virtually) embedding the unknown groups into some
RAAG’s (such a program is outlined in [Wis09, Section 6], see also the references
over there), or by finding embedded copies of RAAG’s in the unknown groups
([CLM10, Kob12b, Tay13, KK13a, BKK14]).
In this paper, we study the asymptotic geometry of RAAG’s and classify a
particular class of RAAG’s by their quasi-isometric types. Previously, the quasiisometric classification of RAAG’s has been done for the following two classes.
(1) Tree groups by Behrstock and Neumann. It is shown in [BN08] that for
any two trees Γ1 and Γ2 with diameter ≥ 3, G(Γ1 ) and G(Γ2 ) are quasiisometric. Higher dimensional analogs of tree groups are studied in [BJN10].
(2) Atomic groups by Bestvina, Kleiner and Sageev. A RAAG is atomic if its
defining graph Γ is connected and does not contain valence one vertices, cycles of length < 5 and separating closed stars. It is shown in [BKS08a] that
two atomic RAAG’s are quasi-isometric if and only if they are isomorphic.
Note that atomic groups are much more “rigid” than tree groups. We define the
dimension of G(Γ) to be the maximal n such that G(Γ) contains a Zn subgroup,
and it coincides with the cohomological dimension of G(Γ). All atomic groups
are 2-dimensional, hence it is natural to ask what are higher dimensional RAAG’s
which satisfy similar rigidity properties as atomic RAAG’s. This is the starting
point of the current paper.
Since we are looking for RAAG’s which are rigid, those ones with small quasiisometry groups would be reasonable candidates. However, even in the atomic case,
the quasi-isometry group QI(G(Γ)) is huge (see the discussion of quasi-isometry
flexibility in [BKS08a, Section 11]). Then we turn to the outer automorphism
group Out(G(Γ)) for guidance.
Now we ask whether those RAAG’s with “small” outer automorphism groups
are also geometrically rigid in an appropriate sense. Actually, “small” outer automorphism groups and (quasi-isometric or commensurability) rigidity results come
together in several other cases, for example, higher rank lattices ([Mos73, KL97,
EF97, Esk98]), mapping class groups ([Ham05, BKMM12]), Out(Fn ) ([FH07]) etc.
Our first result is about the quasi-isometric classification for RAAG’s with finite
outer automorphism group.
Theorem 1.1. Pick G(Γ1 ) and G(Γ2 ) such that Out(G(Γi )) is finite for i = 1, 2.
Then they are quasi-isomeric if and only if they are isomorphic.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
3
This theorem is proved in Section 4. See Theorem 4.18 for a more detailed
version of Theorem 1.1.
The collection of RAAG’s with finite outer automorphism group is a reasonably
large class. Recall that there is a 1-1 correspondence between finite simplicial graphs
and RAAG’s ([Dro87]), thus it makes sense to talk about a random RAAG by considering the Erdős–Rényi model for random graphs. If the parameters of the model
are in the right range, then almost all RAAG’s have finite outer automorphism
group ([CF12, Day12]).
The class of 2-dimensional RAAG’s with finite outer automorphism group is
strictly larger than the class of atomic RAAG’s, moreover, there are plenty of
higher dimensional RAAG’s with finite outer automorphism group.
Whether Out(G(Γ)) is finite or not can be easily read from Γ. We defined the
closed star of a vertex v in Γ, denoted by St(v), to be the full subgraph (see Section
2.1) spanned by v and vertices adjacent to v. Similarly, lk(v) is defined to be the
full subgraph spanned by vertices adjacent to v. Note that this definition is slightly
different from the usual one.
By results in [Ser89, Lau95], Out(G(Γ)) is generated by the following four types
of elements (we identify the vertex set of Γ with a standard generating set of G(Γ)):
(1) Given vertex v ∈ Γ, sending v → v −1 and fixing all other generators.
(2) Graph automorphisms of Γ.
(3) If lk(w) ⊂ St(v) for vertices w, v ∈ Γ, sending w → wv and fixing all other
generators induces a group automorphism. It is called a transvection. When
d(v, w) = 1, it is an adjacent transvection, otherwise it is a non-adjacent
transvection.
(4) Suppose Γ \ St(v) is disconnected. Then one obtains a group automorphism by picking a connected component C and sending w → vwv −1 for
each vertex w ∈ C (all other generators are fixed). It is called a partial
conjugation.
Elements of type (3) or (4) have infinite order in Out(G(Γ)) while elements of
type (1) or (2) are of finite order. Out(G(Γ)) is finite if and only if Γ does not
contain any separating closed star, and there do not exist distinct vertices v, w ∈ Γ
such that lk(w) ⊂ St(v).
Theorem 1.2. Suppose Out(G(Γ1 )) is finite. Then the following are equivalent:
(1) G(Γ2 ) is quasi-isometric to G(Γ1 ).
(2) G(Γ2 ) is isomorphic to a subgroup of finite index in G(Γ1 ).
(3) Γe2 is isomorphic to Γe1 .
Here Γe denotes the extension graph introduced by Kim and Koberda in [KK13b]
(see Definition 2.11). Extension graphs can be viewed as “curve graphs” for RAAG’s
([KK14]). This analog carries on to the aspect of quasi-isometric rigidity. Namely,
if G is a mapping class group and q : G0 → G is a quasi-isometry, then it is shown
in [BKMM12] that G0 naturally acts on the curve graph associated with G. This is
still true if G is a RAAG with some restriction on its outer automorphism group,
for example, Out(G) is finite.
However, in general, there exists a pair of commensurable RAAG’s with different
extension graphs, see Example 3.29. There also exists a pair of non-quasi-isometric
RAAG’s with isomorphic extension graphs, see [Hua16, Section 5.3].
4
JINGYIN HUANG
Motivated by Theorem 1.2 (2), we now look at finite index RAAG subgroups (i.e.
subgroups which are also RAAG’s) of G(Γ1 ).
Given a RAAG G(Γ) (not necessarily having a finite outer automorphism group)
and pick a standard generating set S for G(Γ). Let dS be the word metric on G(Γ)
with respect to S. A subset K ⊂ G(Γ) is S-convex if for any three points x, y ∈ K
and z ∈ G(Γ) such that dS (x, y) = dS (x, z) + dS (z, y), we must have z ∈ K. Every
finite S-convex subset K naturally gives rise to a finite index RAAG subgroup
G ≤ G(Γ) such that K is the fundamental domain of the left action G y G(Γ).
For example, if G(Γ) = Z ⊕ Z and pick K to be a rectangle of size n by m, then the
corresponding subgroup is of form nZ ⊕ mZ. The detailed construction in the more
general case is given in Section 6.1. G is called an S-special subgroup of G(Γ). A
subgroup of G(Γ) is special if it is S-special for some standard generating set S. A
similar construction in the case of right-angled Coxeter groups is in [Hag08].
Here is an alternating description in terms of the canonical completion introduced
by [HW08]. Let S(Γ) be the Salvetti complex of G(Γ) (see Section 2.4) and let X(Γ)
be the universal cover. We pick an identification between G(Γ) and the 0-skeleton
of X(Γ). The above subset K gives rise to a convex subcomplex K̄ ⊂ X(Γ).
Then the corresponding special subgroup is the fundamental group of the canonical
completion with respect the local isometry K̄ → S(Γ).
Our next result says if Out(G(Γ)) is finite, then this is the only way to obtain
finite index RAAG subgroups of G(Γ).
Theorem 1.3. Suppose Out(G(Γ)) is finite and let S be a standard generating set
for G(Γ). Then all finite index RAAG subgroups are S-special. Moreover, there is
a 1-1 correspondence between non-negative finite S-convex subsets of G(Γ) based at
the identity and finite index RAAG subgroups of G(Γ).
See Theorem 6.20 for a slight reformulation of Theorem 1.3.
We need to explain two terms: non-negative and based at the identity. For
example, take G = nZ ⊕ mZ inside Z ⊕ Z, then any n by m rectangle could be
the fundamental domain for the action of G. We naturally require the rectangle
to be in the first quadrant and contain the identity, which would give us a unique
choice. Similar things can be done in all RAAG’s and this two terms will be defined
precisely in Section 6.
The most simple example is when G(Γ) = Z, we have a 1-1 correspondence
between finite index subgroups of form nZ and the intervals of form [0, n − 1].
Corollary 1.4. If Out(G(Γ1 )) is finite, then G(Γ2 ) is quasi-isometric to G(Γ1 ) if
and only if G(Γ2 ) is isomorphic to a special subgroup of G(Γ1 ).
It turns out that there is an algorithm to enumerate the defining graphs of all
special subgroups of a RAAG, so
Theorem 1.5. If Out(G(Γ)) is finite, then G(Γ0 ) is quasi-isometric to G(Γ) if and
only if Γ0 can be obtained from Γ by finitely many GSE’s. In particular, there is an
algorithm to determine whether G(Γ0 ) and G(Γ) are quasi-isometric by looking at
the graphs Γ and Γ0 .
GSE is a generalized version of star extension in [BKS08a, Example 1.4], see also
[KK13b, lemma 50]. It will be defined in Section 6.
A question motivated by Theorem 1.2 is the following:
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
5
Question 1.6. Let G(Γ) be a RAAG such that Out(G(Γ)) is finite. And let H be
a finite generated group quasi-isometric to G(Γ). What can we say about H?
As a partial answer to this question, we prove the following result in [HK16].
Theorem 1.7. ([HK16]) Let G(Γ) and H be as in Question 1.6. Then the induced
quasi-action H y X(Γ) is quasi-isometrically conjugate to a geometric action H y
X 0 . Here X 0 is a CAT (0) cube complex which is closely related to X(Γ).
1.2. Comments on the Proof.
1.2.1. Proof of Theorem 1.1. We start with several notations. The Salvetti complex
of G(Γ) is denoted by S(Γ), the universal cover of S(Γ) is denoted by X(Γ), and
flats in X(Γ) that cover standard tori in S(Γ) are called standard flats. See Section
2.4 for precise definition of these terms.
Let q : X(Γ) → X(Γ0 ) be a quasi-isometry. The proof of Theorem 1.1 follows
the scheme of the proof of the main theorem in [BKS08a]. Similar schemes can
also be found in [KL97, BKMM12]. There are three steps in [BKS08a]. First they
show q maps top dimensional flats to top dimensional flats up to finite Hausdorff
distance. However, the collection of all top dimensional flats is too large to be
linked directly to the combinatorics of RAAG’s, so the second step is to show
quasi-isometries preserve standard flats up to finite Hausdorff distance. The third
step is to straighten the quasi-isometry such that it actually maps standard flats to
standard flats exactly, not just up to finite Hausdorff distance, and the conclusion
follows automatically.
In our cases, the first step has been done in [Hua14], where we show q still
preserves top dimensional flats up to finite Hausdorff distance in higher dimensional
case. No assumption on the outer automorphism group is needed for this step.
The second step consists of two parts. First we show q preserves certain top
dimensional maximal products up to finite Hausdorff distance. Then one wish to
pass to standard flats by intersecting these top dimensional objects. However, in
the higher dimensional case, a lower dimensional standard flat may not be the
intersection of top dimensional objects, and even in the case it is an intersection,
one may not be able to read this information directly from the defining graph Γ.
This is quite different from the 2-dimensional situation in [BKS08a] and relies on
several new ingredients.
A necessary condition for q to preserve the standard flats is that every elements
in Out(G(Γ)) should do so, which implies there could not be any transvections in
Out(G(Γ)). This condition is also sufficient.
Theorem 1.8. Suppose Out(G(Γ)) is transvection free. Then there exists a positive
constant D = D(L, A, Γ) such that for any standard flat F ⊂ X(Γ), there exists a
standard flat F 0 ⊂ X(Γ0 ) such that dH (q(F ), F 0 ) < D.
Here dH denotes the Hausdorff distance.
In step 3, we introduce an auxiliary simplicial complex P(Γ), which serves as
a link between the asymptotic geometry of X(Γ) and the combinatorial structure
of X(Γ). More precisely, on one hand, P(Γ) can be viewed as a simplified Tits
boundary for X(Γ), on the other hand, one can read certain information about
the wall space structure of X(Γ) from P(Γ). This complex turns out to coincide
with the extension graph introduced in [KK13b], where it was motivated from the
viewpoint of mapping class group.
6
JINGYIN HUANG
Denote the Tits boundary of X(Γ) by ∂T (X(Γ)), and let T (Γ) ⊂ ∂T (X(Γ)) be
the union of Tits boundaries of standard flats in X(Γ). Then T (Γ) has a natural
simplicial structure. However, T (Γ) contains redundant information, this can be
seen in the similar situation where the link of the base point of S(Γ) looks more
complicated than Γ, but they essentially contain the same information.
This redundancy can be resolved by replacing the spheres in T (X) that arise from
standard flats by simplexes of the same dimension. This gives rise to a well defined
simplicial complex P(Γ), since for any standard flats F1 and F2 with ∂T F1 ∩∂T F2 6=
∅, there exists a standard flat F such that ∂T F = ∂T F1 ∩ ∂T F2 . See Section 4.1 for
more properties of P(Γ).
By Theorem 1.8, if both Out(G(Γ)) and Out(G(Γ0 )) are transvection free, then
q induces a boundary map ∂q : P(Γ) → P(Γ0 ), which is a simplicial isomorphism.
Next we want to consider the converse and reconstruct a map X(Γ) → X(Γ0 )
from the boundary map ∂q in the following sense. Pick vertex p ∈ X(Γ), let
{Fi }ni=1 be the collection of maximal standard flats containing p. By Theorem
1.8, for each i, there exists a unique maximal standard flat Fi0 ⊂ X(Γ0 ) such that
dH (q(Fi ), Fi0 ) < ∞. One may wish to map p, which turns out to be the intersection
of Fi ’s, to the intersection of all Fi0 ’s. However, in general ∩ni=1 Fi0 may be empty,
or contain more than one point, hence our map may not be well-defined.
It turns out that if we also rule out partial conjugations in Out(G(Γ)), then
∩ni=1 Fi0 is exactly a point. This give rises to a well-defined map q̄ : X(Γ)(0) →
X(Γ0 )(0) which maps vertices in a standard flat to vertices in standard flat. If
Out(G(Γ0 )) is also finite, then we can define an inverse map of q̄ and this is enough
to deduce Theorem 1.1.
1.2.2. Proof of Theorem 1.2. If only Out(G(Γ)) is assumed to be finite, we can
still recover the fact that ∂q is a simplicial isomorphism (this is non-trivial, since
Theorem 1.8 does not say for any standard flat F 0 ⊂ X(Γ0 ), we can find a standard
flat F ⊂ X(Γ) such that dH (q(F ), F 0 ) < ∞). Hence we can define q̄ as before.
However the inverse of q̄ does not exist in general.
The next step is to trying to extend q̄ to a cubical map (Definition 2.1) from X(Γ)
to X(Γ0 ). There are obvious obstructions: though q̄ maps vertices in a standard
geodesic to vertices in a standard geodesic, q̄ may not preserve the order of these
vertices. A typical example is given in the following picture, where one can permute
the green level and the red level in a tree, then the order of vertices in the black
line is not preserved.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
7
A remedy is to “flip backwards”. Namely we will pre-compose q̄ with a sequence
of permutations of “levels” such that the resulting map restricted to each standard
geodesic respects the order. Then we can extend q̄ to a cubical map. This argument
relies on the understanding of quasi-isometric flexibility, namely how much room
we have to perform these flips. One formulation of this aspect is the following.
Theorem 1.9. If Out(G(Γ)) is finite, then Aut(P(Γ)) ∼
= Isom(G(Γ), dr ).
Here dr denotes the syllable metric, see [KK14, Section 5.2].
Theorem 1.2 - Theorem 1.5 will rely on the cubical map q̄. In particular, q̄ −1 (x)
(x ∈ X(Γ0 ) is a vertex) is a compact convex subcomplex and this is how we obtain
the S-convex subset in Theorem 1.3.
1.3. Organization of the Paper. Section 2 contains basic notations used in this
paper and some background material about CAT (0) cube complexes and RAAG’s.
In particular, Section 2.3 collects several technical lemmas about CAT (0) cube
complex. One can skip Section 2.3 on first reading and come back when needed.
In Section 3 we prove Theorem 1.8. Section 3.1 is about the stability of top
dimensional maximal product subcomplexes under quasi-isometries and Section 3.2
deals with lower dimensional standard flats. In Section 4 we prove Theorem 1.1.
We will construct the extension complex from our viewpoint in Section 4.1 and
explain how is this object related to Tits boundary, flat space and contact graph.
In Section 4.2, we describe how to reconstruct the quasi-isometry.
Section 5.1 and Section 5.2 are devoted to proving Theorem 1.2. We prove
Theorem 1.3, Corollary 1.4 and Theorem 1.5 in Section 6.
1.4. Acknowledgement. I would like to thank Bruce Kleiner and Robert Young,
for reading parts of the paper and give valuable comments. I would thank Jason
Behrstock, Ruth Charney, and Walter Neumann for suggestion interesting questions, which leads to Section 6.3 of this paper. I would thank Pallavi Dani, Mark
Hagen, Thomas Koberda, Do-gyeom Kim and Sang-hyun Kim for related discussions. I also thank the referee for carefully reading the paper and providing helpful
comments.
2. Preliminaries
2.1. Notation and Conventions. All graphs in this paper are simple.
The flag complex of a graph Γ is denoted by F (Γ), i.e. F (Γ) is a flag complex
such that its 1-skeleton is Γ.
A subcomplex K 0 in a combinatorial polyhedral complex K is full if K 0 contains
all the subcomplexes of K which have the same vertex set as K 0 . If K is 1dimensional, then we also call K 0 a full subgraph.
We use “∗” to denote the join of two simplicial complexes and “◦” to denote the
join of two graphs. Let K be a simplicial complex or a graph. By viewing the
1-skeleton as a metric graph with edge length = 1, we obtain a metric defined on
the 0-skeleton, which we denote by d. Let N ⊂ K be a subcomplex. We define the
orthogonal complement of N , denoted by N ⊥ , to be the set {w ∈ K (0) | d(w, v) = 1
for any vertex v ∈ N }; define the link of N , denoted by lk(N ), to be the full
subcomplex spanned by N ⊥ ; and define the closed star of N , denoted by St(N ),
to be the full subcomplex spanned by N ∪ lk(N ). Suppose L is a subcomplex such
that N ⊂ L ⊂ K. We denote the closed star of N in L by St(v, L). If L is a full
8
JINGYIN HUANG
subcomplex, then St(N, L) = St(N ) ∩ L. We can define lk(N, L) in a similar way.
Let M ⊂ K be an arbitrary subset. We denote the collection of vertices inside M
by v(M ).
We use id to denote the identity element of a group, and use Id to denote the
identity map from a space to itself.
Let (X, d) be a metric space. The open ball of radius r centred at p in X will be
denoted by B(p, r). Given subsets A, B ⊂ X, the open r-neighborhood of a subset
A is denoted by Nr (A). The diameter of A is denoted by diam(A). The Hausdorff
distance between A and B is denoted by dH (A, B). We will also use the following
adapted notation of coarse set theory introduced in [MSW04].
Symbol
A ⊂r B
A ⊂∞ B
r
A=B
∞
A=B
A ∩r B
Meaning
A ⊂ Nr (B)
∃r > 0 such that A ⊂ Nr (B)
dH (A, B) ≤ r
dH (A, B) < ∞
Nr (A) ∩ Nr (B)
2.2. CAT (0) space and CAT (0) cube complex. The standard reference for
CAT (0) spaces is [BH99].
Let (X, d) be a CAT (0) space. Pick x, y ∈ X, we denote the unique geodesic
segment joining x and y by xy. For y, z ∈ X \ {x}, denote the comparison angle
between xy and xz at x by ∠x (y, z) and the Alexandrov angle by ∠x (y, z).
The boundary of X, denoted by ∂X, is the collection of asymptotic classes of
geodesic rays. ∂X has an angular metric, which is defined by
∠(ξ1 , ξ2 ) = lim
∠p (l1 (t), l2 (t0 ))
0
t,t →∞
where l1 and l2 are unit speed geodesic rays emanating from a base point p such
that li (∞) = ξi for i = 1, 2. This metric does not depend on the choice of p, and
the length metric associated to the angular metric, denoted by dT , is called the
Tits metric. The Tits boundary, denoted by ∂T X, is the CAT (1) space (∂X, dT )
(see Chapter II.8 and II.9 of [BH99]).
Given two metric spaces (X1 , p
d1 ) and (X2 , d2 ), denote the Cartesian product of
X1 and X2 by X1 × X2 , i.e. d = d21 + d22 on X1 × X2 . If X1 and X2 are CAT (0),
then so is X1 × X2 .
An n-flat in a CAT (0) space X is the image of an isometric embedding En → X.
Note that any flat is convex in X.
Pick a convex subset C ⊂ X, then C is also CAT (0). We use πC to denote
the nearest point projection from X to C. πC is well-defined and is 1-Lipschitz.
Moreover, pick x ∈ X \ C, then ∠πC (x) (x, y) ≥ π/2 for any y ∈ C such that
y 6= πC (x) (see Proposition II.2.4 of [BH99]).
If C 0 ⊂ X is another convex set, then C 0 is parallel to C if d(·, C)|C 0 and d(·, C 0 )|C
are constant functions. In this case, there is a natural isomorphism between C ×
[0, d(C, C 0 )] and the convex hull of C and C 0 . We define the parallel set of C,
denoted by PC , to be the union of all convex subsets of X parallel to C. If C has
geodesic extension property, or more generally, C is boundary-minimal (Section 3.C
of [CM09]), then PC is a convex subset in X. Moreover, PC admits a canonical
splitting PC = C × C ⊥ , where C ⊥ is also a CAT (0) space.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
9
Now we turn to CAT (0) cube complexes. All cube complexes in this paper are
assumed to be finite dimensional.
A cube complex X is obtained by gluing a collection of unit Euclidean cubes
isometrically along their faces, see II.7.32 of [BH99] for a precise definition. Then
the cube complex has a natural piecewise Euclidean metric. This metric is complete
and geodesic since X is finite dimensional (I.7.19 of [BH99]) and is non-positively
curved if the link of each vertex is a flag complex ([Gro87]). If in addition X is
simply connected, then this metric is CAT (0) and X is said to be a CAT (0) cube
complex. We can put a different metric on the 1-skeleton X (1) by considering it as
a metric graph with all edge lengths 1. This is called the `1 metric. We use d for
the CAT (0) metric on X and d`1 for the `1 metric on X (1) . The natural injection
(X (1) , d`1 ) ,→ (X, d) is a quasi-isometry (I.7.31 of [BH99] or Lemma 2.2 of [CS11]).
In this paper, we will mainly use the CAT (0) metric unless otherwise specified.
Also any notions which depend on the metric, like geodesic, convex subset, convex
hull etc, will be understood automatically with respect to the CAT (0) metric unless
otherwise specified.
Definition 2.1. (cf. [CS11, Section 2.1]) A cellular map between CAT (0) cube
complexes is cubical if its restriction σ → τ between cubes factors as σ → η → τ ,
where the first map σ → η is a natural projection onto a face of σ and the second
map η → τ is an isometry.
A geodesic segment, geodesic ray or geodesic in X is an isometric embedding of
[a, b], [0, ∞) or R into X with respect to the CAT (0) metric. A combinatorial geodesic segment, combinatorial geodesic ray or combinatorial geodesic is a `1 -isometric
embedding of [a, b], [0, ∞) or R into X (1) such that its image is a subcomplex.
Let X be a CAT (0) cube complex and let Y ⊂ X be a subcomplex. Then the
following are equivalent (see [Hag08]):
(1) Y is convex with respect to the CAT (0) metric.
(2) Y is a full subcomplex and Y (1) ⊂ X (1) is convex with respect to the `1
metric.
(3) Lk(p, Y ) (the link of p in Y ) is a full subcomplex of Lk(P, X) for every
vertex p ∈ Y .
The collection of convex subcomplexes in a CAT (0) cube complex enjoys the
following version of Helly’s property ([Ger98]):
Lemma 2.2. Let X be as above and {Ci }ki=1 be a collection of convex subcomplexes.
If Ci ∩ Cj 6= ∅ for any 1 ≤ i 6= j ≤ k, then ∩ki=1 Ci 6= ∅.
Lemma 2.3. Let X1 and X2 be two CAT (0) cube complexes and let K ⊂ X1 × X2
be a convex subcomplex. Then K admits a splitting K = K1 × K2 where Ki is a
convex subcomplex of Xi for i = 1, 2.
The lemma is clear when X1 ∼
= [0, 1], and the general case follows from this
special case.
Now we come to the notion of hyperplane, which is the cubical analog of “track”
introduced in [Dun85]. A hyperplane h in a cube complex X is a subset such that
(1) h is connected.
(2) For each cube C ⊂ X, h ∩ C is either empty or a union of mid-cubes of C.
(3) h is minimal, i.e. if there exists h0 ⊂ h satisfying (1) and (2), then h = h0 .
10
JINGYIN HUANG
Recall that a mid-cube of C = [0, 1]n is a subset of form fi−1 (1/2), where fi is one
of the coordinate functions.
If X is a CAT (0) cube complex, then the following are true (see [Sag95]):
(1) Each hyperplane is embedded, i.e. h ∩ C is either empty or a mid-cube of
C (in more general cube complexes, it is possible that h ∩ C contains two
or more mid-cubes of C);
(2) h is a convex subset in X and h with the induced cell structure from X is
also a CAT (0) cube complex;
(3) X \ h has exactly two connected components, they are called halfspaces.
The closure of a halfspace is called closed halfspace, which is also convex in
X with respect to the CAT (0) metric.
(4) Let Nh be the smallest subcomplex of X that contains h. Then Nh is a
convex subcomplex of X and there is a natural isometry i : Nh → h × [0, 1]
such that i(h) = h × {1/2}. Nh is called the carrier of h.
(5) For every edge e ⊂ X, there exists a unique hyperplane he which intersects
e in its midpoint. In this case, we say he is the hyperplane dual to e and e
is an edge dual to the hyperplane he .
(6) Lemma 2.2 is also true for a collection of hyperplanes.
Now it is easy to see an edge path ω ⊂ X is a combinatorial geodesic segment
if and only if there do not exist two different edges of ω such that they are dual
to the same hyperplane. Moreover, for two vertices v, w ∈ X, their `1 distance is
exactly the number of hyperplanes that separate v from w.
Pick an edge e ⊂ X and let πe : X → e ∼
= [0, 1] be the CAT (0) projection. Then
(1) The hyperplane dual to e is exactly πe−1 (1/2).
(2) πe−1 (t) is convex in X for any 0 ≤ t ≤ 1, moreover, if 0 < t < t0 < 1, then
πe−1 (t) and πe−1 (t0 ) are parallel.
(3) Let Nhe be the carrier of the hyperplane dual to e. Then Nhe is the closure
of πe−1 (0, 1). Alternatively, we can describe Nhe as the parallel set of e.
2.3. Coarse intersections of convex subcomplexes.
Lemma 2.4 (Lemma 2.10 of [Hua14]). Let X be a CAT (0) cube complex of dimension n and let C1 , C2 be convex subcomplexes. Put ∆ = d(C1 , C2 ). Let
Y1 = {y ∈ C1 | d(y, C2 ) = ∆} and Y2 = {y ∈ C2 | d(y, C1 ) = ∆}. Then:
(1) Y1 and Y2 are not empty.
(2) Y1 and Y2 are convex; πC1 maps Y2 isometrically onto Y1 and πC2 maps
Y1 isometrically onto Y2 ; the CAT (0) convex hull of Y1 ∪ Y2 is isometric to
Y1 × [0, ∆] (since we are taking the CAT (0) convex hull, it does not has to
be a subcomplex).
(3) Y1 and Y2 are subcomplexes, and πC2 |Y1 is a cubical isomorphism from Y1
to Y2 with its inverse given by πC1 |Y2 .
(4) For any > 0, there exists A = A(∆, n, ) such that if p1 ∈ C1 , p2 ∈ C2
and d(p1 , Y1 ) ≥ > 0, d(p2 , Y2 ) ≥ > 0, then
(2.5)
d(p1 , C2 ) ≥ ∆ + Ad(p1 , Y1 );
d(p2 , C1 ) ≥ ∆ + Ad(p2 , Y2 )
Remark 2.6. Equation (2.5) implies for any r > 0, (C1 ∩r C2 ) ⊂r0 Yi (i = 1, 2),
where r0 = min(1, (2r − ∆)/A) + r and A = A(∆, n, 1). Moreover, ∂T C1 ∩ ∂T C2 =
∂T Y1 = ∂T Y2 .
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
∞
11
∞
The remark implies Y1 = Y2 = (C1 ∩r C2 ) for r large enough. We use I(C1 , C2 ) =
(Y1 , Y2 ) to describe this situation, where I stands for the word “intersect”. The next
lemma gives a combinatorial description of Y1 and Y2 .
Lemma 2.7. Let X, C1 , C2 , Y1 and Y2 be as above. Pick an edge e in Y1 (or Y2 )
and let h be the hyperplane dual to e. Then h ∩ Ci 6= ∅ for i = 1, 2. Conversely, if
a hyperplane h0 satisfies h0 ∩ Ci 6= ∅ for i = 1, 2, then h0 is the dual hyperplane of
some edge e0 in Y1 (or Y2 ). Moreover, I(h0 ∩ C1 , h0 ∩ C2 ) = (h0 ∩ Y1 , h0 ∩ Y2 ).
Proof. The first part of the lemma follows from Lemma 2.4. Let I(h0 ∩C1 , h0 ∩C2 ) =
(Y10 , Y20 ). Pick x ∈ Y10 and let x0 = πh0 ∩C2 (x) ∈ Y20 . Then πh0 ∩C1 (x0 ) = x. Let Nh0 =
h0 ×[0, 1] be the carrier of h0 . Then (h0 ∩Ci )×( 21 −, 12 +) = Ci ∩(h0 ×( 12 −, 12 +))
for i = 1, 2 and < 12 . Thus for any y ∈ C2 , ∠x0 (x, y) ≥ π/2, which implies
x0 = πC2 (x). Similarly, x = πC1 (x0 ) = πC1 ◦ πC2 (x), hence x ∈ Y1 and Y10 ⊂ Y1 .
By the same argument, Y20 ⊂ Y2 , thus Yi0 = Yi ∩ h0 for i = 1, 2 and the lemma
follows.
Lemma 2.4, Remark 2.6 and Lemma 2.7 can also be applied to CAT (0) rectangle complexes of finite type, whose cells are of form Πni=1 [0, ai ]. “Finite type”
means there are only finitely many isometry types of rectangle cells in the rectangle
complex.
Lemma 2.8. Let X, C1 , C2 , Y1 and Y2 be as above. If h is a hyperplane separating
C1 from C2 , then there exists a convex set Y ⊂ h such that Y is parallel to Y1 (or
Y2 ).
Proof. Let ∆ = d(C1 , C2 ) and let M = Y1 × [0, ∆] be the convex hull of Y1 and Y2 .
We want to prove M ∩ h = Y1 × {t} ⊂ Y1 × [0, ∆] for some t ∈ [0, ∆]. It suffices to
show for any edge e ⊂ Y1 , (e × [0, ∆]) ∩ h = e × {t} for some t.
Pick point x ∈ e and let {x} × {t} be the point in M = Y1 × [0, ∆]. Since e × {t}
is parallel to e, e × {t} sits inside a cube and e × {t} is parallel to an edge of this
cube. Thus either e × {t} ⊂ h or e × {t} is parallel to some edge dual to h, but
the second case implies that h is dual to e and h ∩ Y1 6= ∅, which is impossible, so
e × {t} ⊂ (e × [0, ∆]) ∩ h. Now we are done since ({x} × [0, ∆]) ∩ h is exactly one
point for each x ∈ e.
2.4. Right-angled Artin groups. Pick a finite simplicial graph Γ. Let G(Γ) be a
RAAG. A generating set S ⊂ G(Γ) is called a standard generating set if all relators
in the associated group presentation are commutators. Each standard generating
set S determines a graph ΓS whose vertices are elements in S and two vertices
are adjacent if the corresponding group elements commute. It follows from [Dro87]
that the isomorphism type of ΓS does not depend on the choice of the standard
generating set S, in particular, ΓS is isomorphic to Γ.
Let S be a standard generating set for G(Γ). We label the vertices of Γ by
elements in S. G(Γ) has a nice Eilenberg-MacLane space S(Γ), called the Salvetti
complex (see [CD95, Cha07]). S(Γ) is a non-positively curved cube complex. The
2-skeleton of S(Γ) is the usual presentation complex of G(Γ). If the presentation
complex contains a copy of 2-skeleton of a 3-torus, then we attach a 3-cell to obtain
a 3-torus. We can build S(Γ) inductively in this manner, and this process will stop
after finitely many steps. The closure of each k-cell in S(Γ) is a k-torus. Torus
of this kind is called a standard torus. There is a 1-1 correspondence between the
k-cells (or standard tori of dimension k) in S(Γ) and k-cliques (complete subgraph
12
JINGYIN HUANG
of k vertices) in Γ, thus dim(S(Γ)) = dim(F (Γ)) + 1. We define the dimension of
G(Γ) to be the dimension of S(Γ).
Denote the universal cover of S(Γ) by X(Γ), which is a CAT (0) cube complex.
Our previous labeling of vertices of Γ induces a labeling of the standard circles
of S(Γ), which lifts to a labeling of edges of X(Γ). We choose an orientation for
each standard circle of S(Γ) and this would give us a directed labeling of the edges
in X(Γ). If we pick a base point v ∈ X(Γ) (v is a vertex), then there is a 1-1
correspondence between words in G(Γ) and edge paths in X(Γ) which start at v.
Each full subgraph Γ0 ⊂ Γ gives rise to a subgroup G(Γ0 ) ,→ G(Γ). Subgroup of
this kind is called a S-standard subgroup and a left coset of an S-standard subgroup
is called a S-standard coset (we will omit S when the generating set is clear).
There is also an embedding S(Γ0 ) ,→ S(Γ) which is locally isometric. Let p :
X(Γ) → S(Γ) be the covering map. Then each connect component of p−1 (S(Γ0 )) is
a convex subcomplex isometric to X(Γ0 ). We will call these components standard
subcomplexes with defining graph Γ0 . A standard k-flat is a standard complex which
covers a standard k-torus in S(Γ). When k = 1, we also call it a standard geodesic.
We pick an identification of the Cayley graph of G(Γ) with the 1-skeleton of
X(Γ), hence G(Γ) is identified with the vertices of X(Γ). Let v ∈ X(Γ) be the base
vertex which corresponds to the identity in the Cayley graph of G(Γ). Then for
any h ∈ G(Γ), the convex hull of {hgv}g∈G(Γ0 ) is a standard subcomplex associated
with Γ0 . Thus there is a 1-1 correspondence between standard subcomplexes with
defining graph Γ0 in X(Γ) and left cosets of G(Γ0 ) in G(Γ).
Note that for every edge e ∈ X(Γ), there is a vertex in Γ which shares the same
label as e, and we denote this vertex by Ve . If K ⊂ X(Γ) is a subcomplex, we define
VK to be {Ve | e is an edge in K} and ΓK to be the full subgraph spanned by VK .
ΓK is called the support of K. In particular, if K is a standard subcomplex, then
the defining graph of K is ΓK .
Every finite simplicial graph Γ admits a canonical join decomposition Γ = Γ1 ◦
Γ2 ◦· · ·◦Γk , where Γ1 is the maximal clique join factor and for 2 ≤ i ≤ k, Γi does not
allow any non-trivial join decomposition and is not a point. Γ is irreducible if this
join decomposition is trivial. This decomposition induces a product decomposition
X(Γ) = En ×Πki=2 X(Γi ), which is called the De Rahm decomposition of X(Γ). This
is consistent with the canonical product decomposition of CAT (0) cube complex
discussed in Section 2.5 of [CS11].
We turn to the asymptotic geometry of RAAG’s. A right-angled Artin group
G(Γ) is one-ended if and only if Γ is connected. Moreover, the n-connectivity at
infinity of G(Γ) can be read off from Γ, see [BM01]. In order to classify all RAAG’s
up to quasi-isometry, it suffices to consider those one-ended RAAG’s. This follows
from the main results in [PW02]. Moreover, Lemma 3.2 of [PW02] implies the
following:
Lemma 2.9. If q : X(Γ) → X(Γ0 ) is an (L, A)-quasi-isometry, then there exists
D = D(L, A) > 0 such that for any connected component Γ1 ⊂ Γ such that Γ1
is not a point and any standard subcomplex K1 ⊂ X(Γ) with defining graph Γ1 ,
there is a unique connected component Γ01 ⊂ Γ0 and a unique standard subcomplex
K10 ⊂ X(Γ0 ) with defining graph Γ01 such that dH (q(K1 ), K10 ) < D.
It is shown in [BC12] and [ABD+ 13] that G(Γ) has linear divergence if and only
if Γ is a join or Γ is one point, which implies Γ being a join is a quasi-isometric
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
13
invariant. Moreover, their results together with Theorem B of [KKL98] implies that
the De Rahm decomposition is stable under quasi-isometry:
Qk
Theorem 2.10. Given X = X(Γ) and X 0 = X(Γ0 ), let X = Rn × i=1 X(Γi )
Qk0
0
and let X 0 = Rn × j=1 X(Γ0j ) be the corresponding De Rahm decomposition. If
φ : X → X 0 is an (L, A) quasi-isometry, then n = n0 , k = k 0 and there exist
constants L0 = L0 (L, A), A0 = A0 (L, A), D = D(L, A) such that after re-indexing
the factors in X 0 , we have (L0 , A0 ) quasi-isometry φi : X(Γi ) → X(Γ0i ) so that
Qk
Qk
Qk
d(p0 ◦ φ, i=1 φi ◦ p) < D, where p : X → i=1 X(Γi ) and p0 : X 0 → i=1 X(Γ0i )
are the projections.
Thus in order to study the quasi-isometric classification of RAAG’s, it suffices
to study those RAAG’s which are one-ended and irreducible, but this will rely on
finer quasi-isometric invariant of RAAG’s.
Recall that in the case of Gromov hyperbolic spaces, quasi-isometries map geodesics
to geodesics up to finite Hausdorff distance, hence induces a well-defined boundary
map. The analog of this fact for 2-dimensional RAAG’s has been established in
[BKS08b], i.e. quasi-isometries map 2-flats to 2-flats up to finite Hausdorff distance.
The following is a higher dimensional generalization of Theorem 3.10 of [BKS08b].
Theorem 2.11 (Theorem 5.20 of [Hua14]). If φ : X(Γ1 ) → X(Γ2 ) is an (L, A)quasi-isometry, then dim(X(Γ1 )) = dim(X(Γ2 )). And there is a constant D =
D(L, A) such that for any top-dimensional flat F1 ⊂ X(Γ1 ), there is a unique flat
F2 ⊂ X(Γ2 ) such that dH (φ(F1 ), F2 ) < D.
For each right-angled Artin group G(Γ), there is a simplicial graph Γe , called the
extension graph, which is introduced in [KK13b]. Extension graphs can be viewed
as “curve graphs” for RAAG’s ([KK14]).
Definition 2.12 (Definition 1 of [KK13b]). The vertex set of Γe consists of words
in G(Γ) that are conjugate to elements in S (recall that S is a standard generating
set for G(Γ)), and two vertices are adjacent in Γe if and only if the corresponding
words commute in G(Γ).
The flag complex of the extension graph is called the extension complex.
Since the curve graph captures the combinatorial pattern of how Dehn twist flats
intersect in a mapping class group, it plays an important role in the quasi-isometric
rigidity of mapping class group ([Ham05, BKMM12]). Similarly, we will see in
Section 4 that the extension graph captures the combinatorial pattern of the coarse
intersection of certain collection of flats in a RAAG, and it is a quasi-isometric
invariant for certain classes of RAAG’s.
3. Stable subgraph
We study the behavior of certain standard subcomplexes under quasi-isometries
in this section.
3.1. Coarse intersection of standard subcomplexes and flats.
Lemma 3.1. Let Γ be a finite simplicial graph and let K1 , K2 be two standard
subcomplexes of X(Γ). If (Y1 , Y2 ) = I(K1 , K2 ), then Y1 and Y2 are also standard
subcomplexes.
14
JINGYIN HUANG
Proof. The lemma is clear if K1 ∩ K2 6= ∅. Now we assume d(K1 , K2 ) = c > 0. Pick
a vertex v1 ∈ K1 , by Lemma 2.4, there exists vertex v2 ∈ K2 such that d(v1 , v2 ) = c.
Let l : [0, c] → X(Γ) be the unit speed geodesic from v1 to v2 . We can find sequence
of cubes {Bi }N
i=1 and 0 = t0 < t1 < ... < tN −1 < tN = c such that each Bi contains
{l(t) | ti−1 < t < ti } as interior points.
Let Vl = ∪N
i=1 VBi (recall that VBi is the collection of the labels of edges in Bi ,
cf. Section 2.4) and let Vi = VKi for i = 1, 2. Put V 0 = V1 ∩ V2 ∩ Vl⊥ (Vl⊥ denotes
the orthogonal complement of V1 , see Section 2.1) and let Γ0 be the full subgraph
spanned by V 0 . Let Y10 be the standard subcomplex that has defining graph Γ0 and
contains v1 (if V 0 is empty, then Y10 = v1 ). We claim Y1 = Y10 .
Pick an edge e ⊂ K1 such that v1 ∈ e and Ve ∈ V 0 . Let h be the hyperplane
dual to e and Nh ∼
= h × [0, 1] be the carrier of h. Since d(Ve , w) = 1 for any w ∈ Vl ,
we can assume l ⊂ h × {1} ⊂ Nh . By our definition of V 0 , there is an edge e0 ∈ K2
with v2 ∈ e0 and h dual to e0 , thus e and e0 cobound an isometrically embedded
flat rectangle (one side of the rectangle is l), which implies e ⊂ Y1 . Let l0 be the
side of the rectangle opposite to l. We can define Vl0 similarly as we define Vl ,
then Vl0 = Vl . Now let ω be any edge path starting at v1 such that Ve0 ∈ V 0 for
any edge e0 ⊂ ω. Then it follows from the above argument and induction on the
combinatorial length of ω that ω ⊂ Y1 , thus Y10 ⊂ Y1 .
For the other direction, since Y1 is a convex subcomplex by Lemma 2.4, it suffices
to prove every vertex of Y1 belongs to Y10 . By the induction argument as above, we
only need to show for edge e1 with v1 ∈ e1 , if e1 ⊂ Y1 , then e1 ⊂ Y10 . Lemma 2.4
implies that there exists edge e2 ⊂ Y2 such that e1 and e2 cobound an isometrically
embedded flat rectangle (one side of the rectangle is l). So l is in the carrier of the
hyperplane dual to e1 . It follows that Ve1 ∈ V 0 and e1 ⊂ Y10 .
Corollary 3.2. Let K1 , K2 , Y1 and Y2 be as above. Let h be a hyperplane separating
K1 and K2 and let e be an edge dual to h. Then Ve ∈ VY⊥1 = VY⊥2 . In particular,
pick vertex v ∈ Γ, then v ∈ VY1 if and only if
(1) v ∈ VK1 ∩ VK2 .
(2) For any hyperplane h0 separating K1 from K2 and any edge e0 dual to h0 ,
d(v, Ve0 ) = 1.
Proof. Let l and Vl be the same as the proof of Lemma 3.1. Let Vl0 be a collection
of vertices of Γ such that v ∈ Vl0 if and only if v = Ve0 for some edge e0 ⊂ X(Γ)
satisfying (2). It suffices to prove Vl0 = Vl .
Vl0 ⊂ Vl is clear since if a hyperplanes h separates K1 from K2 , then l intersects h
transversally at one point. To see Vl ⊂ Vl0 , it suffices to show h∩Ki = ∅ for i = 1, 2,
where h is a hyperplane that intersects l transversally. Let x = l ∩ h. Suppose
h ∩ K1 6= ∅ and let x0 = πh∩K1 (x). Now consider the triangle ∆(v1 , x, x0 ) (recall
that v1 = l(0)), we have ∠v1 (x, x0 ) ≥ π/2 (since πK1 (x) = v1 ), ∠x0 (v1 , x) ≥ π/2 (see
the proof of Lemma 2.7) and ∠x (v1 , x0 ) > 0, which is a contradiction, so h∩K1 = ∅,
similarly h ∩ K2 = ∅.
Remark 3.3. Recall that a standard coset of G(Γ) is a left coset of a standard
subgroup of G(Γ). Lemma 3.1 implies that for each pair of standard cosets of G(Γ),
we can associated another standard coset, which captures the coarse intersection of
the pair. Moreover, we can also define a notion of distance between two standard
cosets, which takes value on G(Γ).
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
15
Lemma 3.4. Let K ⊂ X(Γ) be a convex subcomplex and let Γ0 = lk(ΓK ) (ΓK is
the support of K, see Section 2.4; lk(ΓK ) is the full subgraph spanned by VK⊥ , see
Section 2.1). Then the parallel set PK of K is a convex subcomplex and canonically
splits as K × X(Γ0 ).
Note that we do not require K to satisfy geodesic extension property.
Proof. Pick a vertex v ∈ K. Let Γ00 = ΓK and let P1 be the unique standard
subcomplex that passes through v and has defining graph Γ0 ◦ Γ00 (recall that ◦
denotes the graph join). Then K ⊂ P1 . Let P 0 be the natural copy of K × X(Γ0 )
inside P1 . It is clear that P 0 ⊂ PK .
Let K 0 be a convex subset parallel to K and let φ : K → K 0 be the isometry
induced by CAT (0) projection onto K 0 . Pick vertex v ∈ K and let l be the geodesic
segment connecting v and φ(v). We define Vl as in the proof of Lemma 3.1 (note
that φ(v) is not necessarily a vertex). Let e be any edge such that v ∈ e ⊂ K. Then
there is a flat rectangle with e, φ(e) and l as its three sides. Thus l is contained
in the carrier of the hyperplane dual to e and Vl ⊂ Ve⊥ . Note that if l0 is the side
opposite to l, then Vl0 = Vl . For any given edge e0 ⊂ K, we can find an edge path
ω ⊂ K such that e is the first and e0 is the last edge in ω. By induction on the
combinatorial length of w and the same argument as above, we can show Vl ⊂ Ve⊥0 ,
thus Vl ⊂ VK⊥ and K 0 ⊂ P 0 . It follows that PK ⊂ P 0 , so PK = P 0 .
Remark 3.5. The following is a generalization of the above lemma for general
CAT (0) cube complexes. Let X be a CAT (0) cube complex. A convex set K ⊂ X
is regular if for any x ∈ K, Σx K (the space of direction of x in K, see Chapter II.3
of [BH99]) satisfies:
(1) Σx K is a subcomplex of Σx X with respect to the canonical all-right spherical complex structure on Σx X.
(2) There exists r > 0 such that B(x, r) ∩ K is isometric to the r-ball centred
at the cone point in the Euclidean cone over Σx K.
If K ⊂ X is a regular convex subset, the PK is convex and admits a splitting
PK ∼
= K × N where N has an induced cubical structure from X (N is CAT (0)).
Lemma 3.6. Let q : X(Γ1 ) → X(Γ2 ) be an (L, A)-quasi-isometry and let F ⊂
X(Γ1 ) be a subcomplex isometric to Ek . Suppose n = dim(X(Γ1 )) = dim(X(Γ2 )).
R
If there exist R1 > 0, R2 > 0 and top dimensional flats F1 , F2 such that F =2
∞
F1 ∩R1 F2 and F = F1 ∩R F2 for any R ≥ R1 , then
(1) There exist a constant D = D(L, A, R1 , R2 , n) and a subcomplex F 0 ⊂
D
X(Γ2 ) isometric to Ek such that q(F ) = F 0 .
D0
(2) There exists a constant D0 = D0 (L, A) such that q(PF ) = PF 0 .
Proof. By Theorem 2.11, there exist top dimensional flats F10 ⊂ X(Γ2 ) and F20 ⊂
D
X(Γ2 ) such that q(Fi ) =1 F10 for D1 = D1 (L, A) and i = 1, 2. Thus there exists
R0 = R0 (L, A, R1 , R2 ) and R3 = R3 (L, A, R1 , R2 ) > R1 such that q(F1 ∩R1 F2 ) ⊂
F10 ∩R0 F20 ⊂ q(F1 ∩R3 F2 ), this and Remark 2.6 imply
(3.7)
D
q(F1 ∩R1 F2 ) =2 F10 ∩R0 F20
for D2 = D2 (n, d(F1 , F2 )) = D2 (L, A, R1 , R2 , n).
16
JINGYIN HUANG
Let (Y1 , Y2 ) = I(F10 , F20 ). Then there exists D3 = D3 (L, A, R1 , R2 , n) such that
D
Y1 =3 F10 ∩R0 F20 .
(3.8)
From (3.7) and (3.8), we have
D
q(F ) =4 Y1
(3.9)
for D4 = D4 (L, A, R1 , R2 , n). By Lemma 2.4, Y1 is a convex subcomplex of F10 .
Qk0
This together with (3.9) imply Y1 = F 0 × i=1 Ii where F 0 is isometric to Ek and
Qk 0
0
{Ii }ki=1 are finite intervals. Moreover, by (3.9), diam( i=1 Ii ) must be bounded in
terms of D4 , L and A, thus (1) follows.
Let {Fλ }λ∈Λ be the collection of top dimensional flats in X(Γ1 ) which are contained in the parallel set PF of F . Lemma 3.4 implies
(3.10)
dH (∪λ∈Λ Fλ , PF ) ≤ 1.
For λ ∈ Λ, there exists Rλ > 0 such that F ⊂Rλ Fλ . Let Fλ0 be the top dimensional
D
flat in X(Γ2 ) such that q(Fλ ) =1 Fλ0 . Then by (1), there exists Rλ0 > 0 such that
F 0 ⊂Rλ0 (Fλ0 ). This and Lemma 2.4 imply Fλ0 ⊂ PF 0 for any λ ∈ Λ. Thus by (3.10),
there exists D0 = D0 (L, A) such that q(PF ) ⊂D0 PF 0 . And (2) follows by running
the same argument for the quasi-isometry inverse of q.
A tree product is a convex subcomplex K ⊂ X(Γ) such that K splits as a product
of trees, i.e. there exists a cubical isomorphism K ∼
= Πni=1 Ti where Ti ’s are trees.
A standard tree product is a tree product which is also a standard subcomplex.
One can check that K is a standard tree product if and only if the defining graph
ΓK of K has a join decomposition ΓK = Γ1 ◦ Γ2 ◦ · · · ◦ Γn where each Γi is discrete.
Thus one can choose the above Ti ’s to be standard subcomplexes of K. Note that
every standard flat is a standard tree product, and every subcomplex isometric to
Ek is a tree product.
0
Lemma 3.11. Suppose
Qn dim(X(Γ)) = n. Let q : X(Γ) → X(Γ ) be a quasiisometry. Let K = i=1 Ti be a top dimensional tree product with its tree factors.
Then there exists a standard tree product K 0 in X(Γ0 ) such that q(K) ⊂∞ K 0 .
The proof essentially follows [BKS08a, Theorem 4.2].
Proof. For 1 ≤ i ≤ n, let Vi = VTi ∈ Γ be the collection of labels of edges in
Ti . The case where all Vi ’s are consist of one point follows from Theorem 2.11. If
each Vi contains at least two points, then by Lemma 3.6, for any geodesic l ⊂ Ti ,
∞
there exists a subcomplex l0 ⊂ X(Γ0 ) isometric to R such that q(l) = l0 . Since l0 is
0
unique up to parallelism, the collection of labels of edges in l does not depend on
the choice of l0 and will be denoted by Vq(l) . For 1 ≤ i ≤ n, define Vi0 = ∪l⊂Ti Vq(l)
where l varies among all geodesics in Ti .
We claim Vi0 ⊂ (Vj0 )⊥ for i 6= j. To see this, pick geodesic li ∈ Ti and let
Qn
F = i=1 li . Then there exist top dimensional flat F 0 and geodesic lines {li0 }ni=1
∞
∞
(li0 is a subcomplex) in X(Γ0 ) such that q(F ) = F 0 and q(li ) = li0 . Since li0 ⊂∞ F 0 ,
by Lemma 2.4, we can assume li0 is a subcomplex of F 0 . Pick i 6= j. Since li and lj
are orthogonal, they have infinite Hausdorff distance. Thus li0 and lj0 have infinite
Hausdorff distance. By our assumption, li0 and lj0 are isometric to E1 , and they
are convex subcomplexes of F 0 ∼
= En . Thus either li0 and lj0 are parallel, or they
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
17
are orthogonal. The former is impossible since li0 and lj0 have infinite Hausdorff
distance. Thus {li0 }ni=1 is a mutually orthogonal collection.
Let Γ01 = V10 ◦V20 ◦· · ·◦Vn0 ⊂ Γ0 . Then each Vi0 has to be a discrete full subgraph by
our dimension assumption. Let {Fλ }λ∈Λ be the collection of top dimensional flats
∞
in K and let Fλ0 be the unique flat such that q(Fλ ) = Fλ0 . Note that for arbitrary
Fλ1 and Fλ2 , there exists a finite chain in {Fλ }λ∈Λ which starts with Fλ1 and ends
with Fλ2 such that the intersection of adjacent elements in the chain contains a top
dimensional orthant. Thus the collection {Fλ0 }λ∈Λ also have this property. Then
∪λ∈Λ Fλ0 is contained in a standard subcomplex of X(Γ0 ) with defining graph Γ01 .
It remains to deal with the case where there exist i 6= j such that |Vi | = 1 and
|Vj | ≥ 2. We suppose |V
Qi |m= 1 for 1 ≤ i ≤ m and |Vi | ≥ 2 for i > m. By applying
Lemma 3.6 with F = i=1 Ti , we can reduce to lower dimensional case and the
lemma follows by induction on dimension.
Corollary 3.12. Let q : X(Γ) → X(Γ0 ) be a quasi-isometry and let K be a top
dimensional maximal standard tree product, i.e. K is not properly contained in
another tree product. Then there exists a standard tree product K 0 ⊂ X(Γ0 ) such
∞
that q(K) = K 0 .
3.2. Standard flats in transvection free RAAG’s. Up to now, we have only
dealt with top dimensional standard subcomplexes. The next goal is to study those
standard subcomplexes which are not necessarily top dimensional. In particular, we
are interested in whether quasi-isometries will preserve standard flats up to finite
Hausdorff distance. The answer turns out to be related to the outer automorphism
group of G(Γ).
One direction is obvious, namely, if for any quasi-isometry q : X(Γ) → X(Γ0 ), q
maps any standard flat in X(Γ) to a standard flat in X(Γ0 ) up to finite Hausdorff
distance, then Out(G(Γ)) must be transvection free (i.e. Out(G(Γ)) does not contain any transvections). The converse is also true. Now we set up several necessary
tools to prove the converse.
In this section, Γ will always be a finite simplicial graph.
Definition 3.13. A subgraph Γ1 ⊂ Γ is stable in Γ if
(1) Γ1 is a full subgraph.
(2) Let K ⊂ X(Γ) be a standard subcomplex such that ΓK = Γ1 and let Γ0
be a finite simplicial graph so that, for some L, A, there is an (L, A)-quasiisometry q : X(Γ) → X(Γ0 ). Then there exists D = D(L, A, Γ1 , Γ) > 0 and
a standard subcomplex K 0 ⊂ X(Γ0 ) such that dH (q(K), K 0 ) < D.
For simplicity, we will also say the pair (Γ1 , Γ) is stable in this case. A standard
subcomplex K ⊂ X(Γ) is stable if it arises from a stable subgraph of Γ.
We claim the defining graph ΓK 0 of K 0 is stable in Γ0 . To see this, pick any
graph Γ00 so that there is an (L, A)-quasi-isometry q 0 : X(Γ0 ) → X(Γ00 ), and pick
any standard subcomplex K10 ⊂ X(Γ0 ) with defining graph ΓK 0 . Note that there
is an isometry i : X(Γ0 ) → X(Γ0 ) such that i(K 0 ) = K10 . Since the map q 0 ◦ i ◦ q
is a quasi-isometry from X(Γ) to X(Γ00 ), then by the stability of Γ1 , q 0 ◦ i ◦ q(K)
is Hausdorff close to a standard subcomplex in X(Γ00 ), hence the same is true for
q 0 (K10 ). It follows from this claim that one can obtain quasi-isometric invariants by
identifying certain classes of stable subgraphs.
18
JINGYIN HUANG
It is immediate from the definition that for finite simplicial graphs Γ1 ⊂ Γ2 ⊂ Γ3 ,
if (Γ1 , Γ2 ) is stable and (Γ2 , Γ3 ) is stable, then (Γ1 , Γ3 ) is stable. However, it is not
necessarily true that if (Γ1 , Γ3 ) and (Γ2 , Γ3 ) are stable, then (Γ1 , Γ2 ) is stable. In
the sequel, we will investigate several other properties of stable subgraph. The
following lemma is an easy consequence of Lemma 3.1 and Remark 2.6:
Lemma 3.14. Suppose Γ1 and Γ2 are stable in Γ. Then Γ1 ∩ Γ2 is also stable in
Γ.
The following result follows from Lemma 2.9.
Lemma 3.15. If Γ1 is stable in Γ, then every connected component of Γ1 that
contains more than one point is also stable in Γ.
Lemma 3.16. Suppose Γ1 is stable in Γ. Let V be the vertex set of Γ1 and let Γ2
be the full subgraph spanned by V and the orthogonal complement V ⊥ . Then Γ2 is
also stable in Γ.
Proof. Let K2 ⊂ X(Γ) be a standard subcomplex with its defining graph ΓK2 = Γ2
and let K1 ⊂ K2 be any standard subcomplex satisfying ΓK1 = Γ1 . Lemma 3.4
implies K2 = PK1 = K1 × K1⊥ . For vertex x ∈ K1⊥ , denote Kx = K1 × {x}.
Let q : X(Γ) → X(Γ0 ) be an (L, A)-quasi-isometry. Then there exists standard
∞
subcomplex Kx0 such that dH (q(Kx ), Kx0 ) < D = D(L, A, Γ1 , Γ). Thus Kx0 = Ky0 for
vertices x, y ∈ K1⊥ . It follows from Lemma 3.1 that Kx0 and Ky0 are parallel. Thus
q(PK1 ) ⊂R PKx0 for R = D+L+A. Moreover, PKx0 is also a standard subcomplex by
Lemma 3.4. By considering the quasi-isometry inverse and repeating the previous
∞
argument, we know q(PK1 ) = PKx0 , thus Γ2 is also stable in Γ.
Lemma 3.17. Suppose Γ1 is stable in Γ. Pick vertex v ∈
/ Γ1 , then the full subgraph
spanned by v ⊥ ∩ Γ1 is stable in Γ.
Proof. We use Γ2 to denote the full subgraph spanned by v ⊥ ∩ Γ1 . Let K2 ⊂ X(Γ)
be a standard subcomplex such that ΓK2 = Γ2 and let K1 ⊂ X(Γ) be the unique
standard subcomplex such that ΓK1 = Γ1 and K2 ⊂ K1 . Pick vertex x ∈ K2
and let e ⊂ X(Γ) to be the edge such that Ve = v and x ∈ e. Suppose x̄ is the
other end point of e. Let K̄i be the standard subcomplex that contains x̄ and has
defining graph Γi for i = 1, 2. Denote the hyperplane dual to e by h. Since v ∈
/ Γ1 ,
h ∩ K1 = ∅ and h ∩ K̄1 = ∅, thus h separates K1 and K̄1 and d(K1 , K̄1 ) = 1. It
D
follows from Corollary 3.2 that I(K1 , K̄1 ) = (K2 , K̄2 ), in particular K2 = K1 ∩R K̄1
for D depending on R and the dimension of X(Γ). Now the lemma follows since
Γ1 is stable.
The next result is a direct consequence of Corollary 3.12.
Lemma 3.18. If Γ1 is stable in Γ, then there exists Γ2 which is stable in Γ1 such
that
(1) Γ2 is a graph join Γ̄1 ◦ Γ̄2 ◦ · · · ◦ Γ̄k where Γ̄i is discrete for 1 ≤ i ≤ k.
(2) k = dim(X(Γ1 )).
Lemma 3.19. Let Γ be a finite simplicial graph such that there do not exist vertices
v 6= w of Γ such that v ⊥ ⊂ St(w). Then every stable subgraph of Γ contains a stable
vertex.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
19
Proof. Let Γ1 be a minimal stable subgraph, i.e. it does not properly contain any
stable subgraph of Γ. It suffices to show Γ1 is a point. We argue by contradiction
and assume Γ1 contains more than one point.
First we claim Γ1 can not be discrete. Suppose the contrary is true. Pick vertices
v, w ∈ Γ1 and pick vertex u ∈ v ⊥ \ St(w). By Lemma 3.17, u⊥ ∩ Γ1 is also stable.
Note that v ∈ u⊥ ∩ Γ1 and w ∈
/ u⊥ ∩ Γ1 , which contradicts the minimality of Γ1 .
We claim Γ1 must be a clique. Since Γ1 is not discrete, by Lemma 3.18, we can
find a stable subgraph
(3.20)
Γ2 = Γ̄1 ◦ Γ̄2 ◦ · · · ◦ Γ̄m ⊂ Γ1
where {Γ̄i }m
i=1 are discrete full subgraphs and m ≥ 2. Then Γ2 = Γ1 . Suppose
some Γ̄i contains more than one point, and let Γ3 be the join of the remaining join
factors. Then Theorem 2.10 implies that Γ3 is stable, contradicting the minimality
of Γ1 . Therefore Γ1 is a clique.
Pick distinct vertices v1 , v2 ∈ Γ1 . By our assumption, there exists vertex w ∈
v1⊥ \ St(v2 ). Since Γ1 is a clique, Γ1 ⊂ St(v2 ), then w ∈
/ Γ1 . Let Γ4 be the full
subgraph spanned by w⊥ ∩ Γ1 . Then Γ4 is stable by Lemma 3.17. Moreover,
Γ4 ( Γ1 (since v2 ∈
/ Γ4 ), which yields a contradiction.
Lemma 3.21. Let Γ be as in Lemma 3.19 and let Γ1 be a stable subgraph of Γ. Then
for any vertex w ∈ Γ1 , there exists a stable vertex v ∈ Γ1 such that d(v, w) ≤ 1.
Proof. Denote the combinatorial distance in Γ and Γ1 by d and d1 respectively.
Since Γ1 is a full subgraph, d(x, y) = 1 if and only if d1 (x, y) = 1 and d(x, y) ≥ 2 if
and only if d1 (x, y) ≥ 2 for vertices x, y ∈ Γ1 . If w is isolated in Γ1 , then we can
use the argument in the second paragraph of the proof of Lemma 3.19 to get rid of
all vertices in Γ1 except w, which implies w is a stable vertex. If w is not isolated,
we can assume Γ1 is connected by Lemma 3.15.
By Lemma 3.19, there exists a stable vertex u ∈ Γ1 . If d1 (u, w) ≤ 1, then we are
done, otherwise let ω be a geodesic in Γ1 connecting u and w (ω might not be a
geodesic in Γ) and let {vi }ni=0 be the consecutive vertices in ω, here v0 = w, vn = u
and n = d1 (w, u).
Since u is stable, by Lemma 3.16, St(u) is also stable. Note that d1 (vn−2 , u) = 2,
⊥
so d(vn−2 , u) = 2 and vn−2 ∈
/ St(u). Lemma 3.17 implies vn−2
∩ St(u) is stable and
⊥
⊥
by Lemma 3.14, vn−2 ∩St(u)∩Γ1 is also stable. Note that vn−2 ∩St(u)∩Γ1 6= ∅ since
⊥
it contains vn−1 . Lemma 3.19 implies there is a stable vertex u0 ∈ vn−2
∩ St(u) ∩ Γ1
0
and it is easy to see d1 (w, u ) = n − 1. Now the lemma follows by induction.
Lemma 3.22. Let Γ be as in Lemma 3.19. Then every vertex of Γ is stable.
Proof. Let Γw be the intersection of all stable subgraphs that contain w. By Lemma
3.14, Γw is the minimal stable subgraph that contains w. It suffices to prove
Γw = {w}. We argue by contradiction and denote the vertices in Γw \ {w} by
{vi }ki=1 . The minimality of Γw implies we can not use Lemma 3.17 to get rid of
some vi while keep w, thus w⊥ \ St(vi ) ⊂ {v1 · · · vi−1 , vi+1 · · · vk } for any i, in other
words
(3.23)
w⊥ ⊂ St(vi ) ∪ {v1 · · · vi−1 , vi+1 · · · vk }
for 1 ≤ i ≤ k. Then there does not exist i such that Γw ⊂ St(vi ), otherwise we
would have w⊥ ⊂ St(vi ) by (3.23).
20
JINGYIN HUANG
On the other hand, Lemma 3.21 implies there exists a stable vertex u ∈ Γw with
d(w, u) = 1. Then St(u) is stable (Lemma 3.16) and St(u) ∩ Γw is stable (Lemma
3.14). Note that w ∈ St(u) ∩ Γw , by the minimality of Γw , Γw ⊂ St(u), which
yields a contradiction.
Lemma 3.24. Let Γ be a finite simplicial graph and pick stable subgraphs Γ1 , Γ2
of Γ. Let Γ̄ be the full subgraph spanned by V and V ⊥ where V = VΓ1 . If Γ2 ⊂ Γ̄,
then the full subgraph spanned by Γ1 ∪ Γ2 is stable in Γ.
To simplify notation, in the following proof, we will denote q(K) ≈ K 0 where
q, K and K 0 are as in Definition 3.13. We will also assume without loss of generality
that q(K) ⊂ K 0 .
Proof. Let q : X(Γ) → X(Γ0 ) be an (L, A)-quasi-isometry. Suppose K1 and K
are standard subcomplexes in X(Γ) such that ΓK1 = Γ1 , ΓK = Γ̄ and K1 ⊂ K.
Put K 0 ≈ q(K), K10 ≈ q(K1 ), K = K1 × K1⊥ and K 0 = K10 × K10⊥ . The proof of
Lemma 3.16 implies there exist a quasi-isometry q 0 : K1⊥ → K10⊥ and a constant
D1 = D1 (L, A, Γ1 , Γ) such that
(3.25)
d(q 0 ◦ p2 (x), p02 ◦ q(x)) < D1
for any x ∈ K where p2 : K → K1⊥ and p02 : K 0 → K10⊥ are projections.
Let Γ2 = Γ21 ◦Γ22 where Γ21 = Γ1 ∩Γ2 and let K22 , K2 be standard subcomplexes
such that ΓK22 = Γ22 , ΓK2 = Γ2 and K22 ⊂ K2 ⊂ K. By (3.25), it suffices to prove
0
there exist a standard subcomplex K22
⊂ K 0 and a constant D = D(L, A, Γ1 , Γ2 , Γ)
0
0
such that dH (p2 ◦ q(K22 ), K22 ) < D. Let K20 ≈ q(K2 ). Then K20 ⊂ K 0 and p02 (K20 )
∞
∞
is a standard subcomplex. By (3.25), p02 ◦ q(K22 ) = p02 ◦ q(K2 ) = p02 (K20 ), thus we
0
can take K22
= p02 (K20 ).
Remark 3.26. In general the full subgraph spanned by Γ1 ∪ Γ2 is not necessarily
stable even if Γ1 and Γ2 are stable, see Remark 3.35.
The next theorem follows from Lemma 3.22 and Lemma 3.24.
Theorem 3.27. Given finite simplicial graph Γ, the following are equivalent:
(1) Out(G(Γ)) is transvection free.
(2) For any (L, A)-quasi-isometry q : X(Γ) → X(Γ0 ), there exists positive constant D = D(L, A, Γ) such that for any standard flat F ⊂ X(Γ), there exists
a standard flat F 0 ⊂ X(Γ0 ) such that dH (q(F ), F 0 ) < D.
3.3. Standard flats in general RAAG’s. At this point, we have the following
natural questions:
(1) In Theorem 3.27, is it true that every standard flat in X(Γ0 ) comes from
some standard flat in X(Γ)? A related question could be, is condition (1)
in Theorem 3.27 a quasi-isometric invariant for right-angled Artin groups?
(2) What can we say about the stable subgraphs of Γ if we drop condition (1)
in Theorem 3.27?
We will first give a negative answer to question (1) in Example 3.29 below. Then
we will prove the Theorem 3.28, which answers question (2). Section 4, and in
particular the proof of Theorem 1.1, will not depend on this subsection. However,
we will need Theorem 3.28 and Lemma 3.30 for Section 5.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
21
Theorem 3.28. Let Γ be an arbitrary finite simplicial graph. A clique Γ1 ⊂ Γ is
stable if and only if there do not exist vertices w ∈ Γ1 and v ∈ Γ \ Γ1 such that
w⊥ ⊂ St(v).
In other words, the clique Γ1 is stable if and only if the corresponding Zn subgroup of G(Γ1 ) is invariant under all transvections.
Example 3.29. Let Γ1 be the graph on the left and let Γ2 be the one on the right.
It is easy to see Out(G(Γ1 )) is transvection free while Out(G(Γ2 )) contains nontrivial transvection (Γ2 has a dead end at vertex u). We claim G(Γ1 ) and G(Γ2 )
are commensurable, and in particular quasi-isometric.
w
z
k
v
u
Let Γ ⊂ Γ1 be the pentagon on the left side and let Y be the Salvetti complex
of Γ. Suppose X1 = Y t Y t (S 1 × [0, 1])/ ∼, here the two boundary circles of the
annulus are identified with two standard circles which are in different copies of Y .
Then π1 (X1 ) = G(Γ1 ). Define homomorphism h1 : G(Γ) → Z/2 by sending w to
the non-trivial element in Z/2 and other generators to the identity element. Let Y 0
be the 2-sheet cover of Y with respect to ker(h1 ).
Define homomorphism h2 : G(Γ1 ) → Z/2 by sending w and k to the non-trivial
element in Z/2 and other generators to the identity element. Let X be the 2-sheet
cover of X1 with respect to ker(h2 ). Then X is made of two copies of Y 0 and two
annuli with the boundaries of the annuli identified with the v-circles in Y 0 (each Y 0
has two v-circles which cover the v-circle in Y ), see the picture below.
X is homotopy equivalent to a Salvetti complex. To see this, let Sw be the circle
in Y 0 which covers the w-circle in Y two times and let Sz ∨ Sv be a wedge of the two
circles in Y 0 which covers the wedge of z-circle and v-circle in Y . There is a copy
of Sw × (Sz ∨ Sv ) inside Y 0 . Let I be a segment in Sw such that its end points are
mapped to the base point of Y under the covering map. We collapse I × (Sz ∨ Sv )
to {pt} × (Sz ∨ Sv ) inside each copy of Y 0 in X, and collapse one of the annuli in X
to a circle by killing the interval factor. Denote the resulting space by X 0 . Then X 0
is homotopy equivalent to X and the un-collapsed annulus in X becomes a torus
in X 0 . It is not hard to see X 0 is a Salvetti complex with defining graph Γ2 .
Any standard geodesic in X(Γ2 ) which comes from vertex u is not Hausdorff
close to a quasi-isometric image of some standard geodesic in X(Γ1 ), since u is not
a stable vertex while every vertex in Γ1 is stable.
22
JINGYIN HUANG
Here is a generalization of the above example. Suppose Γ is a finite simplicial
graph such that there exist vertices v1 , v2 ∈ Γ with d(v1 , v2 ) = 2 so that they
are separated by the intersection of links lk(v1 ) ∩ lk(v2 ). Define a homomorphism
h : G(Γ) → Z/2 by sending v1 and v2 to the non-trivial element in Z/2 and killing
all other generators. Then ker(h) is also a right-angled Artin group by the same
argument as before. To find its defining graph, let {Ci }ni=1 be the components
of Γ \ (lk(v1 ) ∩ lk(v2 )) and suppose v1 ∈ C1 . Define Γ1 = C1 ∪ (lk(v1 ) ∩ lk(v2 ))
and Γ2 = (∪ni=2 Ci ) ∪ (lk(v1 ) ∩ lk(v2 )). Then Γ1 and Γ2 are full subgraphs of Γ,
moreover, St(vi ) ∈ Ci . For i = 1, 2, let Γ0i be the graph obtained by gluing two
copies of Γi along St(vi ) and let Γ03 be the join of one point and lk(v1 ) ∩ lk(v2 ).
Then the defining graph of ker(h) can be obtained by gluing Γ01 , Γ02 and Γ03 along
lk(v1 ) ∩ lk(v2 ).
Note that we are taking advantage of separating closed stars while constructing
the counterexample. If separating closed stars are not allowed in Γ, then we have
a positive answer to question (1) (see Section 5).
In the rest of this subsection, we will prove Theorem 3.28. Γ will be an arbitrary
finite simplicial graph in the rest of this subsection. Theorem 3.28 is actually a
consequence of the following more general result.
Lemma 3.30. Pick a vertex w ∈ Γ and let Γw be the intersection of all stable
subgraphs of Γ that contain w. Define W = {w0 ∈ Γ | w⊥ ⊂ St(w0 )}. Then Γw is
the full subgraph spanned by W .
In other words, G(Γw ) ≤ G(Γ) is the minimal standard subgroup containing w
with the property that G(Γw ) is invariant under any transvection.
Now we show how to deduce Theorem 3.28 from Lemma 3.30
Proof of Theorem 3.28. The only if part can be proved by contradiction (choose a
transvection which does not preserve the subgroup G(Γ1 )). For the converse, let
{vi }ni=1 be the vertex set of Γ1 and let Γvi be the minimal stable subgraph that
contains vi for 1 ≤ i ≤ n. By our assumption and Lemma 3.30, Γvi ⊂ Γ1 . Thus
the full subgraph spanned by ∪ni=1 Γvi is stable by Lemma 3.24, which means Γ1 is
stable.
It remains to prove Lemma 3.30. We first set up several auxiliary lemmas.
Lemma 3.31. Let v ∈ Γ be a vertex which is not isolated. Then at least one of the
following is true:
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
23
(1) v is contained in a stable discrete subgraph with more than one vertex.
(2) v is contained in a stable clique subgraph.
(3) There is a stable discrete subgraph with more than one vertex whose vertex
set is in v ⊥ .
(4) There is a stable clique subgraph whose vertex set is in v ⊥ .
Proof. Since v is not isolated, we can assume Γ is connected by Lemma 3.15. By
Lemma 3.18, we can find a stable subgraph Γ1 = Γ̄1 ◦ Γ̄2 ◦ · · · ◦ Γ̄n where {Γ̄i }ni=1 are
discrete full subgraphs and n = dim(X(Γ)). If v ∈ Γ1 , then by the third paragraph
of the proof of Lemma 3.19, we know either (1), (2) or (4) is true.
Suppose d(v, Γ1 ) = 1. Let Γ2 be the full subgraph spanned by v ⊥ ∩ Γ1 . Then Γ2
is stable by Lemma 3.17. The proof of Lemma 3.19 implies every stable subgraph
of Γ contains either a stable discrete subgraph or a stable clique subgraph (this
does not depend on the v ⊥ * St(w) assumption), thus either (3) or (4) is true.
Suppose d(v, Γ1 ) ≥ 2. Pick vertex u ∈ Γ1 such that d(v, u) = d(v, Γ1 ) = n and
let ω be a geodesic connecting v and u. Suppose {vi }ni=0 are the consecutive vertices
⊥
in ω such that v0 = v and vn = u. Let Γ0 be the full subgraph spanned by vn−1
∩Γ
00
⊥
and let Γ be the full subgraph spanned by V and V where V = VΓ0 (the vertex
set of Γ0 ). Then Γ0 is stable by Lemma 3.17 and Γ00 is stable by Lemma 3.16. Note
that d(v, x) ≥ n for any vertex x ∈ V , so d(v, y) ≥ n − 1 for any vertex y ∈ V ⊥ .
Thus d(v, Γ00 ) ≥ n − 1. However, vn−1 ∈ Γ00 . So d(v, Γ00 ) = n − 1. Now we can
induct on n and reduce to the d(v, Γ1 ) = 1 case.
It is interesting to see that if Γ has large diameter, then there are a lot of nontrivial stable subgraphs.
We record the following lemma which is an easy consequence of Theorem 2.10.
Lemma 3.32. Suppose Γ = Γ1 ◦ Γ2 where Γ1 is the maximal clique join factor of
Γ. If Γ02 is stable in Γ2 , then Γ1 ◦ Γ02 is stable in Γ.
Now we are ready to prove Lemma 3.30.
Proof of Lemma 3.30. By Lemma 3.14, Γw is the minimal stable subgraph that
contains w. If there exists vertex w0 ∈ W such that w0 ∈
/ Γw , then sending w → ww0
and fixing all other vertices would induce a group automorphism, which gives rise to
a quasi-isometry from X(Γ) to X(Γ). The existence of such quasi-isometry would
contradict the stability of Γw , thus W ⊂ Γw .
Let W 0 be the vertex set of Γw . It remains to prove W 0 ⊂ W . Suppose W ( W 0
and let u ∈ W 0 \ W . Then ∅ 6= w⊥ \ St(u). The minimality of Γw implies we can
not use Lemma 3.17 to get rid of u while keep w, then w⊥ \ St(u) ⊂ W 0 \ {u, w}.
In summary,
(3.33)
∅=
6 w⊥ \ St(u) ⊂ W 0 \ {u, w}.
In particular, w is not isolated in Γw and
(3.34)
Γw * St(u).
Now we apply Lemma 3.31 to Γw and w, and recall that if a subgraph is stable
in Γw , then it is stable in Γ. If case (1) in Lemma 3.31 is true, then we will get a
contradiction since w is not isolated in Γw . If case (2) is true, then Γw sits inside
some clique, which is contradictory to (3.34).
If case (3) is true, let Γ1 ⊂ Γw be the corresponding stable discrete subgraph.
Let V1 = VΓ1 and let V10 = {u ∈ Γw | d(u, v) = 1 for any v ∈ V1 }. Suppose Γ0w is
24
JINGYIN HUANG
the full subgraph spanned by V1 and V10 . Then Γ0w is stable by Lemma 3.16, hence
Γ0w = Γw . Let Γw = Γ̄1 ◦ Γ̄2 ◦ · · · ◦ Γ̄k be the join decomposition induced by the De
Rahm decomposition of X(Γw ). Then k ≥ 2 and u does not sit inside the clique
factor by (3.34).
If there is no clique factor, then each join factor is stable by Theorem 2.10 and
w is inside one of the join factors, which contradict the minimality of Γw . If the
clique factor exists and w sits inside the clique factor, then by Theorem 2.10, the
clique factor is stable and we have the same contradiction as before. If the clique
factor exists and w sits outside the clique factor, this reduces to the next case.
If case (4) is true, let Γ2 ⊂ Γw be the corresponding stable clique subgraph.
We can also assume without loss of generality that w is not contained in a stable
clique. Let V2 = VΓ2 and V20 = {u ∈ Γw | d(u, v) = 1 for any v ∈ V2 }. Suppose
Γ00w is the full subgraph spanned by V2 and V20 . Then Γ00w = Γw as before. Let
Γw = Γ01 ◦ Γ02 where Γ01 corresponds to the Euclidean De Rahm factor of X(Γw ).
Note that Γ02 is non-trivial and w, u ∈ Γ02 as in the discussion of case (3). Equation
(3.33) implies that w⊥ * St(u) is still true if we take the orthogonal complement
of w and the closed star of u in Γ02 , in particular, w is not isolated in Γ02 . Moreover,
dim(X(Γ02 )) < dim(X(Γw )) ≤ dim(X(Γ)).
If dim(X(Γ)) = 2, then Γ02 has to be discrete, which is contradictory to that
w is not isolated in Γ02 . If dim(X(Γ)) = n > 2, by induction we can assume the
lemma is true for all lower dimensional graphs. Then there exists Γ̄w stable in Γ02
such that w ∈ Γ̄w and u ∈
/ Γ̄w . By Lemma 3.32, Γ̄w ◦ Γ01 is stable in Γw , hence in
Γ, which contradicts the minimality of Γw .
Remark 3.35. It is nature to ask whether Theorem 3.28 is still true if we do not
require Γ1 to be a clique. There turns out to be counterexamples. Let Γ be the
graph as below and Γ1 ⊂ Γ be the disjoint union of v and w. It is easy to check
there do not exist v1 ∈ Γ1 and v2 ∈ Γ \ Γ1 such that v1⊥ ⊂ St(v2 ). Note that St(u)
separates Γ, then we get a partial conjugation that sends v → v and w → u−1 wu,
which implies Γ1 is not stable.
v
u
w
A more interesting example (but of the same nature) is the following. Let Γ1
be the graph in the left side as below and Γ2 be the graph in the right side. Then
G(Γ1 ) is quasi-isometric to G(Γ2 ) by the discussion in Section 11 of [BKS08a]. Let
q : X(Γ2 ) → X(Γ1 ) be a quasi-isometry and let K be a standard subcomplex
in X(Γ2 ) such that its defining graph ΓK is a pentagon in Γ2 . Suppose q(K)
is Hausdorff close to a standard subcomplex K 0 in X(Γ). Then ΓK 0 must be a
connected proper subgraph of Γ1 , hence is a tree. But this is impossible by the
results in [BN08].
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
25
4. From quasi-isometries to isomorphisms
4.1. The extension complexes.
4.1.1. Extension complexes and standard flats. Let q : X(Γ) → X(Γ0 ) be a quasiisometry. Usually q does not induce a well-defined boundary map, see [CK00].
However, Theorem 3.27 implies we still have control on a subset of the Tits boundaries when Out(G(Γ)) and Out(G(Γ0 )) are transvection free. In this subsection, we
will reorganize this piece of information in terms of extension complexes.
Recall that we identify the vertex set of Γ with a standard generating set S of
G(Γ). And we also label the standard circles in the Salvetti complex by elements
in S. By choosing an orientation for each standard circle, we obtain a directed
labeling of edges in X(Γ).
Denote the extension complex of Γ by P(Γ). We give an alternative definition
of P(Γ) here, which is natural for our purposes. The vertices of P(Γ) are in 1-1
correspondence with the parallel classes of standard geodesics in X(Γ) (two standard geodesics are in the same parallel class if they are parallel). Two distinct
vertices v1 , v2 ∈ P(Γ) are connected by an edge if and only if we can find standard
geodesic li in the parallel class associated with vi (i = 1, 2) such that l1 and l2 span
a standard 2-flat. The next observation follows from Lemma 3.1 and Lemma 2.4:
Observation 4.1. If v1 6= v2 , then v1 and v2 are joined by an edge if and only if
there exist li0 in the parallel class associated with vi (i = 1, 2) and R > 0 such that
l10 ⊂ NR (Pl20 ).
P(Γ) is defined to be the flag complex of its 1-skeleton.
Lemma 4.2. P(Γ) is isomorphic to the extension complex of Γ.
Proof. It suffices to show the 1-skeleton of P(Γ) is isomorphic to the extension
graph Γe . Pick vertex v ∈ P(Γ) and let l be a standard geodesic in the parallel
class associated with v. We identify l with R in an orientation-preserving way (the
orientation in l is induced by the directed labeling). Recall that G(Γ) y X(Γ)
by deck transformations. Let αv ∈ G(Γ) be the element such that αv (l) = l and
αv (x) = x + 1 for any x ∈ l. It is easy to see αv is conjugate to an element in
S, thus αv gives rise a vertex αv ∈ Γe by Definition 2.12. Note that αv does not
depend the choice of l in the parallel class, so we have a well-defined map from the
vertex set of P(Γ) to the vertex set of Γe . Moreover, if v1 and v2 are adjacent, then
αv1 and αv2 commute.
Now we define an inverse map. Pick α = gsg −1 ∈ Γe (s ∈ S). Then all standard
geodesics which are stabilized by α are in the same parallel class. Let vα be the
vertex in P(Γ) associated with this parallel class. We map the vertex α of Γe to
the vertex vα . Now we show this map extends to the 1-skeleton. For i = 1, 2, let
αi = gi si gi−1 ∈ Γe . By the centralizer theorem of [SDS89], α1 and α2 commute if
26
JINGYIN HUANG
and only if [s1 , s2 ] = 1 and there exists g ∈ G(Γ) such that αi = gsi g −1 . Thus vα1
and vα2 are adjacent in P(Γ).
Since every edge in the standard geodesics of the same parallel class has the
same label, the labeling of the edges of X(Γ) induces a labeling of the vertices of
P(Γ). Moreover, since G(Γ) y X(Γ) by label-preserving cubical isomorphisms, we
obtain an induced action G(Γ) y P(Γ) by label-preserving simplicial isomorphisms.
Moreover, the unique label-preserving map from the vertices of P(Γ) to the vertices
of F (Γ) extends to a simplicial map
(4.3)
π : P(Γ) → F (Γ).
Pick arbitrary vertex p ∈ X(Γ), one can obtain a simplicial embedding ip from
the flag complex F (Γ) of Γ to P(Γ) by considering the collection of standard
geodesics passing through p. We will denote the image of ip by (F (Γ))p . Note
that for each vertex p ∈ X(Γ), π ◦ ip : F (Γ) → F (Γ) is the identity map.
Pick (k − 1)-simplex in P(Γ) with vertex set {vi }ki=1 and pick standard geodesic
li in the parallel class associated with vi for 1 ≤ i ≤ k. Since Pli ∩ Plj 6= ∅ for
1 ≤ i 6= j ≤ k, by Lemma 2.2, ∩ki=1 Pli 6= ∅. By Corollary 3.2 and Lemma 3.4, there
exist standard geodesics {li0 }ki=1 satisfying
(1) li0 is parallel to li for each i.
(2) The convex hull of {li0 }ki=1 is a standard k-flat, denoted by Fk .
(3) ∩ki=1 Pli = PFk .
Thus we have a 1-1 correspondence between the (k − 1)-simplexes of P(Γ) and
parallel classes of standard k-flats in X(Γ). In particular, maximal simplexes in
P(Γ), namely those simplexes which are not properly contained in some larger
simplexes of P(Γ), are in 1-1 correspondence with maximal standard flats in X(Γ).
For standard flat F ⊂ X(Γ), we denote the simplex in P(Γ) associated with the
parallel class containing F by ∆(F ).
Observation 4.4. Let ∆1 , ∆2 be two simplexes in P(Γ) such that ∆ = ∆1 ∩ ∆2 6=
∅. For i = 1, 2, let Fi ⊂ X(Γ) be a standard flat such that ∆(Fi ) = ∆i . Set
(F10 , F20 ) = I(F1 , F2 ). Then ∆(F10 ) = ∆(F20 ) = ∆.
We define the reduced Tits boundary, denoted ∂¯T (X(Γ)), to be the subset of
∂T (X(Γ)) which is the union of Tits boundaries of standard flats in X(Γ). For
standard flat F ⊂ X(Γ), we triangulate ∂T F into all-right spherical simplexes which
are the Tits boundaries of orthant subcomplexes in F . Pick another standard flat
F 0 ⊂ X(Γ), then ∂T F ∩∂T F 0 is a subcomplex in both ∂T F and ∂T F 0 by Lemma 3.1
and Remark 2.6. Thus we can endow ∂¯T (X(Γ)) with the structure of an all-right
spherical complex.
Now we look at the relation between ∂¯T (X(Γ)) and P(Γ). For each standard
flat F ⊂ X(Γ), we can associate ∂T F with ∆(F ) ⊂ P(Γ). This induces a surjective
simplicial map s : ∂¯T (X(Γ)) → P(Γ) (s can be defined by induction on dimension).
Note that the inverse image of each simplex in P(Γ) under s is a sphere in ∂¯T (X(Γ)).
Then one can construct ∂¯T (X(Γ)) from P(Γ) as follows. We start with a collection
of S0 ’s which are in 1-1 correspondence to vertices of P(Γ) and form a join of n
copies of S0 ’s if and only if the corresponding n vertices in P(Γ) span a (n − 1)simplex. In other words, ∂¯T (X(Γ)) is obtained by applying the spherical complex
construction in the sense of [BRS07, Definition 2.1.22] to P(Γ).
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
27
Let K1 ⊂ X(Γ) be a standard subcomplex. We define ∂¯T (K1 ) to be the union
of Tits boundaries of standard flats in K1 . Note that ∂¯T (K1 ) = ∂¯T (X(Γ)) ∩ ∂T K1 .
∂¯T (K1 ) descends to a subcomplex in P(Γ), which will be denoted by ∆(K1 ).
Lemma 4.5. Let K1 and K2 be two standard subcomplexes of X(Γ). Put (K10 , K20 ) =
I(K1 , K2 ). Then ∆(K10 ) = ∆(K20 ) = ∆(K1 ) ∩ ∆(K2 ).
Proof. By Remark 2.6, we have ∂T K10 = ∂T K20 = ∂T K1 ∩ ∂T K2 , hence ∂¯T K10 =
∂¯T K20 = ∂¯T K1 ∩ ∂¯T K2 and ∆(K10 ) = ∆(K20 ) = ∆(K1 ) ∩ ∆(K2 ).
Now we study how the extension complexes behave under quasi-isometries.
Lemma 4.6. Pick Γ1 and Γ2 such that Out(G(Γi )) is transvection free for i = 1, 2.
Then any quasi-isometry q : X(Γ1 ) → X(Γ2 ) induces a simplicial isomorphism
q∗ : P(Γ1 ) → P(Γ2 ). If only Out(G(Γ1 )) is assumed to be transvection free, we still
have a simplicial embedding q∗ : P(Γ1 ) → P(Γ2 ).
Proof. We only proof the case when both Out(G(Γ1 )) and Out(G(Γ2 )) are transvection free. The other case is similar. By Theorem 3.27, every vertex in Γ1 is stable,
thus q sends any parallel class of standard geodesics in X(Γ1 ) to another parallel
class of standard geodesics in X(Γ2 ) up to finite Hausdorff distance. This induces
a well-defined map q∗ from the 0-skeleton of P(Γ1 ) to the 0-skeleton of P(Γ2 ). The
map q∗ is a bijection by considering the quasi-isometry inverse. Moreover, Observation 4.1 implies two vertices in P(Γ1 ) are adjacent if and only if their images
under q∗ are adjacent. So we can extend q∗ to be a graph isomorphism between
the 1-skeleton of P(Γ1 ) and the 1-skeleton of P(Γ2 ). Since both P(Γ1 ) and P(Γ2 )
are flag complexes, we can extend the isomorphism to the whole complex.
4.1.2. Extension complexes and their relatives. Now we discuss the relation between
P(Γ) with several other objects in the literature. The material in this subsection
will not be used later.
We can endow F (Γ) with the structure of complex of groups, which gives us an
alternative definition of P(Γ). More specifically, P(Γ) = F (Γ) × G(Γ)/ ∼, here
St(v) × g1 and St(v) × g2 (v ∈ F (Γ) is a vertex) are identified if and only if there
exists an integer m such that g1−1 g2 = v m (we also view v as one of the generators
of G(Γ)). Hence for k-simplex ∆k ⊂ F (Γ) with vertex set {vi }ki=1 , St(∆k ) × g1
and St(∆k ) × g2 are identified if and only if g1−1 g2 belongs to the Zk subgroup of
G(Γ) generated by {vi }ki=1 . One can compare this with a similar construction for
Coxeter group in [Dav83].
There is another important object associated with a right-angled Artin group,
called the modified Deligne complex in [CD95] and the flat space in [BKS08a].
Definition 4.7. Let P(Γ) be poset of left cosets of standard abelian subgroups
of G(Γ) (include the trivial subgroup) such that the partial order is induced by
inclusion of sets. Then the modified Deligne complex is defined to be the geometric
realization of the derived poset of P(Γ).
Recall that elements in the derived poset of a poset P are totally ordered finite
chains in P. It can be viewed as an abstract simplex.
The extension complex P(Γ) can be viewed as a coarse version of the modified
Deligne complex. Let A, B be two subsets of a metric space. We say A and B are
∞
coarsely equivalent if A = B, and A are coarsely contained in B if A ⊂∞ B. Let
0
P (Γ) be the poset whose elements are coarsely equivalent classes of left cosets of
28
JINGYIN HUANG
non-trivial standard abelian subgroups of G(Γ), and the partial order is induced by
coarse inclusion of sets.
Observation 4.8. The poset P0 (Γ) is an abstract simplicial complex and it is isomorphic to P(Γ).
Roughly speaking, P(Γ) captures the combinatorial pattern of how standard flats
in X(Γ) intersect with each other, and P(Γ) is about how they coarsely intersect
with each other, thus P(Γ) contains more information than P(Γ). However, in
certain cases, it is possible to recover information about P(Γ) from P(Γ), and this
enable us to prove quasi-isometric classification/rigidity results for RAAG’s.
We can define the poset P0 (Γ) for arbitrary Artin group by considering the collection of coarse equivalent classes of spherical subgroups in an Artin group under
coarse inclusion. Then the geometric realization of the derived poset of P0 (Γ) would
be a natural candidate for the extension complex of an Artin group. It is interesting
to ask how much of the results in [KK14] can be generalized to this context.
There is also a link between P(Γ) and the structure of hyperplanes in X(Γ).
Recall that for every CAT (0) cube complex X, the crossing graph of X, denoted
by C(X), is a graph whose vertices are in 1-1 correspondence to the hyperplanes
in X, and two vertices are adjacent if and only if the corresponding hyperplanes
intersect. The contact graph, introduced in [Hag14] and denoted by C(X), has the
same vertex set as C(X), and two vertices are adjacent if and only if the carriers
of the corresponding hyperplanes intersect.
There is a natural surjective simplicial map p : C(X(Γ)) → Γe defined as follows.
Pick a vertex v ∈ C(X(Γ)) and let h be the corresponding hyperplane. Since all
standard geodesics which intersect h at one point are in the same parallel class, we
define p(v) to be the vertex in Γe associated with this parallel class (see Lemma
4.2). It is clear that if v1 , v2 ∈ C(X(Γ)) are adjacent vertices, then p(v1 ) and p(v2 )
are adjacent, so p extends to a simplicial map. Pick vertex w ∈ Γe , then p−1 (e) is
the collection of hyperplanes dual to a standard geodesic.
Theorem 4.9 ([KK13b, Hag14, HK14]). If Γ is connected, then C(X(Γ)), C(X(Γ))
and P(Γ) are quasi-isometric to each other, moreover, they are quasi-isometric to
a tree.
From this viewpoint, P(Γ) captures both the geometric information of X(Γ) (the
standard flats) and the combinatorial information (the hyperplanes).
4.2. Reconstruction of quasi-isometries. We show the boundary map q∗ :
P(Γ) → P(Γ0 ) in Lemma 4.6 induces a well defined map from G(Γ) to G(Γ0 ).
Lemma 4.10. Let F1 and F2 be two maximal standard flats in X(Γ) and let ∆1 ,
∆2 be the corresponding maximal simplexes in P(Γ). If F1 and F2 are separated
by a hyperplane h, then there exist vertices vi ∈ ∆i for i = 1, 2 and v ∈ P(Γ) such
that v1 and v2 are in different connected component of P(Γ) \ St(v).
Proof. Let e be an edge dual to h and let l be the standard geodesic that contains e.
Set v = ∆(l) ∈ P(Γ). By Lemma 3.4, the parallel set Pl of l is isometric to h × E1 .
Thus every standard geodesic parallel to l must have non-trivial intersection with
h. Since F1 ∩ h = ∅, F1 can not contain any standard geodesic parallel to l, which
means v ∈
/ ∆1 . Moreover, ∆1 * St(v) since ∆1 is a maximal simplex. Similarly,
∆2 * St(v), thus we can find vertices vi ∈ ∆i \ St(v) for i = 1, 2. We claim v1 , v2
and v are the vertices we are looking for.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
29
If there is a path ω ⊂ P(Γ) \ St(v) connecting v1 and v2 , we can assume ω is
consist of a sequence of edges {ei }ki=1 with v1 ∈ e1 and v2 ∈ ek . For each ei , pick a
maximal simplex ∆0i that contains ei and let Fi0 be the maximal standard flat such
that ∆(Fi0 ) = ∆0i . Then v ∈
/ ∆0i for 1 ≤ i ≤ k, hence Fi0 ∩ h = ∅.
0
0
0
Set ∆0 = ∆1 , ∆k+1 = ∆2 , F00 = F1 and Fk+1
= F2 . Since ∆0i ∩ ∆0i+1 contains a
vertex in ω, we have
(4.11)
(∆0i ∩ ∆0i+1 ) \ St(v) 6= ∅
0
for 0 ≤ i ≤ k. Since F00 and Fk+1
are in different sides of h, there exists i0 such that
0
0
h separates Fi0 and Fi0 +1 . Let (Fi000 , Fi000 +1 ) = I(Fi00 , Fi00 +1 ). By Observation 4.4,
∆(Fi000 ) = ∆(Fi000 +1 ) = ∆0i ∩ ∆0i+1 . However, by Lemma 2.8, there exists a convex
subset of h parallel to Fi000 , thus Fi000 ⊂∞ h ⊂ Pl . It follows from Observation 4.1
that ∆0i ∩ ∆0i+1 ⊂ St(v), which contradicts (4.11).
Denote the Cayley graph of G(Γ) with respect to the standard generating set S
by C(Γ). We pick an identification between C(Γ) and the 1-skeleton of X(Γ). Thus
G(Γ) is identified with the vertex set of X(Γ).
Lemma 4.12. Let Γ1 be a finite simplicial such that
(1) There is no separating closed star in F (Γ1 ).
(2) F (Γ1 ) is not contained in a union of two closed stars.
Then any simplicial isomorphism s : P(Γ1 ) → P(Γ2 ) induces a unique map s0 :
G(Γ1 ) → G(Γ2 ) such that for any maximal standard flat F1 ⊂ X(Γ1 ), vertices in
F1 are mapped by s0 to vertices lying in a maximal standard flat F2 ⊂ X(Γ2 ) with
∆(F2 ) = s0 (∆(F1 )).
Proof. Pick vertex p ∈ G(Γ1 ). Let {Fi }ki=1 be the collection of maximal standard
flats containing p. For 1 ≤ i ≤ k, define ∆i = ∆(Fi ) and ∆0i = s(∆i ). Let
Fi0 ⊂ X(Γ2 ) be the maximal standard flat such that ∆(Fi0 ) = ∆0i . Let Kp =
(F (Γ1 ))p = ∪ki=1 ∆i (recall that Kp ∼
= F (Γ1 )). We claim
(4.13)
∩ki=1 Fi0 6= ∅.
The lemma will then follows from (4.13). To see this, we deduce from condition (2)
that ∩ki=1 ∆i = ∅. Hence ∩ki=1 Fi = {p}. It follows that ∩ki=1 ∆0i = ∅. This together
with (4.13) imply that ∩ki=1 Fi0 is exactly one point. We define s0 by sending p to
this point. One readily verifies that s0 has the required properties.
It remains to prove (4.13).
Suppose (4.13) is not true. Then by Lemma 2.2, there exist i1 and i2 such that
Fi01 ∩ Fi02 = ∅. Thus Fi01 and Fi02 are separated by a hyperplane. It follows from
Lemma 4.10 that there exist vertices v 0 ∈ P(Γ2 ), v10 ∈ ∆0i1 and v20 ∈ ∆0i2 such that
v10 and v20 are in different connected components of P(Γ2 ) \ St(v 0 ). Let v = s−1 (v 0 ),
v1 = s−1 (v10 ) and v2 = s−1 (v20 ). Then Kp \ (Kp ∩ St(v)) is disconnected (since
v1 , v2 ∈ Kp and they are separated by St(v)).
If v ∈ Kp , then Kp would contain a separating closed star, which yields a contradiction, thus (4.13) is true in this case.
Suppose v ∈
/ Kp . Pick a standard geodesic l such that ∆(l) = v and let {hi }ni=1
be the collection of hyperplanes in X(Γ) such that each hi separates p from the
parallel set Pl of l (note that p ∈
/ Pl ). For 1 ≤ i ≤ n, pick an edge ei dual to hi and
let wi be the unique vertex in Kp that has the same label as ei . Let w0 ∈ Kp be
30
JINGYIN HUANG
the unique vertex which has the same label as v. We claim
(4.14)
St(v) ∩ Kp = ∩ni=0 (St(wi ) ∩ Kp ).
For every u ∈ Kp , let lu be the unique standard geodesic such that ∆(lu ) = u
and p ∈ lu .
Pick u ∈ St(v) ∩ Kp . Observation 4.1 implies I(lu , Pl ) = (lu , lu0 ), where lu0
is some standard geodesic in Pl . Then for 1 ≤ i ≤ n, hi separates lu from Pl ,
otherwise hi ∩ lu 6= ∅ and Lemma 2.7 implies hi ∩ Pl 6= ∅, which is a contradiction.
It follows from Corollary 3.2 that u and wi are adjacent for 0 ≤ i ≤ n, thus
u ∈ ∩ni=0 (St(wi ) ∩ Kp ). Therefore St(v) ∩ Kp ⊂ ∩ni=0 (St(wi ) ∩ Kp ).
Pick u ∈ ∩ni=0 (St(wi ) ∩ Kp ). First we show lu ∩ Pl = ∅. Suppose there is a
vertex z in lu ∩ Pl . Since v and w0 have the same label and u ∈ St(w0 ), it follows
that the edge in lu which contains z belongs to the parallel set Pl . Then lu ⊂ Pl ,
contradicting the fact that p ∈
/ Pl . Therefore lu ∩ Pl = ∅.
Now we pick an edge path ω of shortest combinatorial length that travels from
lu to Pl . Let {fj }m
j=1 be the consecutive edges in ω such that f1 ∩ lu 6= ∅. For each
fj , let h̄j be the hyperplane dual to fj . Then h̄j separates lu from Pl (otherwise
ω would not be the shortest edge path), hence separates p from Pl . This and
u ∈ ∩ni=0 (St(wi ) ∩ Kp ) imply that d(π(u), Vfj ) ≤ 1 for each j, where π is the map
in (4.3) and Vfj is the label of the edge fj . It follows that ω is contained in the
parallel set Plu , and hence the intersection Plu ∩ Pl contains some vertex z. Again,
since u ∈ St(w0 ), and w0 has the same label as v, we find that the standard geodesic
lu0 ⊂ Plu that is parallel to lu and passes through z, is contained in Pl . Therefore
u ∈ St(v) ∩ Kp . Then (4.14) follows.
By condition (2) of Lemma 4.12, we have
(4.15)
(St(w0 ) ∩ Kp ) ∪ (∩ni=1 (St(wi ) ∩ Kp )) ( Kp .
Let A = Kp \ (St(w0 ) ∩ Kp ) and let B = Kp \ (∩ni=1 (St(wi ) ∩ Kp )). Then (4.15)
implies A ∩ B 6= ∅. Thus we have the following Mayer-Vietoris sequence for reduced
homology.
· · · → H̃0 (A ∩ B) → H̃0 (A) ⊕ H̃0 (B) → H̃0 (A ∪ B) → 0.
Recall that Kp \(Kp ∩St(v)) is disconnected, we deduce that H̃0 (A∪B) is nontrivial
from (4.14). Thus H̃0 (A) ⊕ H̃0 (B) is nontrivial, which implies either ∩ni=1 (St(wi ) ∩
Kp ) or St(w0 ) ∩ Kp would separate Kp . Thus we can induct on n to deduce that
there exists i0 such that St(wi0 ) ∩ Kp separates Kp . This yields a contradictory to
condition (1) of Lemma 4.12.
There are counterexamples if we only assume (1) in Lemma 4.12. For example,
let Γ1 and Γ2 be discrete graphs made of two points. Then P(Γ1 ) and P(Γ2 ) are
discrete sets. Now it is not hard to construct a permutation of a discrete set to
itself which does not satisfy the conclusion of Lemma 4.12. If we go back to the
proof of Lemma 4.12, then the step using Mayer-Vietoris sequence will fail, since
we need A ∩ B 6= ∅ in order to use the reduced version of Mayer-Vietoris sequence.
Corollary 4.16. Suppose G(Γ1 ) and G(Γ2 ) both satisfy the assumption of Lemma
4.12. Then they are isomorphic if and only if P(Γ1 ) and P(Γ2 ) are isomorphic as
simplicial complexes.
Proof. The only if direction follows from the fact that G(Γ1 ) and G(Γ2 ) are isomorphic if and only if Γ1 and Γ2 are isomorphic, see [Dro87]. It remains to prove the
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
31
if direction. Pick an isomorphism s : P(Γ1 ) → P(Γ2 ), and let s0 : G(Γ1 ) → G(Γ2 )
be the map in Lemma 4.12. Pick vertex p ∈ G(Γ1 ) and let q = s(p). We define
(F (Γ1 ))p ⊂ P(Γ1 ) and (F (Γ2 ))q ⊂ P(Γ2 ) as in the first paragraph of the proof of
Lemma 4.12. Then (4.13) implies s((F (Γ1 ))p ) ⊂ (F (Γ2 ))q . This induces a graph
embedding Γ1 ,→ Γ2 . By repeating the previous discussion for s−1 , we obtain another graph embedding Γ2 ,→ Γ1 . Since both Γ1 and Γ2 are finite simplicial graphs,
they are isomorphic. Hence G(Γ1 ) ∼
= G(Γ2 ).
Lemma 4.17. Let G(Γ) be a RAAG such that Out(G(Γ)) is finite and G(Γ) Z.
Then F (Γ) satisfies the assumption of Lemma 4.12.
Proof. It is clear that F (Γ) should satisfy condition (1) of Lemma 4.12 since no
nontrivial partial conjugation is allowed. If F (Γ) is contained in a closed star, then
Γ is a point. So if (2) is not true, then F (Γ) = St(v) ∪ St(w) for distinct vertices
v, w ∈ Γ. Since the orthogonal complement v ⊥ satisfies v ⊥ * St(w), there exists
u ∈ v ⊥ such that d(u, w) ≥ 2. Pick any edge e such that u ∈ e, then e * St(w),
so e ⊂ St(v). This implies u⊥ ⊂ St(v), hence Out(G(Γ)) is infinite, which yields a
contradiction.
By Lemma 4.6, Lemma 4.17 and Corollary 4.16, we have following result, which
in particular establishes Theorem 1.1 of the introduction.
Theorem 4.18. Let Γ1 and Γ2 be two finite simplicial graphs such that Out(G(Γi ))
is finite for i = 1, 2. Then G(Γ1 ) and G(Γ2 ) are quasi-isometric if and only if they
are isomorphic. Moreover, for any (L, A)-quasi-isometry q : X(Γ1 ) → X(Γ2 ), there
exist a bijection q 0 : G(Γ1 ) → G(Γ2 ) and a constant D = D(L, A, Γ1 ) such that
(1) d(q(v), q 0 (v)) < D for any v ∈ G(Γ1 ).
(2) For any standard flat F1 ⊂ X(Γ1 ), there exists a standard flat F2 ⊂ X(Γ2 )
such that q 0 induces a bijection between F1 ∩ G(Γ1 ) and F2 ∩ G(Γ2 ).
If G(Γ1 ) 6= Z, then such q 0 is unique.
Proof. It suffices to look at the case G(Γ1 ) 6= Z. Then G(Γ2 ) 6= Z. In this case,
every vertex v in Γ1 or Γ2 is the intersection of maximal cliques that contain v
(otherwise there exist vertex w such that w 6= v and v ⊥ ⊂ St(w)). It follows
that every standard geodesic in X(Γ1 ) or X(Γ2 ) is the intersection of finitely many
maximal standard flats, so is every standard flat. Let q∗ : P(Γ1 ) → P(Γ2 ) be the
map in Lemma 4.6. We apply Lemma 4.12 to q∗ and q∗−1 to obtain q 0 with the
required properties. Note that each vertex of X(Γ) is the intersection of maximal
standard flats that contain it, thus q 0 is unique.
4.3. The automorphism groups of extension complexes. Suppose Out(G(Γ))
is finite, by Theorem 4.18, each element in the simplicial automorphism group
Aut(P(Γ)) of P(Γ) induces a bijection G(Γ) → G(Γ). However, this bijection does
not extend to an isomorphism from X(Γ) to itself in general. We start by looking
at the following example which was first pointed out in [BKS08a, Section 11] in a
slightly different form.
Example 4.19. Let l ⊂ X(Γ) be a standard geodesic let πl : X(Γ) → l be the
CAT(0) projection. We identify the vertex set of l with Z. Let X (0) (Γ) be the
vertex set of X(Γ). Then the above projection induces a map πl : X (0) (Γ) → Z.
32
JINGYIN HUANG
Recall that each edge of X(Γ) is oriented and labeled, and G(Γ) acts on X(Γ) by
transformations that preserve labels and orientations. There is a unique element
α ∈ G(Γ) such that α translates l one unit towards the positive direction.
We want to define a bijection q : X (0) (Γ) → X (0) (Γ) which basically flips πl−1 (0)
and πl−1 (1). More precisely:
if πl (x) 6= 0, 1
x
q(x)= α(x)
if x ∈ πl−1 (0)
−1
α (x) if x ∈ πl−1 (1)
One can check the following:
(1) q is a quasi-isometry.
(2) q does not respect the word metric.
(3) q maps vertices in a standard flat to vertices in another standard flat. Thus
q induces an element in Aut(P(Γ)).
The above example implies that in general, elements in Aut(P(Γ)) do not respect
the order along the standard geodesics of X(Γ). There is another metric on G(Γ)
which “forgets about” the ordering. Following [KK14], we define the syllable length
of a word ω to be minimal l such that ω can be written as a product of l elements
of form viki , where vi is a standard generator and ki is an integer.
An alternative definition is the following. Let {hi }ki=1 be the collection of hyperplanes separating ω ∈ G(Γ) and the identity element (recall that we have identified
G(Γ) with the 0-skeleton of X(Γ)). For each i, pick a standard geodesic li dual to
hi . Then the syllable length of ω is the number of elements in {∆(li )}ki=1 . The
syllable length induces a left invariant metric on G(Γ), which will be denoted by
dr . Note that the map in Example 4.19 is an isometry with respect to dr .
Denote the word metric on G(Γ) with respect to the standard generators by dw .
Corollary 4.20. Let Γ be a graph such that Out(G(Γ)) is finite and denote the
simplicial automorphism group of P(Γ) by Aut(P(Γ)). Then
Aut(P(Γ)) ∼
= Isom(G(Γ), dr )
Proof. We have a group homomorphism h1 : Aut(P(Γ)) → P erm(G(Γ)) by Lemma
4.12, here P erm(G(Γ)) is the permutation group of elements in G(Γ). Take φ ∈
Aut(P(Γ)), by Lemma 4.17, ϕ = h1 (φ) and ϕ−1 = h1 (φ−1 ) satisfy the conclusion
of Lemma 4.12. Since every standard geodesic is the intersection of finitely many
maximal standard flats, points in a standard geodesic are mapped to points in a
standard geodesic by φ, which implies dr (ϕ(v1 ), ϕ(v2 )) ≤ dr (v1 , v2 ) if dr (v1 , v2 ) ≤ 1.
By triangle inequality, we have dr (ϕ(v1 ), ϕ(v2 )) ≤ dr (v1 , v2 ) for any v1 , v2 ∈ G(Γ).
Similarly, dr (ϕ−1 (v1 ), ϕ−1 (v2 )) ≤ dr (v1 , v2 ). Thus ϕ ∈ Isom(G(Γ), dr ) and we have
a homomorphism h1 : Aut(P(Γ)) → Isom(G(Γ), dr ).
Now pick ϕ ∈ Isom(G(Γ), dr ). Let v1 , v2 , v3 ∈ G(Γ) such that dr (v1 , vi ) = 1 for
i = 2, 3. We claim
(4.21)
∠v1 (v2 , v3 ) = π/2 ⇔ ∠ϕ(v1 ) (ϕ(v2 ), ϕ(v3 )) = π/2.
If ∠v1 (v2 , v3 ) = π/2, then we can find v4 ∈ G(Γ) such that {vi }4i=1 are the vertices
of a flat rectangle in X(Γ). Note that dr (v1 , v4 ) = dr (v2 , v3 ) = 2 and dr (v4 , v2 ) =
dr (v4 , v3 ) = 1, so dr (ϕ(v1 ), ϕ(v4 )) = dr (ϕ(v2 ), ϕ(v3 )) = 2 and dr (ϕ(v4 ), ϕ(v2 )) =
dr (ϕ(v4 ), ϕ(v3 )) = 1. Now we consider the 4-gon formed by ϕ(v1 )ϕ(v2 ), ϕ(v2 )ϕ(v4 ),
ϕ(v4 )ϕ(v3 ) and ϕ(v3 )ϕ(v1 ). Then the angles at the four vertices of this 4-gon are
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
33
bigger or equal to π/2. It follows from CAT (0) geometry that the angles are exactly
π/2 and the 4-gon actually bounds a flat rectangle. Thus one direction of (4.21) is
proved, the other direction is similar.
We need another observation as follows. If three points v1 , v2 , v3 ∈ G(Γ) satisfies
dr (vi , vj ) = 1 for 1 ≤ i 6= j ≤ 3, then the angle at each vertex of the triangle
∆(v1 , v2 , v3 ) could only be 0 or π, thus {vi }3i=1 are inside a standard geodesic. It
follows from this observation that points in a standard geodesic are mapped by ϕ
to points in a standard geodesic.
We define φ : P(Γ) → P(Γ) as follows. For vertex w ∈ P(Γ), let l be a standard
geodesic such that ∆(l) = w. Suppose l0 ⊂ X(Γ) is the standard geodesic such
that φ(v(l)) ⊂ l0 (v(l) denotes the vertex set of l). Suppose w0 = ∆(l0 ). We
define w0 = φ(w). (4.21) implies w0 does not depend on the choice of l, and
φ(w1 ) and φ(w2 ) are adjacent if vertices w1 , w2 ∈ P(Γ) are adjacent. Thus φ is
a well-defined simplicial map. Note that ϕ−1 also induces a simplicial map from
P(Γ) to itself in a similar way, so φ ∈ Aut(P(Γ)). We define φ = h2 (ϕ). One
readily verify that h2 : Isom(G(Γ), dr ) → Aut(P(Γ)) is a group homomorphism
and h2 ◦ h1 = h1 ◦ h2 = Id. Thus the corollary follows.
Remark 4.22. If we drop the assumption in the above corollary about Γ, then
there is still a monomorphism h : Isom(G(Γ), dr ) → Aut(P(Γ)), moreover, any
ϕ ∈ Isom(G(Γ), dr ) maps vertices in a standard flat to vertices in a standard flat
of the same dimension. The homomorphism h is surjective if Out(G(Γ)) is finite.
Remark 4.23. For any finite simplicial graphs Γ1 and Γ2 , G(Γ1 ) ∼
= G(Γ2 ) if
and only if (G(Γ1 ), dr ) and (G(Γ2 ), dr ) are isometric as metric spaces. The if only
direction follows from [Dro87, Lau95]. For the other direction, let ϕ : (G(Γ1 ), dr ) →
(G(Γ2 ), dr ) be an isometry. Pick v ∈ G(Γ1 ) and let {li }ki=1 be the collection of
standard geodesics passing through v. Pick vi ∈ G(Γ1 ) such that vi ∈ li \ {v}.
Then dr (v, vi ) = 1 for 1 ≤ i ≤ k and dr (vi , vj ) = 2 for 1 ≤ i 6= j ≤ k. So
dr (ϕ(v), ϕ(vi )) = 1 for 1 ≤ i ≤ k and dr (ϕ(vi ), ϕ(vj )) = 2 for 1 ≤ i 6= j ≤ k,
and ∠v (vi , vj ) = π/2 if and only if ∠ϕ(v) (ϕ(vi ), ϕ(vj )) = π/2 by (4.21). This
induces a graph embedding Γ1 → Γ2 . By considering ϕ−1 , we obtain another
graph embedding Γ2 → Γ1 . Hence Γ1 and Γ2 are isomorphic.
Corollary 4.24. If Out(G(Γ)) is finite, then we have the following commutative
diagram, where i1 , i2 and i3 are injective homomorphisms:
Isom(G(Γ), dw )
i1
QI(G(Γ))
i2
Isom(G(Γ), dr )
i3
QI(G(Γ)) is the quasi-isometry group of G(Γ).
Proof. The homomorphism i1 and i3 are obvious and i2 is given by Lemma 4.6
and Corollary 4.20. It is clear that i2 is a group homomorphism and i3 = i2 ◦ i1 .
Note that i3 is injective, so i1 is injective. Pick α ∈ QI(G(Γ)), by Corollary 4.20,
we know i2 (α) = Id implies the image of every standard flat under α is uniformly
Hausdorff close to itself, thus α is of bounded distance from the identity map.
34
JINGYIN HUANG
5. Quasi-isometries and special subgroups
Let G(Γ) be a RAAG with finite outer automorphism group. In this section we
characterize all other RAAG’s quasi-isometric to G(Γ).
5.1. Preservation of extension complex.
Lemma 5.1. Let Γ be a finite simplicial graph. Pick a vertex w ∈ Γ and let Γw
be the minimal stable subgraph containing w. Denote Γ1 = lk(w) and Γ2 = lk(Γ1 )
(see Section 2.1 for definition of links). Then either of the following is true:
(1) Γw is a clique. In this case St(w) is a stable subgraph.
(2) Both Γ1 and Γ1 ◦Γ2 are stable subgraphs of Γ. Moreover, Γ2 is disconnected.
Recall that we use (Γ0 )⊥ to denote the orthogonal complement of the subgraph
Γ ⊂ Γ (see Section 2.1) and we assume (∅)⊥ = Γ.
0
Proof. If Γw ⊂ St(w), then Γw is a clique by Lemma 3.30. We also deduce from
⊥
⊥
since
Lemma 3.30 that each vertex of St(w) \ Γw is in Γ⊥
w . Moreover, Γw ⊂ w
w ∈ Γw . Thus St(w) is the full subgraph spanned by vertices in Γw and Γ⊥
w . So
St(w) is stable by Lemma 3.16.
If Γw * St(w), let Γ11 be the full subgraph spanned by vertices in Γw ∩ lk(w)
and let Γ02 be the full subgraph spanned by vertices in Γw \ Γ11 . By Lemma 3.30,
Γw = Γ11 ◦ Γ02 and Γ02 = Γ2 . Note that Γ2 is disconnected with isolated point
w ∈ Γ2 , and Γ11 may be empty.
Let Vw = v(Γw ) be the vertex set of Γw and let Γ12 be the full subgraph spanned
by Vw⊥ . Then Γw ◦ Γ12 = Γ11 ◦ Γ2 ◦ Γ12 is stable by Lemma 3.16. Pick vertex
v ∈ Γ1 \ Γ11 , then v ∈ w⊥ ⊂ St(u) for any vertex u ∈ Γw by Lemma 3.30, thus
v ∈ Γ12 and Γ1 ⊂ Γ11 ◦ Γ12 . On the other hand, w ∈ Γ2 , so Γ11 ◦ Γ12 ⊂ Γ1 and
Γ1 = Γ11 ◦Γ12 . Since Γ2 does not contain any clique factor and Γ11 ◦Γ2 ◦Γ12 = Γ1 ◦Γ2
is stable, we know Γ1 is stable in Γ by Theorem 2.10.
Remark 5.2. In the above proof, Γ12 may be empty. But if Γ12 6= ∅, then it does
not contain any clique join factor. Thus Γ11 is the maximal clique join factor of
Γ11 ◦ Γ2 ◦ Γ12 .
The next result answers the question at the end of Example 3.29.
Theorem 5.3. Suppose Out(G(Γ)) is finite and let q : X(Γ) → X(Γ0 ) be a quasiisometry. Then q induces a simplicial isomorphism q∗ : P(Γ) → P(Γ0 ), in particular, Out(G(Γ0 )) is transvection free.
In the following proof, we identify Γ with the one-skeleton of F (Γ), which is
the flag complex of Γ. Also recall that there are label-preserving projections π :
P(Γ) → F (Γ) and π : P(Γ0 ) → F (Γ0 ).
Proof. By Lemma 4.6, there is a simplicial embedding q∗ : P(Γ) → P(Γ0 ). Note
that q∗ (P(Γ)) is a full subcomplex in P(Γ0 ). To see this, pick a simplex ∆0 ⊂ P(Γ0 )
with its vertices in q∗ (P(Γ)). Then each vertex of ∆0 comes from a stable standard
geodesic line in X(Γ0 ). Thus there exists a stable standard flat F 0 ⊂ X(Γ0 ) such
that ∆(F 0 ) = ∆0 by Lemma 3.24. By considering the quasi-isometry inverse of q,
we know F 0 is Hausdorff close to the q-image of a stable standard flat in X(Γ).
Thus ∆(F 0 ) = ∆0 ⊂ q∗ (P(Γ)).
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
35
Pick vertex p ∈ X(Γ) and let {∆i }ki=1 , {Fi }ki=1 , {∆0i }ki=1 and {Fi0 }ki=1 be as in
the proof of Lemma 4.12. We claim
(5.4)
∩ki=1 Fi0 6= ∅.
Suppose (5.4) is not true. Then there exist 1 ≤ i1 6= i2 ≤ k and hyperplane
h0 ⊂ X(Γ) such that h0 separates Fi01 and Fi02 . Let l0 be a standard geodesic that
intersects h0 transversely and let v 0 = ∆(l0 ). By the discussion in Lemma 4.10, we
can find vertices v10 ∈ ∆0i1 and v20 ∈ ∆0i2 such that v10 and v20 are separated by St(v 0 ).
If there exists i0 such that Fi00 ∩ h 6= ∅, then v 0 ∈ q∗ (P(Γ0 )) and we can prove (5.4)
as in Lemma 4.12. Now we assume Fi0 ∩ h0 = ∅ for any i. Let w0 = π(v 0 ) ∈ Γ0 and
let Γw0 be the minimal stable subgraph of Γ0 that contains w0 .
We apply Lemma 5.1 to w0 ∈ Γ0 , if case (1) is true, let F 0 be the standard flat
in X(Γ0 ) such that l0 ⊂ F 0 and ΓF 0 = Γw0 . Since Γw0 is stable, ∆(F 0 ) ⊂ q∗ (P(Γ0 )),
in particular, v 0 ∈ q∗ (P(Γ0 )) and we can prove (5.4) as in Lemma 4.12.
If case (2) is true, let Γ01 = lk(w0 ) and let Γ02 = lk(Γ01 ). Take K10 and K 0 to be
the standard subcomplexes in X(Γ0 ) such that (1) the defining graphs ΓK10 , ΓK 0 of
K10 and K 0 satisfy ΓK10 = Γ01 and ΓK 0 = Γ01 ◦ Γ02 ; (2) l0 ⊂ K 0 and K10 ⊂ K 0 . Set
M10 = ∆(K10 ) and M 0 = ∆(K 0 ). Let K20 be an orthogonal complement of K10 in K 0 ,
i.e. K20 is a standard subcomplex such that ΓK20 = Γ02 and K 0 = K10 × K20 . It follows
that M 0 = M10 ∗ M20 for M20 = ∆(K20 ). By construction, v 0 ∈ M 0 and lk(v 0 ) = M10 .
Since K 0 and K10 are stable, there exist stable standard subcomplexes K and K1
∞
∞
in X(Γ) such that q(K) = K 0 and q(K1 ) = K10 . Moreover, by applying Theorem
2.10 to the quasi-isometry between K and K 0 , there exists a standard subcomplex
K2 ⊂ K such that K = K1 × K2 and K2 is quasi-isometric to K20 . Thus ΓK2 is also
disconnected. Let Mi = ∆(Ki ) ⊂ P(Γ) for i = 1, 2 and M = M1 ∗ M2 = ∆(K).
Then q∗ (M1 ) ⊂ M10 (at this stage we may not know q∗ (M1 ) = M10 ) and
(5.5)
q∗−1 (M10 ) = M1 .
To see this, pick a simplex ∆ ⊂ P(Γ) with q∗ (∆) ⊂ M10 . Suppose ∆ = ∆(F ) for a
stable standard flat F ⊂ X(Γ). Then q(F ) ⊂∞ K10 , hence F ⊂∞ K1 and ∆ ⊂ M1 .
Let L = ∪ki=1 ∆i and L0 = ∪ki=1 ∆0i . By the proof of Lemma 4.12, L0 \ (St(v 0 ) ∩ L0 )
is disconnected, thus L \ q∗−1 (St(v 0 ) ∩ L0 ) is disconnected. Recall that lk(v 0 ) = M10 ,
and we are assuming v 0 ∈
/ L0 . Thus (St(v 0 ) ∩ L0 ) ⊂ M10 . Then q∗−1 (St(v 0 ) ∩ L0 ) ⊂
−1
0
−1
q∗ (M1 ), hence q∗ (St(v 0 ) ∩ L0 ) ⊂ M1 by (5.5).
Let N = π(q∗−1 (St(v 0 ) ∩ L0 )) and let Ni = π(Mi ) for i = 1, 2. Then N separates
F (Γ), N ⊂ N1 and N2 is disconnected. Pick vertices u1 , u2 in different connected
components of N2 , then d(u1 , u2 ) ≥ 2 (since N2 is the full subcomplex spanned by
ΓK2 ). Since π(M ) = N1 ∗ N2 ⊂ F (Γ), N ⊂ St(ui ) \ {ui } for i = 1, 2. Let {Cj }dj=1
be the connected components of F (Γ) \ N . Then at most one of Cj is contained in
St(u1 ). If d ≥ 3, then St(u1 ) would separate F (Γ), which is a contradiction. Now
we suppose d = 2. Note that for i = 1, 2, there must exist j such that Cj ⊂ St(ui ),
otherwise St(ui ) would separate F (Γ). Moreover, if Cj ⊂ St(ui ), then ui ∈ Cj . So
we can assume without loss of generality that C1 ⊂ St(u1 ) and C2 ⊂ St(u2 ), which
implies F (Γ) = St(u1 ) ∪ St(u2 ), contradiction again (Lemma 4.17). Thus case (2)
is impossible and (5.4) is true.
Let {Fλ }λ∈Λ be the collection of maximal standard flats in X(Γ). Then X(Γ) =
∪λ∈Λ Fλ . For each λ, let Fλ0 be the unique maximal standard flat in X(Γ0 ) such
36
JINGYIN HUANG
∞
that q(Fλ ) = Fλ0 . Then
(5.6)
∞
X(Γ0 ) = ∪λ∈Λ Fλ0 .
Let h ⊂ X(Γ0 ) be an arbitrary hyperplane. Then h ∩ (∪λ∈Λ Fλ0 ) 6= ∅, otherwise
∪λ∈Λ Fλ0 would stay on one side of the hyperplane since it is a connected set by
(5.4), and this contradicts (5.6). Pick any standard geodesic r ⊂ X(Γ0 ) and let hr
be a hyperplane dual to r. Then there exists λ ∈ Λ such that Fλ0 ∩hr 6= ∅. It follows
that r ⊂∞ Fλ0 . So ∆(r) ∈ ∆(Fλ0 ) ⊂ q∗ (P(Γ)), which implies q∗ is surjective on the
vertices. However, q∗ (P(Γ)) is a full subcomplex in P(Γ0 ), so q∗ is surjective.
5.2. Coherent ordering and coherent labeling. Throughout this section, we
assume Out(G(Γ)) is finite and G(Γ) Z. If q : G(Γ) → G(Γ0 ) is a quasi-isometry,
then G(Γ0 ) has a quasi-action (cf. [KL01, Definition 2.2]) on G(Γ), which induces
a group homomorphism:
H : G(Γ0 ) → QI(G(Γ))
On the other hand, since G(Γ) acts by isometries on X(Γ), we can identify G(Γ)
as a subgroup of QI(G(Γ)) (more precisely, we embed G(Γ) into Isom(G(Γ), dw )
and embed Isom(G(Γ), dw ) into QI(G(Γ)) by Corollary 4.24). In this subsection,
we will understand the following question.
Does there exist g ∈ QI(G(Γ)) such that g · H(G(Γ0 )) · g −1 ⊂ G(Γ)?
Recall that we have picked an identification between G(Γ) and the 0-skeleton of
X(Γ). Each circle in the 1-skeleton of the Salvetti complex of G(Γ) is labeled by
an element in the standard generating set S of G(Γ). Moreover, we have chosen
an orientation for each such circle. By pulling back the labeling and orientation
of edges to the universal cover X(Γ), we obtain a G(Γ)-invariant directed labeling
of edges in X(Γ). Moreover, both the labeling and orientation of edges in X(Γ)
are compatible with parallelism between edges. This also induces an associated
G(Γ)-invariant labeling of vertices in P(Γ).
Let {lλ }λ∈Λ be the collection of standard geodesics in X(Γ) and let Vλ = v(lλ )
be the vertex set of lλ . A coherent ordering of G(Γ) is obtained by assigning a
collection of bijections fλ : Vλ → Z for each λ ∈ Λ such that if lλ1 and lλ2 are
parallel, then the fλ2 ◦ p ◦ fλ−1
: Z → Z is a translation, where p : Vλ1 → Vλ2 is the
1
map induced by parallelism. The map fλ pulls back the total order on Z to Vλ ,
which we denote by ≤λ . Then p : Vλ1 → Vλ2 is order-preserving.
Two coherent ordering Ω1 and Ω2 are equivalent, denoted by Ω1 = Ω2 , if their
collections of bijections agree up to a translation of Z. Recall that we have a G(Γ)invariant orientation of edges in X(Γ) which is compatible with parallelism between
edges. This induces a unique coherent ordering Ω of G(Γ) up to the equivalence
relation defined before. Moreover, for any element g ∈ G(Γ), the pull back g ∗ (Ω)
is also a coherent ordering, moreover, g ∗ (Ω) = Ω.
Recall that for any vertex v ∈ X(Γ), there is a label-preserving simplicial embedding iv : F (Γ) → P(Γ) by considering the standard geodesics passing through
v. A coherent labeling of G(Γ) is a simplicial map a : P(Γ) → F (Γ) such that
a ◦ iv : F (Γ) → F (Γ) is a simplicial isomorphism for every vertex v ∈ X(Γ).
The label-preserving projection L : P(Γ) → F (Γ) gives rise to a coherent labeling
of G(Γ). Recall that G(Γ) acts on P(Γ) by simplicial automorphisms, and the
labeling of vertices in P(Γ) is G(Γ)-invariant. Thus for any element g ∈ G(Γ), the
pull back g ∗ (L) is also a coherent labeling and g ∗ (L) = L.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
37
We have the following alternative characterization of elements in Isom(G(Γ), dr ).
Lemma 5.7. There is a 1-1 correspondence which associates each element of
Isom(G(Γ), dr ) to a triple consisting of:
(1) A point v ∈ G(Γ).
(2) A coherent ordering of G(Γ) (up to the equivalence relation defined above).
(3) A coherent labeling of G(Γ).
Proof. Pick φ ∈ Isom(G(Γ), dr ) and let ϕ = h(φ) : P(Γ) → P(Γ), where h is the
monomorphism in Remark 4.22. Then ϕ∗ L = L ◦ ϕ : P(Γ) → F (Γ) is a coherent
labeling of G(Γ). Pick a standard geodesic l1 ⊂ X(Γ). Then the parallel set Pl1
admits a splitting Pl1 = l1 × l1⊥ . Since φ maps vertices in a standard flat bijectively
to vertices in a standard flat, there exists a standard geodesic l2 ⊂ X(Γ) such that
φ(v(l1 )) = v(l2 ) and φ(v(Pl1 )) = Pl2 , moreover, φ respects the product structure
on Pl1 . Thus the pull-back φ∗ Ω is a coherent ordering of G(Γ). Now we can set up
the correspondence in one direction:
φ
φ(id), φ∗ Ω and ϕ∗ L
here id denotes the identity element of G(Γ).
Conversely, given a point v ∈ G(Γ), a coherent ordering Ω0 and a coherent
labeling L0 , we can construct a map φ as follows. Set φ(id) = v. For u ∈ G(Γ), pick a
word wu = a1 a2 · · · an representing u. Let ui be the point in G(Γ) represented by the
word a1 a2 · · · ai for 1 ≤ i ≤ n and let u0 = id. We define qi = φ(a1 a2 · · · ai ) ∈ G(Γ)
inductively as follows. Set q0 = v and suppose qi−1 is already defined. Denote
the standard geodesic containing ui−1 and ui by li . Let vi = L0 (∆(li )) which is
a vertex of Γ, and let li0 be the standard line that contains qi−1 and is labeled by
vi . Denote the vertex set of li with the order from Ω0 by (v(li ), ≤Ω0 ). Suppose
k : (v(li ), ≤Ω0 ) → (v(li0 ), ≤Ω ) is the unique order preserving bijection such that
k(ui−1 ) = qi−1 . Then we define qi = k(ui ).
We claim that for any other word wu0 representing u, φ(wu ) = φ(wu0 ), hence there
is a well-defined map φ : G(Γ) → G(Γ). To see this, recall that one can obtain wu
from wu0 by performing the following two basic moves:
(1) w1 aa−1 w2 → w1 w2 .
(2) w1 abw2 → w1 baw2 when a and b commute.
It is clear that φ(w1 aa−1 w2 ) = φ(w1 w2 ). For the second move, let ui−1 , ui , u0i ,
and ui+1 be points in G(Γ) represented by w1 , w1 a, w1 b and w1 ab = w1 ba respectively. Define qi−1 = φ(w1 ), qi = φ(w1 a), qi0 = φ(w1 b), qi+1 = φ(w1 ab) and
0
qi+1
= φ(w1 ba). Since L0 is a coherent labeling, ∠qi (qi+1 , qi−1 ) = ∠qi−1 (qi , qi0 ) =
0
∠qi0 (qi−1 , qi+1
) = π/2, moreover, the standard geodesic containing qi and qi+1 is
parallel to the standard geodesic containing qi−1 and qi0 . Since Ω0 is a coherent
ordering, d(qi , qi+1 ) = d(qi−1 , qi0 ), thus qi qi+1 and qi−1 qi0 are parallel. Similarly,
0
0
qi−1 qi and qi0 qi+1
are parallel, thus qi+1 = qi+1
.
0
Now we define another map φ : G(Γ) → G(Γ), which serves as the inverse
of φ. Set φ0 (v) = id and pick word w = a1 a2 · · · an . Let ri be the point in G(Γ)
represented by va1 a2 · · · ai for 1 ≤ i ≤ n and r0 = v. We define pi = φ0 (va1 a2 · · · ai )
inductively as follows. Put p0 = id and suppose pi−1 is already defined. Since L0
is a coherent labeling, there exists a unique standard geodesic li containing pi−1
such that L0 (∆(li )) and the edge ri−1 ri share the same label. Let li0 be the unique
standard geodesic containing ri−1 and ri and let k 0 : (v(li0 ), ≤Ω ) → (v(li ), ≤Ω0 ) be
38
JINGYIN HUANG
the unique order preserving bijection such that k 0 (ri−1 ) = pi−1 . Put pi = k 0 (ri ).
By a similar argument as above, φ0 : G(Γ) → G(Γ) is well-defined. It is not hard
to deduce the following properties from our construction:
(1) φ0 ◦ φ = φ ◦ φ0 = Id.
(2) dr (φ(v1 ), φ(v2 )) ≤ dr (v1 , v2 ) and dr (φ0 (v1 ), φ0 (v2 )) ≤ dr (v1 , v2 ) for any vertices v1 , v2 ∈ G(Γ).
(3) If L0 = L and Ω0 = Ω, then φ is a left translation. If in addition v = id,
then φ = Id.
It follows from (1) and (2) that φ ∈ Isom(G(Γ), dr ). Moreover, v = φ(id), L0 = ϕ∗ L
(ϕ = h(φ) where h is the monomorphism in Remark 4.22) and Ω0 = φ∗ Ω, thus we
have established the required 1-1 correspondence.
Pick finite simplicial graphs Γ and Γ0 such that (1) Out(G(Γ)) is finite; (2) there
exists a simplicial isomorphism s : P(Γ) → P(Γ0 ). By Lemma 4.12, s induces a
map φ : G(Γ) → G(Γ0 ). For every g 0 ∈ G(Γ0 ), there is a left translation φ̄g0 :
G(Γ0 ) → G(Γ0 ), which gives rise to a simplicial isomorphism s̄g0 : P(Γ0 ) → P(Γ0 ).
Let sg0 = s−1 ◦ s̄g0 ◦ s. Then sg0 gives rise to a map φg0 ∈ Isom(G(Γ), dr ) by
Corollary 4.20, moreover, by Lemma 4.12,
φ̄g0 ◦ φ = φ ◦ φg0
(5.8)
0
0
0
for any g ∈ G(Γ ). So G(Γ ) acts on G(Γ), and we can define a homomorphism
Φ : G(Γ0 ) → Isom(G(Γ), dr ) by sending g 0 to φg0 . Φ is injective since each step in
defining Φ is injective.
Lemma 5.9. In the above setting, there exists an element φ1 ∈ Isom(G(Γ), dr )
such that it conjugates the image of Φ to a finite index subgroup of G(Γ).
We identify G(Γ) as a subgroup of Isom(G(Γ), dr ) via the left action of G(Γ) on
itself.
Proof. Pick a reference point q ∈ Im φ and let Kq = (F (Γ0 ))q . Denote the points
in φ−1 (q) by {pλ }λ∈Λ and let Kpλ = (F (Γ))pλ . Since {φ(Kpλ )}λ∈Λ are distinct
subcomplexes of Kq , Λ is a finite set.
Let L : P(Γ) → F (Γ) and Ω be the coherent labeling and coherent ordering
induced by the G(Γ)-invariant labeling of X(Γ) and P(Γ). We can obtain a coherent
labeling L0 : P(Γ0 ) → F (Γ0 ) and a coherent ordering Ω0 for G(Γ0 ) in a similar fashion
which are invariant under the G(Γ0 )-action, i.e.
(5.10)
(s̄g0 )∗ L0 = L0 and (φ̄g0 )∗ Ω0 = Ω0 .
Our goal is to find a coherent labeling L1 and a coherent ordering Ω1 of G(Γ) such
that (sg0 )∗ L1 = L1 and (φg0 )∗ Ω1 = Ω1 for any g 0 ∈ G(Γ0 ).
Let iq : F (Γ0 ) → P(Γ0 ) be the canonical embedding and let
L1 = L ◦ s−1 ◦ iq ◦ L0 ◦ s
be the simplicial map from P(Γ) to F (Γ). Pick arbitrary p ∈ G(Γ) and let ip :
F (Γ) → P(Γ) be the canonical embedding. We need to show L1 ◦ ip is a simplicial
isomorphism. Let Kp = ip (F (Γ)) and let g10 ∈ G(Γ0 ) such that g10 · φ(p) = q. Then
iq ◦ L0 |s(Kp ) = s̄g10 |s(Kp ) . Thus
L1 ◦ ip = L ◦ s−1 ◦ iq ◦ L0 ◦ s ◦ ip = L ◦ s−1 ◦ s̄g10 ◦ s ◦ ip = L ◦ sg10 ◦ ip ,
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
39
which is a simplicial isomorphism by Lemma 4.12. It follows that L1 is a coherent
labeling, moreover,
(sg0 )∗ L1 = (L ◦ s−1 ◦ iq ◦ L0 ◦ s) ◦ (s−1 ◦ s̄g0 ◦ s) = L ◦ s−1 ◦ iq ◦ L0 ◦ s̄g0 ◦ s
= L ◦ s−1 ◦ iq ◦ L0 ◦ s = L1
for any g 0 ∈ G(Γ0 ), where the third equality follows from (5.10). So L1 is the
required coherent labeling.
To simplify notation, we will write x <Ω y if x < y under the ordering Ω.
We define Ω1 as follows. Let p1 , p2 ∈ G(Γ) be two distinct points in a standard
geodesic line. If φ(p1 ) 6= φ(p2 ), then we set p1 <Ω1 p2 if and only if φ(p1 ) <Ω0
φ(p2 ). If φ(p1 ) = φ(p2 ), then by (5.8), there exists a unique g 0 ∈ G(Γ0 ) such that
φg0 (pi ) ∈ φ−1 (q) for i = 1, 2 and we set p1 <Ω1 p2 if and only if φg0 (p1 ) <Ω φg0 (p2 ).
It follows from (5.10), (5.8) and our construction that p1 <Ω1 p2 if and only if
φg0 (p1 ) <Ω1 φg0 (p2 ) for any p1 , p2 in the same standard geodesic line and any
g 0 ∈ G(Γ0 ), thus (φg0 )∗ Ω1 = Ω1 .
To verify Ω1 is coherent, pick parallel standard geodesics l1 and l2 in X(Γ) and
pick distinct vertices p11 , p12 ∈ l. Let p21 , p22 be the corresponding vertices in l2
via parallelism. We assume p11 <Ω1 p12 , it suffices to prove p21 <Ω1 p22 .
Case 1: We assume φ(p11 ) 6= φ(p12 ). Recall that l1 can be realized as an intersection of finitely many maximal standard flats, so by Lemma 4.12, there exists a
standard geodesic line l10 ⊂ X(Γ0 ) such that φ(v(l1 )) ⊂ v(l10 ) and φ(v(Pl1 )) ⊂ v(Pl10 ),
moreover, φ respects the product structures of Pl1 and Pl10 . Thus φ(p11 )φ(p21 ) and
φ(p21 )φ(p22 ) are the opposite sides of a flat rectangle in X(Γ0 ). Now p21 <Ω1 p22
follows since Ω0 is coherent.
Case 2: We assume φ(p11 ) = φ(p12 ) 6= φ(p21 ). In this case, we can assume
without loss of generality that φ(p11 ) = φ(p12 ) = q (since (φg0 )∗ Ω1 = Ω1 ) and
the points p11 and p21 stay in the same standard geodesic. For i = 1, 2, let ri be
the standard geodesic passing p1i and p2i . Take ri0 ⊂ X(Γ0 ) and li0 ⊂ X(Γ0 ) to be
the standard geodesics such that φ(v(ri )) ⊂ v(ri0 ) and φ(v(li )) ⊂ v(li0 ) respectively.
Denote q 0 = φ(p21 ). Since φ restricted on v(Pl1 ) respects the product structure,
φ(p21 ) = φ(p22 ) = q 0 and r10 = r20 .
Let φg0 be the left translation such that φg0 (q 0 ) = q. Since q 0 ∈ r10 and q ∈ r10 ,
φg0 is a translation along r10 and sg0 fixes every point in St(∆(r10 )), hence sg0 fixes
every point in s−1 (St(∆(r10 ))) = St(∆(r1 )) and
(5.11)
φg0 (ri ) = ri
for i = 1, 2. Let l3 = φg0 (l2 ). Then l3 is parallel to l1 (or l2 ). To see this, note
that ∆(l1 ) ∈ St(∆(r1 )), hence ∆(l1 ) is fixed by sg0 . Put p3i = φg0 (p2i ) for i = 1, 2.
Then p3i ∈ ri by (5.11), hence p11 p12 and p31 p32 are the opposite sides of a flat
rectangle. Moreover p3i ∈ φ−1 (q) for i = 1, 2 by (5.8), then p31 <Ω1 p32 since Ω is
coherent and Ω = Ω1 while restricted on φ−1 (q). Now the G(Γ0 )-invariance of Ω1
implies p21 <Ω1 p22 .
Case 3: If φ(p11 ) = φ(p12 ) = φ(p21 ), then we can assume without loss of generality that they all equal to q. It follows that φ(p22 ) = q since φ respects the product
structure while restricted on v(Pl1 ). Thus p21 <Ω1 p22 by definition.
40
JINGYIN HUANG
By Lemma 5.7, there exists φ1 ∈ Isom(G(Γ), dr ) such that φ∗1 Ω = Ω1 and s∗1 L =
L1 (s1 = h(φ1 ) where h is the monomorphism in Remark 4.22). Thus
−1 ∗
−1 ∗
−1 ∗
∗
∗
∗
∗
(φ1 ◦ φg0 ◦ φ−1
1 ) Ω = (φ1 ) ◦ (φg 0 ) ◦ (φ1 Ω) = (φ1 ) ◦ (φg 0 ) Ω1 = (φ1 ) Ω1 = Ω
∗
0
0
for any g 0 ∈ G(Γ0 ). Similarly, (s1 ◦ sg0 ◦ s−1
1 ) L = L for any g ∈ G(Γ ). Note that
−1
−1
0
s1 ◦ sg0 ◦ s1 = h(φ1 ◦ φg0 ◦ φ1 ), thus by Lemma 5.7, G(Γ ) acts on G(Γ) by left
0
translations via g 0 → φ1 ◦ φg0 ◦ φ−1
1 . This induces a monomorphism G(Γ ) → G(Γ).
−1
Moreover, by (5.8) and the fact that φ (q) is finite, this action has finite quotient,
thus we can realize G(Γ0 ) as a finite index subgroup of G(Γ).
The next result basically says under suitable conditions, if there exists a quasiisometry q : G(Γ) → G(Γ0 ), then there exists a very “nice” quasi-isometry q 0 :
G(Γ) → G(Γ0 ). However, we do not insist that q 0 is of bounded distance away from
q (compared to Theorem 4.18).
Theorem 5.12. Let Γ and Γ0 be finite simplicial graphs such that Out(G(Γ)) is
finite and G(Γ0 ) is quasi-isometric to G(Γ). Then there exists a cubical map (cf.
Definition 2.1) ϕ : X(Γ) → X(Γ0 ) such that
(1) The map ϕ is onto, and ϕ maps any standard flat in X(Γ) onto a standard
flat in X(Γ0 ) of the same dimension.
(2) The map ϕ maps combinatorial geodesics in the 1-skeleton of X(Γ) to combinatorial geodesics in the 1-skeleton of X(Γ0 ).
(3) The map ϕ is a quasi-isometry.
Proof. Let f : G(Γ) → G(Γ0 ) be a quasi-isometry. By Theorem 5.3, f induces
a simplicial isomorphism s : P(Γ) → P(Γ0 ). By Lemma 4.12, s induces a map
φ : G(Γ) → G(Γ0 ) such that dw (f (x), φ(x)) < D for any x ∈ G(Γ). Let φ1 be the
map in Lemma 5.9 and let ϕ = φ ◦ φ−1
1 . We will use the same notation as in the
proof of Lemma 5.9.
We claim that if F = ∩hi=1 Fi where each Fi is a maximal standard flat, then there
exists a unique standard flat F 0 ⊂ G(Γ0 ) such that φ(v(F )) = v(F 0 ). To see this, let
Fi0 be the maximal standard flat in X(Γ0 ) such that ∆(Fi0 ) = s(∆(Fi )) for 1 ≤ i ≤ h
and let F 0 = ∩hi=1 Fi0 . Then it follows from Lemma 4.12 that φ(v(F )) ⊂ v(F 0 ).
Recall that G(Γ0 ) acts on G(Γ0 ), P(Γ0 ), G(Γ) and P(Γ). The stabilizer Stab(v(F 0 ))
fixes ∆(Fi0 ) for all i, hence it fixes ∆i for all i and Stab(v(F 0 )) ⊂ Stab(v(F )).
Since Stab(v(F 0 )) acts on v(F 0 ) transitively, (5.8) implies φ(v(F )) = v(F 0 ) and
|φ−1 (y) ∩ F | = |φ−1 (y 0 ) ∩ F | for any y, y 0 ∈ v(F 0 ). It also follows that Stab(v(F )) ⊂
Stab(v(F 0 )), thus Stab(v(F 0 )) = Stab(v(F )).
Note that the above claim is also true for ϕ, and any standard geodesic satisfies
the assumption of the claim. Moreover, ϕ is surjective since φ1 is surjective by
(5.8). Pick standard geodesic l ⊂ X(Γ) and l0 ⊂ X(Γ0 ) such that v(l0 ) = ϕ(v(l)),
and we identify v(l) and v(l0 ) with Z in an order-preserving way. Then the above
claim and the construction of φ1 imply that ϕ|v(l) is of form
(5.13)
ϕ(a) = ba/dc + r
for some integers r and d (d ≥ 1). In particular, ϕ can be extended to a simplicial
map from the Cayley graph C(Γ) of G(Γ) to C(Γ0 ).
Pick a combinatorial geodesic ω ⊂ C(Γ) connecting vertices x and y, we claim
that ω 0 = φ(ω) is also a geodesic in C(Γ0 ) (it could be a point). Let {vi }ni=0 be
vertices in ω such that for 0 ≤ i ≤ n−1, [vi , vi+1 ] is a maximal sub-segment of ω that
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
41
is contained in a standard geodesic (v0 = x and vn = y). Denote the corresponding
standard geodesic by li . For 0 ≤ i ≤ n − 1, let li0 ⊂ X(Γ0 ) be the standard geodesic
such that v(li0 ) = ϕ(v(li )) and ωi0 = φ([vi , vi+1 ]). Then ωi0 is a (possibly degenerate)
n−1
segment in li0 by (5.13). Since ω is a geodesic, none of two geodesics in {li }i=0
are parallel. Note that ϕ is induced by a simplicial isomorphism between P(Γ)
and P(Γ0 ), thus the same property is true for the collection of geodesics {li0 }n−1
i=0 .
It follows that no hyperplane in X(Γ0 ) could intersect ω 0 at more than one point,
hence ω 0 is a combinatorial geodesic.
Let ui = ϕ(vi ). Then dw (ui , ui+1 ) ≤ dw (vi , vi+1 ) by (5.13) (recall that dw
denotes the word metric on the corresponding group). Thus
(5.14)
dw (ϕ(x), ϕ(y)) =
n−1
X
dw (ui , ui+1 ) ≤
i=0
n−1
X
dw (vi , vi+1 ) = dw (x, y)
i=0
for any x, y ∈ G(Γ).
Pick p ∈ G(Γ0 ) and let k = |ϕ−1 (p)|. Then k does not depend on p by (5.8). It
follows that dw (ϕ(x), ϕ(y)) ≥ 1 whenever dw (x, y) ≥ k + 1. Now we can cut ω into
pieces of length k + 1. Since ϕ(ω) is a combinatorial geodesic,
dw (x, y)
− 1.
k+1
Note that ϕ naturally extends to a cubical map from X(Γ) to X(Γ0 ), which satisfies
all the required properties.
dw (ϕ(x), ϕ(y)) ≥
Theorem 5.15. If Γ and Γ0 are finite simplicial graphs such that Out(G(Γ)) is
finite, then the following are equivalent:
(1) G(Γ0 ) is quasi-isometric to G(Γ).
(2) P(Γ0 ) is isomorphic to P(Γ) as simplicial complexes.
(3) G(Γ0 ) is isomorphic to a subgroup of finite index in G(Γ).
Proof. (1) ⇒ (2) follows from Theorem 5.3. (2) ⇒ (3) follows from Lemma 5.9.
(3) ⇒ (1) is trivial.
This establishes Theorem 1.2 in the introduction.
6. The geometry of finite index RAAG subgroups
Throughout this section, we assume G(Γ) Z, since the main results of this
section (Theorem 6.20 and Theorem 6.26) are trivial when G(Γ) ∼
= Z.
6.1. Constructing finite index RAAG subgroups. A right-angled Artin subgroup is a subgroup which is also a right-angled Artin group. In this section, we
introduce a process to obtain finite index RAAG subgroups of an arbitrary RAAG.
Lemma 6.1. Let X be a CAT (0) cube complex, let l ⊂ X be a geodesic in the
1-skeleton and let {hi }i∈Z be consecutive hyperplanes dual to l. Let πl : X → l be
the CAT (0) projection. Then
(1) For every edge e ⊂ X, if e ∩ hi = ∅ for all i, then πl (e) is a vertex in l, if
e ∩ hi 6= ∅ for some i, then πl (e) is an edge in l.
(2) If K is any connected subcomplex such that e ∩ hi = ∅ for all i, then πl (K)
is a vertex in l, moreover, if K stays between hi and hi+1 , then πl (K) is
the vertex in l that stays between hi and hi+1 .
42
JINGYIN HUANG
(3) For every interval [a, b] ⊂ l, πl−1 ([a, b]) is a convex set in X. In particular,
if x ∈ l is a vertex, then πl−1 (x) is a convex subcomplex of X.
(4) If K is a convex subcomplex such that K ∩ l 6= ∅, then πl (K) = K ∩ l.
Proof. Here (1) and (3) follow from the fact the every hyperplane has a carrier,
and (2) follows from (1). To see (4), it suffices to show for i such that hi ∩ l 6= ∅
and hi ∩ K 6= ∅, we have ei ⊂ K (ei is the edge in l dual to hi ). Let Nhi be
the carrier of hi . By Lemma 2.3, d(x, Nhi ∩ K) ≡ c for any x ∈ ei . Moreover,
d(x, Nhi ∩ K) = d(x, K) for x in the interior of ei , so we must have c = 0, otherwise
the convexity of d(·, K) would imply K ∩ l = ∅.
Lemma 6.2. Let l ⊂ X(Γ) be a standard geodesic. Then there is a map π∆(l) :
v(P(Γ) \ St(∆(l))) → v(l) (recall that v(P(Γ) \ St(∆(l)) is the collection of vertices
in P(Γ) \ St(∆(l)) such that if v1 and v2 are in the same connected component of
P(Γ) \ St(∆(l)), then π∆(l) (v1 ) = π∆(l) (v2 ).
Proof. Let πl : X(Γ) → l be the CAT(0) projection and let l1 ⊂ X(Γ) be a standard
geodesic such that d(∆(l1 ), ∆(l)) ≥ 2. Then πl (l1 ) is a vertex in l by Lemma 3.1
and Corollary 3.2. Moreover, we claim πl (l1 ) = πl (l2 ) if l2 is a standard geodesic
parallel to l1 . It suffices to prove the case when there is a unique hyperplane h
separating l1 from l2 . Note that d(∆(l1 ), ∆(l)) ≥ 2 yields h ∩ l = ∅, so l1 and l2 are
pinched by two hyperplanes dual to l, then the claim follows from Lemma 6.1. Thus
πl induces a well-defined map π∆(l) : v(P(Γ) \ St(∆(l))) → v(l). If ∆(l1 ) and ∆(l2 )
are connected by an edge, then there exist standard geodesics l10 and l20 such that
l10 ∩ l20 6= ∅ and li0 is parallel to li for i = 1, 2. Thus πl (l1 ) = πl (l10 ) = πl (l20 ) = πl (l2 )
and π∆(l) (∆(l1 )) = π∆(l) (∆(l2 )).
Pick a standard generating set S of G(Γ) and let C(Γ, S) be the Cayley graph.
We identify G(Γ) as a subset of C(Γ, S) and attach higher dimensional cubes to
C(Γ, S) to obtain a CAT (0) cube complex X(Γ, S), which is basically the universal
cover of the Salvetti complex. Here we would like to think G(Γ) as a fixed set
and C(Γ, S), X(Γ, S) as objects formed by adding edges and cubes to G(Γ) in a
particular way determined by S, so we write S explicitly. We will choose a G(Γ)equivariant orientation for edges in X(Γ, S) as before.
An S-flat (or an S-geodesic) in G(Γ) is defined to be the vertex set of a standard
flat (or geodesic) in X(Γ, S). We define P(Γ, S) as before such that its vertices
correspond to coarse equivalence classes of S-geodesics.
We define an isometric embedding I : G(Γ) → `1 (v(P(Γ, S))) which depends
on S and the orientation of edges in X(Γ, S). Pick standard geodesic l ⊂ X(Γ, S)
and let πl : X(Γ, S) → l be the CAT (0) projection. We identify v(l) with Z∆(l)
in an orientation preserving way such that πl (id) = 0 (id is the identity element in
G(Γ)). Then πl induces a coordinate function I∆(l) : G(Γ) → Z∆(l) . If we change l
to a standard geodesic l1 parallel to l, then I∆(l) and I∆(l1 ) are identical by Lemma
6.1. Thus for every vertex v ∈ P(Γ), there is a well-defined coordinate function
Iv : G(Γ) → Zv . These coordinate functions induce a map I : G(Γ) → Z(v(P(Γ))) .
I is an embedding since every two points in G(Γ) are separated by some hyperplane. I(G(Γ)) ⊂ `1 (v(P(Γ))) since for any g ∈ G(Γ), there are only finitely
many hyperplanes separating id and g. I naturally extends to a map I : X(Γ, S) →
`1 (v(P(Γ))) and it maps combinatorial geodesics to geodesics by the argument in
Theorem 5.12. Thus I is an isometric embedding with respect to the `1 metric on
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
43
X(Γ, S). We say a convex subcomplex K ⊂ X(Γ, S) is non-negative if each point in
I(K) has non-negative coordinates (this notion depends on the orientation of edges
in X(Γ, S)). Let CN (Γ, S) be the collection of compact, convex, non-negative
subcomplexes of X(Γ, S) that contain the identity.
For any K ∈ CN (Γ, S), we find a maximal collection of standard geodesics
{ci }si=1 such that ci ∩ K 6= ∅ for all i and ∆(ci ) 6= ∆(cj ) for any i 6= j. Let
gi ∈ S be the label of edges in ci and let αi = πci (id). Put ni = |v(K ∩ ci )| and
vi = αi gini αi−1 . Let G0 be the subgroup generated by {vi }si=1 . It follows from the
convexity of K that if a standard geodesic c is parallel to ci and c ∩ K 6= ∅, then
|v(K ∩ ci )| = |v(K ∩ c)|. Thus {vi }si=1 and G0 do not depend on the choice of ci ’s.
Lemma 6.3. G0 is a finite index subgroup of G(Γ).
Proof. We prove this by showing G0 · v(K) = G(Γ). Let dr be the syllable metric
on G(Γ) defined in Section 4.3. Pick word α ∈ G(Γ) and assume α ∈ G0 · v(K)
when dr (α, id) ≤ k − 1. If dr (α, id) = k, then there exists β ∈ G(Γ) such that
dr (id, β) = k − 1 and dr (β, α) = 1. Let β = β1 β2 for β1 ∈ G0 and β2 ∈ v(K). Then
dr (β2 , β1−1 α) = 1. Suppose c is the standard geodesic containing β2 and β1−1 α.
Then there exists i such that ci and c are parallel. Note that Pc ∩ K is a convex set
in the parallel set Pc , hence respects the natural splitting Pc = c × c⊥ , moreover,
the left action of vi translates the c factor by ni units and fixes the other factor.
Thus there exists d ∈ Z and β20 ∈ K ∩ c such that vid β20 = β1−1 α, which implies
α = β1 vid β20 ∈ G0 · v(K).
Let Γ0 be the full subgraph of P(Γ) spanned by points {∆(ci )}si=1 . Then there
is a natural homomorphism G(Γ0 ) → G0 .
Lemma 6.4. The homomorphism G(Γ0 ) → G0 is actually an isomorphism. Hence
G0 is a finite index RAAG subgroup of G(Γ).
We will follow the strategy in [Kob12a], where the following version of ping-pong
lemma for right-angled Artin groups was used.
Theorem 6.5 (Theorem 4.1 of [Kob12a]). Let G = G(Γ) and let X be a set with
a G-action. Suppose the following hold.
(1) For each vertex vi of Γ, there exists subset Xi ⊂ X, such that the union of
all Xi ’s is properly contained in X.
(2) For each nonzero k ∈ Z and vertices vi , vj joined by en edge, vik (Xj ) ⊂ Xj .
(3) For each nonzero k ∈ Z and vertices vi , vj not joined by en edge, vik (Xj ) ⊂
Xi .
(4) There exists x0 ∈ X \ ∪i∈V Xi (V is the vertex set of Γ) such that for each
nonzero k ∈ Z, vik (x0 ) ∈ Xi .
Then the G-action is faithful.
Proof of Lemma 6.4. We will apply Theorem 6.5 with X = X(Γ, S) and G = G(Γ0 ).
For 1 ≤ i ≤ s, we identify ci and R in an orientation preserving way such that πci (id)
corresponds to 0 ∈ R. Define Xi+ = πc−1
([ni − 1/2, ∞)), Xi− = πc−1
((−∞, −1/2])
i
i
+
−
and Xi = Xi ∪ Xi . It clear that the identity element id ∈
/ Xi for all i, so (1) of
Theorem 6.5 is true. Each vi = αi gini αi−1 translates ci by ni units, so (4) is also
true with x0 = id.
44
JINGYIN HUANG
If ∆(ci ) and ∆(cj ) are connected by an edge in P(Γ), then vi stabilizes every
hyperplane dual to vj , thus vik (Xj ) = Xj and (2) is true. If
(6.6)
d(∆(ci ), ∆(cj )) ≥ 2,
then πcj (ci ) is a point. Lemma 6.1 and ci ∩ K 6= ∅ yield that πcj (ci ) ⊂ πcj (K) =
cj ∩ K = [0, nj − 1], thus
(6.7)
ci ∩ Xj = ∅.
(−1/2) be the boundary of Xj− and let Nh be the
Similarly ci ∩Xj = ∅. Let h = πc−1
j
carrier of h. Then (6.6) implies that h has empty intersection with any hyperplane
dual to ci , so is Nh . It follows from Lemma 6.1 that πci (h) = πci (Nh ) = p is a vertex
in ci . If h1 = πc−1
(p − 1/2) and h2 = πc−1
(p + 1/2) are two hyperplanes that pinch p,
i
i
then h ∩ hk = ∅ for k = 1, 2. This and (6.7) yield Xj− ∩ hk = ∅, hence πci (Xj− ) = p
by Lemma 6.1. Similarly πci (Xj+ ) = p, so p = πci (Xj ) = πci (cj ) ⊂ πci (K) =
ci ∩ K = [0, ni − 1]. Note that (πci ◦ vik )(Xj ) = (vik ◦ πci )(Xj ) = vik (p) = p + kni ,
so vik (Xj ) ⊂ Xi for k 6= 0.
The discussion in this subsection yields a well-defined map
ΘS : CN (Γ, S) → {Finite index RAAG subgroups of G(Γ)}.
The images of ΘS are called S-special subgroups of G(Γ). A subgroup of G(Γ)
is special if it is S-special for some standard generating set S of G(Γ).
6.2. Rigidity of RAAG subgroups. In this subsection, we will assume G(Γ0 ) is
finite index RAAG subgroup in G(Γ) and Out(G(Γ)) is finite. We will show that
under such condition, G(Γ0 ) must arise from the process described in the previous
subsection. We will prove this in 3 steps. First we produce a convex subcomplex
of X(Γ, S) from G(Γ0 ). Then we will modify this convex subcomplex such that it
is an element in CN (Γ, S). Thus we have defined a map from finite index RAAG
subgroups of G(Γ) to elements in CN (Γ, S). In the last step, we show the map
defined in step 2 is an inverse to the map ΘS defined in Section 6.1.
Also near the end of this subsection, we will leave several relatively long remarks
which discuss relevant material in the literature. The reader can skip these remarks
at first reading.
Recall that Out(G(Γ)) is finite and Out(G(Γ0 )) is transvection free (Theorem
5.3), so any two standard generating sets of G(Γ) (or G(Γ0 )) differ by a sequence
of conjugations or partial conjugations. Then given any two standard generating
sets S and S1 for G(Γ), there is a canonical way to identify P(Γ, S) and P(Γ, S1 )
(every S-geodesic is Hausdorff close to an S1 -geodesic). Thus we will write P(Γ)
and P(Γ0 ) and omit the generating set.
Lemma 6.8. Let φ, s be as in the discussion before Lemma 5.9. Let l ⊂ X(Γ) and
l0 ⊂ X(Γ0 ) be standard geodesics such that φ(v(l)) = v(l0 ). Then φ◦π∆(l) = π∆(l0 ) ◦s.
Proof. Pick standard geodesics r ⊂ X(Γ) and r0 ⊂ X(Γ0 ) such that φ(v(r)) = v(r0 ),
then s(∆(r)) = ∆(r0 ) by Lemma 4.12 (recall that r is the intersection of maximal
standard flats). Therefore, by the definition of π∆(l) , it suffices to show φ ◦ πl (x) =
πl0 ◦ φ(x) for any vertex x ∈ X(Γ). Let y be a vertex such that y ∈
/ l and let
x = πl (y). By Lemma 6.1, we can approximate xy by a combinatorial geodesic
ω in the 1-skeleton of πl−1 (y), then no hyperplane could intersect both l and ω.
Let {vi }ni=0 be vertices in ω such that for 0 ≤ i ≤ n − 1, [vi , vi+1 ] is a maximal
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
45
sub-segment of ω that is contained in a standard geodesic (v0 = x and vn = y).
Denote the corresponding standard geodesic by li . Then ∆(l) 6= ∆(li ) for all i. Let
ui = φ(vi ) and let li0 be the standard geodesic such that φ(v(li )) = v(li0 ). Then
ui ui+1 ⊂ li0 and ∆(l0 ) 6= ∆(li0 ) for all i, thus πl0 (li0 ) is a point by Corollary 3.2 and
πl0 (ui ) = πl0 (uj ) for all 1 ≤ i, j ≤ n.
Step 1: We produce a convex subcomplex of X(Γ, S) from G(Γ0 ).
The left action G(Γ) y G(Γ) induces G(Γ0 ) y G(Γ) and G(Γ0 ) y X(Γ, S).
By choosing a standard generating set S 0 of G(Γ0 ), we have left action G(Γ0 ) y
X(Γ0 , S 0 ). For h ∈ G(Γ0 ), we use φh , φ̄h , sh and s̄h to denote the action of h on
G(Γ), G(Γ0 ), P(Γ) and P(Γ0 ) respectively. Pick a G(Γ0 )-equivariant quasi-isometry
q : X(Γ, S) → X(Γ0 , S 0 ) such that q|G(Γ0 ) = Id. By Theorem 5.3 and Lemma 4.12, q
induces surjective G(Γ0 )-equivariant maps φ : G(Γ) → G(Γ0 ) and s : P(Γ) → P(Γ0 ).
Note that φ depends on the choice of generating set S and S 0 , and this flexibility
comes from the automorphism groups of G(Γ) and G(Γ0 ).
The key of step 1 is to choose a “nice” standard generating set S 0 of G(Γ0 ) such
that φ behaves like ϕ in Theorem 5.12.
Lemma 6.9. By choosing a possibly different standard generating set S 0 for G(Γ0 ),
we can assume the map φ satisfies φ(id) = id, where id denotes the identity element
in the corresponding group.
Proof. Assume φ(id) = a 6= id, we claim if we change the generating set from S 0
to aS 0 a−1 , then the resulting φ will satisfy our requirement. By the construction
of φ, it suffices to show for any maximal S 0 -flat F10 such that a ∈ F10 , there exists
a maximal aS 0 a−1 -flat F20 such that id ∈ F20 and dH (F10 , F20 ) < ∞. Let us assume
F10 = {ag k }k∈Z for some g ∈ S 0 . Then F20 = {(aga−1 )k }k∈Z would satisfy the
required condition. We can prove the general case in a similar way.
Pick a standard geodesic l ⊂ X(Γ, S), we want to flip the order of points of l
in a G(Γ0 )-equivariant way such that (5.13) is true. We choose an order preserving
identification of v(l) and Z. Let d = |φ−1 (φ(p)) ∩ v(l)| where p is a vertex in v(l).
Let Stab(v(l)) be the stabilizer of v(l) under the action G(Γ0 ) y G(Γ). By the
second paragraph of the proof of Theorem 5.12, d does not depend on the choice
of p in v(l), and Stab(v(l)) acts on v(l) in the same way as dZ acts on Z (recall
that φ is G(Γ0 )-equivariant and the action of G(Γ0 ) on G(Γ) is induced from the
left action of G(Γ) on itself).
We will write χ(l) = d. If ¯l and l are parallel, then χ(l) = χ(¯l). Thus χ : P(Γ) →
Z is well-defined. Since χ(l) only depends on how Stab(v(l)) acts on v(l), χ does
not depend on the standard generating set S 0 . However, for any choice of S 0 , χ
descends to χ : S 0 → Z by the G(Γ0 )-equivariance of φ.
Let φ(0) = a. Then Stab(v(l)) is generated by aha−1 for some h ∈ S 0 . By the
same reasoning as Lemma 6.9, we can assume a = id. Let S 0 = {hλ }λ∈Λ . For each
hλ ∈ S 0 , we associated an integer nλ as follows. If hλ h = hhλ , we set nλ = 0. Now
we consider the case hλ h 6= hhλ . Let lλ0 ⊂ X(Γ0 , S 0 ) be the standard geodesic that
contains all powers of hλ , and let bλ = π∆(l) ◦s−1 (∆(lλ0 )) (π∆(l) is the map in Lemma
6.2). Then nλ is defined to be the unique integer such that bλ + nλ d ∈ [0, d − 1]
(recall that d = χ(l)). Define f : S 0 → G(Γ0 ) by sending hλ to hnλ hλ h−nλ , then
f extends to an automorphism of G(Γ0 ) and S 00 = {f (hλ )}λ∈Λ is also a standard
generating set. Indeed, if ∆(lλ0 1 ) and ∆(lλ0 2 ) stay in the same connected component
46
JINGYIN HUANG
of P(Γ0 ) \ St(∆(l0 )), then bλ1 = bλ2 by Lemma 6.2, hence nλ1 = nλ2 . It follows
that f can be realized as a composition of partial conjugations.
Lemma 6.10. We replace S 0 by S 00 in the definition of φ and denote the new map
by φ1 . Then φ1 |v(l1 ) satisfies (5.13) for any standard geodesic l1 ⊂ X(Γ, S) with
∆(l1 ) ∈ {sh (∆(l))}h∈G(Γ0 ) .
Recall that for any h ∈ G(Γ0 ), we use sh to denote the action of h on P(Γ).
Proof. It suffices to show φ1 |v(l) satisfies (5.13), and the rest follows from the G(Γ0 )equivariance of φ1 . To show this, we only need to prove φ1 (i) = id for any i ∈
[0, d − 1]. Let Λ, bλ and nλ be as above.
We pick i ∈ [0, d − 1]. Then there exists λ ∈ Λ such that bλ + nλ d = i. By
Lemma 6.8, φ(bλ ) = id, hence φ(i) = hnλ . Let li be a standard geodesic such that
bλ ∈ li and d(∆(li ), ∆(l)) ≥ 2. Then there exists hλ0 ∈ S 0 with bλ0 = bλ such that
φ(v(li )) = {hkλ0 }k∈Z . Then (φh )nλ (v(li )) is an S-geodesic passing through i, and
(φ ◦ (φh )nλ )(v(li )) = ((φ̄h )nλ ◦ φ)(v(li )) = {hnλ hkλ0 }k∈Z . Note that
(6.11)
dH ({hnλ hkλ0 }k∈Z , {(f (hλ0 ))k }k∈Z ) < ∞
Now we look at the new map φ1 . Note that φ1 (0) = id is still true. Moreover,
(6.11) and Lemma 6.8 imply φ1 (i) = id. Thus the lemma follows.
The next lemma basically says the above change of basis process does not affect
other geodesics in an essential way.
Lemma 6.12. Let r be a standard geodesic in X(Γ, S) which satisfies the condition
that ∆(r) ∈
/ {sh (∆(l))}h∈G(Γ0 ) . Pick two different vertices x1 , x2 ∈ r. If φ(x) =
φ(y), then φ1 (x) = φ1 (y).
Proof. For i = 1, 2, let ri ⊂ X(Γ, S) be a standard geodesic containing xi such
that d(∆(ri ), ∆(r)) ≥ 2 for i = 1, 2. Let r0 (resp. r00 ) be an S 0 -geodesic (resp. S 00 geodesic) such that φ(v(r)) = v(r0 ) (resp. φ1 (v(r)) = v(r00 )). Let α = φ(x) = φ(y).
Then there exist elements hλ , hλ1 and hλ2 in S 0 such that φ(v(ri )) = {αhkλi }k∈Z for
i = 1, 2, and r0 = {αhkλ }k∈Z . Note that
(6.13)
h 6= hλ , hλ1 6= hλ , hλ2 6= hλ
Recall that h is the generator of Stab(v(l)). The first inequality of (6.13) follows
from ∆(r) ∈
/ {sh (∆(l))}h∈G(Γ0 ) .
It suffices to show there exist S 00 -geodesics r100 and r200 such that
(6.14)
dH (φ(v(ri )), ri00 ) < ∞
for i = 1, 2 and
(6.15)
π∆(r00 ) (∆(r100 )) = π∆(r00 ) (∆(r200 )),
then φ1 (x) = φ1 (y) follows from Lemma 6.8. Define ri00 = {αh−nλi (f (hλi ))k }k∈Z ,
then (6.14) is immediate. Note that for any a ∈ r10 and b ∈ r20 , we have
b = a · (f (hλ1 ))k1 · hnλ1 −nλ2 · (f (hλ2 ))k2
for some k1 , k2 ∈ Z, then (6.15) follows from (6.13) and the definition of π∆(r00 ) .
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
47
Similarly, we can prove that if we change φ with respect to the conjugation
S 0 → aS 0 a−1 , then Lemma 6.12 is still true with r being an arbitrary standard
geodesic.
By Lemma 6.10 and Lemma 6.12, we can apply the above change-of-basis procedure for finitely many times to find appropriate standard generating set S 0 of G(Γ0 )
such that the corresponding map φ satisfies (5.13) when restricted to any standard
geodesic in X(Γ, S). By the proof of Theorem 5.12, we can extend φ to a cubical
map φ : X(Γ, S) → X(Γ0 , S 0 ) such that combinatorial geodesics in C(Γ, S) are
mapped to combinatorial geodesics in C(Γ0 , S 0 ). Thus φ−1 (id) is a combinatorially
convex subcomplex. The subcomplex φ−1 (id) is also compact since φ−1 (id) contains
finitely many vertices. Recall that combinatorial convexity in `1 -metric and convexity in CAT (0) metric are the same for subcomplexes of CAT (0) cube complexes
([Hag07]), so we have constructed a compact convex subcomplex φ−1 (id) ⊂ X(Γ, S)
from a given finite index RAAG subgroup G(Γ0 ) ≤ G(Γ).
Step 2: We show φ−1 (id) can be assumed to be an element in CN (Γ, S).
For K ⊂ G(Γ), denote the union of all standard geodesics in X(Γ, S) that have
non-trivial intersection with K by K ∗ . K is S-convex if and only if K is the vertex
set of some convex subcomplex in X(Γ, S). Now we return to φ. By step 1, we can
assume φ(id) = id, and φ−1 (y) is S-convex for any y ∈ G(Γ0 ).
Step 2.1. Let {li }qi=1 be the collection of standard geodesics passing through id
and Λ1 = {id}. Let I : G(Γ) → `1 (v(P(Γ, S))) and I∆(l) : G(Γ) → Z∆(l) be the
map defined in Section 6.1. Since v(li ) and v(lj ) are in different G(Γ0 )-orbits for
i 6= j, by Lemma 6.10 and Lemma 6.12, we can apply the change-of-basis procedure
in step 1 to find a standard generating set S 0 for G(Γ0 ) such that for each 1 ≤ i ≤ q,
(6.16)
−1
I∆(l
([0, χ(li ) − 1]) ∩ v(li ) ⊂ φ−1 (id).
i)
Step 2.2. Let Λ2 = Λ∗1 ∩ φ−1 (id). Pick a vertex x ∈ Λ2 \ Λ1 (if such x does not
exist, then our process terminate). Let l be a standard geodesic such that x ∈ l. If
l is parallel to some li in step 2.1, then (6.16) with li replaced by l is automatically
true without any modification on S 0 , because both I and φ respect the product
structure of Pli . If l is not parallel to any li , then I∆(l) (x) = 0. Moreover, ∆(l)
is not in the G(Γ0 )-orbits of ∆(li )’s, so we can modify S 0 as before such that both
−1
(6.16) and I∆(l)
([0, χ(l) − 1]) ∩ v(l) ⊂ φ−1 (id) are true. We deal with other standard
geodesics passing through x and other points in Λ2 \ Λ1 in a similar way.
Step 2.3. Let Λ3 = Λ∗2 ∩ φ−1 (id). For each vertex in Λ3 \ Λ2 , we repeat the
procedure in step 2.2. Then we can define Λ4 , Λ5 , · · · . Since |φ−1 (id)| is finite
and this number does not change after adjusting S 0 , our procedure must terminate
after finitely many steps. Since φ−1 (id) remains connected in each step, once the
procedure terminates, we must have already dealt with each point in φ−1 (id) and
each standard geodesic passing through each point in φ−1 (id). By construction, the
−1
resulting φ satisfies id ∈ φ−1 (id) and I∆(l)
([0, χ(l) − 1]) ∩ v(l) ⊂ φ−1 (id) for each
standard geodesic l which intersects φ−1 (id). Thus φ−1 (id) is non-negative.
Note that the sets Λi ’s actually do not depend on the map φ from step i−1. They
only depend on the map χ : v(P(Γ)) → Z. Thus non-negative subset φ−1 (id) ⊂
G(Γ) produced above depends only on S and the subgroup G(Γ0 ) ≤ G(Γ). Then
we have a well-defined map
ΞS : {Finite index RAAG subgroups of G(Γ)} → CN (Γ, S)
48
JINGYIN HUANG
Step 3: We show ΞS is an inverse to the map ΘS defined in Section 5.2.
First we prove ΘS ◦ ΞS = Id. Let K = ΞS (G(Γ0 )). Let S 0 be the corresponding
standard generating set for G(Γ0 ) and let φ : G(Γ) → G(Γ0 ) be the corresponding
map. We find a maximal collection of standard geodesics {ci }si=1 such that ci ∩K 6=
∅ for all i and ∆(ci ) 6= ∆(cj ) for any i 6= j. Let ni = χ(ci ) and let gi ∈ S be the
label of edges in ci . Suppose αi = πci (id) where πci : X(Γ, S) → ci is the CAT (0)
projection. Then it suffices to prove the following lemma.
Lemma 6.17. S 0 = {αi gini αi−1 }si=1 .
Proof. Pick h ∈ S 0 and let ch ⊂ X(Γ0 , S 0 ) be the standard geodesic containing id
and h. Then there exists a unique i such that φ(v(ci )) = ch . To see this, let c
be a standard geodesic in X(Γ, S) such that s(∆(c)) = ∆(ch ). Then φ(v(c)) and
ch are parallel and there exists u ∈ G(Γ0 ) which sends φ(v(c)) to v(ch ). Thus
φ ◦ φu (v(c)) = v(ch ) by (5.8), where φu is defined in the beginning of step 1. Note
that φu (v(c)) has nontrivial intersection with K. We choose ci to be the geodesic
parallel to φu (v(c)). Then φ(v(ci )) = v(ch ).
For any standard geodesic c0i parallel to ci , φ(c0i ) is parallel to ch , so h ∈
Stab(v(φ(c0i ))) = Stab(v(c0i )). It follows that φh stabilizes the parallel set Pci and
acts by translation along the ci -direction. Note that (I∆(ci ) ◦ φh )(x) = I∆(ci ) (x) +
χ(ci ) for any x ∈ v(Pci ), so h = φh (id) = αi gini αi−1 and the claim follows.
It remains to show ΞS ◦ ΘS = Id. This follows from the following result.
Lemma 6.18. Let Γ be an arbitrary finite simplicial graph. Pick a standard generating set S for G(Γ) and K ∈ CN (Γ, S). Let G(Γ0 ) = ΘS (K) and let S 0 be the
corresponding generating set. Suppose q : G(Γ) → G(Γ0 ) is a G(Γ0 )-equivariant
quasi-isometry such that q|G(Γ0 ) is the identity map. Then
(1) q induces a simplicial isomorphism q∗ : P(Γ, S) → P(Γ0 , S 0 ).
(2) q∗ induces a G(Γ0 )-equivariant retraction r : G(Γ) → G(Γ0 ) such that r
sends every S-flat to an S 0 -flat.
(3) r extends to a surjective cubical map r : X(Γ, S) → X(Γ0 , S 0 ) such that
r−1 (id) = K. In particular, the vertex set of K is the strict fundamental
domain for the left action G(Γ0 ) y G(Γ).
Proof. It suffices to prove the case when Γ does not admit a nontrivial join decomposition and Γ is not a point.
By the construction of ΘS , we know the q-image of any S-flat which intersects K
is Hausdorff close to an S 0 -flat which contains the identity. Moreover, if the S-flat
is maximal, then the corresponding S 0 -flat is unique. Since G(Γ0 ) · v(K) = G(Γ),
so the equivariance of q implies the q-image of every S-flat is Hausdorff close to an
S 0 -flat. Since q is a quasi-isometry, so images of parallel S-geodesics are Hausdorff
closed to each other. This induces q∗ : P(Γ, S) → P(Γ0 , S 0 ). q∗ is injective since q
is a quasi-isometry and q∗ is surjective by the G(Γ0 )-equivariance.
Pick x ∈ G(Γ), let {Fi }i∈I be the collection of maximal S-flats containing x. For
each i, let Fi0 be the unique maximal S 0 -flat such that dH (q(Fi ), Fi0 ) < ∞. Note
that ∩i∈I Fi = x by our assumption on Γ. So ∩i∈I Fi0 is either empty or one point.
Note that if x ∈ K, then ∩i∈I Fi0 = id. The equivariance of q∗ implies for every
x, ∩i∈I Fi0 is a point, which is defined to be r(x). It is clear that v(K) ⊂ r−1 (id),
but |G(Γ) : G(Γ0 )| ≤ |v(K)|, so v(K) = r−1 (id). It follows that v(K) is the strict
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
49
fundamental domain for the left action of G(Γ0 ), and r is a G(Γ0 )-equivariant map
which maps v(K) to id.
Note that r(id) = id. Then the G(Γ0 )-equivariance of r implies r(g) = g for
any g ∈ G(Γ0 ) ⊂ G(Γ). Thus r is a retraction. Similarly, by using the G(Γ0 )equivariance of r, we deduce that r sends every S-flat that intersects K to an
S 0 -flat passing through the identity element of G(Γ0 ). Thus r sends every S-flat to
an S 0 -flat by the equivariance of r. It is easy to see r extends to a cubical map
r : X(Γ, S) → X(Γ0 , S 0 ) such that r−1 (id) = K.
Remark 6.19. We can generalize some of the results in Lemma 6.18 to infinite
convex subcomplexes of X(Γ, S). A convex subcomplex K ⊂ X(Γ, S) is admissible
if for any standard geodesic l, the CAT (0) projection πl (K) is either a finite interval
or the whole l (a ray is not allowed). Let {lλ }λ∈Λ be a maximal collection of
standard geodesics such that (1) lλ ∩ K 6= ∅; (2) lλ and lλ0 are not parallel for
λ 6= λ0 ; (3) πlλ (K) is a finite interval. For each lλ , let αλ ∈ G(Γ) be an element
which translates along lλ with translation length = 1 + length(πlλ (K)). Let GK
be the subgroup generated by S 0 = {αλ }λ∈Λ . If K is admissible, we can prove
GK · v(K) = G(Γ) as before. Moreover, for any finite subset S10 ⊂ S 0 , the subgroup
G1 generated by S10 is a right-angled Artin group, and G1 ,→ GK is an isometric
embedding with respect to the word metric. We can define S 0 -flat as before and
view each vertex of GK as a 0-dimensional S 0 -flat.
Now we show v(K) is a strict fundamental domain for the action GK y G(Γ).
It suffices to show α(K) ∩ K = ∅ for each nontrivial α ∈ GK . We can assume there
is a right-angled Artin group G1 such that α ∈ G1 ⊂ GK . Let α = w1 w2 · · · wn be
a canonical form of α (see [Cha07, Section 2.3]). Then
(1) Each wi belongs to an abelian standard subgroup of G1 .
ki,n
k
k
(2) For each i, let wi = ri,1i,1 ri,2i,2 · · · ri,ni i (ri,j ∈ S 0 ). Then for each ri+1,j
(1 ≤ j ≤ ni+1 ), there exists ri,j 0 which does not commute with ri+1,j .
We associate each generator ri,j with a subset Xi,j ⊂ X(Γ, S) as in the proof
of Lemma 6.4, and claim there exists j with 1 ≤ j ≤ n1 such that α(K) ⊂
X1,j , then α(K) ∩ K = ∅ follows. We prove by induction on n and assume
w2 w3 · · · wn (K) ⊂ X2,j 0 . By (2), there is r1,j such that r1,j and r2,j 0 does not
k1,h
k1,j
commute, so r1,j
(X2,j 0 ) ⊂ X1,j . Moreover, by (1), r1,h
(X1,j ) = X1,j for h 6= j, so
α(K) ⊂ w1 (X2,j 0 ) ⊂ X1,j .
Now we can define a GK -equivariant map r : G(Γ) → GK by sending v(K)
to the identity of GK . We prove as before that r maps S-flats to (possibly lower
dimensional or 0-dimensional) S 0 -flats, thus r is 1-Lipschitz with respect to the
word metric. Let i : GK ,→ G(Γ) be the inclusion. Then by the equivariance of r,
r ◦ i is a left translation of GK . In particular, if K contains the identity, then r is
a retraction. It follows that if S 0 is finite, then i is a quasi-isometric embedding.
Note that a related construction in the case of right-angled Coxeter groups has
been discussed in [Hag08]. By taking larger and larger convex compact subcomplexes of X(Γ, S), we know G(Γ) is residually finite. Moreover, pick β ∈ Stab(K) ⊂
G(Γ), by definition of S 0 , S 0 = βS 0 β −1 , so Stab(K) normalize GK . Now we have
obtained a direct proof of the fact that every word-quasi-convex subgroup of a finite
generated right-angled Artin group is separable (Theorem F of [Hag08]) by using
the above discussion together with the outline in Section 1.5 of [Hag08].
The following result follows readily from the above discussion.
50
JINGYIN HUANG
Theorem 6.20. Let G(Γ) be a RAAG with Out(G(Γ)) finite. We pick a standard
generating set S for G(Γ). Then there is a 1-1 correspondence between non-negative
convex compact subcomplexes of X(Γ, S) that contain the identity and finite index
RAAG subgroups of G(Γ). In particular, these subgroups are generated by conjugates of powers of elements in S.
In particular, Theorem 1.3 in the introduction follows from Theorem 6.20.
Remark 6.21. If we drop the finite automorphism group assumption in the above
theorem, then there exist a RAAG G(Γ1 ) and its finite index RAAG subgroup
G(Γ2 ) such that G(Γ2 ) is not isomorphic to any special subgroup of G(Γ1 ). To see
this, let G(Γ1 ) be a right-angled Artin group such that Out(G(Γ1 )) is transvection
free. Then Lemma 6.18 and Theorem 3.27 imply each special subgroup of G(Γ1 )
does not admit non-trivial transvection in its outer automorphism group. Let Γ1
and Γ2 be the graphs in Example 3.29. Then G(Γ2 ) is a right-angled Artin subgroup
of G(Γ1 ) and there are non-trivial transvections in Out(G(Γ2 )). Thus G(Γ2 ) is not
isometric to any special subgroup of G(Γ1 ).
Remark 6.22. Pick G(Γ) such that Out(G(Γ)) is finite, then Theorem 6.20 can
be used to show certain subgroup of G(Γ) is not a RAAG. For example, let {vi }ki=1
be a subset of some standard generating set for G(Γ). We define homomorphism
h : G(Γ) → Z/2 by sending each vi to the non-trivial element in Z/2 and killing all
other generators. Then ker(h) is a RAAG if and only if k = 1. One can compare
this example to Example 3.29.
Remark 6.23. It is shown in [KK13b, Theorem 2] that if F (Γ0 ) embeds into P(Γ)
as a full subcomplex, then there exists a monomorphism G(Γ0 ) ,→ G(Γ). This result
can be recovered by the our previous discussion as follows. Let Γ be an arbitrary
finite simplicial graph. Let S be a standard generating set for G(Γ). For any vertex
w ∈ P(Γ), let αw ∈ G(Γ) be a conjugate of some element in S such that αw (l) = l
for every standard geodesic l ⊂ X(Γ, S) with ∆(l) = w.
Suppose M ⊂ P(Γ, S) is a compact full subcomplex and Γ0 is the 1-skeleton of
M . Denote the vertex set of M by {wi }ni=1 and let li be a standard geodesic with
∆(li ) = wi . We identify each li in an orientation-preserving way with R such that
0 ∈ R is identified with πli (id) ⊂ li , whereπli is the CAT (0) projection to li and id
is the identity element of G(Γ).
For 1 ≤ i ≤ n, define Λi = {1 ≤ j ≤ n | d(wi , wj ) ≥ 2}. For each i, we define
a pair of integers ai and ki as follows. If Λi 6= ∅, then let [ai , ai + ki ] ⊂ R be the
minimal interval such that ∪j∈Λi πli (lj ) ⊂ [ai , ai + ki ] (recall that li is identified
with R). If Λi = ∅, then we pick an arbitrary ai and set ki = 0. Define Xi =
πc−1
((−∞, ai − 1/2]) ∪ πc−1
([ai + ki + 1/2, ∞)). Then by construction, Xi ∩ Xj = ∅
i
i
for i, j satisfying d(wi , wj ) ≥ 2. Using the argument in Section 6.1, we can show
ki +1 n
the subgroup generated by S 0 = {αw
}i=1 is a RAAG with defining graph Γ0 .
i
At this point it is natural to ask the following question.
Question 6.24. Let S be a standard generating set of G(Γ) and let S 0 be a finite
collection of elements of form αrk α−1 , where r ∈ S, k ∈ Z and α ∈ G(Γ). Suppose
G is the subgroup generated by S 0 . Is G a right-angled Artin group?
6.3. Generalized star extension. Our goal in this subsection is to find an algorithm to determine whether G(Γ) and G(Γ0 ) are quasi-isometric or not, given
Out(G(Γ)) is finite.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
51
For convex subcomplex E ⊂ X(Γ), we denote the full subcomplex in P(Γ, S)
spanned by {∆(lλ )}λ∈Λ by Ê, where {lλ }λ∈Λ is the collection of standard geodesics
in X(Γ) with lλ ∩ E 6= ∅.
Now we describe a process to construct a graph Γ0 from Γ such that G(Γ0 ) is
isomorphic to a special subgroup of G(Γ). Let Γ1 = Γ and let K1 be one point. We
will construct a pair (Γi , Ki ) inductively such that
(1) Ki is a compact CAT (0) cube complex and there is a cubical embedding
f : Ki → X(Γ) such that f (Ki ) is convex in X(Γ).
(2) Γi is a finite simplicial graph and there is a simplicial isomorphism g :
F (Γi ) → f\
(Ki ).
Note that these assumptions are true for i = 1.
We associate each edge e ⊂ Ki with a vertex in Γi , denoted by ve , as follows. Let
le be the standard geodesic in X(Γ) that contains f (e). We define ve := g −1 (∆(le )).
Each vertex x ∈ Ki can be associated with a full subcomplex Φ(x) ⊂ F (Γi ) defined
by Φ(x) = g −1 (x̂).
To define (Γi+1 , Ki+1 ), pick a vertex v ∈ Γi and let {xj }m
j=1 be the collection
of vertices in Ki such that v ∈ Φ(xj ). Then {f (xj )}kj=1 are exactly the vertices
in Pl ∩ f (Ki ), where l is a standard geodesic such that ∆(l) = g(v). Let L be the
convex hull of {xj }m
j=1 in Ki . Then e ⊂ L for any edge e ⊂ Ki with ve = v.
Since f (L) = Pl ∩ f (Ki ), the natural product decomposition Pl ∼
= l × l⊥ induces
a product decomposition of L = h × [0, a]. Note that it is possible that a = 0, and
a > 0 if and only if there exists an edge e ⊂ Ki with ve = v. When a > 0, h is
isomorphic to the hyperplane dual to e, and for any edge e0 ∈ Ki with ve0 = v, the
projection of e0 to the interval factor [0, a] is an edge.
Let Li = h×{a} ⊂ L and let Mi = ∪x∈Li Φ(x) (x is a vertex). We define F (Γi+1 )
to be the simplicial complex obtained by gluing F (Γi ) and Mi along St(v, Mi ) (see
Section 2.1 for the notation), and define Ki+1 to be the CAT (0) cube complex
obtained by gluing Ki and Li × [0, 1] along Li . One readily verifies that one can
extend f to a cubical embedding f 0 : Ki+1 → X(Γ) such that f 0 (Ki+1 ) is convex.
This also induces an isomorphism g 0 : F (Γi+1 ) → K̂i+1 which is an extension of g.
By construction, each G(Γi ) is isomorphic to a special subgroup of G(Γ), moreover, the associated convex subcomplex of this special subgroup is Ki . Also note
that the above induction process actually does not depend on knowing what X(Γ)
is. Thus it also provides a way to construct convex subcomplexes of X(Γ) by hand.
The above process of obtaining (Γi+1 , Ki+1 ) from (Γi , Ki ) is called a generalized
star extension (GSE) at v. Note that the following are equivalent.
(1) Γi ( Γi+1 .
(2) Pl ( X(Γ), where l is the standard geodesic in X(Γ) such that ∆(l) = g(v).
(3) St(π(g(v))) ( F (Γ), where π : P(Γ) → F (Γ) is the natural label-preserving
projection defined in (4.3).
A GSE is nontrivial if Γi ( Γi+1 . If Γ is not a clique, then at each stage, there
exists a vertex v ∈ Γi such that the GSE at v is nontrivial.
Lemma 6.25. Suppose G(Γ0 ) is isomorphic to a special subgroup of G(Γ). Then
we can construct Γ0 from Γ by using finitely many GSE’s.
52
JINGYIN HUANG
Proof. Let ΘS and CN (Γ, S) be the objects defined in Section 6.1. Suppose G(Γ0 )
is isomorphic to ΘS (K) for K ∈ CN (Γ, S). We define a sequence of convex subcomplexes in K by induction. Let K1 be the identity element in G(Γ). Suppose Ki
is already defined. If Ki = K, then the induction terminates. If Ki ( K, pick an
edge ei ⊂ K such that ei ∩ Ki is a vertex and let Ki+1 be the convex hull of Ki ∪ ei .
Let {Ki }si=1 be the resulting collection of convex subcomplexes. An alternative
way of describing Ki+1 is the following. If hi is the hyperplane in K dual to ei
and Ni is the carrier of hi in K, then hi ∩ Ki = ∅ by the convexity of Ki . Thus
Ki ∩ Ni is disjoint from hi . Hence there is a copy of (Ki ∩ Ni ) × [0, 1] inside Ni ,
which is denoted by Mi . Then Ki+1 = Ki ∪ Mi . Now one readily verifies that one
can obtain (K̂i+1 , Ki+1 ) from (K̂i , Ki ) by a GSE.
The above construction gives rise to an algorithm to detect whether G(Γ0 ) is
isomorphic to a special subgroup of G(Γ). If there are n vertices in Γ0 , then Γ0
can be obtained from Γ by at most n nontrivial GSE’s. So we can start with Γ,
enumerate all possible n-step nontrivial GSE’s from Γ, and compare each resulting
graph with Γ0 . By Theorem 5.15 and Theorem 6.20, we have the following result.
Theorem 6.26. If Out(G(Γ)) is finite, then G(Γ0 ) is quasi-isometric to G(Γ) if
and only if Γ0 can be obtained from Γ by finitely many GSE’s. In particular, there
is an algorithm to determine whether G(Γ0 ) and G(Γ) are quasi-isometric.
Note that a GSE gives rise to a pair (Γi , Ki ). If one does not care about the
associated convex subcomplex Ki , then there is a simpler description of GSE when
Out(G(Γ)) is finite. Suppose we have already obtained F (Γi ) together with a finite
collection of full subcomplexes {Gλ }λ∈Λi such that
(1) {Gλ }λ∈Λi is a covering of F (Γi ).
(2) Each Gλ is isomorphic to F (Γ).
When i = 1, we pick the trivial cover of F (Γ) by itself. To construct Γi+1 , pick
vertex v ∈ F (Γi ), let Λv = {λ ∈ Λi | v ∈ Gλ } and let Γv = ∪λ∈Λv Gλ . Suppose
{Cj }m
j=1 is the collection of connected components of Γv \ St(v, Γv ) and suppose
Cj0 = Cj ∪ St(v, Γv ). Then F (Γi+1 ) is defined by gluing C10 and F (Γi ) along
St(v, Γv ), and Γi+1 is the 1-skeleton of F (Γi+1 ).
Lemma 6.27. Suppose Out(G(Γ)) is finite. Then the above simplified process is
consistent with GSE.
Proof. We assume inductively that there is a CAT (0) cube complex Ki such that
the two induction assumptions for GSE are satisfied, moreover, {Gλ }λ∈Λi coincides
with {Φ(x)}x∈Ki (x is a vertex). Let L = h × [0, a] be as before and let Lj = h ×
{j} ⊂ L for each integer j ∈ [0, a]. It suffices to show there is a 1-1 correspondence
0
between {Lj }aj=0 and {Cj0 }m
j=1 such that for each j, there exists a unique j with
f\
(Lj ) = g(C 0 0 ). Pick adjacent vertices x1 , x2 ∈ f (Lj ) and let w̄ ∈ Γ be the label
j
of edge x1 x2 . Suppose v̄ = π(g(v)). Then d(w̄, v̄) = 1. Since Out(G(Γ)) is finite,
the orthogonal complement of w̄ satisfies w̄⊥ * St(v̄). Then there is a vertex
ū ∈ w̄⊥ such that d(ū, v̄) = 2. The lifts of ū in x̂1 and x̂2 are the same point, so
(x̂1 ∩ x̂2 ) \ St(g(v)) contains a vertex. Since F (Γ) does not have separating closed
stars, x̂i \ St(g(v)) is connected for i = 1, 2. Thus (x̂1 ∩ x̂2 ) \ St(g(v)) is connected.
It follows that f\
(Lj ) \ St(g(v)) is connected. Moreover, Lemma 4.10 implies that
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
53
f\
(Lj1 )\St(g(v)) and f\
(Lj2 )\St(g(v)) are in different components of P(Γ)\St(g(v))
when j1 6= j2 , so there exists a unique j 0 such that f\
(Lj ) = g(Cj0 0 ).
References
[ABD+ 13]
Aaron Abrams, Noel Brady, Pallavi Dani, Moon Duchin, and Robert Young. Pushing
fillings in right-angled Artin groups. Journal of the London Mathematical Society,
87(3):663–688, 2013.
[BB97]
Mladen Bestvina and Noel Brady. Morse theory and finiteness properties of groups.
Inventiones mathematicae, 129(3):445–470, 1997.
[BC12]
Jason Behrstock and Ruth Charney. Divergence and quasimorphisms of right-angled
Artin groups. Mathematische Annalen, 352(2):339–356, 2012.
[BH99]
Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature, volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
[BJN10]
Jason A. Behrstock, Tadeusz Januszkiewicz, and Walter D. Neumann. Quasi-isometric
classification of some high dimensional right-angled Artin groups. Groups Geom.
Dyn., 4(4):681–692, 2010.
[BKK14]
Hyungryul Baik, Sang-hyun Kim, and Thomas Koberda. Right-angled Artin subgroups of the C ∞ interval diffeomorphism group. arXiv preprint arXiv:1404.5559,
2014.
[BKMM12] Jason Behrstock, Bruce Kleiner, Yair Minsky, and Lee Mosher. Geometry and rigidity
of mapping class groups. Geom. Topol., 16(2):781–888, 2012.
[BKS08a] Mladen Bestvina, Bruce Kleiner, and Michah Sageev. The asymptotic geometry of
right-angled Artin groups. I. Geom. Topol., 12(3):1653–1699, 2008.
[BKS08b] Mladen Bestvina, Bruce Kleiner, and Michah Sageev. Quasiflats in CAT (0) complexes. arXiv preprint arXiv:0804.2619, 2008.
[BM01]
Noel Brady and John Meier. Connectivity at infinity for right angled Artin groups.
Transactions of the American Mathematical Society, 353(1):117–132, 2001.
[BN08]
Jason A Behrstock and Walter D Neumann. Quasi-isometric classification of graph
manifold groups. Duke Mathematical Journal, 141(2):217–240, 2008.
[BRS07]
Noel Brady, Tim Riley, and Hamish Short. The geometry of the word problem for
finitely generated groups. Springer Science & Business Media, 2007.
[CD95]
Ruth Charney and Michael W. Davis. Finite K(π, 1)s for Artin groups. In Prospects
in topology (Princeton, NJ, 1994), volume 138 of Ann. of Math. Stud., pages 110–124.
Princeton Univ. Press, Princeton, NJ, 1995.
[CF12]
Ruth Charney and Michael Farber. Random groups arising as graph products. Algebr.
Geom. Topol, 12(2):979–995, 2012.
[Cha07]
Ruth Charney. An introduction to right-angled Artin groups. Geometriae Dedicata,
125(1):141–158, 2007.
[CK00]
Christopher B Croke and Bruce Kleiner. Spaces with nonpositive curvature and their
ideal boundaries. Topology, 39(3):549–556, 2000.
[CLM10]
Matt Clay, Christopher J Leininger, and Johanna Mangahas. The geometry of right
angled Artin subgroups of mapping class groups. arXiv preprint arXiv:1007.1129,
2010.
[CM09]
Pierre-Emmanuel Caprace and Nicolas Monod. Isometry groups of non-positively
curved spaces: structure theory. Journal of Topology, page jtp026, 2009.
[CS11]
Pierre-Emmanuel Caprace and Michah Sageev. Rank rigidity for CAT (0) cube complexes. Geometric and functional analysis, 21(4):851–891, 2011.
[Dav83]
Michael W Davis. Groups generated by reflections and aspherical manifolds not covered by Euclidean space. Annals of Mathematics, pages 293–324, 1983.
[Day12]
Matthew B. Day. Finiteness of outer automorphism groups of random right-angled
Artin groups. Algebr. Geom. Topol., 12(3):1553–1583, 2012.
[Dro87]
Carl Droms. Isomorphisms of graph groups. Proceedings of the American Mathematical Society, 100(3):407–408, 1987.
[Dun85]
Martin J Dunwoody. The accessibility of finitely presented groups. Inventiones mathematicae, 81(3):449–457, 1985.
54
[EF97]
[Esk98]
[FH07]
[Ger98]
[Gro87]
[Hag07]
[Hag08]
[Hag14]
[Ham05]
[HK14]
[HK16]
[Hua14]
[Hua16]
[HW08]
[KK13a]
[KK13b]
[KK14]
[KKL98]
[KL97]
[KL01]
[Kob12a]
[Kob12b]
[Lau95]
[Mos73]
[MSW04]
[PW02]
[Sag95]
[SDS89]
JINGYIN HUANG
Alex Eskin and Benson Farb. Quasi-flats and rigidity in higher rank symmetric spaces.
Journal of the American Mathematical Society, 10(3):653–692, 1997.
Alex Eskin. Quasi-isometric rigidity of nonuniform lattices in higher rank symmetric
spaces. Journal of the American Mathematical Society, 11(2):321–361, 1998.
Benson Farb and Michael Handel. Commensurations of Out(Fn ). Publications mathématiques, 105(1):1–48, 2007.
V Gerasimov. Fixed-point-free actions on cubings. Siberian Advances in Mathematics,
8(3):36–58, 1998.
M. Gromov. Hyperbolic groups. In Essays in group theory, volume 8 of Math. Sci.
Res. Inst. Publ., pages 75–263. Springer, New York, 1987.
Frédéric Haglund. Isometries of CAT (0) cube complexes are semi-simple. arXiv
preprint arXiv:0705.3386, 2007.
Frédéric Haglund. Finite index subgroups of graph products. Geometriae Dedicata,
135(1):167–209, 2008.
Mark F Hagen. Weak hyperbolicity of cube complexes and quasi-arboreal groups.
Journal of Topology, 7(2):385–418, 2014.
Ursula Hamenstaedt. Geometry of the mapping class groups III: Quasi-isometric rigidity. arXiv preprint math/0512429, 2005.
Mark F Hagen and Thomas Koberda. Personal communication. July 2014.
Jingyin Huang and Bruce Kleiner. Groups quasi-isometric to raag’s. arXiv preprint
arXiv:1601.00946, 2016.
Jingyin Huang. Top dimensional quasiflats in cat(0) cube complexes. arXiv preprint
arXiv:1410.8195, 2014.
Jingyin Huang. Quasi-isometry classification of right-angled artin groups ii: several
infinite out cases. arXiv preprint arXiv:1603.02372, 2016.
Frédéric Haglund and Daniel T. Wise. Special cube complexes. Geom. Funct. Anal.,
17(5):1551–1620, 2008.
Sang-hyun Kim and Thomas Koberda. Anti-trees and right-angled Artin subgroups
of braid groups. arXiv preprint arXiv:1312.6465, 2013.
Sang-hyun Kim and Thomas Koberda. Embedability between right-angled Artin
groups. Geometry & Topology, 17(1):493–530, 2013.
Sang-Hyun Kim and Thomas Koberda. The geometry of the curve graph of a rightangled Artin group. International Journal of Algebra and Computation, 24(02):121–
169, 2014.
Michael Kapovich, Bruce Kleiner, and Bernhard Leeb. Quasi-isometries and the de
Rham decomposition. Topology, 37(6):1193–1211, 1998.
Bruce Kleiner and Bernhard Leeb. Rigidity of quasi-isometries for symmetric spaces
and Euclidean buildings. Comptes Rendus de l’Académie des Sciences-Series IMathematics, 324(6):639–643, 1997.
Bruce Kleiner and Bernhard Leeb. Groups quasi-isometric to symmetric spaces. Communications in analysis and geometry, 9(2):239–260, 2001.
Thomas Koberda. Ping-pong lemmas with applications to geometry and topology.
Geometry, Topology and Dynamics of Character Varieties, 23:139, 2012.
Thomas Koberda. Right-angled Artin groups and a generalized isomorphism problem
for finitely generated subgroups of mapping class groups. Geometric and Functional
Analysis, 22(6):1541–1590, 2012.
Michael R Laurence. A generating set for the automorphism group of a graph group.
Journal of the London Mathematical Society, 52(2):318–334, 1995.
George D Mostow. Strong rigidity of locally symmetric spaces, volume 78. Princeton
University Press, 1973.
Lee Mosher, Michah Sageev, and Kevin Whyte. Quasi-actions on trees II: Finite depth
Bass-Serre trees. arXiv preprint math/0405237, 2004.
Panos Papasoglu and Kevin Whyte. Quasi-isometries between groups with infinitely
many ends. Commentarii Mathematici Helvetici, 77(1):133–144, 2002.
Michah Sageev. Ends of group pairs and non-positively curved cube complexes. Proc.
London Math. Soc. (3), 71(3):585–617, 1995.
Herman Servatius, Carl Droms, and Brigitte Servatius. Surface subgroups of graph
groups. Proceedings of the American Mathematical Society, 106(3):573–578, 1989.
QUASI-ISOMETRIC CLASSIFICATION OF RIGHT-ANGLED ARTIN GROUPS I
[Ser89]
[Tay13]
[Wis09]
55
Herman Servatius. Automorphisms of graph groups. Journal of Algebra, 126(1):34–60,
1989.
Samuel J Taylor. Right-angled Artin groups and Out(Fn ) I: quasi-isometric embeddings. arXiv preprint arXiv:1303.6889, 2013.
Daniel T Wise. Research announcement: the structure of groups with a quasiconvex
hierarchy. Electron. Res. Announc. Math. Sci, 16:44–55, 2009.
Courant Institute of Mathematical Science, New York University, 251 Mercer
Street, New York, NY, USA.
E-mail address: [email protected]
| 4math.GR
|
(This is a shortened workshop version of the paper at https://arxiv.org/abs/1608.04428).
arXiv:1612.00817v1 [cs.LG] 2 Dec 2016
T ERPRE T: A Probabilistic Programming Language
for Program Induction
Alexander L. Gaunt1 , Marc Brockschmidt1 , Rishabh Singh1 ,
Nate Kushman1 , Pushmeet Kohli1 , Jonathan Taylor2∗, Daniel Tarlow1
1
2
Microsoft Research
perceptiveIO
{t-algaun, mabrocks, risin, nkushman, pkohli, dtartlow}@microsoft.com
[email protected]
Abstract
We study machine learning formulations of inductive program synthesis; that is,
given input-output examples, synthesize source code that maps inputs to corresponding outputs. Our key contribution is T ERPRE T, a domain-specific language
for expressing program synthesis problems. A T ERPRE T model is composed of
a specification of a program representation and an interpreter that describes how
programs map inputs to outputs. The inference task is to observe a set of inputoutput examples and infer the underlying program. From a T ERPRE T model we
automatically perform inference using four different back-ends: gradient descent
(thus each T ERPRE T model can be seen as defining a differentiable interpreter),
linear program (LP) relaxations for graphical models, discrete satisfiability solving,
and the S KETCH program synthesis system. T ERPRE T has two main benefits. First,
it enables rapid exploration of a range of domains, program representations, and
interpreter models. Second, it separates the model specification from the inference
algorithm, allowing proper comparisons between different approaches to inference.
We illustrate the value of T ERPRE T by developing several interpreter models and
performing an extensive empirical comparison between alternative inference algorithms on a variety of program models. To our knowledge, this is the first work
to compare gradient-based search over program space to traditional search-based
alternatives. Our key empirical finding is that constraint solvers dominate the gradient descent and LP-based formulations.
1
Introduction
Learning computer programs from input-output examples, or Inductive Program Synthesis (IPS), is
a fundamental problem in computer science, dating back at least to Summers [1977] and Biermann
[1978]. The field has produced many successes, with perhaps the most visible example being the
FlashFill system in Microsoft Excel [Gulwani, 2011, Gulwani et al., 2012].
There has also been significant recent interest in neural network-based models with components
that resemble computer programs [Giles et al., 1989, Joulin and Mikolov, 2015, Grefenstette et al.,
2015, Graves et al., 2014, Weston et al., 2014, Kaiser and Sutskever, 2016, Reed and de Freitas,
2016, Neelakantan et al., 2016, Kurach et al., 2015, Andrychowicz and Kurach, 2016]. These models
combine neural networks with external memory, external computational primitives, and/or built-in
structure that reflects a desired algorithmic structure in their execution. However, none produce
programs as output. Instead, the program is hidden inside “controllers” composed of neural networks
that decide which operations to perform, and the learned program can only be understood in terms
of the executions that it produces on specific inputs.
∗
Work done while author was at Microsoft Research.
1st Workshop on Neural Abstract Machines & Program Induction (NAMPI), @NIPS 2016, Barcelona, Spain.
(a)
(b)
(c)
Outputs
Outputs
Execution
Outputs
Observed
element
Execution
TERPRET interpreter model
Interpreter
TERPRET interpreter model
Inference
Source Code
Inputs
Params
Params
Inputs
Unobserved
element
Inputs
Figure 1: A high level view of the program synthesis task. (a) Forward execution of a traditional
interpreter. (b) Forward execution of a T ERPRE T model. (c) Inference in T ERPRE T model.
Technique name
Optimizer/Solver
Description
FMGD
(Forward marginals, gradient descent)
(I)LP
((Integer) linear programming)
SMT
TensorFlow
A gradient descent based approach which generalizes the approach used by Kurach et al. [2015].
Gurobi
A novel linear program relaxation that supports Gates [Minka
and Winn, 2009].
Z3
Translation of the problem into a first-order logical formula
with existential constraints.
Cast the T ERPRE T model as a partial program (the interpreter) containing holes (the source code) to be inferred from
a specification (input-output examples).
S KETCH
[Solar-Lezama, 2008]
S KETCH
Table 1: T ERPRE T back-end inference algorithms.
In this work we focus on models that represent programs as simple, natural source code [Hindle et al.,
2012], i.e., the kind of source code that people write. There are two main advantages to representing
programs as source code rather than weights of a neural network controller. First, source code is
interpretable; the resulting models can be inspected by a human and debugged and modified. Second,
programming languages are designed to make it easy to express the algorithms that people want to
write. By using these languages as our model representation, we inherit inductive biases that can lead
to strong generalization (i.e., generalizing successfully to test data that systematically differs from
training). Of course, natural source code is likely not the best representation in all cases. For example,
programming languages are not designed for writing programs that classify images, and there is no
reason to expect that natural source code would impart a favorable inductive bias in this case.
Optimization over program space is known to be a difficult problem. However, recent progress in
neural networks showing that it is possible to learn models with differentiable computer architectures,
along with the success of gradient descent-based optimization, raises the question of whether gradient
descent could be a powerful new technique for searching over program space.
These issues motivate the main questions in this work. We ask (a) whether new models can be
designed specifically to synthesize interpretable source code that may contain looping and branching structures, and (b) how searching over program space using gradient descent compares to the
combinatorial search methods from traditional IPS.
To address the first question we develop models inspired by intermediate representations used in
compilers like LLVM [Lattner and Adve, 2004] that can be trained by gradient descent. These models
interact with external storage, handle non-trivial control flow with explicit if statements and loops,
and, when appropriately discretized, a learned model can be expressed as interpretable source code.
We note two concurrent works, Adaptive Neural Compilation [Bunel et al., 2016] and Differentiable
Forth [Riedel et al., 2016], which implement similar models.
To address the second question, concerning the efficacy of gradient descent, we need a way of
specifying many IPS problems such that the gradient based approach can be compared to alternative
approaches in a like-for-like manner across a variety of domains. The alternatives originate from rich
histories of IPS in the programming languages and inference in discrete graphical models. To our
knowledge, no such comparison has previously been performed.
Both of the above benefit from being formulated in the context of T ERPRE T. T ERPRE T provides
a means for describing an execution model (e.g., a Turing Machine, an assembly language, etc.)
by defining a a program representation and an interpreter that maps inputs to outputs using the
2
T = 5
# ################################################
# Source code parametrisation
#
# ################################################
ruleTable = Param (2)[2 , 2]
1
2
3
4
5
tape[1] [0]
0
0
copy
1
6
# ################################################
# Interpreter model
#
# ################################################
tape = Var (2)[T]
initial_tape = Input (2)[2]
final_tape = Output (2)
tape [0]. set_to ( initial_tape [0])
tape [1]. set_to ( initial_tape [1])
7
8
9
10
11
12
13
14
15
for t in range (1, T − 1):
with tape[t] as x1:
with tape[t − 1] as x0:
tape[t + 1]. set_to ( ruleTable [x0 , x1 ])
16
17
18
19
20
0
ruleTable
[0,0]
1
tape[2]
[0,1]
[1,0]
0
[1,1]
0
0
1
1
1
0
0
copy
1
1
0
1
final_tape . set_to (tape[T − 1])
21
1
tape[T - 1]
Figure 2: Illustrative example of a T ERPRE T model and corresponding factor graph which describe
a toy automaton that updates a binary tape according to the previous two entries and a rule (refer
to long version for definition of graphical symbols).
parametrized program. The T ERPRE T description is independent of any particular inference algorithm. The IPS task is to infer the execution model parameters (the program) given an execution
model and pairs of inputs and outputs. An overview of the synthesis task appears in Fig. 1. To
perform inference, T ERPRE T is automatically compiled into an intermediate representation which
can be fed to a particular inference algorithm. Table 1 describes the inference algorithms currently
supported by T ERPRE T. Interpretable source code can be obtained directly from the inferred model
parameters. The driving design principle for T ERPRE T is to strike a subtle balance between the
breadth of expression needed to precisely capture a range of execution models, and the restriction of
expression needed to ensure that automatic compilation to a range of different back-ends is tractable.
2
T ERPRE T Language
Due to space, we give just a simple example of how T ERPRE T models are written. For a full grammar of the language and several longer examples, see the long version Gaunt et al. [2016b]. T ER PRE T obeys Python syntax, and we use the Python ast library to parse and compile T ERPRE T models.
The model is composed of Param variables that define the program, Var variables that represent the
intermediate state of computation, and Inputs and Outputs. Currently, all variables must be discrete
with constant sizes. T ERPRE T supports loops with ranges defined by constants2 , if statements, arrays,
array indexing, and user-defined functions that map a discrete set of inputs to a discrete output. In
total, the constraints imply that a model can be converted into a gated factor graph [Minka and Winn,
2009]. The details of how this works are in the longer version. A very simple T ERPRE T model
appears in Fig. 2. In this model, there is a binary tape that the program writes to. The program is
a rule table that specifies which state to write at the current position of a tape, conditional on the
pattern that appeared on the two previous tape locations. The first two tape positions are initialized
as Inputs, and the final tape position is the program output.
We have used T ERPRE T to build much more complicated models, including a Turing Machine,
boolean circuits, a Basic Block model that is similar to the intermediate representation used by
LLVM, and an assembly-like model.
2
These constants define maximum quantities, like the maximum number of timesteps that a program can
execute for, or the maximum number of instructions that can be in a program. Variable-length behavior can be
achieved by defining an absorbing end-state, by allowing a no-op instruction, etc.
3
Turing machine code
state read write
0
Head
0110_
Tape
1
move new state
0
??
??
??
1
??
??
??
_
??
??
??
0
??
??
??
1
??
??
??
…
*T
wire 1
??
wire 3
??
??
wire 4
…
??
Logic gate
wire 2
regout
??
Current Block
regin1 regin2
instr
(
=
??
??
>0
??
,
??
)
regcond
if
then
??blockthen
else
Current Line
instr
if
??
regin1
is JZ:
if
instr
elif
??
== 0 : goto
regin1
is JNZ: if
regout
else:
??
??
instr
??
!= 0 : goto
regin1
regin2
= ?? ( ?? ,
goto next
Next Line
??
)
??blockelse
TURING MACHINE
Description
Invert
Prepend zero
Binary decrement
Perfom bitwise inversion of a binary string on the tape.
Right shift all symbols and insert a “0" in the first cell.
Decrement a binary representation on the tape by 1.
BOOLEAN CIRCUITS
Description
2-bit controlled shift register
full adder
2-bit adder
Swap bits on wires 2 & 3 iff wire 1 is on.
Perform binary addition of two bits including carries.
Perform binary addition on two-bit numbers.
BASIC BLOCK
Description
Access
Decrement
List-K
Access the k th element of a contiguous array.
Decrement all elements in a contiguous array.
Access the k th element of a linked list.
ASSEMBLY
Description
branchAddr
??
branchAddr
Access
Decrement
List-K
As above.
Table 2: Overview of benchmark problems, grouped by execution model. We illustrate the basic
structure of each model, with parameters of the model to be inferred denoted by ??.
3
Experimental Results
Table 2 gives an overview of the benchmark programs which we attempt to synthesize. We created T ERPRE T descriptions for each execution model, we designed three synthesis tasks per model
which are specified by up to 16 input-output examples. We measure the time taken by the inference
techniques listed in Table 1 to synthesize a program for each task. Results are summarized in Table 3.
With the exception of the FMGD algorithm, we perform a single run with a timeout of 4 hours
for each task. For the FMGD algorithm we run both a Vanilla form and an Optimized form with
additional heuristics such as gradient clipping, gradient noise and an entropy bonus (Kurach et al.
[2015]) to aid convergence. Even with these heuristics we observe that several random initializations
of the FMGD algorithm stall in an uninterpretable local optimum rather than finding an interpretable
discrete program in a global optimum. In the Vanilla case we report the fraction of 20 different
random initializations which lead to any globally optimal solution consistent with the input-output
specification and also the wall clock time for 1000 epochs of the gradient descent algorithm (the
typical number required to reach convergence on a successful run). In the Optimized FMGD case,
we randomly draw 120 sets of hyperparameters from a manually chosen distribution and run the
learning with 20 different random initializations for each setting. We then report the success rate for
the best hyperparameters found and also for the average across all runs in the random search.
Our results show that traditional techniques employing constraint solvers (SMT and Sketch) outperform other methods (FMGD and ILP), with S KETCH being the only system able to solve all of
these benchmarks before timeout3 . Furthermore, Table 3 highlights that the precise formulation of
the interpreter model can affect the speed of synthesis. Both the Basic Block and Assembly models
are equally expressive, but the Assembly model is biased towards producing straight line code with
minimal branching. In cases where synthesis was successful the Assembly representation is seen to
outperform the Basic Block model in terms of synthesis time.
In addition, we performed a separate experiment to investigate the local optima arising in the FMGD
formulation. Using T ERPRE T we describe the task of inferring the values of a bit string of length K
from observation of the parity of neighboring variables. It is possible to show analytically that the
number of local optima grow exponentially with K, and Table 4 provides empirical evidence that
these minima are encountered in practice and they hinder the convergence of the FMGD algorithm.
3
Additional (pre)processing performed by S KETCH makes it slower than a raw SMT solver on the smaller
tasks but allows it to succeed on the larger tasks
4
4
Conclusion
We presented T ERPRE T, a probabilistic programming language for specifying IPS problems. The
flexibility of the T ERPRE T language in combination with the four inference backends allows likefor-like comparison between gradient-based and constraint-based techniques for inductive program
synthesis. The primary take-away from the experiments is that constraint solvers outperform other
approaches in all cases studied. However, our work is (intentionally) only measuring the ability
to efficiently search over program space. We remain optimistic about extensions to the T ERPRE T
framework which allow differentiable interpreters to handle problems involving perceptual data
[Gaunt et al., 2016a], and about using machine learning techniques to guide search-based techniques
[Balog et al., 2016].
References
Marcin Andrychowicz and Karol Kurach. Learning efficient algorithms with hierarchical attentive
memory. arXiv preprint arXiv:1602.03218, 2016.
Matej Balog, Alexander L Gaunt, Marc Brockschmidt, Sebastian Nowozin, and Daniel Tarlow. Deepcoder: Learning to write programs. arXiv preprint arXiv:1611.01989, 2016.
Alan W Biermann. The inference of regular lisp programs from examples. IEEE transactions on
Systems, Man, and Cybernetics, 8(8):585–600, 1978.
Rudy Bunel, Alban Desmaison, Pushmeet Kohli, Philip H. S. Torr, and M. Pawan Kumar. Adaptive
neural compilation. CoRR, abs/1605.07969, 2016. URL http://arxiv.org/abs/1605.07969.
Alexander L Gaunt, Marc Brockschmidt, Nate Kushman, and Daniel Tarlow. Lifelong perceptual
programming by example. arXiv preprint arXiv:1611.02109, 2016a.
Alexander L Gaunt, Marc Brockschmidt, Rishabh Singh, Nate Kushman, Pushmeet Kohli, Jonathan
Taylor, and Daniel Tarlow. Terpret: A probabilistic programming language for program induction.
arXiv preprint arXiv:1608.04428, 2016b.
C. Lee Giles, Guo-Zheng Sun, Hsing-Hen Chen, Yee-Chun Lee, and Dong Chen. Higher order
recurrent networks and grammatical inference. In Advances in Neural Information Processing
Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pages 380–387,
1989.
Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. CoRR, abs/1410.5401, 2014.
URL http://arxiv.org/abs/1410.5401.
Edward Grefenstette, Karl Moritz Hermann, Mustafa Suleyman, and Phil Blunsom. Learning to
transduce with unbounded memory. In Advances in Neural Information Processing Systems,
pages 1828–1836, 2015.
Sumit Gulwani. Automating string processing in spreadsheets using input-output examples. In ACM
SIGPLAN Notices, volume 46, pages 317–330. ACM, 2011.
Sumit Gulwani, William Harris, and Rishabh Singh. Spreadsheet data manipulation using examples.
Communications of the ACM, Aug 2012.
Abram Hindle, Earl T Barr, Zhendong Su, Mark Gabel, and Premkumar Devanbu. On the naturalness
of software. In 2012 34th International Conference on Software Engineering (ICSE), pages 837–
847. IEEE, 2012.
Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-augmented recurrent nets. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver,
Colorado, USA, November 27-30, 1989], pages 190–198, 2015.
Łukasz Kaiser and Ilya Sutskever. Neural gpus learn algorithms. In Proceedings of the 4th International Conference on Learning Representations., 2016.
Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever. Neural random-access machines. In
Proceedings of the 4th International Conference on Learning Representations 2016, 2015. URL
http://arxiv.org/abs/1511.06392.
5
log10 (D)
T
N
Time
Vanilla
FMGD
Best Hypers
Average Hypers
ILP
Time
SMT
Time
S KETCH
Time
Invert
Prepend zero
Binary decrement
BOOLEAN CIRCUITS
4
9
9
6
6
9
5
5
5
76.5
98
163
100%
60%
5%
100%
100%
25%
51%
37%
2%
0.6
17.0
191.9
0.7
0.9
1.6
3.1
2.6
3.3
2-bit controlled shift register
Full adder
2-bit adder
BASIC BLOCK
10
13
22
4
5
8
8
8
16
-
-
-
-
2.5
38
13076.5
0.7
1.9
174.4
2.7
3.5
355.4
Access
Decrement
List-K
ASSEMBLY
14
19
33
5
18
11
5
5
5
173.8
811.7
-
15%
-
50%
5%
-
1.1%
0.04%
-
98.0
-
14.4
-
4.3
559.1
5493.0
Access
Decrement
List-K
13
20
29
5
27
16
5
5
5
134.6
-
20%
-
90%
-
16%
-
3.6
-
10.5
-
3.8
69.4
16.8
TURING MACHINE
6
Table 3: Benchmark results. For FMGD we present the time in seconds for 1000 epochs and the success rate out of {20, 20, 2400} random restarts in the {Vanilla,
Best Hypers and Average Hypers} columns respectively. For other back-ends we present the time in seconds to produce a synthesized program. The symbol indicates timeout (> 4h) or failure of any random restart to converge. N is the number of provided input-output examples used to specify the task in each case.
Vanilla FMGD
Best Hypers
Average Hypers
K=4
K=8
K = 16
K = 32
K = 64
K = 128
100%
100%
84%
53%
100%
42%
14%
100%
21%
0%
70%
4%
0%
80%
1%
0%
0%
0%
Table 4: Percentage of runs that converge to the global optimum for FMGD on the Parity Chain example of length K.
Chris Lattner and Vikram Adve. Llvm: A compilation framework for lifelong program analysis &
transformation. In Code Generation and Optimization, 2004. CGO 2004. International Symposium
on, pages 75–86. IEEE, 2004.
Tom Minka and John Winn. Gates. In Advances in Neural Information Processing Systems, pages
1073–1080, 2009.
Arvind Neelakantan, Quoc V. Le, and Ilya Sutskever. Neural programmer: Inducing latent programs with gradient descent. In Proceedings of the 4th International Conference on Learning
Representations 2016, 2016.
Scott E. Reed and Nando de Freitas. Neural programmer-interpreters. In Proceedings of the 4th
International Conference on Learning Representations 2016, 2016.
Sebastian Riedel, Matko Bosnjak, and Tim Rocktäschel. Programming with a differentiable forth
interpreter. CoRR, abs/1605.06640, 2016. URL http://arxiv.org/abs/1605.06640.
Armando Solar-Lezama. Program Synthesis By Sketching. PhD thesis, EECS Dept., UC Berkeley,
2008.
Phillip D Summers. A methodology for lisp program construction from examples. Journal of the
ACM (JACM), 24(1):161–175, 1977.
Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Proceedings of the 3rd
International Conference on Learning Representations 2015, 2014. URL http://arxiv.org/
abs/1410.3916.
7
| 9cs.NE
|
Multi-Label Classification of Patient Notes: Case Study on ICD Code Assignment
Tal Baumel
arXiv:1709.09587v3 [cs.CL] 20 Nov 2017
Ben-Gurion University
Beer-Sheva, Israel
Jumana Nassour-Kassis Raphael Cohen
Michael Elhadad
Noémie Elhadad
Chorus.ai
San Francisco, CA
Ben-Gurion University
Beer-Sheva, Israel
Columbia University
New York, NY
Ben-Gurion University
Beer-Sheva, Israel
Abstract
The automatic coding of clinical documentation according
to diagnosis codes is a useful task in the Electronic Health
Record, but a challenging one due to the large number
of codes and the length of patient notes. We investigate
four models for assigning multiple ICD codes to discharge
summaries, and experiment with data from the MIMIC II
and III clinical datasets. We present Hierarchical Attentionbidirectional Gated Recurrent Unit (HA-GRU), a hierarchical approach to tag a document by identifying the sentences
relevant for each label. HA-GRU achieves state-of-the art results. Furthermore, the learned sentence-level attention layer
highlights the model decision process, allows for easier error
analysis, and suggests future directions for improvement.
Introduction
In Electronic Health Records (EHRs), there is often a need
to assign multiple labels to a patient record, choosing from
a large number of potential labels. Diagnosis code assignment is such a task, with a massive amount of labels to chose
from (14,000 ICD9 codes and 68,000 ICD10 codes). Largescale multiple phenotyping assignment, problem list identification, or even intermediate patient representation can
all be cast as a multi-label classification over a large label
set. More recently, in the context of predictive modeling,
approaches to predict multiple future healthcare outcomes,
such as future diagnosis codes or medication orders have
been proposed in the literature. There again, the same setup
occurs where patient-record data is fed to a multi-label classification over a large label set.
In this paper, we investigate how to leverage the unstructured portion of the EHR, the patient notes, along a novel application of neural architectures. We focus on three characteristics: (i) a very large label set (6,500 unique ICD9 codes
and 1,047 3-digit unique ICD9 codes); (ii) a multi-label setting (up to 20 labels per instance); (iii) instances are long
documents (discharge summaries on average 1,900-word
long); and (iv) furthermore, because we work on long documents, one critical aspect of the multi-label classification is
transparency—to highlight the elements in the documents
that explain and support the predicted labels. While there
has been much work on each of these characteristics, there
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
has been limited work to tackle all at once, particularly in
the clinical domain.
We experiment with four approaches to classification: an
SVM-based one-vs-all model, a continuous bag-of-words
(CBOW) model, a convolutional neural network (CNN)
model, and a bidirectional Gated Recurrent Unit model
with a Hierarchical Attention mechanism (HA-GRU).
Among them, the attention mechanism of the HA-GRU
model provides full transparency for classification decisions. We rely on the publicly available MIMIC datasets to
validate our experiments. A characteristic of the healthcare
domain is long documents with a large number of technical
words and typos/misspellings. We experiment with simple
yet effective preprocessing of the input texts.
Our results show that careful tokenization of the input
texts, and hierarchical segmentation of the original document allow our Hierarchical Attention GRU architecture to
yield the most promising results, over the SVM, CBOW, and
CNN models, while preserving the full input text and providing effective transparency.
Previous Work
We review previous work in the healthcare domain as well
as recent approaches to extreme multi-label classification,
which take place in a range of domains and tasks.
Multi-label Patient Classifications
Approaches to classification of patient records against multiple labels fall into three types of tasks: diagnosis code assignment, patient record labeling, and predictive modeling.
Diagnosis Code Assignment. Automated ICD coding is a
well established task, with several methods proposed in the
literature, ranging from rule based (Crammer et al. 2007;
Farkas and Szarvas 2008) to machine learning such as support vector machines, Bayesian ridge regression, and Knearest neighbor (Larkey and Croft 1995; Lita et al. 2008).
Some methods exploit the hierarchical structure of the ICD
taxonomy (Perotte et al. 2011; Perotte et al. 2014), while
others incorporated explicit co-occurrence relations between
codes (Kavuluru, Rios, and Lu 2015). In many cases, to handle the sheer amount of labels, the different approaches focus on rolled-up ICD codes (i.e., 3-digit version of the codes
and their descendants in the ICD taxonomy) or on a subset
of the codes, like in the shared community task for radiology
code assignment (Pestian et al. 2007).
It is difficult to compare the different methods proposed,
since each relies on different (and usually not publicly available) datasets. We experiment with the MIMIC dataset, since
it is publicly available to the research community. Methodswise, our approach departs from previous work in two important ways: we experiment with both massively large and
very large label sets (all ICD9 code and rolled-up ICD9
codes), and we experiment with transparent models that
highlight portions of the input text that support the assigned
codes.
Patient Record Labeling. Other than automated diagnosis coding, most multi-label patient record classifiers fall in
the tasks of phenotyping across multiple conditions at once.
For instance, the UPhenome model takes a probabilistic generative approach to assign 750 latent variables (Pivovarov et
al. 2015). More recently, in the context of multi-task learning, Harutyunyan and colleagues experimented with phenotyping over 25 critical care conditions (Harutyunyan et al.
2017).
Predictive Modeling. Previous work in EHR multi-label
classification has mostly focused on predictive scenarios.
The size of the label set varies from one approach to another, and most limit the label set size however: DeepPatient (Miotto et al. 2016) predicts over a set of 78 condition codes. (Lipton et al. 2015) leverage an LSTM model
to predict over a vocabulary of 128 diagnosis codes. DoctorAI (Choi, Bahadori, and Sun 2015) predicts over a set of
1,183 3-digit ICD codes and 595 medication groups. The
Survival Filter (Ranganath et al. 2015) predicts a series of
future ICD codes across approximately 8,000 ICD codes.
Inputs to Multi-Label Classifications. Most work in
multi-label classification takes structured input. For instance, the Survival Filter expects ICD codes as input to
predict the future ICD codes. DoctorAI takes as input medication orders, ICD codes, problem list, and procedure orders at a given visit. Deep Patient does take the content of
notes as input, but the content is heavily preprocessed into a
structured input to their neural network, by tagging all texts
with medical named entities. In contrast, our approach is
to leverage the entire content of the input texts. Our work
contributes to clinical natural language processing (Demner Fushman and Elhadad 2016), which only recently investigated neural representations and architectures for traditional tasks such as named entity recognition (Jagannatha
and Yu 2016).
Multi-label Extreme Classification
In extreme multi-label learning, the objective is to annotate each data point with the most relevant subset of labels
from an extremely large label set. Much work has been carried outside of the healthcare domain on tasks such as image classification (Tsoumakas and Katakis 2006; Weston,
Bengio, and Usunier 2011), question answering (Choi et al.
2016), and advertising (Jain, Prabhu, and Varma 2016). In
(Weston, Bengio, and Usunier 2011), the task of annotating a
very large dataset of images (> 10M ) with a very large label
set (> 100K) was first addressed. The authors introduced
the WSABIE method which relies on two main features: (i)
records (images) and labels are embedded in a shared lowdimension vector space; and (ii) the multi-label classification task is modeled as a ranking problem, evaluated with a
Hamming Loss on a P@k metric. The proposed online approximate WARP loss allowed the algorithm to perform fast
enough on the scale of the dataset. We found that in our case,
the standard Micro-F measure is more appropriate as we do
not tolerate approximate annotations to the same extent as in
the image annotation task.
The SLEEC method (Bhatia et al. 2015) also relies on
learning an embedding transformation to map label vectors
into a low-dimensional representation. SLEEC learns an ensemble of local distance preserving embeddings to accurately predict infrequently occurring labels. This approach
attempts to exploit the similarity among labels to improve
classification, and learns different representations for clusters of similar labels. Other approaches attempt to reduce
the cost of training over very large datasets by considering
only part of the labels for each classification decision (Yen
et al. 2016). SLEEC was later improved in (Jain, Prabhu,
and Varma 2016) with the PfastreXML method which also
adopted P@k loss functions aiming at predicting tail labels.
In (Joulin et al. 2016), the FastText method was introduced as a simple and scalable neural bag of words approach
for assigning multiple labels to text. We test a similar model
(CBOW) in our experiments as one of our baselines.
Dataset and Preprocessing
We use the publicly available de-identified MIMIC dataset
of ICU stays from Beth Israel Deaconess Medical Center (Saeed et al. 2011; Johnson et al. 2016).
MIMIC Datasets
To test the impact of training size, we relied on both the
MIMIC II (v2.6) and MIMIC III (v1.4) datasets. MIMIC
III comprises records collected between 2001 and 2012, and
can be described as an expansion of MIMIC II (which comprises records collected between 2001 and 2008), along with
some edits to the dataset (including de-identification procedures).
To compare our experiments to previous work in ICD coding, we used the publicly available split of MIMIC II from
(Perotte et al. 2014). It contains 22,815 discharge summaries
divided into a training set (20,533 summaries) and a testset of unseen patients (2,282 summaries). We thus kept the
same train and the test-set from MIMIC II, and constructed
an additional training set from MIMIC III. We made sure
that the test-set patients remained unseen in this training set
as well. Overall, we have two training sets, which we refer
to as MIMIC II and MIMIC III, and a common test-set comprising summaries of unseen patients.
While there is a large overlap between MIMIC II and
MIMIC III, there are also marked differences. We found
Nb of records
Nb of unique tokens
Avg nb of tokens / record
Avg nb of sentences / record
Nb of full labels
Nb of rolled-up labels
Label Cardinality
Label Density
% labels with 50+ records
MIMIC II
MIMIC III
Test Set
20,533
69,248
1,529
90
4,847
948
9.24
0.0019
11.33%
49,857
119,171
1,947
112
6,527
1,047
11.48
0.0018
18.19%
2,282
33,958
1,893
104
2,451
684
11.42
0.0047
4.08%
Table 1: Datasets descriptive statistics.
many cases where discharge summaries from 2001-2008 are
found in one dataset but not in the other. In addition, MIMIC
III contains addenda to the discharge summaries that were
not part of MIMIC II. After examining the summaries and
their addenda, we noticed that the addenda contain vital information for ICD coding that is missing from the main discharge summaries; therefore, we decided to concatenate the
summaries with their addenda.
Table 1 reports some descriptives statistics regarding the
datasets. Overall, MIMIC III is larger than MIMIC II from
all standpoints, including amounts of training data, vocabulary size, and overall number of labels.
ICD9 Codes
Our label set comes from the ICD9 taxonomy. The International Classification of Diseases (ICD) is a repository maintained by the World Health Organization (WHO) to provide
a standardized system of diagnostic codes for classifying
diseases. It has a hierarchical structure, connecting specific
diagnostic codes through is-a relations. The hierarchy has
eight levels, from less specific to more specific. ICD codes
contain both diagnosis and procedure codes. In this paper,
we focus on diagnosis codes only. ICD9 codes are conveyed
as 5 digits, with 3 primary digits and 2 secondary ones.
Table 1 provides the ICD9 label cardinality and density
as defined by (Tsoumakas and Katakis 2006). Cardinality
is the average number of codes assigned to records in the
dataset. Density is the cardinality divided by the total number of codes. For both training sets, the number of labels is of
the same order as the number of records, and the label density is extremely low. This confirms that the task of code assignment belongs to the family of extreme multi-label classification.
We did not filter any ICD code based on their frequency.
We note, however that there are approximately 1,000 frequent labels (defined as assigned to at least 50 records) (Table 1). We experimented with two versions of the label set:
one with all the labels (i.e., 5-digit), and one with the labels
rolled up to their 3-digit equivalent.
Input Texts
Tokenization. Preprocessing of the input records comprised the following steps: (i) tokenize all input texts us-
ing spaCy library 1 ; (ii) convert all non-alphabetical characters to pseudo-tokens (e.g., “11/2/1986” was mapped to
“dd/d/dddd”); (iii) build the vocabulary as tokens that appear
at least 5 times in the training set; and (iv) map any out-ofvocabulary word to its nearest word in the vocabulary (using
the edit distance). This step is simple, yet particularly useful
in reducing the number of misspellings of medical terms.
These preprocessing steps has a strong impact on the vocabulary. For instance, there were 1,005,489 unique tokens
in MIMIC III and test set before preprocessing, and only
121,595 remaining in the vocabulary after preprocessing (an
88% drop). This step improved F-measure performance by
˜0.5% when tested on the CBOW and CNN methods (not
reported).
Hierarchical Segmentation. Besides tokenization of the
input texts, we carried one more level of segmentation, at the
sentence level (using the spaCy library as well). There are
two reasons for preprocessing the input texts with sentence
segmentation. First, because we deal with long documents,
it is impossible and ineffective to train a sequence model like
an GRU on such long sequences. In previous approaches in
document classification, this problem was resolved by truncating the input documents. In the case of discharge summaries, however, this is not an acceptable solution: we want
to preserve the entire document for transparency. Second, we
are inspired by the moving windows of (Johnson and Zhang
2014) and posit that sentences form linguistically inspired
windows of word sequences.
Beyond tokens and sentences, discharge summaries exhibit strong discourse-level structure (e.g., history of present
illness and past medical history, followed by hospital course,
and discharge plans) (Li, Lipsky Gorman, and Elhadad
2010). This presents an exciting opportunity for future work
to exploit discourse segments as an additional representation
layer of input texts.
Methods
We describe the four models we experimented with. ICD
coding has been evaluated in the literature according to different metrics: Micro-F, Macro-F, a variant of Macro-F that
takes into account the hierarchy of the codes (Perotte et al.
2014), Hamming and ranking loss (Wang et al. 2016), and
a modified version of mean reciprocal rank (MRR) (Subotin
and Davis 2014). We evaluate performance using the MicroF metric, since it is the most commonly used metric.
SVM. We used Scikit Learn (Pedregosa et al. 2011) to implement a one-vs-all, multi-label binary SVM classifier. Features were bag of words, with tf*idf weights (determined
from the corpus of release notes) for each label. Stop words
were removed using Scikit Learn default English stop-word
list. The model fits a binary SVM classifier for each label
(ICD code) against the rest of the labels. We also experimented with χ2 feature filtering to select the top-N words
1
https://spacy.io/
Binary Cross-Entropy
Loss
Binary Cross-Entropy
Loss
Fully Connected
(Sigmoid Activation)
Fully Connected
(Sigmoid Activation)
Max Pooling
Average
Convolution Layer
Embedding
Embedding
𝑤" 𝑤# 𝑤$ …
𝑤%
𝑤" 𝑤# 𝑤$ …
𝑤%
diagnoses that share lexical words will not be distinguished
by the model.
CNN. To address the problems of the CBOW model, the
next model we investigate is a convolutional neural network
(CNN). A one dimensional convolution applied on list of
embedded words could be considered as a type of n-gram
model, where n is the convolution filter size.
The architecture of this model is very similar to the
CBOW model, but instead of averaging the embedded words
we apply a one dimensional convolution layer with filter f ,
followed by a max pooling layer. On the output of the max
pool layered a fully connected layer was applied, like in the
CBOW model. We also experimented with deeper convolution networks and inception module (LeCun 2015), but they
did not yield improved results.
Embedding = E · [w1 , w2 , ..., wn ]
Figure 1: CBOW architecture on the left and CNN model
architecture on the right.
according to their mutual information with each label, but
this did not improve performance.
CBOW. The continuous-bag-of-words (CBOW) model is
inspired by the word2vec CBOW model (Mikolov et al.
2013) and FastText (Joulin et al. 2016). Both methods use
a simple neural-network to create a dense representation of
words and use the average of this representation for prediction. The word2vec CBOW tries to predict a word from the
words that appear around it, while our CBOW model for
ICD classification predicts ICD9 codes from the words of
its input discharge summary.
The model architecture consists of an embedding layer
applied to all the words in a given input text [w1 , w2 , ..., wn ],
where wi is a one-hot encoding vector of the vocabulary. E
is the embedding matrix with dimension nemb × V , where
V is the size of the vocabulary and nemb is the embedding
size (set to 100).
The embedded words are averaged into a fixed-size vector and are fed to a fully connected layer with a matrix
W and bias b, where the output dimension is the number
of labels. We use a sigmoid activation on the output layer
so all values are in the range of [0 − 1] and use a fixed
threshold (0.5) to determine whether to assign a particular
label. To train the model, we used binary cross-entropy loss
(loss(target, output) = −(target · log(output) + (1 −
target) · log(1 − output)).
Embedding = E · [w1 , w2 , ..., wn ]
Averaged = 1/nΣe∈Embedding (e)
P rob = sigmoid(W · Averaged + b)
While the model is extremely lightweight and fast it suffers from known bag-of-words issues: (i) it ignores word order; i.e., if negation will appear before a diagnosis mention,
the model would not be able to learn this; (ii) multi-wordexpressions cannot be identified by the model, so different
Conved =
max
i∈channels
(Embedding ∗ f )
P rob = sigmoid(W · Conved + b)
In our experiments, we used the same embedding parameter
as in the CBOW model. In addition, we set the number of
channels to 300, and the filter size to 3.
HA-GRU. We now introduce the Hierarchical Attentionbidirectional Gated Recurrent Unit model (HA-GRU) an
adaptation of a Hierarchical Attention Networks (Yang et
al. 2016) to be able to handle multi-label classification. A
Gated Recurrent Unit (GRU) is a type of Recurrent Neural
Network. Since the documents are long (see Table 1 ad up
to 13,590 tokens in the MIMIC III training set), a regular
GRU applied over the entire document is too slow as it requires number layers of the document length. Instead we apply a hierarchal model with two levels of bidirectional GRU
encoding. The first bidirectional GRU operates over tokens
and encodes sentences. The second bidirectional GRU encodes the document , applied over all the encoded sentences.
In this architecture, each GRU is applied to a much shorter
sequence compared with a single GRU.
To take advantage of the property that each label is invoked from different parts of the text, we use an attention
mechanism over the second GRU with different weights for
each label. This allows the model to focus on the relevant
sentences for each label (Choi et al. 2016). To allow clarity
into what the model learns and enable error analysis attention is also applied over the first GRU with the same weights
for all the labels.
Each sentence in the input text is encoded to a fixed length
vector (64) by applying an embedding layer over all the inputs, applying a bidirectional GRU layer on the embedded
words, and using a neural attention mechanism to encode
the bidirectional GRU outputs (size of 128). After the sentences are encoded into a fixed length vector, we apply a
second bidirectional GRU layer over the sentences using different attention layers to generate an encoding specified to
each class (128 × #labels). Finally we applied a fully connected layer with softmax for each classifier to determine
if the label should be assigned to the document. Training is
ICD9 codes
MIMIC II MIMIC III
Classifiers
Classifiers
Classifiers
SVM
CBOW
CNN
HA-GRU
Document GRU
Layer
Sentence Encoder
Sentence Encoder
Sentence Encoder
𝑤","𝑤",$ 𝑤",% … 𝑤",&'
…
𝑤$,"𝑤$,$ 𝑤$,% … 𝑤$,&)
𝑤*," 𝑤*,$𝑤*,%… 𝑤*,&+
28.13%
30.60%
33.25%
36.60%
22.25%
30.02%
40.72%
40.52%
Rolled-up ICD9 codes
MIMIC II MIMIC III
32.50%
42.06%
46.40%
53.86%
53.02%
43.30%
52.64%
55.86%
Table 2: Micro-F on two settings (full and rolled-up ICDs)
and for the four models when trained on MIMIC II or
MIMIC III datasets.
Figure 2: HA-GRU model architecture overview.
Results
achieved by using categorical cross-entropyP
on every classifier separately (loss(target, output) = − x ouput(x) ·
log(target(x)))
AttW eight(ini , v, w) = v · tanh(w · (ini ))
AttW eight(ini , v, w) =
eAttW eight(ini ,v,w)
e
P
j AttW eightj (v,w)
Attend(in, v, w) = sum(ini · AttW eight(ini , v, w))
Embedding = E · [w1 , w2 , ..., wn ]
EncSentsj = Attend(GRU words(Embedding), vwords , wwords )
EncDoclabel = Attend(GRUsents (EncSents, vlabel , wlabel ), )
P roblabel = sof tmax(pwlabel · EncDoclabel + pblabel )
Where wi is a one-hot encoding vector of the vocabulary size V , E is an embedding matrix size of nemb × V ,
GRUwords is a GRU layer with state size hstate , wwords
is a square matrix (hstate × hstate ) and vwords is a vector (hstate ) for the sentence level attention. GRUsents is a
GRU layer with state size of hstate . wlabel is a square matrix (hstate × hstate ) and vlabel is a vector (hstate ) for the
document level attention for each class, pwlabel is a matrix
(hstate × 2) and pblabel is a bias vector with a size of 2 for
each label. We implemented the model using DyNet (Neubig et al. 2017)2 .
𝑐"
𝑠"
𝑤" ,$ 𝑤" ,% 𝑤" ,& … 𝑤" ,'(
Softmax
Fully Connected
Layer
Classifier
Embedding
Sentence Encoder
Sentence GRU Layer
To evaluate the proposed methods on the MIMIC datasets,
we conducted the following experiments. In the first setting
we considered all ICD9 codes as our label set. We trained
the SVM, CBOW, and CNN on the MIMIC II and on the
MIMIC III training sets separately. All models were evaluated on the same test set according to Micro-F. In the second
setting, we only considered the rolled-up ICD9 codes to their
3-digit codes. There (Table 2).
HA-GRU gives the best results in the rolled-up ICD9 setting, with a 7.4% and 3.2% improvement over the CNN and
SVM, the second best methods, in MIMIC II and MIMIC
III respectively. In the full ICD-9 scenario, all methods yield
better results when trained on MIMIC III rather than on
MIMIC II. This is expected considering the larger size of
MIMIC III over II. We note that our CNN yields the best
Micro-F when trained on MIMIC III passing the HA-GRU
by a small margin.
In comparison to the previous work of (Perotte et al.
2014), our one-vs-all SVM yielded better results than their
flat and hierarchy classifiers. This trend was confirmed when
training on the new MIMIC III set, as well as when using the
same evaluation metrics of (Perotte et al. 2014). We attribute
these improved results both to the one-vs-all approach as
well as our tokenization approach.
Label Frequency
Categorical CrossEntropy loss
Weighted Average
Sentence Attention
Model Comparison
Weighted Average
Label Attention
𝑠$ 𝑠% 𝑠&
… 𝑠'
Figure 3: Zoom-in of the sentence encoder and classifier.
We also tested the effect label frequency on the performance
of the HA-GRU classifier. We recalculated precision and recall scores on subsets of labels. The subsets were created
by sorting the labels by frequency they appear in MIMIC-III
dataset and binning them to groups of 50 labels. As such, bin
1 comprises the 50 most frequent ICD9 codes in the training
set (with an average 12% frequency over the records in the
training set), codes in bin 2 had an average 1.9% frequency,
codes in bin 3 appeared in 1.1% of the records, up to bin 8
which 0.2% of the records in the training set. The effect can
be seen in Figure 5. We note that the recall score drops much
more dramatically than the precision as the label frequency
decreases.
Model Explaining Power
2
Code available at https://github.com/talbaumel/
MIMIC.
We discuss how the CNN and HA-GRU architectures can
support model explaining power.
Figure 4: Sample text of a patient note (one sentence per line). On the left, visualization for the with attention weights at the
sentence and word levels associated with the ICD9 codes, on the left sentence level attention weights for ICD9 code “Heart
failure”, on the the right for code “Traumatic pneumothorax and hemothorax”.
find what word received the highest attention score. For example, in our experiments for label “428-Heart failure” we
found that the sentence with the highest attention score was
“d . congestive heart failure ( with an ejection fraction of
dd % to dd % ) .”, while the token “failure” was found
most relevant across all labels. Figure 4 provides additional
examples. Note that the “d” and “dd” tokens are from the
pre-procecssing step, which mapped all numbers to pseudotokens.
Figure 5: Effect label frequency on HA-GRU performance
when trained on MIMIC III. X-axis represents the bins of
labels ranked by their frequency in the training set.
CNN. To analyze the CNN prediction we can test which
n-grams triggered the max-pooling layer. Given a sentence
with n words we can feed it forward through the embedding
layer and the convolution layer. The output of the convolution a list of vectors each the size of the number of channels of the convolution layer where vector corresponds to
an n-gram. We can identify what triggered the max pooling
layer by finding the maximum value of each channel. Thus,
for predicted labels, one of the activated n-grams does include information relevant for that label (whether correct for
true positive labels or incorrect for false positive labels). For
example in our experiments, for the label: “682.6-Cellulitis
and abscess of leg, except foot” one of the activated n-gram
detected was “extremity cellulitis prior”.
This transparency process can also be useful for error
analysis while building a model, as it can highlight True Positive and False Positive labels. However, it is difficult in the
CNN to trace back the decisions for False Negatives predictions.
HA-GRU For the HA-GRU model we can use attention
weights to better understand what sentences and what words
in that sentence contributed the most to each decision. We
can find which sentence had the highest attention score for
each label, and given the most important sentence, we can
Like in the CNN, we can use this process for error analysis. In fact, the HA-GRU model explains prediction with
greater precision, at the sentence level. For instance, we
could explore the following False Positive prediction: the
model assigned the label “331-Other cerebral degenerations” to the sentence:“alzheimer ’s dementia .”. We can see
that the condition was relevant to the medical note, but was
mentioned under the patient’s past medical history (and not
a current problem). In fact, many of the False Positive labels
under the HA-GRU model were due to mentions belonging
to the past medical history section. This suggests that the
coding task would benefit from a deeper architecture, with
attention to discourse-level structure.
In contrast to the CNN, the HA-GRU model can also
help analyze False Negative label assignments. When we
explored the False Negative labels, we found that in many
cases the model found a relevant sentence, but failed to classify correctly. This suggests the document-level attention
mechanism is successful. For instance, for the False Negative “682-Other cellulitis and abscess”, the most attended
sentence was “... for right lower extremity cellulitis prior
to admission ...”. The false positive codes for this sentence
included “250-Diabetes mellitus” and “414-Other forms of
chronic ischemic heart disease”. We note that in the case of
cellulitis, it is reasonable that the classifier preferred other,
more frequent codes, as it is a common comorbid condition
in the ICU.3
3
Full visualizations of sample discharge summaries are
provided at https://www.cs.bgu.ac.il/˜talbau/
mimicdemo
Conclusion
We investigate four modern models for the task of extreme
multi-label classification on the MIMIC datasets. Unlike
previous work, we evaluate our models on all ICD9 codes
thus making sure our models could be used for real world
ICD9 tagging. The tokenization step, mapping rare variants
using edit distance, improved results for CBOW and CNN
models by ˜0.5%, highlighting the importance of preprocessing data noise problems in real-world settings. The HA-GRU
model not only achieves the best performance on the task
of rolled-up codes (55.86% F 1 on MIMIC III, ˜2.8% absolute improvement on the best SVM baseline) but is able
to provide insight on the task for future work such as using discourse-level structure available in medical notes yet
never used before. The ability to highlight the decision process of the model is important for adoption of such models
by medical experts. On the sub-task of MIMIC II, which
includes a smaller training dataset, HA-GRU achieved ˜7%
absolute F 1 improvement, suggesting it requires less training data to achieve top performance, which is important for
domain adaptation efforts when applying such models to patient records from other sources (such as different hospitals).
Acknowledgements
This work is supported by National Institute of General
Medical Sciences Grant R01GM114355 (NE) and Frankel
Center for Computer Science .
References
[Bhatia et al. 2015] Bhatia, K.; Jain, H.; Kar, P.; Varma, M.;
and Jain, P. 2015. Sparse local embeddings for extreme
multi-label classification. In Advances in Neural Information Processing Systems (NIPS), 730–738.
[Choi, Bahadori, and Sun 2015] Choi, E.; Bahadori, M. T.;
and Sun, J.
2015.
Doctor AI: Predicting clinical
events via recurrent neural networks.
arXiv preprint
arXiv:1511.05942.
[Choi et al. 2016] Choi, E.; Hewlett, D.; Lacoste, A.; Polosukhin, I.; Uszkoreit, J.; and Berant, J. 2016. Hierarchical question answering for long documents. arXiv preprint
arXiv:1611.01839.
[Crammer et al. 2007] Crammer, K.; Dredze, M.; Ganchev,
K.; Talukdar, P. P.; and Carroll, S. 2007. Automatic code assignment to medical text. In Proceedings of the ACL Workshop on BioNLP 2007: Biological, Translational, and Clinical Language Processing, 129–136.
[Demner Fushman and Elhadad 2016] Demner
Fushman,
D., and Elhadad, N. 2016. Aspiring to unintended consequences of natural language processing: A review of recent
developments in clinical and consumer-generated text processing. Yearbook of Medical Informatics 10(1):224–233.
[Farkas and Szarvas 2008] Farkas, R., and Szarvas, G. 2008.
Automatic construction of rule-based ICD-9-CM coding
systems. BMC bioinformatics 9(3):S10.
[Harutyunyan et al. 2017] Harutyunyan, H.; Khachatrian,
H.; Kale, D. C.; and Galstyan, A. 2017. Multitask learn-
ing and benchmarking with clinical time series data. arXiv
preprint arXiv:1703.07771.
[Jagannatha and Yu 2016] Jagannatha, A. N., and Yu, H.
2016. Structured prediction models for RNN based sequence labeling in clinical text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 856.
[Jain, Prabhu, and Varma 2016] Jain, H.; Prabhu, Y.; and
Varma, M. 2016. Extreme multi-label loss functions for
recommendation, tagging, ranking & other missing label applications. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, 935–944.
[Johnson and Zhang 2014] Johnson, R., and Zhang, T. 2014.
Effective use of word order for text categorization
with convolutional neural networks.
arXiv preprint
arXiv:1412.1058.
[Johnson et al. 2016] Johnson, A. E.; Pollard, T. J.; Shen, L.;
Lehman, L.-w. H.; Feng, M.; Ghassemi, M.; Moody, B.;
Szolovits, P.; Celi, L. A.; and Mark, R. G. 2016. MIMIC-III,
a freely accessible critical care database. Scientific Data 3.
[Joulin et al. 2016] Joulin, A.; Grave, E.; Bojanowski, P.;
Douze, M.; Jégou, H.; and Mikolov, T. 2016. Fasttext.zip:
Compressing text classification models. arXiv preprint
arXiv:1612.03651.
[Kavuluru, Rios, and Lu 2015] Kavuluru, R.; Rios, A.; and
Lu, Y. 2015. An empirical evaluation of supervised learning
approaches in assigning diagnosis codes to electronic medical records. Artificial Intelligence in Medicine 65(2):155–
166.
[Larkey and Croft 1995] Larkey, L. S., and Croft, W. B.
1995. Automatic assignment of ICD9 codes to discharge
summaries. Technical report, University of Massachusetts
at Amherst, Amherst, MA.
[LeCun 2015] LeCun, Y. 2015. LeNet-5, convolutional
neural networks. http://yann.lecun.com/exdb/
lenet.
[Li, Lipsky Gorman, and Elhadad 2010] Li, Y.; Lipsky Gorman, S.; and Elhadad, N. 2010. Section classification in
clinical notes using supervised hidden Markov model. In
Proceedings of the 1st ACM International Health Informatics Symposium, 744–750. ACM.
[Lipton et al. 2015] Lipton, Z. C.; Kale, D. C.; Elkan, C.; and
Wetzell, R. 2015. Learning to diagnose with LSTM recurrent neural networks. arXiv preprint arXiv:1511.03677.
[Lita et al. 2008] Lita, L. V.; Yu, S.; Niculescu, R. S.; and
Bi, J. 2008. Large scale diagnostic code classification
for medical patient records. In Proceedings of the International Joint Conference on Natural Language Processing
(IJCNLP), 877–882.
[Mikolov et al. 2013] Mikolov, T.; Chen, K.; Corrado, G.;
and Dean, J. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781.
[Miotto et al. 2016] Miotto, R.; Li, L.; Kidd, B. A.; and Dudley, J. T. 2016. Deep Patient: An unsupervised represen-
tation to predict the future of patients from the electronic
health records. Scientific Reports 6.
[Neubig et al. 2017] Neubig, G.; Dyer, C.; Goldberg, Y.;
Matthews, A.; Ammar, W.; Anastasopoulos, A.; Ballesteros,
M.; Chiang, D.; Clothiaux, D.; Cohn, T.; et al. 2017.
Dynet: The dynamic neural network toolkit. arXiv preprint
arXiv:1701.03980.
[Pedregosa et al. 2011] Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.;
Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. 2011. Scikitlearn: Machine learning in python. Journal of Machine
Learning Research 12(Oct):2825–2830.
[Perotte et al. 2011] Perotte, A. J.; Wood, F.; Elhadad, N.;
and Bartlett, N. 2011. Hierarchically supervised latent
dirichlet allocation. In Advances in Neural Information Processing Systems (NIPS), 2609–2617.
[Perotte et al. 2014] Perotte, A.; Pivovarov, R.; Natarajan,
K.; Weiskopf, N.; Wood, F.; and Elhadad, N. 2014. Diagnosis code assignment: models and evaluation metrics.
Journal of the American Medical Informatics Association
21(2):231–237.
[Pestian et al. 2007] Pestian, J. P.; Brew, C.; Matykiewicz, P.;
Hovermale, D. J.; Johnson, N.; Cohen, K. B.; and Duch, W.
2007. A shared task involving multi-label classification of
clinical free text. In Proceedings of the ACL Workshop on
BioNLP: Biological, Translational, and Clinical Language
Processing, 97–104.
[Pivovarov et al. 2015] Pivovarov, R.; Perotte, A. J.; Grave,
E.; Angiolillo, J.; Wiggins, C. H.; and Elhadad, N. 2015.
Learning probabilistic phenotypes from heterogeneous EHR
data. Journal of Biomedical Informatics 58:156–165.
[Ranganath et al. 2015] Ranganath, R.; Perotte, A. J.; Elhadad, N.; and Blei, D. M. 2015. The Survival Filter: joint
survival analysis with a latent time series. In UAI, 742–751.
[Saeed et al. 2011] Saeed, M.; Villarroel, M.; Reisner, A. T.;
Clifford, G.; Lehman, L.-W.; Moody, G.; Heldt, T.; Kyaw,
T. H.; Moody, B.; and Mark, R. G. 2011. Multiparameter intelligent monitoring in intensive care II (MIMIC-II):
a public-access intensive care unit database. Critical Care
Medicine 39(5):952.
[Subotin and Davis 2014] Subotin, M., and Davis, A. R.
2014. A system for predicting ICD-10-PCS codes from electronic health records. In Proceedings of the ACL Workshop
on Biomedical Natural Language Processing (BioNLP), 59–
67. Citeseer.
[Tsoumakas and Katakis 2006] Tsoumakas, G., and Katakis,
I. 2006. Multi-label classification: An overview. International Journal of Data Warehousing and Mining 3(3).
[Wang et al. 2016] Wang, S.; Chang, X.; Li, X.; Long, G.;
Yao, L.; and Sheng, Q. Z. 2016. Diagnosis code assignment using sparsity-based disease correlation embedding.
IEEE Transactions on Knowledge and Data Engineering
28(12):3191–3202.
[Weston, Bengio, and Usunier 2011] Weston, J.; Bengio, S.;
and Usunier, N. 2011. WSABIE: Scaling up to large vocabulary image annotation. In Proceedings of the Interna-
tional Joint Conference on Artificial Intelligence (IJCAI),
volume 11, 2764–2770.
[Yang et al. 2016] Yang, Z.; Yang, D.; Dyer, C.; He, X.;
Smola, A.; and Hovy, E. 2016. Hierarchical attention
networks for document classification. In Proceedings of
NAACL-HLT, 1480–1489.
[Yen et al. 2016] Yen, I. E.; Huang, X.; Zhong, K.; Ravikumar, P.; and Dhillon, I. S. 2016. PD-Sparse: A primal and
dual sparse approach to extreme multiclass and multilabel
classification. In Proceedings of the International Conference on Machine Learning (ICML).
| 2cs.AI
|
Submitted to the Brazilian Journal of Probability and Statistics
arXiv: arXiv:1308.6626
Searching for the core variables in principal
components analysis
arXiv:1308.6626v4 [math.ST] 30 Jan 2017
Yanina Gimeneza , and Guido Giussania
a
Universidad de San Andrés and CONICET
Abstract. In this article, we introduce a procedure for selecting variables in principal components analysis. It is developed to identify a small
subset of the original variables that best explain the principal components through nonparametric relationships. There are usually some
noisy uninformative variables in a dataset, and some variables that are
strongly related to one another because of their general dependence. The
procedure is designed to be used following the satisfactory initial principal components analysis with all variables, and its aim is to help to
interpret the underlying structures. We analyze the asymptotic behavior
of the method and provide some examples.
1 Introduction
Principal components analysis (PCA) is the best known dimensional reduction procedure for multivariate data. An important drawback of PCA is
that it sometimes provides poor quality interpretation of data for practical
problems, because the final output is a linear combination of the original
variables. The aim of the present study is to identify a small subset of the
original variables in a dataset, whilst retaining most of the information related to the first k principal components.
There is a large body of literature that focuses on trying to interpret
principal components. Jolliffe (1995) has introduced rotation techniques.
Vines (2000) proposed to restrict the value of the loadings for PCA to a
small set of allowable integers such as {-1, 0, 1}. McCabe (1984) presented a
different strategy that aims to select a subset of the original variables with
a similar criterion to PCA.
A few years later a whole literature of variable selection appeared, inspired by the LASSO (least absolute shrinkage and selection operator). This
technique was introduced by Tibshirani (1996). The way the LASSO works
was described thus: “ It shrinks some coefficients and sets others to 0, and
hence tries to retain the good features of both subset selection and ridge
Primary 62H25; secondary 62G08
Keywords and phrases. Informative Variables, Multivariate Analysis, Principal Components, Selection of Variables
1
2
Gimenez and Giussani
regression... The ‘LASSO’ minimizes the residual sum of squares subject to
the sum of the absolute value of the coefficients being less than a constant.
Because of the nature of this constraint it tends to produce some coefficients
that are exactly 0 and hence gives interpretable models.”
Jolliffe et al. (2003) proposed SCoTLASS that imposes a bound on the
sum of the absolute values of the loadings in the component, using a similar
idea to the one that LASSO used in regression. Zou et al. (2006) presented
SPCA (sparse PCA) that extends the “elastic net” of Zou & Hastie (2005)
that was a generalization of the LASSO. As they mentioned “SPCA is built
on the fact that PCA can be written as a regression-type optimization problem, with quadratic penalty; the LASSO penalty (via the elastic net) can
then be directly integrated into the regression criterion, leading to a modified
PCA with sparse loadings.”
Luss & d’Aspremont (2010) studied the application of sparse PCA to clustering and problems of feature selection. Sparse PCA seeks sparse factors, or
linear combinations of variables in the dataset, to explain as much variance
in the data as possible while limiting the number of nonzero coefficients as
far as possible. The authors applied their results to some classic biological
clustering and feature selection problems.
Recently, Witten & Tibshirani (2008) introduced the notion of Lassoed
principal components for identifying differentially-expressed genes, and considered the problem of testing the significance of features in high dimensional
data.
Our approach is rather different and is designed to be used after satisfactory PCA has been achieved rather than, as in other methods, to produce
principal components with particular characteristics (e.g., some coefficients
that are zero) so that only interpretable principal components are produced.
We first perform classical PCA and then look for a small subset of the original variables that contain almost all the relevant information to explain the
principal components. However, our method can also be used after performing any sparse PCA method as those described previously.
To develop our method, we borrowed some ideas for selecting variables
from Fraiman et al. (2008), who introduced two procedures for selecting
variables in cluster analysis, and classification rules. Both of these procedures are based on the idea of “blinding” unnecessary variables. To cancel
the effect of a variable, they substituted all its values with the marginal
mean (in the first procedure) or with the conditional mean (in the second).
The marginal mean approach was mainly intended to identify “noisy” uninformative variables, but the conditional mean approach could also deal with
dependence. We adapted the idea behind the second procedure to PCA.
Searching for the core variables in principal components analysis
3
In Section 2, we introduce the main definitions, a population version of
our proposed method, and the empirical version; we also present our main
results. In Section 3, a simulation study is conducted, where the results are
compared with other well-known PCA variable selection procedures. Finally,
in Section 4, we study a real data example. Proofs are given in the Appendix.
2 Our method: Definitions and properties
We begin by defining our notation and stating the problem in terms of the
underlying distribution of the random vector X. Then we give our estimates
based on the sample data, via the empirical distribution.
2.1 Population version
We define X ∈ Rp as a random vector with distribution P. The coordinates
of the vector X are defined as X[i], i = 1, . . . , p. The covariance matrix of
X is denoted Σ.
For a given random vector X̃, we say that it satisfy the Assumption H1
if:
i) E(kX̃k2 ) < ∞;
ii) The covariance matrix of X̃ is positive definite.
iii) All the eigenvalues of the covariance matrix are different.
Throughout the manuscript, X fulfills Assumption H1.
As is well known, the first principal component associated with the vector
X is defined as
α1 (P) := α1 = arg max V ar α0 X
kαk=1
= arg max α0 Σα,
kαk=1
and the next principal components are defined as
αk (P) := αk =
arg max
V ar α0 X
kαk=1, α⊥[α1 ,...,αk−1 ]
=
arg max
kαk=1,
α0 Σα ∀ 2 ≤ k ≤ p,
α⊥[α1 ,...,αk−1 ]
where [α1 , . . . , αk−1 ] is the subspace generated by the vectors {α1 , . . . , αk−1 }.
From the spectral theorem, it follows that, if λ1 > λ2 > . . . > λp are the
Σ eigenvalues, the solutions to the PCA are the corresponding eigenvectors,
αk , k = 1, . . . , p.
4
Gimenez and Giussani
Given a subset of indices I ⊂ {1, . . . , p} with cardinality d ≤ p, we define
X[I] as the subset of random variables {X[i], i ∈ I}. With a slight abuse
of notation, if I = {i1 < . . . < id }, we can also denote X[I] to the vector
(X[i1 ], . . . , X[id ]), and define the vector YI := (Y I [1], . . . , Y I [p]), where
X[i]
if i ∈ I
I
Y [i] =
E(X[i]|X[I])
if i ∈
/ I.
YI ∈ Rp depends only on the {X[i], i ∈ I} variables and then the principal
components associated with the vector YI depend only on the variables in
X[I]. The distribution of YI is denoted PYI , the covariance matrix is ΣYI
and g i (z) = E(X[i]|X[I] = z) for i ∈
/ I is the regression function.
In what follows we assume that YI fulfills the Assumption H1.
Remark: This assumptions will not hold if there is a variable X[i] such
that g i (z) is a linear combination of the variables in I. For instance, if X is
gaussian or X[i] is independent of the variables in I. In practice we will see
how this problem can be tackled in the example 2 of section 3.
We are looking for a subset I of small cardinality that minimizes the
distance between the original principal components and the principal components that are function of the variables in the subset I. This can be done
following two different approaches.
Local approach
We define the objective function h1 (I) as
h1 (I, P, PYI ) := h1 (I, P) := h1 (I) = kα1 (P) − α1 (PYI )k2 ,
(2.1)
which measures the squared distance between the first original principal
component and the first principal component that is a function of the variables in the subset I.
Given a fixed integer d, 1 ≤ d << p, Id is the family of all subsets of
{1, . . . , p} with cardinality d and I1,0 ⊂ Id is the family of subsets in which
the minimum h1 (I) is attained for I ∈ Id , i.e.,
I1,0 = argminI∈Id h1 (I),
or, equivalently,
h1 (I1 ) = minI∈Id h1 (I) for all I1 ∈ I1,0 .
(2.2)
Analogously, we define
hk (I, P, PYI ) := hk (I, P) := hk (I) = kαk (P) − αk (PYI )k2 ,
(2.3)
Searching for the core variables in principal components analysis
5
and
Ik,0 = argminI∈Id hk (I),
for k = 2, . . . , p.
For k = 1, . . . , p (2.3) measures the squared distance between the k-th
original principal component and the k-th principal component that is a
function of the variables in the subset I.
Principal components may, in some cases, be rather difficult to interpret,
which is an important issue in practice. If we can find a small cardinal
subset I, for which the objective function (2.3) is small enough, then the
k-th principal component will be well explained by a subset of the original
variables.
Global approach
In this case, the objective function is
h(I) :=
q
X
pk hk (I),
(2.4)
k=1
Pq
with pk ≥ 0,
k=1 pk = 1, 2 ≤ q < p,
and we define
Ĩq,0 = argminI∈Id h(I).
This time the objective function (2.4) deals with finding a unique subset
I to explain the first q principal components at once. If we think that the
q components are equally important then we choose pk = 1q . Otherwise, if
some components are more important than others then we can put different weights. As a default we suggest choosing weights proportional
to the
P
variance that each component is explaining, i.e. pk = λk / qk=1 λk .
On the Local approach, we consider a different subset for each principal
component, using the objective function (2.3) for each k. On the Global approach, we seek a unique subset for all the first q principal components using
(2.4). In practice, we can choose q, so that the first q principal components
explain a high percent of the total variance and consider a subset for this
principal components.
These subsets tell us which are the original variables that “best explain”
the first q principal components. In what follows we refer to this method as
the Blinding Procedure (BP).
An important issue is how to choose d, the cardinality of the set I. On
the one hand we have that the objective function (2.3) decreases when d
6
Gimenez and Giussani
increases for every k. On the other hand we look for a subset I with small
cardinality, so we have a constant tug of war. Since the components have
unitary norms there is a direct relationship between (2.3) and the angle
between the two vectors. It is clear that the smaller the angle is the closer
the components are.
Hence, for the k-th component we propose to fix an angle, γ, and choose d
as the smallest value for which Ik,0 (d) makes the angle between the blinding
and the original component smaller than γ.
If we want “to explain” the first q principal components at once, we
propose to fix an angle, γ, and choose d as the smallest value for which
Ĩq,0 (d) makes the largest of the q angles smaller than γ. In both cases, the
angle γ must be chosen by the user. As default, we propose to fix an angle
not larger than 25 degrees.
2.2 Empirical version. Consistent estimates of the optimal subset
We aimed to consistently estimate the sets I1 , . . . , Iq from a sample X1 , . . . , Xn
of i.i.d. random vectors with distribution P.
Given a subset I ∈ Id , the first step is to build a sample Y1I , . . . , YnI of
random vectors in Rp , which only depends on X[I], using nonparametric
estimates of the conditional expectation (the regression function). Below,
we assume that
Assumption H2 For all i ∈
/ I, gni (z) is a strongly consistent estimate
of g i (z) uniformly in z. Conditions under which H2 holds can be found in
Hansen (2008).
First, we define the empirical version of the blinded observations. As an
example, we consider the nearest neighbours estimator. We therefore set an
integer value r (the number of nearest neighbours that we are going to use)
with respect to an appropriate metric (typically Euclidean or Mahalanobis
distance), only considering the coordinates on I.
More precisely, for each j ∈ {1, . . . , n}, we find the set of indices Cj =:
Cj (Xj [I]) of the r-nearest neighbours of Xj [I] among {X1 [I], . . . , Xn [I]},
where Xj [I] = {Xj [i], i ∈ I}.
Next, we define the random vectors YjI , 1 ≤ j ≤ n, verifying:
YjI [i]
=
1
r
P Xj [i]
m∈Cj Xm [i]
if i ∈ I
otherwise,
where Xj [i] stands for the ith-coordinate of the vector Xj .
(2.5)
7
Searching for the core variables in principal components analysis
This corresponds to use for i ∈
/ I the nonparametric estimate
gni (z) = gni (z[I]) =
1
r
X
Xm [i],
m∈Cj (z[I])
since gni (z) is a strongly consistent estimate of g i (z), where g i (z) =
E(X[i]|X[I] = z). Moreover, observe that the notation z[I] is to emphasize that the function depends on the coordinates indicated by the subset
I.
Pn stands for the empirical distribution associated with {Xj , 1 ≤ j ≤ n}
and Pn,YI stands for the empirical distribution of {YjI , 1 ≤ j ≤ n}.
In our examples to consistently estimate the optimal number of nearest
neighbours we use the generalized cross validation procedure proposed by
Li & Gong (1987), i.e.
2
1 Pn
j=1 (Xj [i] − Yj [i])
n
rbopt (i, I) = arg min
, i∈
/ I.
2
r
1 − 1r
When d is large, nonparametric estimators perform poorly due to the
curse of dimensionality. In this case, a semi-parametric approach should be
used like that proposed by He & Shi (1996). Also a recent proposal by Biau
et al. (2013) called COBRA can be used to avoid the curse of dimensionality.
The consistency result given in the following theorem will still be valid as
long as the semi-parametric estimates verify the consistency assumptions
required for the purely nonparametric estimates.
We will be mainly interested in cases where d is small. In our experience a
good idea is to start the search with some genetic algorithm which provides
an initial solution with a few variables (d small) and then improve the result
working with card(I) ≤ d. Also a forward-backward algorithm as the one
proposed in Fraiman et al. (2008) can be used.
Finally, we define, for each I ∈ Id , the corresponding empirical versions
α̂nk (I) := αk (Pn,YI ), ĥkn (I) = hk (I, Pn , Pn,YI ), Iˆkn = argminI∈Id ĥkn (I) (2.6)
and
ĥn (I) =
q
X
k=1
respectively.
pk ĥkn (I), Iˆn = argminI∈Id ĥn (I)
(2.7)
8
Gimenez and Giussani
Let us observe that α̂nk (I) corresponds to the principal component associated with the sample {YjI , 1 ≤ j ≤ n}. ĥkn (I) corresponds to the objective
function of the k-th principal component, it measures the square distance
between the k-th original principal component and the k-th principal component that is a function of the variables in the subset I. The function ĥkn (I)
determines which are the variables that retain most of the information related to the k-th principal component. The variables we are choosing are
the ones that minimize the function ĥkn (I). The subset Iˆkn indicates which
this variables are. In case we are looking for a unique subset I to explain
the first q principal components, we consider the function ĥn (I).
A robust version can be obtained using robust principal components (see
for instance Maronna et al. (2006)) and replacing in (2.5) the local mean by
local median.
Theorem 2.1. Under assumptions H1 for X and YI and H2 we have that
Iˆkn ∈ Ik,0 ultimately for k = 1, . . . , q, i.e., Iˆkn = Ik with Ik ∈ Ik,0 ∀n >
n0 (ω), with probability one. We also have that Iˆn ∈ Ĩq,0 ultimately.
Proof. The proof is given in the Appendix.
3 Some simulated examples
In this section we consider two simulated experiments to analyze the behavior of our method and compare it with other methods proposed in the
literature.
3.1 Example 1
To better understand the heart of our procedure we start with a simple
simulation example in dimension four.
There are two “hidden” factors:
V1 ∼ N (0, 1.252 ),
V2 ∼ N (0, 0.552 ),
where V1 , V2 are independent.
Searching for the core variables in principal components analysis
9
We construct the 4 observable variables as follows:
X1 = V1 + 1 ,
(3.1)
X2 = |V1 | + 2 ,
X3 = V2 + 3 ,
X4 = V1 V2 + 4 ,
where εj , 1 ≤ j ≤ 4 are i.i.d. N (0, σ 2 ), with σ = 0.01 (Case 1), σ = 0.1
(Case 2) and σ = 0.25 (Case 3).
When we perform the variable selection it is clear that two variables containing the information given by V1 and V2 should be kept. This means that
{X1 , X3 }, {X1 , X4 } or {X3 , X4 } would be “good” choices while {X1 , X2 }
is clearly a “bad” choice since only retains information of V1 . Additionally,
{X2 , X3 } (respec. {X2 , X4 }) is not a “good” choice since can not recover all
the information of V1 (respec. V1 and V2 ).
In this example, in all the cases, we compare our procedure with other
proposals: algorithms B2 and B4 (Jolliffe (2002)), a variable selection approach proposed by McCabe (1984) and sparce PCA introduced by Zou
et al. (2006). To retain q variables, algorithm B2 associates one of the original variables to each of the p−q last PCA vectors and deletes those variables,
while B4 associates one of the original variables to each of the first q PCA
vectors and retains those variables.
For each case, we perform 500 replicates and on each of them we generate
samples of size 100 from model (3.1) and compute the covariance matrix.
• Case 1: The first two principal components explain between 70 and
85 percent of the total variance. For the BP 85% of the times the
maximum angle conformed by the first two principal components is
smaller than 25 degrees, hence we keep two variables.
The results exhibitted in Table 1 on page 9 show that 78% of the
times the BP makes a “good” choice of variables, B2 only does it 27%
of the times, while the other methods fail in more than 75% of the
replications.
• Case 2: The first two principal components explain between 69 and 85
percent of the total variance. For the BP more than 78% of the times
the maximum angle conformed by the first two principal components
is smaller than 25 degrees, hence we keep two variables.
The results exhibitted in Table 2 on page 10 show that that 75% of
the times the BP makes a “good” choice of variables, B2 only does
it 29% of the times, B4 and SPCA 27%, while McCabe fails in more
than 75% of the replications.
10
Gimenez and Giussani
Table 1 Proportion of times where each method selects each pair of variables.(Case 1)
X1 , X2
X1 , X3
X1 , X4
X2 , X3
X2 , X4
X3 , X4
BP
0.22
0.37
0.366
0
0
0.044
B2
0.732
0.004
0.264
0
0
0
B4
0.76
0.002
0.238
0
0
0
McCabe
0.782
0.002
0.216
0
0
0
SPCA
0.76
0.002
0.238
0
0
0
Table 2 Proportion of times where each method selects each pair of variables.(Case 2)
X1 , X2
X1 , X3
X1 , X4
X2 , X3
X2 , X4
X3 , X4
BP
0.248
0.354
0.368
0
0
0.03
B2
0.714
0.002
0.284
0
0
0
B4
0.728
0.002
0.27
0
0
0
McCabe
0.752
0.002
0.246
0
0
0
SPCA
0.728
0.002
0.27
0
0
0
• Case 3: The first two principal components explain between 67 and
82 percent of the total variance. For the BP, 69% of the times the
maximum angle conformed by the first two principal components is
smaller than 25 degrees, hence we keep two variables.
The results exhibitted in Table 3 on page 10 show that 74% of the
times the BP makes a “good” choice of variables, B2 does it 51% of
the times, B4 and SPCA 48%, while McCabe 46%.
Table 3 Proportion of times where each method selects each pair of variables.(Case 3)
X1 , X2
X1 , X3
X1 , X4
X2 , X3
X2 , X4
X3 , X4
BP
0.264
0.192
0.54
0
0
0.004
B2
0.488
0.002
0.51
0
0
0
B4
0.516
0
0.484
0
0
0
McCabe
0.536
0
0.464
0
0
0
SPCA
0.516
0
0.484
0
0
0
From the definition of X1 and X2 we have that X2 is a function of X1 plus
an error. We can see from Table 1 on page 9, Table 2 on page 10 and Table
3 on page 10 that our procedure selects X1 and X2 only between 0.22 and
0.264 proportion of the times, when the other procedures select it in around
a 0.5 proportion of the times or more. In most of the cases our procedure
detects that X2 is a “function” of X1 . That is, if X1 is selected to be one
of the two explanatory variables, for the other one, the BP, select between
Searching for the core variables in principal components analysis
11
X3 and X4 . This way, the BP gains information about V2 instead of getting
redundant information by choosing X2 . Note that {X3 , X4 } is also a “good”
choice.
To make the problem a bit more challenging we enlarge the dimension of
the dataset repeating the same variables plus noisy noninformative errors
and also add some i.i.d. noisy variables. More precisely we consider the
following model:
V1 + ε j ,
|V1 | + εj ,
V 2 + εj ,
Xj =
V V + εj ,
1 2
εj ,
if 1 ≤ j ≤ 5,
if 6 ≤ j ≤ 10,
if 11 ≤ j ≤ 15,
if 16 ≤ j ≤ 20,
if 21 ≤ j ≤ 23,
where εj , 1 ≤ j ≤ 23 are i.i.d. N (0, σ 2 ), with σ = 0.01 (Case 1), σ = 0.1
(Case 2) and σ = 0.25 (Case 3).
We keep the first and second principal component because on 98% of the
500 replicates they explain between 70% and 85% of the total variance in
Case 1, between 70% and 83% in Case 2 and between 65% and 78% in Case
3. Again we require all the methods to select two variables. In the three
cases, at least 93% of the times the largest angle is smaller than 25 degrees,
hence we consider it is a good decision to look for two variables.
There are five groups of variables A1 , . . . , A5 . Within each group all the
variables only differ on a noisy noninformative error. More precisely, the
first four groups are Aj = {Xk , k = 5j − i, i = 0, . . . , 4} (j = 1, . . . , 4) and
the last one is A5 = {Xk , k = 21, 22, 23}. Clearly {{Aj , Aj }, j = 1, . . . 5},
{A1 , A2 } and {{Aj , A5 }, j = 1, . . . 4} are “bad” choices and also {A2 , A3 }
and {A2 , A4 } are not good choices either.
Table 4 on page 11, Table 5 on page 12, and Table 6 on page 12
exhibit the proportion of replicates where each method selects one variable
per group of variables. In the three cases the BP achieves the desired results
more than 90% of the times, while the other methods fail more than 73% of
the times.
3.2 Example 2
Zou et al. (2006) introduced the following simulation example. They have
two “hidden” factors:
V1 ∼ N (0, 290),
V2 ∼ N (0, 300),
12
Gimenez and Giussani
Table 4 Proportion of replicates where each method selects one variable per group of
variables.(Case1)
A1 , A1
A1 , A2
A1 , A3
A1 , A4
A1 , A5
A2 , A2
A2 , A3
A2 , A4
A2 , A5
A3 , A3
A3 , A4
A3 , A5
A4 , A4
A4 , A5
A5 , A5
BP
0.012
0.026
0.214
0.57
0.028
0
0
0
0
0
0.124
0
0
0
0.026
B2
0.184
0.202
0.08
0.114
0.086
0.076
0.048
0.058
0.032
0.016
0.02
0.016
0.036
0.026
0.006
B4
0
0.744
0.004
0.252
0
0
0
0
0
0
0
0
0
0
0
McCabe
0
0.784
0
0.216
0
0
0
0
0
0
0
0
0
0
0
SPCA
0
0.744
0.004
0.252
0
0
0
0
0
0
0
0
0
0
0
Table 5 Proportion of replicates where each method selects one variable per group of
variables.(Case2)
A1 , A1
A1 , A2
A1 , A3
A1 , A4
A1 , A5
A2 , A2
A2 , A3
A2 , A4
A2 , A5
A3 , A3
A3 , A4
A3 , A5
A4 , A4
A4 , A5
A5 , A5
BP
0.006
0.024
0.254
0.598
0.026
0
0
0
0
0
0.092
0
0
0
0
B2
0.188
0.218
0.07
0.114
0.076
0.074
0.044
0.072
0.03
0.014
0.026
0.012
0.034
0.022
0.006
B4
0
0.744
0.006
0.25
0
0
0
0
0
0
0
0
0
0
0
McCabe
0
0.79
0
0.21
0
0
0
0
0
0
0
0
0
0
0
SPCA
0
0.74
0.006
0.25
0
0
0
0
0
0
0
0
0
0
0
and a lineal combination of them,
V3 = −0.3V1 + 0.925V2 + ε
where V1 , V2 and ε are independent, ε ∼ N (0, 1).
Then 10 observable variables are constructed as follows:
Searching for the core variables in principal components analysis
13
Table 6 Proportion of replicates where each method selects one variable per group of
variables.(Case3)
A1 , A1
A1 , A2
A1 , A3
A1 , A4
A1 , A5
A2 , A2
A2 , A3
A2 , A4
A2 , A5
A3 , A3
A3 , A4
A3 , A5
A4 , A4
A4 , A5
A5 , A5
BP
0.028
0.004
0.292
0.604
0.026
0
0
0
0
0
0.046
0
0
0
0
B2
0.17
0.234
0.094
0.104
0.084
0.072
0.032
0.07
0.03
0.022
0.024
0.012
0.04
0.006
0.006
V1 + εj ,
V2 + εj ,
Xj =
V3 + εj ,
B4
0
0.736
0.008
0.256
0
0
0
0
0
0
0
0
0
0
0
McCabe
0
0.762
0.004
0.234
0
0
0
0
0
0
0
0
0
0
0
SPCA
0
0.728
0.008
0.256
0
0
0
0
0
0
0
0
0
0
0
if 1 ≤ j ≤ 4,
if 5 ≤ j ≤ 8,
if j = 9, 10,
where εj , 1 ≤ j ≤ 10 are i.i.d. N (0, 1).
Zou et al. (2006) used the true covariance matrix of (X1 , . . . , X10 ) to
perform PCA, SPCA and Simple Thresholding.
There are three groups of variables that share the same information. The
first group, which we denote by A1 corresponds to the variables X1 to X4 ,
the second one, A2 , to the variables X5 to X8 , and the third group, A3 , to
the variables X9 and X10 . By definition V3 is also a function of V1 and V2 ,
up to a noisy noninformative error.
Zou et al. (2006) show that SPCA correctly identifies the sets of important
variables, the first SPCA identifies the group of the variables A2 and the
second SPCA the group of the variables A1 . Simple Thresholding incorrectly
mixes two variables of A2 with two of A3 .
We first consider the population version, using the true covariance matrix,
where the first two principal components explain 99.6% of the total variance.
This example is outside of the theoretical framework stated in section 2
but this will not be a problem for the implemetation of our procedure.
Clearly it is not adecuate to keep one variable because when the cardinality of I is 1, ΣYI has only one positive eigenvalue and then we can not
14
Gimenez and Giussani
recover two principal components.
If we choose a subset I of cardinal 2, then ΣYI has two positive eigenvalues
of multiplicity one and then, we can recover two principal components. If we
select one variable of each of the three groups will be a “good” solution. A
“bad” solution should be to choose the 2 variables within the same group.
Table 7 Theoretical objective’s value h(I) when two variables are chosen (one of each
group).
A1 , A1
A1 , A2
A1 , A3
A2 , A2
A2 , A3
A3 , A3
h(I)
1.656
2.257 × 10−6
2.536 × 10−5
1.028
0.001
1
In Table 7 on page 14 we can see that h(I) takes large values (larger
than one) for the “bad” choices and small values (smaller than 0.001) for the
“good” choices. Moreover, for the “good” choices, the largest angle between
the original and the blinded components is less than two dergees in all cases.
Then, the BP is going to select a “good” choice. So our method performs
perfectly well when the cardinal of I is 2.
Next, we consider a more realistic situation and perform a small simulation, where we estimate the covariance matrix generating a sample of size
50 which we iterate 500 times. We perform the analysis considering the two
first principal components, which explain more than the 99% of the variance.
We apply our procedure to select two variables for the first two principal
components and in only 4.8% of the cases the larger angle was above 15
degrees and in 2.2% of the cases larger than 20 degrees.
In Table 8 on page 14 we can see that 99.2% of the times the BP makes a
“good” choice. In addition McCabe and SPCA always make a “good” choice,
B4 on 99% of the times and B2 only makes a “good” choice in 76.6% of the
times.
4 A real data example
We consider a dataset obtained from the University of California, Irvine
repository (Frank & Asuncion (2010)) as an example. This dataset contains
the values of six bio-mechanical characteristic that were used to classify
orthopaedic patients into three classes (normal, disc hernia and spondylolisthesis groups), and includes data for one hundred normal patients, sixty
Searching for the core variables in principal components analysis
15
Table 8 Proportion of times where each method selects one variable per group of
variables.
BP
B2
B4
McCabe SPCA
A1 , A1 0
0.134 0.01
0
0
A1 , A2 0.426 0.57
0.562 1
0.736
A1 , A3 0.222 0.108 0.428 0
0.264
A2 , A2 0
0.096 0
0
0
A2 , A3 0.344 0.088 0
0
0
A3 , A3 0.008 0.004 0
0
0
patients with a disc hernia, and one hundred and fifty patients with spondylolisthesis. Each patient is represented in the dataset in terms of the six
bio-mechanical attributes associated with the shape and orientation of the
pelvis and the lumbar spine, namely (a) pelvic incidence, (b) pelvic tilt, (c)
lumbar lordosis angle, (d) sacral slope, (e) pelvic radius and (f) grade of
spondylolisthesis.
We use our procedure to select one variable for the first two principal
components, which explain 85% of the variance. We give the same weight
to all components, that is pk = 1/2 in equation (2.4), and the “grade of
spondylolisthesis”, variable (f) is selected. In this case the largest angle is
7.5 degrees, hence we decide to keep one variable. Cross-validation finds
55 nearest neighbours for variables (a) and (b), 70 for variable (c), 102 for
variable (d) and 39 for variable (e), while the empirical objective function
(2.7) attains the value 0.017. We use the Euclidean distance in our procedure,
but with the Mahalanobis distance we get the same results.
50
Second component
Second component
50
0
−50
−100
0
−50
−100
−150
−100
First component
−50
0
−150
−100
−50
0
First component
Figure 1 Left: Map of the projection of the data on the original principal components.
Right: Map of the projection of the data on the blinded based principal components. Blue:
Disk Hernia patients. Black: Normal patients. Green: Spondylolisthesis patients.
16
Gimenez and Giussani
−100
−120
−140
−80
Second component
−80
Second component
Second component
−80
−100
−120
−140
−140
−120
−100
−80
−60
First component
−40
−20
0
−100
−120
−140
−140
−120
−100
−80
−60
First component
−40
−20
0
−140
−120
−100
−80
−60
−40
−20
0
First component
Figure 2 Left: Map of the projection of the data on the original principal components.
Middle: Map of the projection of the data on the blinded based principal components using
Euclidean distance. Right: Map of the projection of the data on the blinded based principal components using Mahalanobis distance. Blue: Disk Hernia patients. Black: Normal
patients.
The plot on the first two principal components for the example data looks
very similar to the plot on the principal components calculated using our
procedure (Figure 1). It also shows that the patients with spondylolisthesis
are separated from the rest of the patients, but the normal patients and the
patients with disc hernias are mixed up together.
Next, we calculate the two principal components using the traditional
method and our procedure for the subset of normal and disc hernia patients,
using now two variables. The first two principal components explain 75% of
the variance then we decide to keep them. We discard to select one variable
because in that case the largest angle is close to 78 degrees. For two variables
the largest angle is close to 21 degrees and for three variables the angle does
not decrease very much (19 degrees), so we decide to keep two variables.
We consider both, the Euclidean and the Mahalanobis distance, and in
both cases the variables selected are “lumbar lordosis angle” and “pelvic
radius”. The value of the objective function (2.7) with the Euclidean distance
is 0.125, while with the Mahalanobis distance is 0.141. Figure 2 shows that
the plots for the principal components produced by these procedures look
very similar.
5 Conclusions
In this paper we introduce a new variable selection procedure for PCA that
aim to gain interpretation in principal components.
On one hand we consider a local approach where we deals with finding a
subset of variables that best explains each principal component, and on the
other hand we analyzed a global approach when the objective is to find a
Searching for the core variables in principal components analysis
17
unique subset to explain the first q principal components at once.
We explained how through the conditional expectation we can choose the
variables.
Under wide regular conditions in the conditional expectation estimates
and on the covariance matrix, strong consistency results are studied and
numerical aspects have also been analysed.
The performance of the procedure has been compared with several wellknown variable selection techniques using real and simulated data sets, showing the strengths of the method.
Appendix: Proof of Theorem 2.1.
First we are going to proof the following Proposition:
Proposition. For k ∈ {1, ..., p}, if
ĥkn (I) → hk (I) a.s. for all I ∈ Id ,
(5.1)
then it converges uniformly, and
arg min ĥkn (I) → arg min hk (I) a.s.
I∈Id
I∈Id
Proof. Since Id is a discrete and finite space , the convergence of (5.1) is
uniform.
For the sake of simplicity, let k = 1 (the proof is analogue for k ∈
{2, ..., p}).
For all δ > 0 there exists n0 (ω) such that for every n ≥ n0 (ω), with
probability one
ĥ1n (I) − h1 (I) < δ for all I ∈ Id ,
then,
h1 (I) − δ < ĥ1n (I) < h1 (I) + δ for all I ∈ Id
(5.2)
ĥ1n (I) − δ < h1 (I) < ĥ1n (I) + δ for all I ∈ Id .
(5.3)
and
From (2.2), we know that exists δ0 > 0 such that
18
Gimenez and Giussani
h1 (I1 ) < h1 (I) − δ0 for all I ∈
/ I1,0 , for all I1 ∈ I1,0 .
By choosing in (5.2) δ =
ĥ1n (I1 ) < h1 (I1 ) +
δ0
2
we obtain
δ0
δ0
δ0
< h1 (I) − δ0 +
= h1 (I) −
< ĥ1n (I)
2
2
2
for all I ∈
/ I1,0 , for all I1 ∈ I1,0 , if n ≥ n0 (ω) with probability one.
That means, that there exists n0 (ω) such that, if n ≥ n0 (ω), with probability one, if we choose I1 ∈ I1,0 , then I1 minimizes ĥ1n (I).
Moreover, from (2.6), we have that for each n fix, exists δ1 > 0, such that
ĥ1n (I 1 ) < ĥ1n (I) − δ1 for all I ∈
/ Iˆ1n , for all I 1 ∈ Iˆ1n .
Let δ =
δ1
2
in (5.3) and n ≥ n1 (ω),
h1 (I 1 ) < ĥ1n (I 1 ) +
δ1
δ1
δ1
< ĥ1n (I) − δ1 +
= ĥ1n (I) −
< h1 (I)
2
2
2
for all I ∈
/ Iˆ1n , and for all I 1 ∈ Iˆ1n with probability one.
That means, that exists n1 (ω) such that, if n ≥ n1 (ω) for all I 1 ∈ Iˆ1n , I 1
minimizes h1 (I) with probability one.
In conclusion, if n > max{n0 , n1 }, for all I1 ∈ I1,0 , I1 minimizes ĥ1n (I)
and for all I 1 ∈ Iˆ1n , I 1 minimizes h1 (I) with probability one.
As the proof is analogue for k ∈ {2, ..., p},
Ik,0 = argminI∈Id hk (I),
and
Iˆkn = argminI∈Id ĥkn (I)
this implies,
arg min ĥkn (I) → arg min hk (I) a.s. for k ∈ {1, ..., p}.
I∈Id
I∈Id
Searching for the core variables in principal components analysis
19
Proof. (of Theorem)
To prove the Theorem 2.1 then it is enough to see that for each I the
empirical objective function (2.6) converges a.s. to the theoretical objective
function (2.3). To prove it, let us see that
α̂nk (Pn ) − α̂nk (Pn,YI ) → αk (P) − αk (PYI )
a.s. for all k.
Dauxois et al. (1982) proved that under H1, if
= max Σ̂n − Σ (u) → 0 a.s.,
Σ̂n − Σ
∞
kuk=1
then
α̂nk (Pn ) − αk (P) → 0 a.s. for all 1 ≤ k ≤ p,
where α̂nk (Pn ) (respec. αk (P)) are the eigenvectors of Σ̂n , that denotes the
empirical covariance matrix associated with Pn (respec. Σ denotes the covariance matrix associated with P).
So, if we prove that
max Σ̂n (I) − Σ(I) (u) → 0 a.s.
kuk=1
then
kα̂nk (Pn,YI ) − αk (PYI )k → 0 a.s. ∀1 ≤ k ≤ p,
where α̂nk (Pn,YI ) (respec. αk (PYI )) are the eigenvectors of Σ̂n (I), that denotes the empirical covariance matrix associated with Pn,YI (respec. Σ(I)
denotes the covariance matrix associated with PYI ).
Note, that I = {1, . . . , p} is the classic case. We will show that this also
holds for any I ⊆ {1, . . . , p}, that is
max Σ̂n (I) − Σ(I) (u) → 0 a.s.
kuk=1
To simplify the notation, we can assume (without losing generality) that
I = {1, . . . , d}, then
t
Yj = Xj [1], . . . , Xj [d], gn1 (Xj [I]), . . . , gnp−d (Xj [I]) ,
where gni (z) is a uniformly consistent nonparametric estimate of g i (z). Specifically, gni will fulfil the following assumption,
gnl (Xj [I]) → g l (Xj [I]) a.s. for any l, for all j, uniformly.
20
Gimenez and Giussani
We define a non observable auxiliary vector,
t
Zj = Xj [1], . . . , Xj [d], g 1 (Xj [I]), . . . , g p−d (Xj [I]) ,
where g i (z) = E(X[d + i]|X[I] = z).
The proof will be complete if we show that
max k(ΣZ − Σ(I)) (u)k → 0 a.s.,
(5.4)
max k(ΣY − ΣZ ) (u)k → 0 a.s.,
(5.5)
kuk=1
and that
kuk=1
where ΣY = Σ̂n (I).
Considering,
X[i] =
n
1X
Xj [i],
n j=1
gnl (X[I]) =
n
1X l
gn (Xj [I]),
n j=1
g l (X[I]) =
n
1X l
g (Xj [I])
n j=1
the three matrices are
(Σ(I))ii0
(ΣY )ii0 =
(ΣZ )ii0 =
cov(X[i], X[i0 ]) if 1 ≤ i, i0 ≤ d,
cov(X[i], E(X[i0 ]|X[I])) if 1 ≤ i ≤ d < i0 ≤ p,
=
cov(E(X[i]|X[I]), X[i0 ]) if 1 ≤ i0 ≤ d < i ≤ p,
cov(E(X[i]|X[I]), E(X[i0 ]|X[I])) if d + 1 ≤ i, i0 ≤ p,
1
n
Pn
j=1 (Xj [i]
− X[i])(Xj [i0 ] − X[i0 ]) if 1 ≤ i, i0 ≤ d,
0
i0 −d
1
(Xj [I]) − gni −d (X[I])) if 1 ≤ i ≤ d < i0 ≤ p,
j=1 (Xj [i] − X[i])(gn
n
Pn
i−d
0
0
i−d
1
0
j=1 (gn (Xj [I]) − gn (X[I]))(Xj [i ] − X[i ]) if 1 ≤ i ≤ d < i ≤ p,
n
P
0
0
n
i−d
i −d
i−d
1
(Xj [I]) − gni −d (X[I])) if d + 1 ≤ i, i0
j=1 (gn (Xj [I]) − gn (X[I]))(gn
n
Pn
Pn
0
0
0
j=1 (Xj [i] − X[i])(Xj [i ] − X[i ]) if 1 ≤ i, i ≤ d,
Pn
0
i0 −d
1
(Xj [i] − X[i])(g
(Xj [I]) − g i −d (X[I])) if 1 ≤ i ≤ d < i0 ≤ p,
n Pj=1
n
i−d
1
(Xj [I]) − g i−d (X[I]))(Xj [i0 ] − X[i0 ]) if 1 ≤ i0 ≤ d < i ≤ p,
j=1 (g
n
P
0
n
i−d
1
(Xj [I]) − g i−d (X[I]))(g i −d (Xj [I]) − g i0 −d (X[I])) if d + 1 ≤ i, i0
j=1 (g
n
1
n
≤ p,
We will now prove the convergence in (5.4) and (5.5).
First, we show that
max k(ΣZ − Σ(I)) (u)k → 0 a.s.
kuk=1
≤ p.
21
Searching for the core variables in principal components analysis
It is sufficient to prove that each of the coordinates of the matrix ΣZ −Σ(I)
converge to zero.
Set 1 ≤ i, i0 ≤ p, then (ΣZ )ii0 → (Σ(I))ii0 a.s. and then (ΣZ − Σ(I))ii0 →
0 a.s. ∀i, i0 = 1, ..., p.
Because we are using a finite dimensional space, and each of the coordinates converge to zero, it holds that
max k(ΣZ − Σ(I)) (u)k → 0 a.s.
kuk=1
Now, we show that
max k(ΣY − ΣZ ) (u)k → 0 a.s.
kuk=1
As before, we are going to prove that each of the coordinates of the matrix
ΣY − ΣZ converge to zero.
For better understanding, we define
G(l, j, n) = gnl (Xj [I]) − gnl (X[I]) − g l (Xj [I]) + g l (X[I]).
• If 1 ≤ i, i0 ≤ d, (ΣY )ii0 = (ΣZ )ii0 then (ΣY − ΣZ )ii0 = 0.
• If 1 ≤ i ≤ d, d + 1 ≤ i0 ≤ p,
n
(ΣY − ΣZ )ii0
=
=
1X
0
0
0
(Xj [i] − X[i])(gni −d (Xj [I]) − gni −d (X[I]) − g i −d (Xj [I]) + g i0 −d (X[I]))
n
1
n
j=1
n
X
(Xj [i] − X[i])G(i0 − d, j, n).
j=1
From the Cauchy-Schwarz inequality, we have that
1
1
2
2
n
n
n
X
X
1X
1
1
0
2
0
2
(Xj [i] − X[i])G(i − d, j, n) ≤
(Xj [i] − X[i])
(G(i − d, j, n))
n
n
n
j=1
j=1
j=1
On the other hand,
n
1X
(Xj [i] − X[i])2 → V ar(X[i]) a.s.
n
j=1
(5.6)
22
Gimenez and Giussani
In addition, since
gnl (Xj [I]) → g l (Xj [I]) a.s. uniformly, for any l, for all j,
then for any l,
n
gnl (X[I])
−
g l (X[I])
1X l
=
gn (Xj [I]) − g l (Xj [I]) ≤ max {gnl (Xj [I]) − g l (Xj [I])} → 0 a.s.
1≤j≤n
n
j=1
and therefore,
G(l, j, n) → 0 a.s. for any l, for all j.
This implies
n
1X
(G(l, j, n))2 → 0 a.s. for any l.
n
(5.7)
j=1
Thus, by (5.6) and (5.7) we have
n
1X
(Xj [i] − X[i])G(i0 − d, j, n) → 0 a.s.
n
(5.8)
j=1
which is what we required for our proof.
• If d + 1 ≤ i ≤ p, 1 ≤ i0 ≤ d, from the simetry of the matrices and (5.8),
we have that (ΣY − ΣZ )ii0 → 0 a.s.
• If d + 1 ≤ i ≤ p, d + 1 ≤ i0 ≤ p,
(ΣY − ΣZ )ii0
=
n
i
1 X h i−d
0
0
(gn (Xj [I]) − gni−d (X[I]))(gni −d (Xj [I]) − gni −d (X[I])) −
n
j=1
h
i
0
(g i−d (Xj [I]) − g i−d (X[I]))(g i −d (Xj [I]) − g i0 −d (X[I])) .
0
0
If we add and subtract (g i−d (Xj [I])−g i−d (X[I]))(gni −d (Xj [I])−gni −d (X[I])),
and rearrange, we get that (ΣY − ΣZ )ii0 is the sum of
n
1 X i0 −d
0
(gn (Xj [I]) − gni −d (X[I]))G(i − d, j, n)
n
j=1
(5.9)
23
Searching for the core variables in principal components analysis
and
n
1 X i−d
(g (Xj [I]) − g i−d (X[I]))G(i0 − d, j, n)
n
j=1
were (5.9) can be rewrited as
n
1X
0
[G(i0 − d, j, n) + (g i −d (Xj [I]) − g i0 −d (X[I]))]G(i − d, j, n)
n
j=1
or
n
n
j=1
j=1
1 X i0 −d
1X
G(i0 − d, j, n)G(i − d, j, n) +
(g
(Xj [I]) − g i0 −d (X[I]))G(i − d, j, n).
n
n
Then we have that
n
(ΣY − ΣZ )ii0
=
1X
G(i0 − d, j, n)G(i − d, j, n)
n
j=1
n
X
+
1
n
+
n
1 X i−d
(g (Xj [I]) − g i−d (X[I]))G(i0 − d, j, n).
n
0
(g i −d (Xj [I]) − g i0 −d (X[I]))G(i − d, j, n)
j=1
j=1
From triangular and Cauchy-Schwarz inequality,
1/2
1/2
n
n
X
X
1
1
|(ΣY − ΣZ )ii0 | ≤
(G(i0 − d, j, n))2
(G(i − d, j, n))2
n
n
j=1
j=1
1/2
1/2
n
n
X
X
1
1
0
+
(g i −d (Xj [I]) − g i0 −d (X[I]))2
(G(i − d, j, n))2
n
n
j=1
n
j=1
1 X i−d
(g (Xj [I]) − g i−d (X[I]))2
+
n
j=1
1/2
n
X
1
(G(i0 − d, j, n))2
n
1/2
j=1
On the other hand, we have that
n
1X l
(g (Xj [I]) − g l (X[I]))2 → V ar(g l (X[I])) a.s. for any l.
n
j=1
24
Gimenez and Giussani
If V ar(g l (X[I])) < ∞ for any l and I, (5.7) entails |(ΣY − ΣZ )ii0 | → 0.
As V ar(X[l]) < ∞, then for any l and any I,
V ar(X[l]) = E(V ar(X[l]|X[I]))+V ar(E(X[l]|X[I])) = E(V ar(X[l]|X[I]))+V ar(g l (X[I]))
We now that E(V ar(X[l]|X[I])) > 0, so we conclude that V ar(g l (X[I])) <
∞
We have proved that all the coordinates of the matrix ΣY − ΣZ converge
to zero, so
sup k(ΣY − ΣZ ) (u)k → 0 a.s.
kuk=1
Now we are ready to complete the proof of Theorem 2.1. From (5.4) and
(5.5) we are able to use the result from Dauxois et al. (1982) to derive that,
ĥkn (I) = kα̂nk (Pn )− α̂nk (Pn,YI )k2 → kαk (P)−αk (PYI )k2 = hk (I) a.s. (5.10)
Indeed, we have that
kα̂nk (Pn ) − α̂nk (Pn,YI )k − kαk (P) − αk (PYI )k ≤
kα̂nk (Pn ) − αk (P)k + kαk (P) − αk (PYI )k +
kαk (PYI ) − α̂nk (Pn,YI )k − kαk (P) − αk (PYI )k =
kα̂nk (Pn ) − αk (P)k + kαk (PYI ) − α̂nk (Pn,YI )k → 0 a.s.
and
kαk (P) − αk (PYI )k − kα̂nk (Pn ) − α̂nk (Pn,YI )k ≤
kαk (P) − α̂nk (Pn )k + kα̂nk (Pn ) − α̂nk (Pn,YI )k +
kα̂nk (Pn,YI ) − αk (PYI )k − kα̂nk (Pn ) − α̂nk (Pn,YI )k =
kαk (P) − α̂nk (Pn )k + kα̂nk (Pn,YI ) − αk (PYI )k → 0 a.s.
which entails (5.10).
Finally, (5.10) and the Proposition implies
arg min ĥkn (I) → arg min hk (I) a.s.
I∈Id
I∈Id
Searching for the core variables in principal components analysis
25
which concludes the proof.
Acknowledgements
This work has been partially supported by Grant pict 2008-0921 from ANPCyT (Agencia Nacional de Promoción Cientı́fica y Tecnológica), Argentina.
We are very grateful to Dr. Ricardo Fraiman for his invaluable assistance
and Dra. Marcela Svarc for her helpful suggestions. The authors would like
to thank the referees for their constructive comments which improve significantly this work.
References
Biau, G., Fischer, A., Guedj, B. and Malley, J.D. (2013), “COBRA: A collective regression
strategy”, arxiv.org/pdf/1303.2236.
Dauxois, J., Pousse, A. and Romain, Y., (1982),“Asymptotic theory for the principal component analysis of a vector random function: some applications to statistical inference”,
Journal of Multivariate Analysis 12, 136–154.
Fraiman, R., Justel, A. and Svarc, M., (2008), “Selection of variables for cluster analysis
and classification rules”, Journal of the American Statistical Association. 103(483),
1294–1303.
Frank, A. and Asuncion, A., (2010), UCI Machine Learning Repository
[http://archive.ics.uci.edu/ml], Irvine, CA: University of California, School of
Information and Computer Science.
Hansen, B.E. (2008), “Uniform convergence rates for kernel estimation with dependent
data”, Econometric Theory. 24, 726–748.
He, X. and Shi, P., (1996), “Bivariate Tensor-Product B-Splines in Partly Linear Models”,
Journal of Multivariate Analysis. 58, 162–181.
Jolliffe, I., (1995), “ Rotation of principal components: choise of normalization constraints”, Journal of Applied Statistics.22, 29–35.
Jolliffe, I. (2002), “Principal Components Analysis. Second Edition”, Ed. Springer.
Jolliffe, I., Trendafilov, N. and Uddin, M. (2003), “ A modified principal component technique based on the LASSO”, Journal of Computational and Graphical Statistics. 12,
531–547.
Li, R. and Gong, G. (1987), “K-nn Nonparametric Estimation Of Regression Functions
In the Presence of Irrelevant Variables”, Econometrics Journal, 00, 1-12.
Luss, R. and d’Aspremont, A., (2010), “Clustering and feature selection using sparse
principal component analysis”, Optimization and Engineering. 11(1), 145–157.
Maronna, R. A., Martin, R. D. and Yohai, V. J., (2006), “Robust Statistics: Theory and
Methods”, Wiley, London.
McCabe, G. P. (1984), “Principal Variables”, Technometrics. 26, 137–144.
26
Gimenez and Giussani
Tibshirani, R., (1996), “Regression shrinkage and selection via the LASSO”, Journal of
the Royal statistical society. Series B 58(1), 267–288.
Vines,S., (2000), “Simple principal components”, Applied Statistics. 49, 441–451.
Witten, D.M. and Tibshirani, R., (2008), “Testing significance of features by lassoed principal components”, Annals of Applied Statistics. 2(3), 986–1012.
Zou, H. and Hastie, T., (2005), “‘Regularizations and variable selection via the elastic
net”, Journal of the Royal Statistical Society. Series B, 67, 301–320.
Zou, H., Hastie, T. and Tibshirani, R. (2006), “Sparse principal component analysis”,
Journal of Computational and Graphical Statistics, 15, 265–286.
Universidad de San Andrés, Vito Dumas 284, 1644 Victoria, Buenos Aires, Argentina and CONICET
E-mail: [email protected]; [email protected]
| 10math.ST
|
A Pyramidal Evolutionary Algorithm with Different Inter-Agent
Partnering Strategies for Scheduling Problems
in GECCO 2001: Proceedings of the Genetic and Evolutionary Computation Conference, latebreaking papers volume, pp 1-8, San Francisco, USA, 2001.
Uwe Aickelin
School of Computer Science
University of Nottingham
NG8 1BB UK
[email protected]
Abstract
This paper combines the idea of a hierarchical
distributed genetic algorithm with different interagent partnering strategies. Cascading clusters of
sub-populations are built from bottom up, with
higher-level sub-populations optimising larger
parts of the problem. Hence higher-level subpopulations search a larger search space with a
lower resolution whilst lower-level subpopulations search a smaller search space with a
higher resolution. The effects of different partner
selection schemes amongst the agents on
solution quality are examined for two multiplechoice optimisation problems. It is shown that
partnering strategies that exploit problemspecific knowledge are superior and can counter
inappropriate (sub-) fitness measurements.
1
INTRODUCTION
When hierarchically distributed evolutionary algorithms
are combined with multi-agent structures a number of
new questions become apparent. One of these questions is
addressed in this paper: the issue of intelligently selecting
mating partners from another population for an agent.
This paper will look at seven different partnering
strategies when combined with a genetic algorithm that
uses a co-operative sub-population structure. We will
evaluate the different strategies according to their
optimisation performance of two scheduling problems.
Genetic algorithms are generally attributed to Holland
[1976] and his students in the 1970s, although
evolutionary computation dates back further (refer to
Fogel [1998] for an extensive review of early
approaches). Genetic algorithms are stochastic metaheuristics that mimic some features of natural evolution.
Canonical genetic algorithms were not intended for
function optimisation, as discussed by De Jong [1993].
However, slightly modified versions proved very
successful. For an introduction to genetic algorithms for
function optimisation, see Deb [1996].
The twist when applying our type of distributed genetic
algorithm lies in its special hierarchical structure. All subpopulations follow different (sub-) fitness functions, so in
effect only searching specific parts of the solution space.
Following special crossover-operators these parts are then
gradually merged to full solutions. The advantage of such
a divide and conquer approach is reduced epistasis within
the lower-level sub-populations which makes the
optimisation task easier for the genetic algorithm.
The paper is arranged as follows: the following section
describes the nurse scheduling and tenant selection
problems. Pyramidal genetic algorithms and their
application to these two problems are detailed in section
3. Section 4 explains the seven partnering strategies
examined in the paper and section 5 describes their use
and computational results. The final section discusses all
findings and draws conclusions.
2
THE NURSE SCHEDULING AND
TENANT SELECTION PROBLEMS
Two optimisation problems are considered in this paper,
the nurse scheduling problem and the tenant selection
problem. Both have a number of characteristics that make
them an ideal testbed for the enhanced genetic algorithm
using partnering strategies. Firstly, they are both in the
class of NP-complete problems [Johnson 1998, Martello
& Toth 1990], hence they are challenging problems.
Secondly, they have proved resilient to optimisation by a
standard genetic algorithm, with good solutions only
found by using a novel strategy of indirectly optimising
the problem with a decoder based genetic algorithm
[Aickelin & Dowsland 2001]. Finally, both problems are
similar multiple-choice allocation problems. For the nurse
scheduling, the choice is to allocate a shift-pattern to each
nurse, whilst for the tenant selection it is to allocate an
area of the mall to a shop. However, as the following
more detailed explanation of the two will show, the two
problems also have some very distinct characteristics
making them different yet similar enough for an
interesting comparison of results.
The nurse-scheduling problem is that of creating weekly
schedules for wards of up to 30 nurses at a major UK
hospital. These schedules have to satisfy working
contracts and meet the demand for given numbers of
nurses of different grades on each shift, whilst at the same
time being seen to be fair by the staff concerned. The
latter objective is achieved by meeting as many of the
nurses’ requests as possible and by considering historical
information to ensure that unsatisfied requests and
unpopular shifts are evenly distributed. Due to various
hospital policies, a nurse can normally only work a subset of the in total 411 theoretically possible shift-patterns.
For instance, a nurse should work either days or nights in
a given week, but not both. The interested reader is
directed to Aickelin & Dowsland [2000] and Dowsland
[1998] for further details of this problem.
For our purposes, the problem can be modelled as
follows. Nurses are scheduled weekly on a ward basis
such that they work a feasible pattern with regards to their
contract and that the demand for all days and nights and
for all qualification levels is covered. In total three
qualification levels with corresponding demand exists. It
is hospital policy that more qualified nurses are allowed
to cover for less qualified one. Infeasible solutions with
respect to cover are not acceptable. A solution to the
problem would be a string, with the number of elements
equal to the number of nurses. Each element would then
indicate the shift-pattern worked by a particular nurse.
Depending on the nurses’ preferences, the recent history
of patterns worked, and the overall attractiveness of the
pattern, a penalty cost is then allocated to each nurseshift-pattern pair. These values were set in close
consultation with the hospital and range from 0 (perfect)
to 100 (unacceptable), with a bias to lower values. The
sum of these values gives the quality of the schedule. 52
data sets are available, with an average problem size of 30
nurses per ward and up to 411 possible shift-patterns per
nurse.
For comparison, all data sets were solved using a standard
IP package [Fuller 1998]. However, some remained
unsolved after each being allowed 15 hours run-time on a
Pentium II 200. Experiments with a number of descent
methods using different neighbourhoods, and a standard
simulated annealing implementation, were even less
successful and frequently failed to find feasible solutions.
The most successful approach to date is based on Tabu
Search [Dowsland 1998]. However, the quality of
solutions relies heavily on data-specific chains of moves
that work well because of the way in which the different
factors affecting the quality of a schedule are combined.
A straightforward genetic algorithm approach failed to
solve the problem [Aickelin & Dowsland 2000]. The best
evolutionary results to date have been achieved with an
indirect genetic approach employing a decoder function
[Aickelin & Dowsland 2001]. However, we believe that
there is further leverage in direct evolutionary approaches
to this problem. Hence we propose to use an enhanced
pyramidal genetic algorithm in this paper.
The second problem is a mall layout and tenant selection
problem; in future mall problem for short. The mall
problem arises both in the planning phase of a new
shopping centre and on completion when the type and
number of shops occupying the mall has to be decided. To
maximise revenue a good mixture of shops that is both
heterogeneous and homogeneous has to be achieved. Due
to the difficulty of obtaining real-life data because of
confidentiality, the problem and data used in this research
are constructed artificially, but closely modelled after the
actual real-life problem as described for instance in Bean
et al. [1988]. In the following, we will briefly outline our
model.
The objective of the mall problem is to maximise the rent
revenue of the mall. Although there is a small fixed rent
per shop, a large part of a shop’s rent depends on the sales
revenue generated by it. Therefore, it is important to
select the right number, size and type of tenants and to
place them into the right locations to maximise revenue.
As outlined in Bean et al. [1988], the rent of a shop
depends on the following factors:
•
The attractiveness of the area in which the shop is
located.
•
The total number of shops of the same type in the
mall.
•
The size of the shop.
•
Possible synergy effects with neighbouring similar
shops, i.e. shops in the same group (not used by Bean
et al.).
•
A fixed amount of rent based on the type of the shop
and the area in which it is located.
This problem can be modelled as follows: Before placing
shops, the mall is divided into a discrete number of
locations, each big enough to hold the smallest shop size.
Larger sizes can be created by placing a shop of the same
type in adjacent locations. Hence, the problem is that of
placing i shop-types (e.g. menswear) into j locations,
where each shop-type can belong to one or more of l
groups (e.g. clothes shops) and each location is situated in
one of k areas. For each type of shop there will be a
minimum, ideal and maximum number allowed in the
mall, as consumers are drawn to a mall by a balance of
variety and homogeneity of shops.
The size of shops is determined by how many locations
they occupy within the same area. For the purpose of this
study, shops are grouped into three size classes, namely
small, medium, and large, occupying one, two, and three
locations in one area of the mall respectively. For
instance, if there are two locations to be filled with the
same shop-type within one area, then this will be a shop
of medium size. If there are five locations with the same
shop-type assigned in the same area, then they will form
one large and one medium shop etc. Usually, there will
also be a maximum total number of small, medium and
large shops allowed in the mall.
To test the robustness and performance of our algorithms
thoroughly on this problem, 50 problem instances were
created. All problem instances have 100 locations
grouped into five areas. However, the sets differ in the
number of shop-types available (between 50 and 20) and
in the tightness of the constraints regarding the minimum
and maximum number of shops of a certain type or size.
Full details on how the data was created, its dimensions,
the differences between the sets can be found in [Aickelin
1999].
3
PYRAMIDAL GENETIC
ALGORITHMS
Both problems failed to be optimised with a standard
genetic algorithm [Aickelin & Dowsland 2000 and 2001].
Our previous research showed that the difficulties were
attributable to epistasis created by the constrained nature
of the optimisation. Briefly, epistasis refers to the ‘nonlinearity’ of the solution string [Davidor 1991], i.e.
individual variable values which were good in their own
right, e.g. a particular shift / location for a particular nurse
/ shop formed low quality solutions once combined. This
effect was created by those constraints that could only be
incorporated into the genetic algorithm via a penalty
function approach. For instance, most nurses preferred
working days; thus partial solutions with many ‘day’
shift-patterns have a higher fitness. However, combining
these shift-patterns leads to shortages at night and
therefore infeasible solutions. The situation for the mall
problem is similar yet more complex, as two types of
constraints have to be dealt with: size constraints and
number constraints.
In Aickelin & Dowsland [2000] we presented a simple
and on its own unsuccessful pyramidal genetic algorithm
for the nurse-scheduling problem. A pyramidal approach
can best be described as a hierarchical distributed genetic
algorithm where cascading clusters of sub-populations are
built from bottom up, with higher-level sub-populations
optimising larger parts of the problem. Thus, the
hierarchy is not within one string but rather between subpopulations which optimise different string-portions.
Hence, higher-level sub-populations search a larger
search space with a lower resolution whilst lower-level
sub-populations search a smaller search space with a
higher resolution. This can be applied to the nursescheduling problem in the following way:
•
Agents in sub-populations 1, 2 and 3 have their
fitness based on cover and requests only for grade 1,
2 and 3 respectively.
•
Agents in sub-populations 4, 5 and 6 have their
fitness based on cover and requests for grades (1+2),
(2+3), (3+1).
•
Agents in sub-population 7 optimise cover and
requests for (1+2+3).
•
Agents in sub-population 8 solve the original (all)
problem.
The full structure is illustrated in figure 1. Sub-solution
strings from lower populations are cascaded upwards
using suitable crossover and selection mechanisms. For
instance, fixed crossover points are used such that an
agent from sub-population (1) combined with one from
(2+3) forms a full solution. Although the full problem is
as epistatic as before, the sub-problems are less so as the
interaction between nurse grades is (partially) ignored.
Compatibility problems of combining the parts are
reduced by the pyramidal structure with its hierarchical
and gradual combining.
Using this approach improved solution quality in
comparison to a standard genetic algorithm was recorded.
However, the quality of solutions was still short of those
produced by Tabu Search. So far roulette wheel selection
based on fitness rank had been used to choose parents.
The fitness of each sub-string is calculated using a
substitute fitness measure based on the requests and cover
as detailed above, i.e. the possibility of more qualified
nurses covering for less-qualified ones is partially
ignored. Unsatisfied constraints are included via a penalty
function. This paper will investigate various partnering
strategies between the agents of the sub-populations to
improve upon these results.
population and pastes them together to form a full
solution.
all
1+2+3
1+2
1
2+3
2
3+1
all
3
1
Figure 1: Nurse Problem Pyramidal Structure.
Similar to the nurse problem, a solution to the mall
problem can be represented by a string with as many
elements as locations in the mall. Each element then
indicates what shop-type is to be located there. The mall
is geographically split into different regions, for instance
north, east, south, west and central. Some of the
objectives are regional; e.g. the size of a shop, the synergy
effects, the attractiveness of an area to a shop-type,
whereas others are global, e.g. the total number of shops
of a certain type or size.
The application of the pyramidal structure to the mall
problem follows along similar lines to that of the nurse
problem. In line with splitting the string into partitions
with nurses of the same grade, the string is now split into
the areas of the mall. Thus, we will have sub-strings with
all the shops in one area in them. These can then be
combined to create larger ‘parts’ of the mall and finally
full solutions.
However, the question arises how to calculate the
substitute fitness measure of the partial strings. The
solution chosen here will be a pseudo measure based on
area dependant components only, i.e. global aspects are
not taken into account when a substitute fitness for a
partial string is calculated. Thus, sub-fitness will be a
measure of the rent revenue created by parts of the mall,
taking into account those constraints that are area based.
All other constraints are ignored. A penalty function is
used to account for unsatisfied constraints.
Due to the complexity of the fitness calculations and the
limited overall population size, we refrained from using
several levels in the hierarchical design as we did with the
nurse scheduling. Instead a simpler two level hierarchy is
used as shown in figure 2: Five sub-populations
optimising the five areas separately (1,2,3,4, 5) and one
main population optimising the original problem (all). A
special crossover then selects one solution from each sub-
2
3
4
5
Figure 2: Mall Problem Pyramidal Structure.
The remainder of this paper will investigate ways to try to
improve on previously found poor results by suggesting
ways of combining partial strings more intelligently. An
alternative, particularly for the mall problem, would be a
more gradual build-up of sub-populations. Without
increasing the overall population size, this would lead to
more and hence smaller sub-populations. However, this
more gradual approach might have enabled the algorithm
to find good feasible solutions by more slowly joining
together promising building blocks. This is in contrast to
the relatively harsh two-level and three-level design
where building blocks had to ‘succeed’ immediately.
Exploring the exact benefits of a gradual build-up of subsolutions would make for another challenging area of
possible future research.
4
PARTNERING STRATEGIES
The problem of how to pick crossover partners has been
noted in both competitive and co-operative coevolutionary algorithms. Many strategies have been
presented in the literature as summarised for instance by
Bull [1997]. In this paper seven such strategies are
compared for their effectiveness in fighting epistasis in
the pyramidal genetic algorithm optimising the nurse
scheduling and the mall problems.
Rank-Selection (S): This is the method used so far in our
algorithms. Agents are assigned a sub-fitness score based
as closely as possible on the contribution of their partial
string to full solutions. All agents are then ranked within
each sub-population and selection follows a roulette
wheel scheme based on the ranks [e.g. Aickelin &
Dowsland 2000].
Random (R): Agents choose their mating partners
randomly from amongst all those in the sub-population
their sub-population is paired with [e.g. Bull & Fogarty
1993].
Best (B): In this strategy each agent is paired with the
currently best agent of the other sub-population(s). In case
of a tie, the agent with the lower population index is
chosen [e.g. Potter & De Jong 1994].
Distributed (D): The idea behind this approach is to match
agents with similar ones to those paired with previously
[e.g. Ackely & Littman 1994]. To achieve this each subpopulation is spaced out evenly across a single toroidal
grid. Subsequently, agents are paired with others on the
same grid location in the appropriate other subpopulations. Children created by this are inserted in an
adjacent grid location. This is said to be beneficial to the
search process because a consistent co-evolutionary
pressure emerges since all offspring appear in their
parents’ neighbourhoods [Husbands 1994]. In our
algorithms we use local mating with the neighbourhood
set to the eight agents surrounding the chosen location.
Joined (J): In nature, some species carry others internally
with the relationship propagated from generation to
generation [e.g. Iba 1996]. Thus, each agent represents a
complete solution; i.e. all the parts have been joined
together. In our case, this results in all sub-populations
solving the original problem, i.e. we have a traditional
parallel genetic algorithm. This means that all subpopulations use the full fitness function for evaluation and
rank-proportional selection.
Attractiveness (A): The five strategies described so far are
general and do not make use of problem specific
knowledge. However, there is a growing body of research
[e.g. Stanley et al. 1994, Wolpert & Macready 1995], as
well as our own previous work, which suggests that
approaches that exploit problem specific knowledge
achieve better results. Here pairing is done as for the
rank-selection strategy (S). However, the pair is only
accepted with a probability proportional to their fitness or
substitute fitness once combined. The probabilities are
scaled such that if the (substitute) fitness fcomb is equal or
greater to the best-known fitness fbest the pairing is
automatically accepted. Otherwise the probability is fcomb /
fbest for the mall problem and the inverse for the nurse
scheduling.
Partner Choice (C): This approach again exploits problem
specific knowledge and was inspired by an idea presented
by Ronald [1995]. He solves Royal Roads and multiobjective optimisation problems using a genetic algorithm
where the first parent is chosen following standard rules,
i.e. proportional to its fitness. However, the second parent
is not chosen according to its fitness, but depending on its
‘attractiveness’ to the first parent, which is measured on a
different scale. Our approach will be slightly different.
The first parent is still chosen according to its rank. But
rather than picking one agent from the appropriate subpopulation as the second parent, ten candidates are chosen
at random. The second parent will then be chosen as the
one that creates the fittest children with the first parent.
5
5.1
EXPERIMENTAL RESULTS
THE MODEL
To allow for fair comparison, the parameters and
strategies used for both problems are kept as similar as
possible. Both have a total population of 1000 agents.
These are split into sub-populations of size 100 for the
lower-levels and a main population of size 300 for the
nurse scheduling and respectively of size 500 for the mall
problem. In principle, two types of crossover take place:
within sub-populations a two-parent-two-children
parameterised uniform crossover with p=0.66 for genes
coming from one parent takes place.
Between sub-populations a fixed-point crossover is used
such that appropriate parts are assembled. For instance, in
the nurse problem agents of sub-population 1 and 2 would
parent a new child for sub-population 1+2. All in all 50%
of children are created via the uniform and the remainder
with the fixed-point crossover. For some of the fixed
point cases a choice exists, e.g. in the nurse problem new
agents of sub-population 1+2+3 can be created in four
different ways, either (1+2) + (3), (2+3) +(1), (3+1)+(2)
or (1)+(2)+(3). In these situations there is an equal
probability for each child to be created in either way.
Parent selection followed the seven strategies outlined
above.
Each new solution created undergoes mutation with a 1%
bit mutation probability, where a mutation would reinitialise the bit in the feasible range. The algorithm is run
in generational mode to accommodate the sub-population
structure better. In every generation the worst 90% of
parents of all sub-populations are replaced. For all fitness
and sub-fitness function calculations a fitness score as
described before is used. Constraint violations are
penalised with a dynamic penalty parameter, which
adjusts itself depending on the (sub)-fitness difference
between the best and the best feasible agent in each (sub-)
population. Full details on this type of weight and how it
was calculated can be found in Smith & Tate [1993] and
Aickelin & Dowsland [2000]. The stopping criterion is
the top sub-population showing no improvement for 50
generations.
To obtain statistically sound results all experiments were
conducted as 20 runs over all problem instances. All
experiments were started with the same set of random
seeds, i.e. with the same initial populations. The results
are presented in feasibility and cost respectively rent
format. Feasibility denotes the probability of finding a
feasible solution averaged over all problem instances.
Cost / Rent refer to the objective function value of the
best feasible solution for each problem instance averaged
over the number of instances for which at least one
feasible solution was found.
Should the algorithm fail to find a single feasible solution
for all 20 runs on one problem instance, a censored
observation of one hundred in the nurse case and zero for
the mall problem is made instead. As we are minimising
the cost for the nurses and maximising the rent of the
mall, this is equivalent to a very poor solution. For the
nurse-scheduling problem, the cost represents the sum of
unfulfilled nurses’ requests and unfavourable shiftpatterns worked. For the mall, the values for the rent are
in thousands of pounds per year.
5.2
RESULTS
Table 1 shows the results found by our algorithms for the
two problems (N = Nurse problem, M = Mall problem)
using the seven different partnering strategies in
combination with the pyramidal structure. The results are
compared to those found by the standard genetic
algorithm (SGA) [Aickelin & Dowsland 2000 and 2001]
and the Tabu Search results [Dowsland 1998] for the
nurse problem and theoretical bounds for the mall
problem (both referred to as ‘bound’). A number of
interesting observations can be made.
In the nurse scheduling case, the SGA approach failed to
find good or even feasible solutions for many data sets.
This can be explained by the high degree of epistasis
present and the inability of the unmodified genetic
algorithm to deal with it. Once the pyramidal structure
with rank-based selection (S) is introduced, results
improve significantly, however they are still below those
found by Tabu Search. For the mall problem, the situation
is different. Results found by the SGA are fairly good
with high feasibility. This indicates the higher number of
feasible solutions for this problem. Solution quality seems
reasonably good, too. However, the addition of the
pyramidal structure (S) results in a marked deterioration
of results.
How can these different results be explained? With the
nurse scheduling, the objective function value of a partial
solution was obtained by summing the cost values of the
nurses and shift-patterns involved. Furthermore, we were
able to define relatively meaningful sub-fitness scores by
exploiting the ‘cumulative’ nature of the covering
constraints due to the grade structure. Hence the substitute
fitness scores calculated allowed for an effective
recombination of partial solutions for the nursescheduling problem. Thus, there is a good correlation
between the sub-fitness of an agent (and hence its rank
and its chance of being selected) and the likelihood that it
will form part of a good solution. This also explains why
the random (R) scheme produces worse results. The best
(B) strategy although giving better results than the
random selection fails to solve many problems. However,
closer observation of experiments showed that it solved
some single data sets well. This indicates that genetic
variety is as important as fitness in the evolution of good
solutions.
Both the distributed (D) and joint (J) strategies again fail
to provide better solutions than the rank-based selection.
The distributed strategy is similar to the random strategy
as it too ignores fitness scores for selection. Choosing
from a fixed pool does have some benefits as the results
are better than for complete random choice. The joint
strategy works almost as well as the rank-selection. This
shows that the principle of the ‘dividing and conquering’
works well with the nurse problem split along the grade
boundaries. The slightly poorer results can be explained
by the ‘full’ evaluation of all sub-strings although only
‘parts’ are passed on. Thus, some of the correlation
described above is lost.
The two best strategies both outperforming (S) are partner
selection based on attractiveness (A) and choice (C).
Again this further confirms that the partial sub-fitness
scores are a good criterion of selection for the pyramidal
algorithm. Overall, (C) is better than (A) which
corresponds to (C) having a higher selection pressure than
(A), which in turn has a higher selection pressure than (S).
To conclude, it seems that for this problem a good
correlation between agents’ sub-fitness, the pyramidal
structure and good full solutions exist. Hence, the scheme
with the highest selection pressure using most problem
specific information scores best. However, the results also
show that even this scheme cannot compete with the Tabu
Search results, something we will discuss in the
concluding section.
With the Mall Problem, the situation is more complicated
since unlike for the nurse problem a large part of the
objective function is a source of epistasis, which the
proposed partitioning of the string will not eliminate fully.
The constraints are a second source for epistasis. In
contrast to the objective function, these depend largely on
the whole string, as for instance the total number of shops
of a particular size allowed. Only after adding up the
shops and sizes for all areas is it known if a solution is
feasible or not. So unsurprisingly, a combination of these
partial solutions is often unsuccessful because it usually
violates the overall constraints.
On their own, solutions of the sub-populations are
extremely unlikely to be feasible for the overall problem,
as they covered only one fifth of the string. It is equally
unlikely for those solutions in the main population, which
are formed from the five sub-populations, to be feasible.
Although these solutions are of high rent, because the
sub-populations ignore the main constraints, their
combination is unlikely to produce an overall feasible
solution.
The situation is only slightly better with those solutions
formed by an agent of the sub-populations and an agent of
the main population. Usually, even if the agent of the
main population is feasible, the children were not. Again,
even though the partial string from the sub-population
agent was of high rent, it was usually incompatible with
the rest of the string, resulting in too many or too few
shops of some types. Thus, in contrast to the nursescheduling problem, their sub-fitness scores are a far
poorer predictor for the compatibility of the parts to form
complete solutions.
This is confirmed by the above average performance of
the random strategy (R) and the extremely poor results
found by the best strategy (B). Similarly to before, the
distributed strategy (D) performs well again giving credit
to the idea of even selection pressure without relying on
fitness scores, whereas the joint strategy (J) performs
poorly suffering both from the unsuitable sub-fitness
scores and the now hindering pyramidal structure.
Overall, the real winners are again the more complex
strategies of choice (C) and attraction (A). At first this
seems contradictory as these rely heavily upon the subfitness scores. However, apart from the rank-based initial
selection of the first parent, subsequent fitness
calculations are made after combining the agents. Since
the mall pyramid only has two layers, these combinations
are always full solution and hence the full fitness score is
used. Thus, the direct link between high fitness and good
solutions is re-established. Of the two, (A) performs better
than (C). This seems to show that a certain amount of
randomness is still important here, which again might be
an indication for the lower predictive quality of the subfitness scores.
N Cost
N Feas
M Rent
M Feas
Bound
8.8
100%
2640
100%
SGA
54.2
33%
1850
94%
S
17.6
75%
1540
78%
R
37.4
54%
1790
86%
B
27.1
57%
1490
70%
D
26.5
61%
1770
84%
J
19.9
71%
1590
78%
A
12.2
83%
1950
98%
C
11.1
87%
1910
94%
Table 1: Experimental Results (N = Nurse, M = Mall).
6
CONCLUSIONS
This paper has shown the effect different partner
strategies have on a pyramidal genetic algorithm solving
two different optimisation problems from the area of
multiple-choice scheduling. The result for the five simple
strategies (S, R, B, D and J) differ for both problems. This
is a reflection of the accurateness of the sub-fitness
measure in the sense of its predictive power for subsolutions to form full solutions following the pyramidal
recombination strategies. Therefore, in the case of the
nurse problem with a good match between sub-fitness and
usefulness for recombination the simple strategies worked
well, whereas for the mall problem with its poorer
correlation between the two it did not.
For both problems the distributed partnering strategy (D)
gives consistent results. This leads to the question
whether some form of fitness sharing might have been
beneficial to some or all of the partnering strategies. We
are taking up this idea in our ongoing research into
pyramidal genetic algorithms.
The two more advanced strategies (A) and (C) use most
problem specific knowledge and work well for both
problems. They worked well for the nurse problem
because the sub-fitness scores are meaningful. They also
worked well for the mall problem because the partners are
chosen based on a fitness score after recombination,
which in this case equals the full original fitness score.
Thus, choosing parents ‘post-birth’ after evaluating
possible children can overcome possible shortcomings in
the sub-fitness measure.
It has to be noted that overall none of the partnering
schemes managed to outperform the Tabu Search
algorithm or come as close to the bounds as we had
hoped. However, one has to remember that the Tabu
Search uses highly problem-specific hill-climbing
routines, some of which also relied heavily on certain
criteria present in the actual data used. For different or
more random data this would probably no longer hold.
Therefore, without adding a more specific hill-climbing
component it would not be possible to reach this level of
solution quality. It is our conclusion that pyramidal
genetic algorithms (with or without hill-climbing) benefit
greatly from the right choice of partnering strategy as this
improves solution quality and negates possible
shortcoming of the chosen sub-fitness scores.
References
Ackley D H & Littman M L (1994), “Altruism in the
Evolution of Communication”, in R Brooks & P Maes
(eds.) Artificial Life IV, MIT Press, Mass., pp40-48.
Aickelin U (1999). “Genetic Algorithms for MultipleChoice Optimisation Problems.” PhD Dissertation,
University of Wales, Swansea, United Kingdom.
Aickelin U and Dowsland K (2000). “Exploiting problem
structure in a genetic algorithm approach to a nurse
rostering problem.” Journal of Scheduling 3, 139-153.
Aickelin U and Dowsland K (2001). “An indirect genetic
algorithm approach to a nurse scheduling problem.”
Under review by the Journal of Computing and
Operational Research.
Bean J, Noon C, Ryan S, Salton G (1988), “Selecting
Tenants in a Shopping Mall,” Interfaces 18, 1-9.
Bull, L. (1997) Evolutionary Computing in Multi-Agent
Environments: Partners. In T.Baeck (ed.) Proceedings of
the Seventh International Conference on Genetic
Algorithms. Morgan
Kaufmann, pp370-377.Bull L & Fogarty T C (1993),
“Coevolving Communicating Classifier Systems for
Tracking”, in R F Albrecht, C R Reeves & N C Steele
(eds.) Artificial Neural Networks and Genetic Algorithms,
Springer-Verlag, New York, pp522-527.
Davidor Y. (1991), Epistasis Variance: A Viewpoint on
GA-Hardness, Foundations of Genetic Algorithms 1, 2335, G Rawlins (Ed), Morgan Kaufmann, 1991.
De Jong K. Genetic Algorithms are NOT Function
Optimisers. In Whitley D. Editor. Foundations of Genetic
Algorithms 2. San Mateo: Morgan Kaufmann Publishers,
1993:5-17.
Deb K. Genetic Algorithms for Function Optimisation.
Genetic Algorithms and Soft Computing 1996:4-31.
Dowsland K.A., Nurse Scheduling with Tabu Search and
Strategic Oscillation, European Journal of Operational
Research 106, 393-407 (1998).
Fogel D. Evolutionary Computation: The Fossil Record.
IEEE Press, 1998.
Fuller E. Tackling Scheduling Problems Using Integer
Programming. Master Thesis, University of Wales
Swansea, United Kingdom, 1998.
Holland J. Adaptation in Natural and Artificial Systems.
Ann Arbor: University of Michigan Press, 1976.
Husbands P (1994), “Distributed Coevolutionary Genetic
Algorithms for Multi-Criteria and Multi-Constraint
Optimisation”, in T C Fogarty (ed.) Evolutionary
Computing, Springer-Verlag, pp150-165.
Iba H (1996), “Emergent Co-operation for Multiple
Agents Using Genetic Programming”, in H-M Voigt, W
Ebeling, I Rechenberg & H-P Schwefel (eds.) Parallel
Problem Solving from Nature - PPSN IV, Springer,
Berlin, pp32-41.
Johnson D.S. private communication 1998.
Martello S. and Toth P., Knapsack Problems, Wiley,
Chichester, 1990
Potter M. & De Jong K. (1994), “A Co-operative
Coevolutionary Approach to Function Optimisation”, in
Y Davidor, H-P Schwefel & R Manner (eds.) Parallel
Problem Solving From Nature - PPSN III, SpringerVerlag, Berlin, pp249-259.
Ronald E (1995), When Selection Meets Seduction, pp
167-173 in Eshelman L, Proceedings of the International
Conference on Genetic Algorithms, Morgan Kaufmann
Publishers, San Francisco, 1995.
Smith A. and Tate D. (1993), Genetic Optimisation Using
a Penalty Function, Proceedings ICGA 5, 499-505,
Forrest S (ed.), Morgan Kaufmann, 1993.
Stanley A E, Ashlock D, Testatsion L (1994), “Iterated
Prisoner’s Dilemma with Choice and Refusal of
Partners”, in C G Langton (ed.) Artificial Life III,
Addison-Wesley, Redwood City, pp131-146.
Wolpert D & Macready W (1995), No Free Lunch
Theorem for Search, SFI-TR-95-02-010, The Santa Fe
Institute, Santa Fe, pp 1-38, 1995.
| 5cs.CE
|
arXiv:1512.05695v1 [math.AC] 17 Dec 2015
EXAMPLES OF LOCAL COHOMOLOGY MODULES FOR
RAMIFIED REGULAR LOCAL RINGS HAVING FINITE SET OF
ASSOCIATED PRIMES
RAJSEKHAR BHATTACHARYYA
Abstract. Lyubeznik’s conjecture, ([Ly1], Remark 3.7) asserts the finiteness
of the set ssociated primes of local cohomology modules for regular rings. But,
in the case of ramified regular local ring, it is open. Recently, in Theorem 1.2
of [Nu], it is proved that in any Noetherian regular local ring S and for a fixed
ideal J ⊂ S, associated primes of local cohomology HJi (S) for i ≥ 0 is finite, if
it does not contain p. In this paper, we use this result to construct examples
of local cohomology modules for ramified regular local ring so that they have
finitely many associated primes.
1. introduction
Consider a Noetherian ring R and an R-module M . For an ideal I ⊂ R and for
some integer i ≥ 0, let HIi (M ) be i-th local cohomology module with support in the
ideal I. In the fourth problem of [Hu], it is asked that whether local cohomology
modules of Noetherian rings have finitely many associated prime ideals. There
are several important examples where we have the finiteness of associated primes of
local cohomology modules: (1) For regular rings of prime characteristic [HS], (2) for
regular local and affine rings of characteristic zero [Ly1], (3) for unramified regular
local rings of mixed characteristic [Ly2] and (4) for smooth algebra over Z and ZpZ
[BBLSZ]. All these results support Lyubeznik’s conjecture, ([Ly1], Remark 3.7):
Conjecture. Let R be a regular ring and I ⊂ R be its ideal, then for each i ≥ 0,
i-th local cohomology module HIi (R) has finitely many associated prime ideals.
Recently, in [Pu], it is shown that for excellent regular ring R of dimension d,
containing a field of characteristic zero, for an ideal I ⊂ R, Ass R HId−1 (R) is finite.
Lyubeznik’s conjecture is open for the ramified regular local ring, but recently,
there is a significant progress (see [Nu]). In fact, in Theorem 1.2 of [Nu], it is proved
that in a Noetherian regular local ring S and for a fixed ideal J ⊂ S, associated
primes of local cohomology HJi (S) for i ≥ 0 is finite, if it does not contain p. In
this paper, we use this result to construct examples of local cohomology modules
of ramified regular local ring so that they have finitely many associated primes.
In section 2, we review some of the basic results which we need in proving our
main results. In section 3, we prove our main results. At first, in Proposition 3.1,
we prove a new variant of the result of Theorem 2.1 of [Bh]. Using Proposition 3.1,
in Theorem 3.2, we prove our first main result: Consider a ramified regular local
ring S of mixed characteristic p > 0 of dimension d and choose a regular system
2010 Mathematics Subject Classification. 13D45.
Key words and phrases. Local Cohomology.
1
2
of parameters such that p is not in the ideal generated by any d − 1 of them. Let
J be an ideal generated by monomials formed out of d − 1 such a regular system
of parameters. Consider an ideal β in a polynomial rings in d − 1 indeterminates
over the prime field, generated by similar monomials but now formed out of d − 1
indeterminates. If β is perfect, then Ass S HJi (S) is finite for every i ≥ 0 except
i = ht J +1, ht J +2. It is important to note that, here the monomials are generated
by regular system of parameters, which is obviously a system of parameter, but it
does not contain p as one of the parameter.
As a consequence of Theorem 3.2, in Corollary 1, we have our next important
result which can be stated as follows: Consider ramified regular local ring S of
dimension d and let there be positive integers m and n such that d − 1 = mn.
Consider a set of regular system of parameters where p is not in the ideal generated
by any d − 1 of them. If J is an ideal, generated by size t minors of m × n matrix
formed by d − 1 such regular system of parameters, then for every i ≥ 0, except
i = ht J + 1, ht J + 2, Ass S HJi (S) is finite. In Theorem 3.3, we present our third
main result. There, we construct examples of ramified regular local rings, so that
for a countable set of ideals, there are countable set of local cohomology modules
with supports in those ideals, having finite set of associated primes.
2. basic results
In this section, we discuss two basic results which we need to prove the results in
the next section. For local cohomology modules we have the following well known
result, regarding the behaviour of associated primes of local cohomology modules
under faithfully flat extension.
Lemma 2.1. Let S be faithfully flat over R. Then, for the local cohomology modules
i
of an arbitrary R-module M , for every ideal I ⊂ R and for i ≥ 0, Ass S HIS
(M ⊗R S)
i
is a finite set if and only if Ass R HI (M ) is a finite set.
Next, we state the following known result (see Theorem 1.2 of [Nu]). We need
it to prove Proposition 3.1 in the next section.
Theorem 2.2. Let S be a regular commutative Noetherian local ring of mixed
characteristic p > 0. Then, the set of associated primes of HJi (S) that do not
contain p is finite for every i ∈ N and every ideal J ⊂ S.
3. main results
We begin the section with the following proposition which is a new variant of
the result of Theorem 2.1 of [Bh] to prove the next theorem. The part of the proof
this proposition is similar to that of Theorem 2.1 of [Bh].
Proposition 3.1. Let S be a ramified regular local ring of mixed characteristic
p > 0 and let J ⊂ S be an ideal. Suppose there exists an unramified regular local
ring R with the following properties:
(1) S is an Eisenstein extension of R by Eisenstein polynomial f (X) ∈ R[X].
(2) I ⊂ R[X] is an ideal such that IS = J and multiplication by p on HIi (R[X])
is an isomorphism.
(3) p ∈
/ Ass Z HIi+1 (R[X])
Then p ∈
/ Ass S HJi (S) and Ass S HJi (S) is finite.
3
Proof. Here p is an R[X]-regular element. So, for short exact sequence
p
0 → R[X] → R[X] → R[X]/pR[X] → 0
we get the following long exact sequence of local cohomologies
p
δ
.. → HIi−1 (R[X]/pR[X]) → HIi (R[X]) → HIi (R[X]) → HIi (R[X]/pR[X]) → ..
Since p is a non zero divisor of both HIi (R[X]) and HIi+1 (R[X]), from above long
exact sequence we get that, HIi (R[X]/pR[X]) = HIi (R[X])/pHIi (R[X]). Moreover,
since multiplication by p on HIi (R[X]) is an isomorphism, HIi (R[X]/pR[X]) =
HIi (R[X])/pHIi (R[X]) = 0.
Here S = R[X]/f (X) where (R, m) is an unramified regular local ring and Eisenstein polynomial f (X) is of the form f (X) = X n + a1 X n−1 + . . . + an with ai ∈ m
for every i ≥ 0 and an ∈
/ m2 . At first, we claim that(p, f (X)) is an R[X]-regular
sequence. To see this, let 0 6= g(X) = r0 + r1 X + . . . + rt X t ∈ (R/p)[X] with
rt 6= 0. Now f (X)g(X) = f (X)g(X) = 0 implies rt X t+n = 0. Thus rt = 0 and
we have a contradiction. Thus (p, f (X)) is a R[X]-regular sequence. Using (b) of
(12.2) Discussion of [Ho] we can see that (f (X), p) is also R[X]-regular sequences.
Set R[X] = R′ and f (X) = f .
Consider the following diagram of short exact sequences where every row and
column is exact.
0
y
f
0 −−−−→
R′
py
−−−−→
0 −−−−→
R′
y
−−−−→
f
f
0
y
0
y
R′
p
y
−−−−→
R′ /f R′
p
y
−−−−→ 0
R′
y
−−−−→
R′ /f R′
y
−−−−→ 0
0 −−−−→ R′ /pR′ −−−−→ R/pR′ −−−−→ R′ /(p, f )R′ −−−−→ 0
y
y
y
0
0
0
This above diagram yields the following diagram of long exact sequences where
all rows and columns are exact.
4
y
HIi−1 (R′ )
πi−1 y
y
f
−−−−→
HIi−1 (R′ )
πi−1 y
f
ψi−1
−−−−→
φi−1
y
HIi−1 (R′ /f R′ )
αi−1 y
HIi (R′ )
py
y
−−−−→
HIi (R′ )
πi y
f
−−−−→
f
HIi (R′ )
πi y
HIi−1 (R′ /pR′ ) −−−−→ HIi−1 (R′ /pR′ ) −−−−→ HIi−1 ((R′ /(p, f )R′ ) −−−−→ HIi (R′ /pR′ ) −−−−→ HIi (R′ /pR′ )
y
y
y
y
f
HIi (R′ )
py
−−−−→
HIi (R′ )
y
−−−−→
f
HIi (R′ )
py
−−−−→
HIi (R′ /f R′ )
py
−−−−→
HIi+1 (R′ )
py
HIi (R′ )
y
−−−−→
HIi (R′ /f R′ )
y
−−−−→
HIi+1 (R′ )
y
From first paragraph of the proof, we already have HIi (R′ /pR′ ) = 0. Thus in the
p
above diagram φi−1 is surjective. Let x be in Ker (HIi (R′ /f R′ ) → HIi (R′ /f R′ )).
Then there exists some y ∈ HIi−1 (R′ /(p, f )R′ ) such that x is the image of y under
the map HIi−1 (R′ /(p, f )R′ ) → HIi (R′ /f R′ ). Since φi−1 is surjective, y = φi−1 (z)
for some z ∈ HIi−1 (R′ /pR′ ). Now, if we come down to HIi (R′ ) via the map
HIi−1 (R′ /pR′ ) → HIi (R′ ), due to injectivity of the multiplication map by p, the
image of z in HIi (R′ ) is zero. Since every square in the diagram is commutative, we
i
(R[X]/f R[X]) =
get that x = 0. So, p is a nonzero divisor of HIi (R′ /f R′ ) = HIR[X]
i
i
HI(R[X]/f R[X]) (R[X]/f R[X]) = HJ (S). Now using Theorem 1.2 of [Nu] we conclude.
Now, we state our first main result.
Theorem 3.2. Let S be a ramified regular local ring of dimension d of mixed
characteristic p > 0 and consider a set of regular system of parameters where p is
not in the ideal generated by any d − 1 of them. Let J ⊂ S be an ideal generated by
monomials formed out of d − 1 such regular system of parameters. Assume that, for
a polynomial ring in d − 1 indeterminates over the prime field of characteristic p,
monomials similar to J which are now formed out of d−1 indeterminates, generates
a perfect ideal. Then for every i ≥ 0 except i = ht J + 1, ht J + 2, Ass S HJi (S) is
finite.
Proof. Since completion is a faithfully flat extension, using Lemma 2.1, it is sufficient to prove the theorem for complete ramified regular local ring. So, we can
assume that S is complete. Fix a regular system of parameters x1 , x2 , . . . , xd in S.
Let J be an ideal generated by d − 1 number of these regular system of parameters
x2 , . . . , xd . Let R0 be its coefficient ring. From hypothesis, we can get a system
of parameters p, x2 , . . . , xd in S. It turns out that R0 [[x2 , . . . , xd ]] is an unramified
regular local ring where x2 , . . . , xd are algebraically independent on R0 , i.e. we can
think R0 [[x2 , . . . , xd ]] as power series ring in d − 1 variables (see part of the proof
5
of (iii) of Theorem 29.4 and (ii) of Theorem 29.8 of [Ma]). Set R = R0 [[x2 , . . . , xd ]]
and there exists Eisenstein polynomial f (X), (which is actually minimal polynomial of x1 over R) such that S = R[X]/f (X) (see proof of (ii) of Theorem 29.8 of
[Ma]).
Set I = J ∩ R[X]. At first we would like to show that ht I − ht J = 1 and for
that we can proceed as follows: Since S = R[X]/f (X), there exists a maximal ideal
m ∈ R[X] such that S = R[X]/f (X) = R[X]m /f (X)R[X]m . Since I = J ∩ R[X],
we get R[X]/I ∼
= S/J. This gives R[X]/I is local and I ⊂ m. Moreover we have
f (X) ∈ I. Since S is regular local ring, dim S = ht J + dim S/J. In the similar way,
dim R[X]m = ht IR[X]m + dim R[X]m /IR[X]m . Now every minimal primes over I
is inside m, otherwise R[X]/I should have more than one maximal ideal. This gives
ht IR[X]m = ht I and our last equation becomes dim R[X]m = ht I + dim R[X]/I.
On the other hand dim R[X]m = ht f (X)R[X]m + dim R[X]m /f (X)R[X]m . Since
R[X]m is a domain, ht f (X)R[X]m = 1. Also we get dim S = dim R[X]/f (X) =
dim R[X]m /f (X)R[X]m = dim R[X]m − ht f (X)R[X]m . Combining above results
we get, dim R[X]m − dim S = ht I − ht J = 1.
Set R0 [x2 , . . . , xd ] = T . Consider the ideal α ⊂ T such that α is generated by
monomials similar to those which generates J. Since p is an T -regular element, for
short exact sequence
p
0 → T → T → T /pT → 0
we get the following long exact sequence of local cohomologies
δ
p
..Hαi−1 (T /pT ) → Hαi (T ) → Hαi (T ) → Hαi (T /pT ) → ..
It is to be noted that, we have the natural map T = R0 [x2 , . . . , xd ] → R0 [[x2 , . . . , xd ]][X] =
R[X] → S, such that αS = J where αR[X] ⊂ I.
Consider polynomial ring (Z/pZ)[x2 , . . . , xd ] of d − 1 variables over prime field
Z/pZ and let β be the ideal generated by monomials similar to those which generates
J. From assumption, β is perfect. Set K = R0 /pR0 . Clearly (Z/pZ)[x2 , . . . , xd ] →
(R0 /pR0 )[x2 , . . . , xd ] = T /pT is faithfully flat extension where (α)(T /pT ) = (β)(T /pT ).
It is easy to see that due to flatness (β)(T /pT ) i.e. (α)(T /pT ) is perfect ideal in
T /pT . Now From Proposition 19 of [HE], we get T /(p + α)T is Cohen-Macaulay.
From Proposition 4.1 of section III of [PS], we get Hαi (T /pT ) = 0 for every i ≥ 0,
except i = ht α. This gives that multiplication by p on Hαi (T ) is an isomorphism
for every i ≥ 0 except i = ht α, ht α + 1.
Next, we observe the following: For faithfully flat extension of Noetherian rings
A → B, let α be an ideal of A. Then grade (α, A) = grade (αB, B). Let a1 , . . . , an
be the maximal length of A-regular sequence which is inside α. It is easy to observe
grade (α, A) ≤ grade (αB, B). Let there be an element an+1 b ∈ αB such that
a1 , . . . , an , an+1 b is a B-regular sequence in αB. Let an+1 a ∈ (a1 , . . . , an ). This
implies an+1 ba ∈ (a1 , . . . , an )B. This gives a ∈ (a1 , . . . , an )B ∩ A = (a1 , . . . , an ).
Moreover A/(a1 , . . . , an+1 )A 6= 0. Thus a1 , . . . , an+1 be the maximal length of Aregular sequence which is inside α, but this is a contradiction. Thus grade (α, A) =
grade (αB, B).
Now, T → R[X] is faithfully flat extension. So tensoring with last long exact
sequence of homology with R[X] we get,
δ
p
i−1
i
i
i
(R[X]/pR[X]) → ..
(R[X]) → HαR[X]
(R[X]) → HαR[X]
..HαR[X]
(R[X]/pR[X]) → HαR[X]
6
Moreover using results of above paragraph we get grade (α, T ) = grade (αR[X], R[X]).
Since we are in Cohen-Macaulay ring, height and grade of any ideal coincides. Thus
i
we get, HαR[X]
(R[X]/pR[X]) = 0 for every i ≥ 0, except i = ht αR[X]. Moreover
i
this gives that multiplication by p on HαR[X]
(R[X]) is an isomorphism every i ≥ 0
except i = ht αR[X], ht αR[X] + 1.
Here αR[X] ⊂ R[X] such that (αR[X])S = J. From above paragraph of the
i
proof we have that multiplication by p on HαR[X]
(R[X]) is an isomorphism every i ≥ 0 except i = ht αR[X], ht αR[X] + 1. Moreover from long exact sei+1
quence we get p is a nonzero divisor of HαR[X]
(R[X]) for every i ≥ 0 except
i = ht αR[X], ht αR[X] + 1. Thus from Proposition 3.1 we get that, for every
i ≥ 0 except i = ht αR[X], ht αR[X] + 1, Ass S HJi (S) is finite.
To finish the proof, we need to show that grade (αR[X], R[X]) = grade (I, R[X]).
We argue in the following way: Here R[X]/f (X) is local, so there exists only one
maximal ideal m which contains f (X) and every primes which contains f (X). Same
is true for I. Thus grade (αR[X], R[X]) = grade (αR[X]m , R[X]m ) and similar is
true for I. So we can pass to local ring R[X]m . Let grade (αR[X]m , R[X]m ) = s and
grade (IR[X]m , R[X]m ) = r. Assume there exist two sets of sequences of maximal
R[X]m -regular elements {z1 , z2 , . . . , zs } and {f (X), y2 , . . . , yr } in ideals αR[X]m
and IR[X]m respectively. We can change the regular sequence {z1 , z2 , . . . , zs } to
{f (X), z2, . . . , zs } where f (X) may not be inside αR[X]m . Passing to the ring
S = R[X]m /f (X)R[X]m we find that, αR[X]m S = IR[X]m S = J and we get two
sets of R[X]m /f (X)R[X]m -regular sequences {z¯2 , . . . , z¯s } and {y¯2 , . . . , y¯r } in J. We
claim that both are maximal.
To prove this claim, let there be yr+1 ∈ IR[X]m ⊂ R[X]m such that yr+1
¯ is
S/(y¯2 , . . . , y¯r )S-regular element. For z ∈ R[X]m , let yr+1 z ∈ (f (X), y2 , . . . , yr )R[X]m .
This gives yr+1
¯ z̄ ∈ (y¯2 , . . . , y¯r )S which in turn implies that z ∈ (f (X), y2 , . . . , yr )R[X]m .
This contradicts the fact that {f (X), y2, . . . , yr } is a maximal R[X]m -regular sequence in IR[X]m . To see that {z¯2 , . . . , z¯s } is a maximal S-regular sequence,
let there exists zs+1 = a + bf (X) with a ∈ αR[X] and b ∈
/ f (X)R[X], such
that zs+1
¯ is S/(z¯2 , . . . , z¯s )S-regular sequence. At first, we observe that a can
not be zero. Otherwise, f (X) and bf (X) becomes part of regular sequence and
bf (X).1 ∈ f (X)R[X]m but 1 ∈
/ f (X)R[X]m . For some z ∈ R[X]m , assume
az ∈ (f (X), z2 , . . . , zs )R[X]m . This gives (a + bf (X))z ∈ (f (X), z2 , . . . , zs ). Thus
passing to the ring S, we get z̄ ∈ (z¯2 , . . . , z¯s )S. Thus z ∈ (f (X), z2 , . . . , zs )R[X]m .
Thus f (X), z2 , . . . , zs , a forms R[X]m -regular sequence. Again replacing f (X) by
z1 , we get (z1 , z2 , . . . , zs , a)R[X] ⊂ αR[X] forms R[X]-regular sequence which contradicts the maximality of R[X]-regular sequence z1 , z2 , . . . , zs in αR[X]m . Thus
both the sets of R/f (X)-regular sequences z¯2 , . . . , z¯s and y¯2 , . . . , y¯r in J are maximal and r = s. Thus ht αR[X] = ht I.
From the second paragraph of the proof, we notice that ht I = ht J + 1. Thus for
every i ≥ 0, except i = ht J +1, ht J +2, Ass S HJi (S) is finite. Thus we conclude.
We observe our next important result as an application of above theorem in the
following corollary.
Corollary 1. Consider ramified regular local ring S of dimension d of mixed characteristic p > 0 and let there be positive integers m and n such that d − 1 = mn.
Consider a set of regular system of parameters where p is not in the ideal generated
by any d − 1 of them. If J is an ideal, generated by size t minors of m × n matrix
7
formed by d − 1 such regular system of parameters, then for every i ≥ 0, except
i = ht J + 1, ht J + 2, Ass S HJi (S) is finite.
Proof. Consider polynomial ring K[{Xij : m ≥ i ≥ 1, m ≥ i ≥ 1}] = K[Xij ] (for
brevity) over field K, in matrix of indeterminates and let It (X)K[Xij ] is an ideal
generated by size t minor of matrix (Xij ). From [HE], we get that for It (X)K[Xij ]
is perfect and result is then immediate from above theorem.
Remark 1. In Theorem 3.2 above, it is to be noted that minimal polynomial f (X)
of x1 in the ring R[X], is not even inside the radical of αR[X]. Let f (X) = X n +
b1 X n−1 + . . . + bn . Raising to a power of integer t, we get (f (X))t = X nt + · · · + btn .
If (f (X))t ∈ αR[X] then 1 ∈ α, which is a contradiction.
Example 3.1. Here, we give examples of local cohomology modules such that p is
a nonzero divisor of it, or more generally, multiplication by p on local cohomology
modules, is an isomorphism.
(a) In section 4 of [Bh], there are examples of local cohomology modules such
that p is a nonzero divisor of it.
(b) In the proof of Theorem 3.2, we have already seen certain examples of ideals,
so that multiplication by p, on the local cohomology modules with support in those
ideals, is an isomorphism.
(c) Here we present another source of examples such that upon multiplication by
p, on the local cohomology modules, is an isomorphism: Consider an ideal J ⊂ pS.
Since p is an S-regular element, for short exact sequence
p
0 → S → S → S/pS → 0
we get the following long exact sequence of local cohomologies
p
δ
.. → HJi−1 (S/pS) → HJi (S) → HJi (S) → HJi (S/pS) → ..
Since expansion of J in S/pS is a zero ideal, for every i ≥ 1, HJi (S/pS) = 0. Thus,
above long exact sequence reduces to
p
δ
p
0 → HJ0 (S) → HJ0 (S) → HJ0 (S/pS) → HJ1 (S) → HJ1 (S) → 0
and
p
0 → HJi (S) → HJi (S) → 0
for every i ≥ 2.
Now we present our third main result of this paper in the following theorem.
It proves that, there always exists a ramified regular local ring, with a countable
collection of ideals, such that we get a countable collection of local cohomology
modules with support in those ideals, having finite set of associated primes.
Theorem 3.3. Let {Iλ } be a countable collection of ideals of an unramified regular
local ring R in mixed characteristic p > 0, such that for every λ, p is a non zero
divisor of HIiλ (R) and HIi+1
(R), for every i ∈ Cλ ⊂ N and for every Iλ , where Cλ is
λ
a countable subset of N. Then, there exists a ramified regular local ring S of mixed
characteristic p > 0, which is a homomorphic image of R, so that for expanded
ideal Jλ = Iλ S of S, and for corresponding Cλ ⊂ N, Ass S HJi λ (S) is finite for every
i ∈ Cλ and this is true for every λ.
8
Proof. At first we observe that, we can always include Iλ = 0 in the countable
collection of ideals as stated in the hypothesis along with corresponding Cλ = N.
Let (R, m) be an unramified regular local ring in mixed characteristic p > 0. Since
p is an R-regular element, for short exact sequence
p
0 → R → R → R/pR → 0
we get the following long exact sequence of local cohomologies
p
δ
.. → HIi−1
(R/pR) → HIiλ (R) → HIiλ (R) → HIiλ (R/pR) → ..
λ
Since p be a non zero divisor of HIiλ (R) and HIi+1
(R) for every i ∈ Cλ and for every
λ
Iλ , from above long exact sequence, we get HIiλ (R/pR) = HIiλ (R)/pHIiλ (R).
We can assume R to be a complete regular local ring. From [Pu], we know
that for every ideal Iλ of R, and for every i ≥ 0, the set Ass R HIiλ (R) is countable. Since p is regular, same is true for Ass R HIiλ (R)/pHIiλ (R). For any RS
module M , set Ass R M = Ass R M − {m}. Now A = λ (∪i∈Cλ Ass R HIiλ (R)) ∪
(∪i∈Cλ Ass R HIiλ (R)/pHIiλ (R)) is countable. From [Bu] we get that there exists an
element g ∈ m such that g does not belong to any element of A. From construction, (p, h) is a HIiλ (R)-regular sequence. Set h = g 2 and thus (p, h) also becomes
a HIiλ (R)-regular sequence where h ∈ m2 .
Now consider the sequence (p, h−p). Let (h−p)x = py for x, y ∈ HIiλ (R) for some
i ∈ Cλ and for some ideal Iλ . Then hx = p(x + y) and since (p, h) is an HIiλ (R)regular sequence we get that (p, h − p) is also an HIiλ (R)-regular sequence. In fact,
for Iλ = 0, we get Ass R H00 (R/pR) = Ass R R/pR. Thus, from second paragraph of
the proof we get that (p, h) as well as (p, h − p) are R-regular sequence.
Set h − p = f . Since h ∈ m2 , we find f ∈ m − m2 . Thus, it is part of a regular
system of parameters. So, R/f R is again a regular local ring such that p ∈ m2 .
Moreover, h ∈
/ pR, otherwise, h = h.1 = pr implies 1 ∈ pR since (p, h) are regular
sequence. Thus, it is a contradiction. This shows p is not a multiple of h − p. So,
R/f R is a ramified regular local ring of mixed characteristic p ≥ 0. Moreover, each
of h and f are also nonzero divisor in R. Thus using (b) of (12.2) Discussion of [Ho],
we find (h, p) as well as (f, p) are also a R-regular sequence. So, we can consider the
following diagram of short exact sequences where every row and column is exact.
0
y
f
0 −−−−→
R
py
−−−−→
0 −−−−→
R
y
−−−−→
f
f
0
y
0
y
R
p
y
−−−−→
R/f R
p
y
−−−−→ 0
R
y
−−−−→
R/f R
y
−−−−→ 0
0 −−−−→ R/pR −−−−→ R/pR −−−−→ R/(p, f )R −−−−→ 0
y
y
y
0
0
0
9
This above diagram yields the following diagram of long exact sequences where
every rows and columns are exact.
y
HIi−1
(R)
λ
πi−1 y
y
f
−−−−→
HIi−1
(R)
λ
πi−1 y
f
ψi−1
−−−−→
φi−1
y
HIi−1
(R/f R)
λ
αi−1
y
HIiλ (R)
py
y
−−−−→
HIiλ (R)
πi y
f
−−−−→
f
HIiλ (R)
πi y
((R/(p, f )R) −−−−→ HIiλ (R/pR) −−−−→ HIiλ (R/pR)
HIi−1
(R/pR) −−−−→ HIi−1
(R/pR) −−−−→ HIi−1
λ
λ
λ
y
y
y
y
f
HIiλ (R)
py
−−−−→
HIiλ (R)
y
−−−−→
f
HIiλ (R)
py
−−−−→
HIiλ (R/f R)
py
−−−−→
HIi+1
(R)
λ
py
HIiλ (R)
y
−−−−→
HIiλ (R/f R)
y
−−−−→
HIi+1
(R)
λ
y
In the above diagram, φi−1 is surjective since f is HIiλ (R/pR) = HIiλ (R)/pHIiλ (R)p
regular. Let x be in Ker (HIiλ (R/f R) → HIiλ (R/f R)). Then there exists some y ∈
i−1
HIλ (R/(p, f )R) such that x is the image of y under the map HIi−1
(R/(p, f )R) →
λ
i−1
i
HIλ (R/f R). Since φi−1 is surjective, y = φi−1 (z) for some z ∈ HIλ (R/pR). Now,
if we come down to HIiλ (R) via the map HIi−1
(R/pR) → HIiλ (R), due to injectivity
λ
of the multiplication map by p, the image of z in HIiλ (R) is zero. Since every square
in the diagram is commutative, we get that x = 0. So, p is a nonzero divisor of
HIiλ (R/f R). For ramified regular local ring R/f R, set S = R/f R. Then, using
Theorem 1.2 of [Nu], we get Ass S HJi λ (S) is finite.
References
[BBLSZ] B. Bhatt, M. Blickle, G. Lyubeznik, A. Singh and W. Zhang, Local Cohomology Modules
of a Smooth Z-Algebra have Finitely Many Associated Primes, Inventiones Mathematicae, 197
(2014) 509-519, arXiv: 1304.4692v2 [math.AC].
[Bh] R. Bhatacharyya, A Note on Associated Primes and Bockstein Homomorphisms of Local
Cohomologies for Ramified Regular Local Rings, Communicated.
[Bu] L. Burch, Codimension and analytic spread, Proc. Cambridge Philos. Soc. 72. (1972), 369373.
[Ho] M. Hochster, Solid Closure In: Commutative algebra: syzygies, multiplicities, and birational
algebra, Contemp. Math., vol. 159, pp. 103-172.
[HE] M. Hochster and J. Eagon, Cohen-Macaulay rings, invariant theory and the generic perfection of determinantal loci, Amer. J. of Math. 93, 1020-1058, 1971.
[HS] Huneke and Sharp, Bass Numbers of Local Cohomology Modules, Transaction of American
Mathematical Society, Vol. 339, No. 2, 765-779, 1993.
[Hu] C. Huneke, Problems on local cohomology, in: Free resolutions in commutative algebra and
algebraic geometry (Sundance, Utah, 1990), 93-108, Res. Notes Math. 2, Jones and Bartlett,
Boston, MA, 1992.
[Ly1] G. Lyubeznik, Finiteness Properties of Local Cohomology modules (an application of Dmodules to commutative algebra), Invent. Math. 113, 41-55, 1993.
10
[Ly2] G. Lyubeznik, F-modules: applications to local cohomology and D-modules in characteristic
p ¿ 0, J. Reine Angew. Math. 491 (1997), 65-130.
[Ma] H. Matsumara, Commutative Ring Theory, Cambridge University Press, 1990.
[Nu] L. Nunez-Betancourt On certain rings of differentiable type and finiteness properties of
local cohomology, Journal of Algebra, Volume 379, April 2013, Pages 1-10, arXiv: 1202.3597
[math.AC].
[PS] C. Peskine and L. Szpiro, Dimensionprojective finie et cohomology locale, I.H.E.S. 42, 47-119,
1973.
[Pu] T. J. Puthenpurakal, Associated primes of local cohomology modules over regular rings,
arXiv: 1411.2734.
Dinabandhu Andrews College, Garia, Kolkata 700084, India
E-mail address: [email protected]
| 0math.AC
|
arXiv:1409.3605v2 [math.RA] 5 Jan 2016
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
OLGUR CELIKBAS, LARS WINTHER CHRISTENSEN, LI LIANG,
AND GREG PIEPMEYER
Abstract. We analyze stable homology over associative rings and obtain results over Artin algebras and commutative noetherian rings. Our study develops similarly for these classes; for simplicity we only discuss the latter here.
Stable homology is a broad generalization of Tate homology. Vanishing of
stable homology detects classes of rings—among them Gorenstein rings, the
original domain of Tate homology. Closely related to gorensteinness of rings is
Auslander’s G-dimension for modules. We show that vanishing of stable homology detects modules of finite G-dimension. This is the first characterization
of such modules in terms of vanishing of (co)homology alone.
Stable homology, like absolute homology, Tor, is a theory in two variables.
It can be computed from a flat resolution of one module together with an
injective resolution of the other. This betrays that stable homology is not
balanced in the way Tor is balanced. In fact, we prove that a ring is Gorenstein
if and only if stable homology is balanced.
Introduction
The homology theory studied in this paper was introduced by P. Vogel in the 1980s.
Vogel did not publish his work, but the theory appeared in print in a 1992 paper by
Goichot [18], who called it Tate–Vogel homology. As the name suggests, the theory
is a generalization of Tate homology for modules over finite group rings. Vogel and
Goichot also considered a generalization of Tate cohomology, which was studied
in detail by Avramov and Veliche [5]. In that paper, the theory was called stable
cohomology, to emphasize a relation to stabilization of module categories. To align
terminology, we henceforth refer to the homology theory as stable homology.
For modules M and N over a ring R, stable homology is a Z-indexed family of
g R (M, N ). These fit into an exact sequence
abelian groups Tor
i
(†)
R
R
gR
gR
· · · → Tor
i (M, N ) → Tori (M, N ) → Tori (M, N ) → Tori−1 (M, N ) → · · ·
R
where the groups TorR
i (M, N ) and Tori (M, N ) form, respectively, the unbounded
homology and the standard absolute homology of M and N . We are thus led to
study stable and unbounded homology simultaneously. Our investigation takes cues
from the studies of stable cohomology [5] and absolute homology, but our results
look quite different. This comes down to an inherent asymmetry in the definition
Date: 29 December 2015.
2010 Mathematics Subject Classification. 16E05, 16E30, 16E10, 13H10.
Key words and phrases. Stable homology, Tate homology, Gorenstein ring, G-dimension.
This research was partly supported by a Simons Foundation Collaboration Grant for Mathematicians, award no. 281886 (L.W.C.), NSA grant H98230-14-0140 (L.W.C.) and NSFC grant
11301240 (L.L.). Part of the work was done during L.L.’s stay at Texas Tech University with
support from the China Scholarship Council; he thanks the Department of Mathematics and
Statistics at Texas Tech for its kind hospitality.
2
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
of stable homology that is not present in either of these precursors. It manifests
itself in different ways, but it is apparent in most of our results.
∗ ∗ ∗
g R (M, N ), the ring acts on M from the right
When we consider stable homology Tor
and on N from the left. In this paper an R-module is a left R-module; to distinguish
right R-modules, we speak of modules over the opposite ring R◦ . We study stable
homology over associative rings and obtain conclusive results for Artin algebras and
commutative noetherian rings. In the following overview, R denotes an Artin algebra or a commutative noetherian local ring. The results we discuss are special cases
of results obtained within the paper; internal references are given in parentheses.
One expects a homology theory to detect finiteness of homological dimensions.
Our first two results reflect the asymmetry in the definition of stable homology.
Right vanishing (3.1, 3.2). For a finitely generated R◦ -module M , the following
conditions are equivalent.
(i) M has finite projective dimension.
g R (M, −) = 0 for all i ∈ Z.
(ii) Tor
i
g R (M, −) = 0.
(iii) There is an i > 0 with Tor
i
Left vanishing (5.1, 5.12). For a finitely generated R-module N , the following
conditions are equivalent.
(i) N has finite injective dimension.
g R (−, N ) = 0 for all i ∈ Z.
(ii) Tor
i
g R (−, N ) = 0.
(iii) There is an i 6 0 with Tor
i
These two vanishing results reveal that stable homology cannot be balanced in the
way absolute homology, Tor, is balanced.
Balancedness (4.5, 4.6, 4.7). The following conditions on R are equivalent.
(i) R has finite injective dimension over R and over R◦ .
(ii) For all finitely generated R◦ -modules M , all finitely generated R-modules N ,
g R◦ (N, M ).
g R (M, N ) ∼
and all i ∈ Z there are isomorphisms Tor
= Tor
i
i
(iii) For all R◦ -modules M , all R-modules N , and all i ∈ Z there are isomorphisms
g R◦ (N, M ).
g R (M, N ) ∼
Tor
= Tor
i
i
Another way to phrase part (i) above is to say that R is Iwanaga-Gorenstein. On
that topic: a commutative noetherian ring A is Gorenstein (regular or Cohen–
Macaulay) if the local ring Ap is Iwanaga-Gorenstein (regular or Cohen–Macaulay)
for every prime ideal p in A. It follows that a commutative noetherian ring of finite
self-injective dimension is Gorenstein, but a Gorenstein ring need not have finite
self-injective dimension; consider, for example, Nagata’s regular ring of infinite
Krull dimension [28, appn. exa. 1].
It came as a surprise to us that balancedness of stable homology detects gorensteinness outside of the local situation (4.6): A commutative noetherian ring A is
g A (N, M ) are isomorphic for all finitely
g A (M, N ) and Tor
Gorenstein if and only if Tor
i
i
generated A-modules M and N and all i ∈ Z. This turns out to be only one of
several ways in which stable homology captures global properties of rings. Indeed,
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
3
vanishing of stable homology detects if a commutative noetherian (not necessarily
local) ring is regular (3.3), Gorenstein (3.12), or Cohen–Macaulay (5.11).
For finitely generated modules over a noetherian ring, Auslander and Bridger [1]
introduced a homological invariant called the G-dimension. A noetherian ring is
Iwanaga-Gorenstein if and only if there is an integer d such that every finitely
generated left or right module has G-dimension at most d; see Avramov and
Martsinkovsky [4, 3.2]. In [18, sec. 3] Tate homology was defined for modules over
Iwanaga-Gorenstein rings and shown to agree with stable homology. Iacob [24]
generalized Tate homology further to a setting that includes the case where M is a
finitely generated R◦ -module of finite G-dimension. We prove:
Tate homology (6.4). Let M be a finitely generated R◦ -module of finite Gg R (M, −) is isomorphic
dimension. For every i ∈ Z the stable homology functor Tor
i
d R (M, −).
to the Tate homology functor Tor
i
There are two primary generalizations of G-dimension to modules that are not
finitely generated: Gorenstein projective dimension and Gorenstein flat dimension.
They are defined in terms of resolutions by Gorenstein projective (flat) modules,
notions which were introduced by Enochs, Jenda, and collaborators; see Holm [21]1.
In general, it is not known if Gorenstein projective modules are Gorenstein flat.
d R (M, −) is for R◦ -modules M of finite
Iacob’s definition of Tate homology Tor
Gorenstein projective dimension. We show (6.7) that Tate homology agrees with
g R (M, −) for all such R◦ -modules if and only if every Gorenstein
stable homology Tor
◦
projective R -module is Gorenstein flat.
Here is an example to illustrate how widely the various homology theories differ.
Example. Let (R, m) be a commutative artinian local ring with m2 = 0, and
assume that R is not Gorenstein. If k denotes a field, then k[x, y]/(x2 , xy, y 2 ) is an
example of such a ring. Let E be the injective hull of the residue field R/m. Then:
(a) Absolute homology TorR
i (E, E) is non-zero for every i ≥ 2.
R
g
(b) Stable homology Tori (E, E) is zero for every i.
R
(c) Unstable homology TorR
i (E, E) agrees with Tori (E, E) for every i.
d R (E, E)” is not defined.
(d) Tate homology “Tor
Since R is not Gorenstein, E has infinite G-dimension; see [22, thm. 3.2]. That
explains (d); see 6.1 and 6.2. It also explains (a), as the first syzygy of E is a
k-vector space, whence vanishing of TorR
i (E, E) for some i > 2 would imply that E
has finite projective dimension and hence finite G-dimension; see [4, thm. 4.9]. Left
vanishing, see p. 2, yields (b), and then the exact sequence (†) on p. 1 gives (c).
Finally we return to the topic of vanishing. A much studied question in Gorenstein homological algebra is how to detect finiteness of G-dimension. A finitely
generated module G over a commutative noetherian ring A has G-dimension zero if
it satisfies ExtiA (G, A) = 0 = ExtiA (HomA (G, A), A) for all i > 0 and the canonical
map G → HomA (HomA (G, A), A) is an isomorphism. This is the original definition [1], and the last requirement cannot be dispensed with, as shown by Jorgensen
1 Enochs and Jenda consolidated their work in [16, chaps. 10–12], which deal almost exclusively
with Gorenstein rings. In [21] Holm does Gorenstein homological algebra over associative rings,
so that is our standard reference for this topic. Basic definitions are recalled in 6.6.
4
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
and Şega [26]. There are other characterizations of modules of finite G-dimension
in the literature, but none that can be expressed solely in terms of vanishing of
(co)homology functors. Hence, we were excited to discover:
G-dimension (3.11). Let A be a commutative noetherian ring. A finitely generg A (M, A) = 0 holds for
ated A-module M has finite G-dimension if and only if Tor
i
all i ∈ Z.
1. Tensor products of complexes
All rings in this paper are assumed to be associative algebras over a commutative
ring k, where k = Z is always possible, and k = R works when R is commutative.
For an Artin algebra R, one can take k to be an artinian ring such that R is finitely
generated as a k-module. We recall that in this situation the functor D(−) =
Homk (−, E), where E is the injective hull of k/Jac(k), yields a duality between the
categories of finitely generated R-modules and finitely generated R◦ -modules.
By M(R) we denote the category of R-modules. We work with complexes, index
these homologically, and follow standard notation (see the appendix in [9] for what
is not covered below).
1.1. For a complex X with differential ∂ X and an integer n ∈ Z the symbol Cn (X)
X
, and Hn (X) denotes the homology of X in degree n,
denotes the cokernel of ∂n+1
X
X
i.e., Ker ∂n /Im ∂n+1 . A complex with H(X) = 0 is called acyclic, and a morphism
of complexes X → Y that induces an isomorphism H(X) → H(Y ) is called a quasiisomorphism. The symbol ≃ is used to decorate quasi-isomorphisms; it is also used
for isomorphisms in derived categories.
1.2. For an R◦ -complex X and an R-complex Y
`, the tensor product X ⊗R Y is
the k-complex with degree n term (X ⊗R Y )n = i∈Z (Xi ⊗R Yn−i ) and differential
Y
given by ∂(x ⊗ y) = ∂iX (x) ⊗ y + (−1)i x ⊗ ∂n−i
(y) for x ∈ Xi and y ∈ Yn−i .
Contrast the standard tensor product complex in 1.2 with the construction in 1.3,
which first appeared in [18]. Similar constructions for Hom were also treated in [18]
and in great detail by Avramov and Veliche [5]. We recall the Hom constructions
in the Appendix, where we also study their interactions with the one below.
1.3 Definition. For an R◦ -complex X and an R-complex Y , consider the graded
k-module X ⊗R Y with degree n component
Y
(X ⊗R Y )n =
Xi ⊗R Yn−i .
i∈Z
Endowed with the degree −1 homomorphism ∂ defined on elementary tensors as in
1.2, it becomes a complex called the unbounded tensor product. It contains the tensor product, X ⊗R Y , as a subcomplex. The quotient complex (X ⊗R Y )/(X ⊗R Y )
e R Y , and it is called the stable tensor product.
is denoted X ⊗
1.4. The definition above yields an exact sequence of k-complexes,
e R Y −→ 0 .
(1.4.1)
0 −→ X ⊗R Y −→ X ⊗R Y −→ X ⊗
Notice that if X or Y is bounded, or if both complexes are bounded on the same
side (above or below), then the unbounded tensor product coincides with the usual
e R Y is zero.
tensor product, and the stable tensor product X ⊗
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
5
Note that the unbounded tensor product X ⊗R Y is the product totalization of
the double complex (Xi ⊗R Yj ).
1.5. We collect some basic results about the unbounded and stable tensor products. For the unbounded tensor product the proofs mimic the proofs for the tensor
product, and one obtains the results for the stable tensor product via (1.4.1). As
above X is an R◦ -complex and Y is an R-complex.
(a) There are isomorphisms of k-complexes
eR Y ∼
e R◦ X .
and
X⊗
X ⊗R Y ∼
= Y ⊗
= Y ⊗R◦ X
(b)
(c)
(d)
(e)
e R − are additive and right exact.
The functors X ⊗R − and X ⊗
e R − preserve degree-wise split exact sequences.
The functors X ⊗R − and X ⊗
e R − preserve homotopy.
The functors X ⊗R − and X ⊗
′
A morphism α : Y → Y of R-complexes yields isomorphisms of k-complexes
e R α) ∼
e R Cone α .
Cone(X ⊗R α) ∼
= X ⊗R Cone α and Cone(X ⊗
= X⊗
It follows that X ⊗R α is a quasi-isomorphism if and only if X ⊗R Cone α is
e R α.
acyclic, and similarly for X ⊗
1.6 Lemma. Let D = (Di,j ) be a double k-complex. Let z be a cycle in the
product totalization of D. Assume z contains a component zm,n that satisfies
zm,n = ∂ h (x′ ) + ∂ v (x′′ ) for some x′ ∈ Dm+1,n and x′′ ∈ Dm,n+1 . If both
(1) Hhm+k (D∗,n−k ) = 0 for every k > 0 and
(2) Hvn+k (Dm−k,∗ ) = 0 for every k > 0,
then z is a boundary in the product totalization of D.
Proof. The goal is to prove the existence of a sequence (xi ) with xm+1 = x′ and
xm = x′′ such that
zm+k,n−k = ∂ h (xm+k+1 ) + ∂ v (xm+k )
holds for all k ∈ Z; i.e., one has z = ∂(x) in the product totalization of D. Assume
k is positive and that xm , . . . , xm+k have been constructed. Then
zm+k−1,n−k+1 = ∂ h (xm+k ) + ∂ v (xm+k−1 )
holds. As z is a cycle in the product totalization of D, one has
0 = ∂ h (zm+k,n−k ) + ∂ v (zm+k−1,n−k+1 )
= ∂ h (zm+k,n−k ) + ∂ v (∂ h (xm+k ) + ∂ v (xm+k−1 ))
= ∂ h (zm+k,n−k ) − ∂ h (∂ v (xm+k )) .
Thus, zm+k,n−k − ∂ v (xm+k ) is a horizontal cycle, and hence by (1) it is a horizontal boundary. There exists, therefore, an element xm+k+1 with ∂ h (xm+k+1 ) +
∂ v (xm+k ) = zm+k,m−k . By induction, this argument yields the elements xm+k for
k > 1; a symmetric argument using (2) yields xm+k for k < 0.
1.7 Proposition. Let X be an R◦ -complex and let Y be an R-complex.
(a) If X is bounded above and Xi ⊗R Y is acyclic for all i, then X ⊗R Y is acyclic.
(b) If Y is bounded below and Xi ⊗R Y is acyclic for all i, then X ⊗R Y is acyclic.
6
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
Proof. The product totalization of the double complex (Xi ⊗ Yj ) is X ⊗R Y .
(a) Let m be an upper bound for X, let n be an integer, and let z = (zm+k,n−k )k
be a cycle in X ⊗R Y of degree m + n. The component zm,n is a cycle in Xm ⊗R Y
and hence a boundary as Xm ⊗R Y is assumed to be acyclic. Moreover, the assumption that Xi ⊗R Y is acyclic for every i ensures that condition (2) in Lemma 1.6 is
met. Condition (1) is satisfied due to the boundedness of X; thus z is a boundary
in X ⊗R Y . As n is arbitrary, it follows that X ⊗R Y is acyclic.
(b) A similar argument with the roles of m and n exchanged handles this case.
2. Homology
Now we consider the homology of unbounded and stable tensor product complexes.
Our notation differs slightly from the one employed by Goichot [18].
≃
2.1 Definition. Let M be an R◦ -module and N be an R-module. Let P −−→ M be
≃
a projective resolution and let N −−→ I be an injective resolution. The k-modules
TorR
i (M, N ) = Hi (P ⊗R I)
are the unbounded homology modules of M and N over R, and the stable homology
modules of M and N over R are
gR
e
Tor
i (M, N ) = Hi+1 (P ⊗R I) .
2.2. As any two projective resolutions of M , and similarly any two injective resolutions of N , are homotopy equivalent, it follows from 1.5(d) that the definitions
in 2.1 are independent of the choices of resolutions.
2.3. Notice from 1.4 that if M has finite projective dimension, or if N has finite
R
injective dimension, then for every i ∈ Z one has TorR
i (M, N ) = Tori (M, N ) and
g R (M, N ) = 0.
hence Tor
i
2.4. The standard liftings of module homomorphisms to morphisms of resolutions
gR
imply that the definitions of TorR
i (M, N ) and Tori (M, N ) are functorial in either
argument; that is, for every i ∈ Z there are functors
◦
gR
TorR
i (−, −), Tori (−, −) : M(R ) × M(R) −→ M(k) .
These functors are homological in the sense that every short exact sequence of R◦ modules 0 → M ′ → M → M ′′ → 0 and every R-module N give rise to a connected
exact sequence of stable homology modules
(2.4.1)
g R (M ′′ , N ) → Tor
g R (M ′ , N ) → Tor
g R (M, N ) → Tor
g R (M ′′ , N ) → · · · ,
· · · → Tor
i+1
i
i
i
and to an analogous sequence of Tor modules. Similarly, every short exact sequence
0 → N ′ → N → N ′′ → 0 of R-modules and every R◦ -module M give rise to a
connected exact sequence,
(2.4.2)
′′
′
′′
gR
gR
gR
gR
· · · → Tor
i+1 (M, N ) → Tori (M, N ) → Tori (M, N ) → Tori (M, N ) → · · · ,
and an analogous sequence of Tor modules. For details see [18, sec. 1].
Stable and unbounded homology entwine with absolute homology.
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
7
2.5. For every R◦ -module M and every R-module N the exact sequence of complexes (1.4.1) yields an exact sequence of homology modules:
g R (M, N ) → · · ·
g R (M, N ) −
→ Tor
(2.5.1) · · · → Tor
→ TorR
→ TorR
i (M, N ) −
i−1
i
i (M, N ) −
Flat resolutions. Just like absolute homology, unbounded and stable homology
can be computed using flat resolutions in place of projective resolutions.
≃
2.6 Proposition. Let M be an R◦ -module and N be an R-module. If F −−→ M is
≃
a flat resolution and N −−→ I is an injective resolution, then there are isomorphisms,
∼
TorR
i (M, N ) = Hi (F ⊗R I)
g R (M, N ) ∼
e R I)
and Tor
= Hi+1 (F ⊗
i
for every i ∈ Z, and they are natural in either argument.
≃
Proof. Let P −−→ M be a projective resolution, then there is a quasi-isomorphism
α : P → F . For every R-complex E the complex Cone(α) ⊗R E is acyclic by [11,
cor. 7.5]. In particular, Cone(α) ⊗R I is acyclic, and it follows from Proposition 1.7
e R I is acyclic in view of (1.4.1). Thus,
that Cone(α) ⊗R I is acyclic; now Cone(α) ⊗
e
the morphisms α ⊗R I and α ⊗R I are quasi-isomorphisms by 1.5(e), and the desired isomorphisms follow from 2.1; it is standard to verify that they are natural.
In view of 1.4 the next result is now immediate.
g R (M, −) = 0
2.7 Corollary. For every R◦ -module M of finite flat dimension, Tor
i
holds for every i ∈ Z.
2.8 Remark. If every flat R◦ -module has finite projective dimension, then Corollary 2.7 is covered by 2.3. There are wide classes of rings over which flat modules
have finite projective dimension: Iwanaga-Gorenstein rings, see [16, thm. 9.1.10],
and more generally rings of finite finitistic projective dimension, see Jensen [25,
prop. 6]; in a different direction, rings of cardinality at most ℵn for some n ∈ N,
see Gruson and Jensen [20, thm. 7.10].
On the other hand, recall that any product of fields is a von Neumann regular
ring, i.e., every module over such a ring is flat. Osofsky [29, 3.1] shows that a
sufficiently large product of fields has infinite global dimension and hence must
have flat modules of infinite projective dimension.
The next result is an analogue for stable homology of [5, thm. 2.2].
2.9 Proposition. Let M be an R◦ -module and let n ∈ Z. The following conditions
are equivalent.
g R (M, −) → TorR (M, −) is an isomorphism for
(i) The connecting morphism Tor
i
i
every i > n.
(ii) TorR
i (M, E) = 0 for every injective R-module E and every i > n.
(iii) TorR
i (M, −) = 0 for every i > n.
g R (M, E) = 0
Proof. For every injective R-module E and for every i ∈ Z one has Tor
i
by 2.3. Thus (i) implies (ii).
In the following we use the notation (−)⊃s for the soft truncation below at s.
≃
(ii) =⇒ (iii): Let N be an R-module and N −−→ I be an injective resolution. Let
≃
+
P −−→ M be a projective resolution and let P denote its mapping cone. Let E be
an injective R-module; by assumption one has TorR
i (M, E) = 0 for all i > n. By
8
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
right exactness of the tensor product, P⊃+n−2 ⊗R E is acyclic and hence P⊃+n−2 ⊗R I
is acyclic by 1.5(a) and Proposition 1.7. Thus, for every i > n one has
+
TorR
i (M, N ) = Hi (P ⊗R I) = Hi (P⊃n−2 ⊗R I) = 0.
(Notice that for all n 6 0 one has P⊃+n−2 = P + .)
(iii) =⇒ (i): This is immediate from the exact sequence (2.5.1).
TorR
≫0 (M, −)
As vanishing of
detects finite flat dimension of M , the next result
is an immediate consequence of Proposition 2.9. In the parlance of [15] it says that
g R (M, −) = 0.
the flat and copure flat dimensions agree for R◦ -modules M with Tor
g R (M, −) = 0 for all i ∈ Z, then
2.10 Corollary. Let M be an R◦ -module. If Tor
i
fdR◦ M = sup{i ∈ Z | TorR
i (M, E) 6= 0 for some injective R-module E } .
Dimension shifting is a useful tool in computations with stable homology.
2.11. For a module X, we denote by Ωm X the mth syzygy in a projective resolution
of X and by Ωm X the mth cosyzygy in an injective resolution of X. By Schanuel’s
lemma a syzygy/cosyzygy is uniquely determined up to a projective/injective summand. In view of 2.3 and the exact sequences in 2.4 one gets
∼ gR
gR
Tor
i (M, N ) = Tori−m (Ωm M, N ) for m > 0
(2.11.1)
(2.11.2)
≃
gR
Tor
i (M, N )
∼
=
m
gR
Tor
i+m (M, Ω N )
and
for m > 0 .
Moreover, if F −−→ M is a flat resolution of M , then there are isomorphisms,
g R (M, N ) ∼
g R (Cm (F ), N ) for m > 0 ;
Tor
= Tor
i
i−m
(2.11.3)
this follows from (2.4.1) and Corollary 2.7.
2.12. Suppose R is commutative and S is a flat R-algebra. For every S ◦ -module
≃
M , a projective resolution P −−→ M over S ◦ is a flat resolution of M as an Rg R (M, −)
module. Thus, it follows from Proposition 2.6 that stable homology Tor
i
≃
◦
is a functor from M(R) to M(S ). Similarly, an injective resolution N −−→ I of an
g R (−, N )
S-module is also an injective resolution of N as an R-module, whence Tor
i
is a functor from M(R) to M(S).
g
3. Vanishing of stable homology Tor(M,
−)
We open with a partial converse to Corollary 2.7; recall, for example from [16,
prop. 3.2.12], that a finitely presented module is flat if and only if it is projective.
3.1 Theorem. Let M be an R◦ -module that has a degree-wise finitely generated
projective resolution. The following conditions are equivalent.
(i) pdR◦ M is finite.
g R (M, −) = 0 for all i ∈ Z.
(ii) Tor
i
g R (M, −) = 0 for some i > 0.
(iii) Tor
i
g R (M, Homk (M, E)) = 0 for some faithfully injective k-module E.
(iv ) Tor
0
Moreover, if R is left noetherian with idR R finite, then (i)–(iv ) are equivalent to
g R (M, −) = 0 for some i ∈ Z.
(iii’ ) Tor
i
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
9
Proof. The implications (i) =⇒ (ii) =⇒ (iii) =⇒ (iii’ ) are clear; see 2.3. Part
(iv ) follows from (iii) by dimension shifting (2.11.2).
For (iv ) =⇒ (i) use Theorem A.8 to get
∼
gR
g0
0 = Tor
0 (M, Homk (M, E)) = Homk (ExtR◦ (M, M ), E) .
g 0 ◦ (M, M ) is zero, hence M has finite projective dimension by [5, prop. 2.2].
Thus Ext
R
Finally, assume that R is left noetherian with idR R finite. Let E be a faithfully
injective k-module; in view of what has already been proved, it is sufficient to show
g R (M, −) = 0 for some i > 0 implies Tor
g R (M, Homk (M, E)) = 0. It follows
that Tor
0
−i
from the assumptions on R that every free R-module has finite injective dimension.
Thus every projective R-module P has finite injective dimension, and 2.3 yields
g R (−, P ) = 0 for all i ∈ Z. From the exact sequence (2.4.2) it follows that there
Tor
i
g R (M, Homk (M, E)) ∼
g R (M, Ωi Homk (M, E)) for i > 0,
are isomorphisms Tor
= Tor
0
−i
R
R
g (M, Homk (M, E)) = 0.
g (M, −) implies Tor
so vanishing of Tor
0
−i
3.2 Corollary. Let R be an Artin algebra with duality functor D(−). For a finitely
generated R◦ -module M the following conditions are equivalent.
(i) pdR◦ M is finite.
g R (M, N ) = 0 for all finitely generated R-modules N .
(ii) There is an i > 0 with Tor
i
g R (M, D(M )) = 0.
(iii) Tor
0
Proof. Part (ii) follows from (i) in view of 2.3. Part (iii) follows from (ii) by
dimension shifting (2.11.2) in a degree-wise finitely generated injective resolution of
D(M ). Finally, one has D(M ) = Homk (M, E), see the first paragraph in Section 1,
so (i) follows from (iii) by Theorem 3.1, as the injective hull E of k/Jac(k) is
faithfully injective.
By a result of Bass and Murthy [6, lem. 4.5], a commutative noetherian ring is
regular if and only if every finitely generated module has finite projective dimension.
3.3 Corollary. For a commutative noetherian ring R, the following are equivalent:
(i) R is regular.
g R (M, −) = 0 for all finitely generated R-modules M and all i ∈ Z.
(ii) Tor
i
g R (M, −) = 0 for all finitely generated R-modules M .
(iii) There is an i ∈ Z with Tor
i
Proof. The implications (i) =⇒ (ii) =⇒ (iii) are clear; see 2.3.
(iii) =⇒ (i): If i > 0, then every finitely generated R-module has finite projective
dimension by Theorem 3.1; i.e., R is regular. If i < 0, then for any finitely generated
∼ gR
gR
R-module M one has Tor
0 (M, −) = Tori (Ω−i M, −) = 0 by dimension shifting
(2.11.1), and as above it follows that R is regular.
Tate flat resolutions. We show that under extra assumptions on M one can
g R (M, N ) without resolving the second argument.
compute Tor
≃
3.4. Let N be an R-module; choose an injective resolution N −−→ I and a projective
≃
resolution P −−→ N . The composite P → N → I is a quasi-isomorphism, so the
N
complex CP I = Cone(P → N → I) is acyclic, and up to homotopy equivalence it
is independent of the choice of resolutions. This construction is functorial in N .
10
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
3.5 Lemma. Let F be a bounded below complex of flat R◦ -modules; let N be an
≃
≃
R-module with a projective resolution P −−→ N and an injective resolution N −−→ I.
N
e R I ≃ F ⊗R CP I in the derived category D(k), and it
There is an isomorphism F ⊗
is functorial in the second argument; cf. 3.4.
Proof. There is by 1.5(c) a short exact sequences of k-complexes
0 −→ F ⊗R I −→ F ⊗R CN
P I −→ F ⊗R Σ P −→ 0 .
With (1.4.1) it forms a commutative diagram whose rows are triangles in D(k),
F ⊗R P
/ F ⊗R I
/ F ⊗R CN
PI
≃
F ⊗R I
/ Σ(F ⊗R P )
≃
/ F ⊗R I
/F⊗
eR I
/ Σ(F ⊗R I) ,
where the top triangle is rotated back once. The isomorphism is induced by the
composite quasi-isomorphism P → N → I, as one has F ⊗R P = F ⊗R P ; cf. 1.4.
Completion of the morphism of triangles yields the desired isomorphism via the
Triangulated Five Lemma; see [23, prop. 4.3]. Functoriality in the second argument
is straightforward to verify.
≃
3.6 Lemma. Let N be an R-module with a projective resolution P −−→ N and
≃
an injective resolution N −−→ I. Let T be a complex of flat R◦ -modules such that
T ⊗R E is acyclic for every injective R-module E. There is then an isomorphism
T ⊗ R CN
P I ≃ Σ(T ⊗R N ) in D(k) which is functorial in the second argument; cf. 3.4.
Proof. By the assumptions and Proposition 1.7(a) the complex T ⊗R I is acyclic,
so application of T ⊗R − to the exact sequence 0 → I → CN
P I → Σ P → 0 yields
→
T
⊗
Σ
P
;
cf.
1.5(c).
Now,
let P + denote the
a quasi-isomorphism T ⊗R CN
R
PI
mapping cone of the quasi-isomorphism P → N . The complex T ⊗R P + is acyclic
by Proposition 1.7(b), so application of T ⊗R − to the mapping cone sequence
0 → N → P + → Σ P → 0 yields an isomorphism T ⊗R Σ P ≃ Σ(T ⊗R N ) in D(k),
and the desired isomorphism follows as N is a module; cf. 1.4. Functoriality in the
second argument is straightforward to verify; cf. 3.4.
3.7 Remark. The quasi-isomorphisms in Lemmas 3.5 and 3.6 hold with CN
P I re≃
N
placed by CF I where F −−→ N is a flat resolution.
The objects discussed in the next paragraph appear in the literature under a
variety of names; here we stick with the terminology from [27].
3.8. An acyclic complex T of flat R◦ -modules is called totally acyclic if T ⊗R E is
acyclic for every injective R-module E. A Tate flat resolution of an R◦ -module M is
≃
a pair (T, F ), where F −−→ M is a flat resolution and T is a totally acyclic complex
◦
of flat R -modules with T>n ∼
= F>n for some n > 0. Tate’s name is invoked here
because these resolutions can be used to compute Tate homology; see [27, thm. A].
3.9. Over a left coherent ring R, an R◦ -module M has a Tate flat resolution if and
only if it has finite Gorenstein flat dimension; see [27, prop. 3.4].
If R is noetherian, then every finitely generated R◦ -module M of finite Gdimension has a Tate flat resolution. Here is a direct argument: By [4, thm. 3.1]
≃
there exists a pair (T, P ) where P −−→ M is a projective resolution and T is an
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
11
acyclic complex of finitely generated projective R◦ -modules with the following properties: the complex HomR◦ (T, R) is acyclic, and there is an n > 0 with T>n ∼
= P>n .
For every injective R-module E one now has
∼ T ⊗R HomR (R, E) =
∼ HomR (HomR◦ (T, R), E)
(3.9.1)
T ⊗R E =
where the last isomorphism holds because T is degree-wise finitely generated; see [7,
prop. VI.5.2]. Thus T is a totally acyclic complex of flat R◦ -modules and (T, P ) is
a Tate flat resolution of M .
The next result shows that stable homology can be computed via Tate flat resolutions. We apply it in Section 6 to compare stable homology to Tate homology,
but before that it is used in the proofs of Theorems 3.11 and 4.2.
3.10 Theorem. Let M be an R◦ -module that has a Tate flat resolution (T, F ).
For every R-module N and every i ∈ Z there is an isomorphism,
g R (M, N ) ∼
Tor
= Hi (T ⊗R N ) ,
i
and it is functorial in the second argument.
Proof. Choose n > 0 such that T>n and F>n are isomorphic, and consider the
degree-wise split exact sequence
0 −→ T6n−1 −→ T −→ F>n −→ 0 .
≃
≃
Let P −−→ N be a projective resolution and N −−→ I be an injective resolution; set
C = CN
P I ; cf. 3.4. There is, by 1.5(c), an exact sequence of k-complexes,
0 −→ T6n−1 ⊗R C −→ T ⊗R C −→ F>n ⊗R C −→ 0 .
As C is acyclic, it follows from Proposition 1.7(a) that T6n−1 ⊗R C is acyclic,
whence the surjective morphism above is a quasi-isomorphism. Since T is a totally acyclic complex of flat modules, there is by Lemma 3.6 an isomorphism
T ⊗R C ≃ Σ(T ⊗R N ) in D(k). Moreover, Lemma 3.5 yields an isomorphism
e R I in D(k).
e R I in D(k). Thus one has Σ(T ⊗R N ) ≃ F>n ⊗
F>n ⊗R C ≃ F>n ⊗
This explains the second isomorphism in the next computation.
g R (M, N ) ∼
g R (Cn (F ), N )
Tor
= Tor
i
i−n
e R I)
= Hi−n+1 ((Σ−n F>n ) ⊗
e R I)
= Hi+1 (F>n ⊗
∼
= Hi+1 (Σ(T ⊗R N ))
= Hi (T ⊗R N )
The first isomorphism holds by (2.11.3); the first equality follows from Proposition 2.6 as the canonical surjection Σ−n F>n → Cn (F ) is a flat resolution.
As discussed in the introduction, the G-dimension is a homological invariant for
finitely generated modules over noetherian rings. Characterizations of modules of
finite G-dimension have traditionally involved both vanishing of (co)homology and
invertibility of a certain morphism; see for example [9, (2.1.6), (2.2.3), (3.1.5), and
(3.1.11)]. More recently, Avramov, Iyengar, and Lipman [3] showed that a finitely
generated module M over a commutative noetherian ring R has finite G-dimension
if and only M is isomorphic to the complex RHomR (RHomR (M, R), R) in the
12
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
derive category D(R). The crucial step in our proof of the next theorem is to show
g R (M, R) implies such an isomorphism.
that vanishing of stable homology Tor
3.11 Theorem. Let R be a commutative noetherian ring. For a finitely generated
R-module M , the following conditions are equivalent.
(i) G-dimR M < ∞.
g R (M, N ) = 0 for every R-module N of finite flat dimension and all i ∈ Z.
(ii) Tor
i
R
g
(iii) Tor (M, R) = 0 for all i ∈ Z.
i
Proof. The implication (ii) =⇒ (iii) is trivial.
(i) =⇒ (ii): By 3.9 the R◦ module M has a Tate flat resolution (T, F ). Theog R (M, −) ∼
rem 3.10 yields isomorphisms Tor
= Hi (T ⊗R −) for all i ∈ Z. It follows
i
by induction on the flat dimension of N that T ⊗R N is acyclic; cf. [10, lem. 2.3].
≃
(iii) =⇒ (i): Let P −−→ M be a degree-wise finitely generated projective reso≃
lution, and let R −−→ I be an injective resolution. By assumption, the complex
e R I is acyclic, which explains the third ≃ below:
P⊗
M ≃ P ⊗R R ≃ P ⊗R I ≃ P ⊗R I ∼
= P ⊗R HomR (R, I) ∼
= HomR (HomR (P, R), I).
The last isomorphism holds by Proposition A.6 as HomR (P, R) equals HomR (P, R).
Now it follows from [3, thm. 2] and [9, thm. (2.2.3)] that G-dimR M is finite.
3.12 Corollary. For a commutative noetherian ring R the following are equivalent.
(i) R is Gorenstein.
g R (M, R) = 0 for all finitely generated R-modules M and all i ∈ Z.
(ii) Tor
i
R
g
(iii) Tori (M, R) = 0 for all finitely generated R-modules M and all i < 0.
Proof. By a result of Goto, R is Gorenstein if and only if every finitely generated
R-module has finite G-dimension [19, cor. 2]. Combining this with Theorem 3.11
one gets the the equivalence of (i) and (ii). The implication (ii) =⇒ (iii) is clear.
Finally, (ii) follows from (iii) by dimension shifting (2.11.1).
4. Balancedness of stable homology
Absolute homology is balanced over any ring: there are always isomorphisms
◦
TorR (M, N ) ∼
= TorR (N, M ). It follows already from Corollary 3.2 that stable
homology can be balanced only over special rings. Indeed, if R is an Artin algebra and stable homology is balanced over R, then the dual module of R has finite
projective dimension over both R and R◦ , whence R is Iwanaga-Gorenstein. The
converse is part of Corollary 4.5.
We open this section with a technical lemma; it is similar to a result of Enochs,
Estrada, and Iacob [14, thm. 3.6]. Recall the notation Cm (−) from 1.1.
4.1 Lemma. Let T and T ′ be acyclic complexes of flat R◦ -modules and flat Rmodules, respectively. For all integers m and n there are isomorphisms in D(k),
T6n−1 ⊗R Cm (T ′ ) ≃ Σn−m (Cn (T ) ⊗R T6′ m−1 )
T>n ⊗R Cm (T ′ ) ≃ Σn−m (Cn (T ) ⊗R T>′ m )
T ⊗R Cm (T ′ ) ≃ Σn−m (Cn (T ) ⊗R T ′ ) .
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
13
Proof. Because the complexes T and T ′ are acyclic, there are quasi-isomorphisms
Cn (T ) → Σ1−n T6n−1 and Cm (T ′ ) → Σ1−m T6′ m−1 . Hence [10, prop. 2.14(b)] yields
T6n−1 ⊗R Cm (T ′ ) ≃ T6n−1 ⊗R Σ1−m T6′ m−1
≃ Σn−m (Σ1−n T6n−1 ⊗R T6′ m−1 )
≃ Σn−m (Cn (T ) ⊗R T6′ m−1 ) .
This demonstrates the first isomorphism in the statement, and the second is proved
similarly. Finally, these two isomorphisms connect the exact sequences
0 → T6n−1 ⊗R Cm (T ′ ) → T ⊗R Cm (T ′ ) → T>n ⊗R Cm (T ′ ) → 0
and
0 → Cn (T ) ⊗R T6′ m−1 → Cn (T ) ⊗R T ′ → Cn (T ) ⊗R T>′ m → 0 ,
and one obtains the last isomorphism via the Triangulated Five Lemma in D(k);
see [23, prop. 4.3].
4.2 Theorem. Let M be an R◦ -module and N be an R-module. If they both have
Tate flat resolutions, then for each i ∈ Z there is an isomorphism
∼ g R◦
gR
Tor
i (M, N ) = Tori (N, M ) .
Proof. Let (T, F ) and (T ′ , F ′ ) be Tate flat resolutions of M and N , respectively.
Choose n ∈ Z such that there are isomorphisms T>n ∼
= F>′ n . For
= F>n and T>′ n ∼
every i ∈ Z one has
∼
gR
Tor
i (M, N ) = Hi (T ⊗R N )
∼
= Hi−n (T ⊗R Cn (T ′ ))
∼
= Hi−n (Cn (T ) ⊗R T ′ )
∼
= Hi (M ⊗R T ′ )
◦
∼
gR
= Tor
i (N, M ) ,
where the first and the last isomorphisms follow from Theorem 3.10, the second
and fourth isomorphisms follow by dimension shifting, and the third isomorphism
holds by the last isomorphism in Lemma 4.1.
4.3 Definition. Let M be an R◦ -module and N be an R-module. Stable homology
g R (M, N ) ∼
g R◦ (N, M ) for all i ∈ Z.
is balanced for M and N if one has Tor
= Tor
i
i
4.4. Theorem 4.2 says that stable homology is balanced for all (pairs of) R◦ - and Rmodules that have Tate flat resolutions. If R is Iwanaga–Gorenstein, then every R◦ module and every R-module has a Tate flat resolution, see 3.9 and [16, thm. 12.3.1],
so stable homology is balanced for all (pairs of) R◦ - and R-modules.
4.5 Corollary. For an Artin algebra R the following conditions are equivalent.
(i) R is Iwanaga-Gorenstein.
(ii) Stable homology is balanced for all R◦ - and R-modules.
(iii) Stable homology is balanced for all finitely generated R◦ - and R-modules.
14
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
Proof. Per 4.4, part (i) implies (ii), which clearly implies (iii). Let E = D(R)
be the dual module of R; it is injective and finitely generated over R◦ and over
R. Thus, if stable homology is balanced for finitely generated R◦ - and R modules,
then it follows from 2.3 and Corollary 3.2 that pdR E as well as pdR◦ E is finite.
By duality, both idR R and idR◦ R are then finite.
4.6 Corollary. A commutative noetherian ring R is Gorenstein if and only if stable
homology is balanced for all finitely generated R-modules.
Proof. Over a Gorenstein ring, the G-dimension of every finitely generated module
is finite by [19, cor. 2], so every finitely generated module has a Tate flat resolution;
see 3.9. Thus, balancedness of stable homology follows from Theorem 4.2. Cong R (M, R) = 0 for every finitely
versely, balancedness of stable homology implies Tor
i
generated R-module M and all i ∈ Z, so R is Gorenstein by Corollary 3.12.
4.7 Corollary. A commutative noetherian ring R of finite Krull dimension is
Gorenstein if and only if stable homology is balanced over R.
Proof. If R is Gorenstein of finite Krull dimension, then it is Iwanaga-Gorenstein.
Therefore, stable homology is balanced over R per 4.4. The converse holds by
Corollary 4.6.
g
5. Vanishing of stable homology Tor(−,
N)
g R (−, N ) over an Artin algebra can by duality be
Vanishing of stable homology Tor
i
g R (M, −).
understood via vanishing of Tor
i
5.1 Proposition. Let R be an Artin algebra with duality functor D(−). For a
finitely generated R-module N the following conditions are equivalent.
(i) idR N is finite.
g R (−, N ) = 0 for all i ∈ Z.
(ii) Tor
i
g R (M, N ) = 0 for all finitely generated
(iii) There is an integer i 6 0 with Tor
i
◦
R -modules M .
g R (D(N ), N ) = 0.
(iv ) Tor
0
Proof. The implications (i) =⇒ (ii) =⇒ (iii) are clear; see 2.3. Part (iv ) follows
from (iii) by dimension shifting (2.11.1), as D(N ) is finitely generated. Finally,
∼ gR
gR
vanishing of Tor
0 (D(N ), N ) = Tor0 (D(N ), D(D(N ))) implies by Corollary 3.2 that
pdR◦ D(N ) is finite, whence idR N is finite, and so (i) follows from (iv ).
g R (−, N ) over commuLocal rings. To analyze vanishing of stable homology Tor
i
tative noetherian rings, we start locally.
5.2. Let R be a commutative noetherian local ring with residue field k. For an
R-module M , the depth invariant can be defined as
depthR M = inf{i ∈ Z | ExtiR (k, M ) 6= 0} ,
and if M is finitely generated, then its injective dimension can be computed as
idR M = sup{i ∈ Z | ExtiR (k, M ) 6= 0} .
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
15
The depth is finite for M 6= 0. The ring R is Cohen–Macaulay if there exists a
finitely generated module M 6= 0 of finite injective dimension; this is a consequence
of the New Intersection Theorem due to Peskine and Szpiro [30] and Roberts [32].
The following is an analogue of [5, thm. 6.1].
5.3 Lemma. Let R be a commutative noetherian local ring with residue field k.
For every finitely generated R-module N and for every i ∈ Z there is an isomorphism
Y
j−i
∼
TorR
Homk (ExtjR (k, R), ExtR
(k, N ))
i (k, N ) =
j∈Z
of R-modules, and in particular, TorR
i (k, N ) is a k-vector space.
≃
Proof. Let P −−→ k be a degree-wise finitely generated projective resolution and
≃
≃
let N −−→ I and R −−→ J be injective resolutions. By definition TorR
i (k, N ) is the
ith homology of the complex P ⊗R I which we compute using Proposition A.6 as
follows P ⊗R I ∼
= P ⊗R HomR (R, I) ∼
= HomR (HomR (P, R), I). Next we simplify
using quasi-isomorphisms and Hom-tensor adjointness:
HomR (HomR (P, R), I) ≃ HomR (HomR (P, J), I)
≃ HomR (HomR (k, J), I)
∼
= HomR (HomR (k, J) ⊗k k, I)
∼
= Homk (HomR (k, J), HomR (k, I)) .
Finally, pass to homology.
g R (k, N ) may be a k5.4 Remark. Lemma 5.3 suggests that stable homology Tor
i
vector space, and indeed it is. If R is a ring and if x annihilates the R◦ -module M ,
or the R-module N , then there are two homotopic lifts—zero and multiplication by
x—to the projective resolution in the case of M or to the injective resolution in the
case of N ; see Definition 2.1 and [34, 22.6 and 2.3.7]. Hence multiplication by x is
zero on unbounded and on stable homology of M against N . In particular, stable
g R (k, N ) is a k-vector space.
homology Tor
i
5.5 Proposition. Let R be a commutative noetherian local ring with residue field
k and let N be a finitely generated R-module. If for some i ∈ Z the k-vector space
g R (k, N ) has finite rank, then N has finite injective dimension, or R is Gorenstein.
Tor
i
R
Proof. Each k-vector space TorR
j (k, N ) has finite rank, so Tori+1 (k, N ) has finite
rank by the assumption and the exact sequence (2.5.1). For a finitely generated
R-module M 6= 0 the vector spaces ExtjR (k, M ) are non-zero for all j between
depthR M < ∞ and idR M ; see Roberts [31, thm. 2]. When TorR
i+1 (k, N ) has finite
rank, it follows from Lemma 5.3 that R or N has finite injective dimension.
Compared to the characterization of globally Gorenstein rings in Corollary 3.12,
condition (iii) below is sharper.
5.6 Theorem. Let R be a commutative noetherian local ring with residue field k.
The following conditions are equivalent.
(i) R is Gorenstein.
g R (−, R) = 0 for all i ∈ Z.
(ii) Tor
i
16
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
g R (k, R) has finite rank for some i ∈ Z.
(iii) Tor
i
g R (k, M ) has finite
(iv ) There exists a finitely generated R-module M such that Tor
i
g R (M, R) = 0 holds for all i ∈ Z.
rank for some i ∈ Z and Tor
i
Proof. The implications (i) =⇒ (ii) =⇒ (iii) =⇒ (iv ) are clear; see 2.3. Let M
be a module as specified in (iv ); Theorem 3.11 yields G-dimR M < ∞. By a result
of Holm [22, thm. 3.2], R is Gorenstein if idR M is also finite. Now it follows from
Proposition 5.5 that R is Gorenstein.
5.7 Remark. Let R be a commutative noetherian local ring with residue field
k, and let E denote the injective hull of k; it is a faithfully injective R-module.
g R (k, k) and
A computation based on Theorem A.8 shows that the ranks of Tor
i
i
g (k, k) are simultaneously finite:
Ext
R
g R (k, HomR (k, E))
g R (k, k) = rankk Tor
rankk Tor
i
i
g iR (k, k), E)
= rankk HomR (Ext
g iR (k, k) .
= rankk Ext
Combined with this equality of ranks, [5, thm. 6.4, 6.5, and 6.7] yield characterizations of regular, complete intersection, and Gorenstein local rings in terms of the
g R (k, k). For example, R is regular if and only
size of the stable homology spaces Tor
i
g R (k, k) = 0 holds for some (equivalently, all) i ∈ Z, and R is Gorenstein if
if Tor
i
g R (k, k) has finite rank for some (equivalently, all) i ∈ Z.
and only if Tor
i
Commutative rings. If R is commutative and p is a prime ideal in R, then the
local ring Rp is a flat R-algebra, and for an Rp -module M it follows from 2.12 that
g R (M, N ) are Rp -modules.
the stable homology modules Tor
i
5.8 Lemma. Let R be a commutative noetherian ring and let N be an R-module;
let p be a prime ideal in R and let M be an Rp -module. For every i ∈ Z there is a
natural isomorphism of Rp -modules,
g R (M, N ) ∼
g Rp (M, Np ) .
Tor
= Tor
i
i
g R (−, N ) = 0 implies Tor
g Rp (−, Np ) = 0 for all prime ideals p in R.
Hence, Tor
i
i
≃
≃
Proof. Let N −−→ I be an injective resolution. Let P −−→ M be a projective resolution over Rp ; it is a flat resolution of M as an R-module. The second isomorphism
below follows from Proposition A.4.
∼ (P ⊗R Rp ) ⊗R I
P ⊗R I =
p
∼
= P ⊗Rp (Rp ⊗R I)
∼
= P ⊗Rp (Rp ⊗R I)
The computation gives P ⊗R I ∼
= P ⊗Rp Ip .
= P ⊗Rp Ip ; similarly one gets P ⊗R I ∼
eR I ∼
e Rp Ip , and the desired isoNow (1.4.1) and the Five Lemma yield P ⊗
=P⊗
morphisms follow as Np → Ip is an injective resolution by Matlis Theory.
The proof of the next result is similar. Compared to Lemma 5.8 the noetherian
hypothesis on R has been dropped, as it was only used to invoke Matlis Theory.
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
17
5.9 Lemma. Let R be a commutative ring and let M be an R-module; let p be a
prime ideal in R and let N be an Rp -module. For every i ∈ Z there is a natural
isomorphism of Rp -modules,
∼ g Rp
gR
Tor
i (M, N ) = Tori (Mp , N ) .
g R (M, −) = 0 implies Tor
g Rp (Mp , −) = 0 for all prime ideals p in R.
Hence, Tor
i
i
5.10 Theorem. Let R be a commutative noetherian ring and let N be a finitely
g R (−, N ) = 0 holds for some i ∈ Z, then idRp Np is
generated R-module. If Tor
i
finite for every prime ideal p in R.
R
g p (−, Np ) = 0. It follows
Proof. From the hypotheses and Lemma 5.8 one has Tor
i
from Proposition 5.5 that the local ring Rp is Gorenstein, or idRp Np is finite.
g Rp (−, Np ) = 0 implies that
However, if Rp is Gorenstein, then vanishing of Tor
i
pdRp Np is finite by Corollary 4.7 and Theorem 3.1, and then idRp Np is finite.
The next corollary is now immediate per the remarks in 5.2.
5.11 Corollary. A commutative noetherian ring R is Cohen–Macaulay if there is
g R (−, N ) = 0 for some i ∈ Z.
a finitely generated R-module N 6= 0 with Tor
i
5.12 Corollary. Let R be a commutative noetherian ring of finite Krull dimension.
For a finitely generated R-module N , the following conditions are equivalent.
(i) idR N is finite.
g R (−, N ) = 0 for all i ∈ Z.
(ii) Tor
i
g R (−, N ) = 0 for some i ∈ Z.
(iii) Tor
i
Proof. The implications (i) =⇒ (ii) =⇒ (iii) are clear; see 2.3. Part (i) follows
from (iii) as idR N equals sup{idRp Np | p is a prime ideal in R} 6 dim R.
5.13 Remark. We do not know if the assumption of finite Krull dimension in
g R (−, N ) implies that
Corollary 5.12 is necessary. By Theorem 5.10, vanishing of Tor
N is locally of finite injective dimension, but that does not imply finite injective
dimension over R: Just consider a Gorenstein ring R of infinite Krull dimension.
g R (−, R) vanishes for such a ring.
On the other hand, we do not know if Tor
6. Comparison to Tate homology
In this section we compare stable homology to Tate homology. We parallel some of
the findings of Avramov and Veliche [5]. First we recall a few definitions.
6.1. An acyclic complex T of projective R◦ -modules is called totally acyclic if
HomR◦ (T, P ) is acyclic for every projective R◦ -module P ; cf. 3.8. A complete
̟
≃
projective resolution of an R◦ -module M is a diagram T −−→ P −−→ M , where
≃
◦
T is a totally acyclic complex of projective R -modules, P −−→ M is a projective
resolution, and ̟i is an isomorphism for i ≫ 0; see [33, sec. 2].
Let M be an R◦ -module with a complete projective resolution T → P → M .
For an R-module N , the Tate homology of M and N over R are the k-modules
d R (M, N ) = Hi (T ⊗R N ) for i ∈ Z; see Iacob [24, sec. 2].
Tor
i
18
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
6.2. An R◦ -module has a complete projective resolution if and only if it has finite
Gorenstein projective dimension; see [33, thm. 3.4].
If R is noetherian and M is a finitely generated R◦ -module with a complete projective resolution, then M has finite G-dimension and it has a complete projective
resolution T → P → M with T and P degree-wise finitely generated and T → P
surjective; see [33, 2.4.1] and [4, thm. 3.1].
6.3 Lemma. Let M be an R◦ -module that has a complete projective resolution
T → P → M . The following conditions are equivalent.
(i) T ⊗R E is acyclic for every injective R-module E.
d R (M, −) for all i ∈ Z.
g R (M, −) ∼
(ii) There are isomorphisms of functors Tor
= Tor
i
i
Proof. Assume that T ⊗R E is acyclic for every injective R-module E. The pair
(T, P ) is then a Tate flat resolution of M , see 3.8, and it follows from Theorem 3.10
d R (M, −) are isomorphic for all i ∈ Z. For
g R (M, −) and Tor
that the functors Tor
i
i
the converse, let E be an injective R-module. By 2.3 one then has
∼ dR
gR
0 = Tor
i (M, E) = Tori (M, E)
for all i ∈ Z, and hence H(T ⊗R E) = 0.
6.4 Theorem. Let R be noetherian, and let M be a finitely generated R◦ -module
that has a complete projective resolution. There are isomorphisms of functors
g R (M, −) ∼
d R (M, −) for all i ∈ Z .
Tor
= Tor
i
i
Proof. The module M has a complete projective resolution T → P → M with T
and P degree-wise finitely generated; see 6.2. The isomorphisms (3.9.1) show that
the complex T ⊗R E is acyclic for every injective R-module E, and Lemma 6.3
finishes the argument.
6.5 Remark. The isomorphisms of homology modules in Theorem 6.4 actually
follow from one isomorphism in D(k), but this is unapparent in the proof, which
rests on Theorem 3.10. The finitely generated module M has a complete projective
resolution T → L → M with T and L degree-wise finitely generated and T → L
surjective; see 6.2. Thus the kernel K = Ker(T → L) is a bounded above complex of
projective modules. Given a module N , let C be the cone as one would construct in
3.4. By Proposition 1.7(a) the complex K ⊗R C is acyclic, so that Lemma 3.5 and
e R I) ≃ Σ−1 (L ⊗R C) ≃ Σ−1 (T ⊗R C). As in the proof above,
1.5(c) yield Σ−1 (L ⊗
T ⊗R E is acyclic, so Lemma 3.6 gives Σ−1 (T ⊗R C) ≃ T ⊗R N , and combining
these isomorphisms in D(k) gives the desired one.
Without extra assumptions on the ring, we do not know if stable homology agrees
with Tate homology whenever the latter is defined. In general, the relation between
stable homology and Tate homology is tied to an unresolved problem in Gorenstein
homological algebra. Theorem 6.7 explains how.
6.6. An R◦ -module G is called Gorenstein projective if there exists a totally acyclic
complex T of projective R◦ -modules with C0 (T ) ∼
= G; see 6.1. Similarly, an R◦ module G is called Gorenstein flat if there exists a totally acyclic complex T of flat
R◦ -modules with C0 (T ) ∼
= G; see 3.8.
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
19
6.7 Theorem. The following conditions on R are equivalent.
(i) Every Gorenstein projective R◦ -module is Gorenstein flat.
(ii) For every R◦ -module M that has a complete projective resolution there are
g R (M, −) ∼
d R (M, −) for all i ∈ Z.
isomorphisms of functors Tor
= Tor
i
i
Proof. Assume that every Gorenstein projective R◦ -module is Gorenstein flat.
Let T → P → M be a complete projective resolution. It follows that T is a
totally acyclic complex of flat modules; see Emmanouil [13, thm. 2.2]. Thus stable
homology and Tate homology coincide by Lemma 6.3.
For the converse, let M be a Gorenstein projective R◦ -module and let T be a
totally acyclic complex of projective R◦ -modules with M ∼
= C0 (T ). Since there are
R
R
∼
g
d
isomorphisms of functors Tori (M, −) = Tori (M, −) for all i ∈ Z, it follows from
2.3 that T ⊗R E is acyclic for every injective R-module E. Thus T is a totally
acyclic complex of flat R◦ -modules, and so M is Gorenstein flat.
6.8 Remark. As Holm notes [21, prop. 3.4], the obvious way to achieve that every
Gorenstein projective R◦ -module is Gorenstein flat is to ensure that (1) the Pontryagin dual of every injective R-module is flat, and (2) that every flat R◦ -module
has finite projective dimension. The first condition is satisfied if R is left coherent,
and the second is discussed in Remark 2.8. A description of the rings over which
Gorenstein projective modules are Gorenstein flat seems elusive; see [13, sec. 2].
Complete homology . In his thesis, Triulzi considers the J-completion of the homological functor TorR (M, −) = {TorR
i (M, −) | i ∈ Z}. His construction is similar
to Mislin’s P-completion of covariant Ext and Nucinkis’ I-completion of contravariant Ext. The resulting homology theory is called complete homology. Like stable
homology, it is a generalization of Tate homology. We compare these two generalizations in [8]. From the point of view of stable homology, it is interesting to know
when it agrees with complete homology, because the latter has a universal property.
In this direction the main results in [8] are that these two homology theories agree
over Iwanaga-Gorenstein rings, and for finitely generated modules over Artin algebras and complete commutative local rings. Moreover, the two theories agree with
Tate homology, whenever it is defined, under the exact same condition; that is, if
and only if every Gorenstein projective module is Gorenstein flat; see Theorem 6.7.
Appendix
We start by recalling the definition of stable cohomology.
A.1. Let X and Y be R-complexes, following [5, 18] we denote by HomR (X, Y )
the subcomplex of HomR (X, Y ) with degree n term
a
HomR (X, Y )n =
HomR (Xi , Yi+n ) .
i∈Z
It is called the bounded Hom complex, and the quotient complex
] R (X, Y ) = HomR (X, Y )/HomR (X, Y )
Hom
is called the stable Hom complex.
≃
≃
For R-modules M and N with projective resolutions PM −−→ M and PN −−→ N ,
the k-modules
ExtiR (M, N ) = H−i (HomR (PM , PN )) ,
20
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
are called the bounded cohomology of M and N over R, and the stable cohomology
modules of M and N over R are
g i (M, N ) = H−i (Hom
] R (PM , PN )) .
Ext
R
d i (M, N ) for the stable cohomolAvramov and Veliche [5] use the notation Ext
R
ogy; this notation is standard for Tate cohomology, which coincides with stable
cohomology whenever the former is defined; see [5, cor. 2.4].
A.2 Proposition. Let X and Y be R-complexes.
(a) If X or Y is bounded above, and HomR (Xi , Y ) is acyclic for all i, then the
complex HomR (X, Y ) is acyclic.
(b) If X or Y is bounded below, and HomR (X, Yi ) is acyclic for all i, then the
complex HomR (X, Y ) is acyclic.
Proof. Similar to the proof of Proposition 1.7.
Standard isomorphisms. We study composites of the functors Hom and ⊗. To
the extent possible, we establish analogs of the standard isomorphisms for composites of Hom and ⊗; see [7, sec. II.5 and VI.5]. There seems to be no analog of
Hom-tensor adjunction [7, prop. II.5.2].
The setup is the same for Propositions A.4–A.6; namely:
A.3. Let X be a complex of R◦ -modules, let Y be a complex of (R, S ◦ )-bimodules,
and let Z be a complex of S-modules.
Under finiteness conditions, the unbounded tensor product is associative.
A.4 Proposition. For complexes as in A.3, under either of the following conditions
• X and Z are complexes of finitely presented modules
• Y is a bounded complex
there is an isomorphism of k-complexes,
(X ⊗R Y ) ⊗S Z −→ X ⊗R (Y ⊗S Z) ,
and it is functorial in X, Y , and Z.
Proof. For every n ∈ Z one has,
((X ⊗R Y ) ⊗S Z)n =
Y Q
j∈Z
Xj ⊗R Yi−j ⊗S Zn−i
i∈Z
and
(X ⊗R (Y ⊗S Z))n =
Y
j∈Z
Xj ⊗ R
Q
i∈Z
Yi−j ⊗S Zn−i ,
and from each of these modules there is a canonical homomorphism to
YY
YY
(Xj ⊗R Yi−j ) ⊗S Zn−i ∼
Xj ⊗R (Yi−j ⊗S Zn−i ) .
=
i∈Z j∈Z
j∈Z i∈Z
If Y is bounded, then these homomorphisms are isomorphisms as the inner most
products are finite. Recall, e.g. from [16, thm. 3.2.22], that the functor M ⊗R −
commutes with products if M is finitely presented. Thus, the canonical homomorphisms are isomorphisms when X and Z are complexes of finitely presented
modules. It is straightforward to verify that these isomorphisms commute with the
differentials and form an isomorphism of complexes.
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
21
The model for the following swap isomorphism is [7, ex. II.4]. The proof is similar
to the proof of Proposition A.4 and uses that the functor HomR (M, −) commutes
with coproducts if M is a finitely generated R-module.
A.5 Proposition. For complexes as in A.3, under either of the following conditions
• X and Z are complexes of finitely generated modules
• Y is a bounded complex
there is an isomorphism of k-complexes,
HomR◦ (X, HomS (Z, Y )) −→ HomS (Z, HomR◦ (X, Y )) ,
and it is functorial in X, Y , and Z.
The next result is an analog of [7, prop. VI.5.2 and VI.5.3]; it is used below to
establish a duality between stable homology and stable cohomology.
A.6 Proposition. Let X, Y , and Z be as in A.3, and assume further that X is a
complex of finitely presented R◦ -modules. There is a morphism of k-complexes,
X ⊗R HomS (Y, Z) −→ HomS (HomR◦ (X, Y ), Z) ,
and it is functorial in X, Y , and Z. Furthermore, it is an isomorphism if X is a
complex of projective modules or Z is a complex of injective modules.
Proof. For every n ∈ Z one can compute as follows,
Y
(Xi ⊗R HomS (Y, Z)n−i )
(X ⊗R HomS (Y, Z))n =
i∈Z
=
Y
Q
(Xi ⊗R
∼
=
HomS (Yj , Zn−i+j ))
j∈Z
i∈Z
YY
(Xi ⊗R HomS (Yj , Zn−i+j )) ,
i∈Z j∈Z
where the isomorphism holds as the module Xi is finitely presented for every i ∈ Z;
see [16, thm. 3.2.22]. On the other hand, for every n ∈ Z one has
Y
HomS (HomR◦ (X, Y ), Z)n =
HomS (HomR◦ (X, Y )h , Zn+h )
h∈Z
=
Y
HomS (
h∈Z
∼
=
`
HomR◦ (Xi , Yi+h ), Zn+h )
i∈Z
YY
HomS (HomR◦ (Xi , Yi+h ), Zn+h )
YY
HomS (HomR◦ (Xi , Yj ), Zn−i+j ) .
h∈Z i∈Z
=
i∈Z j∈Z
Now set (θXY Z )n =
Q
i∈Z
Q
i(n−j)
θXi Yj Zn−i+j ,
j∈Z (−1)
where
θXi Yj Zn−i+j : Xi ⊗R HomS (Yj , Zn−i+j ) −→ HomS (HomR◦ (Xi , Yj ), Zn−i+j )
is the homomorphism of k-modules given by θXi Yj Zn−i+j (x ⊗ ψ)(φ) = ψφ(x). It is
straightforward to verify that θXY Z is a morphism of k-complexes and functorial
in X, Y , and Z.
22
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
Finally, if each module Xi is projective or each module Zn−i+j is injective, then
it follows from [7, prop. VI.5.2 and VI.5.3] that θXi Yj Zn−i+j is invertible for all
i, j, n ∈ Z, and so the morphism θXY Z is invertible.
A.7 Lemma. Let P be a bounded below complex of finitely generated projective
R◦ -modules and let X be a complex of R◦ -modules with H(X) bounded. For every
injective k-module E, there is an isomorphism in the derived category D(k):
] R◦ (P, X), E) ,
e R Homk (X, E) −→ Σ Homk (Hom
P⊗
and it is functorial in P , X, and E.
Proof. Set (−)∨ = Homk (−, E). In the commutative square of k-complexes
P ⊗R X ∨
τ
/ P ⊗R X ∨
∼
= α
θ
HomR◦ (P, X)∨
ϑ∨
/ HomR◦ (P, X)∨
each horizontal morphism is (the dual of) a canonical embedding. The vertical
map α is the isomorphism from Proposition A.6. The morphism θ is the standard
evaluation map; it is a quasi-isomorphism by [2, 4.4(I)]. The square induces a
morphism of triangles in the homotopy category:
τ
P ⊗R X ∨
/ P ⊗R X ∨
γ
≃ θ
∼
= α
HomR◦ (P, X)∨
/ HomR◦ (P, X)∨
∨
ϑ
/ Σ(P ⊗R X ∨ )
/ Cone τ
≃ Σθ
/ Σ HomR◦ (P, X)∨
/ Cone ϑ∨
The construction of γ is functorial in all three arguments, and it is a quasi-isomorphism
because α and θ are quasi-isomorphisms. Recall, say from [17, III.3.4–5], that there
are natural quasi-isomorphisms
∨
Cone ϑ
eR X∨
Cone τ ≃ Coker τ = P ⊗
∼
= Σ(Cone ϑ)
∨
∨
≃ Σ(Coker ϑ)
and
] R◦ (P, X)∨ .
= Σ Hom
They yield the desired isomorphism in the derived category
] R◦ (P, X), E) .
e R Homk (X, E) −→ Σ Homk (Hom
P⊗
With regard to functoriality of γ, notice that a morphism between arguments,
P → P1 say, induces the solid commutative square in the following diagram.
❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴/ Coker τ1
Coker
✤ e❑τ❑
8 ✤
❑❑
rrr ✤
✤
❑❑
r
r
r
≃ ❑❑❑
✤
✤
rrr ≃
/
✤
✤
Cone τ
Cone τ1
✤
✤
γ
γ1
✤
✤
✤
✤
∨
∨
/ Cone(ϑ1 )
Cone ϑ
✤
✤
▲▲▲
tt
✤
✤
≃ tt
▲▲≃
t
▲▲▲
tt
✤
✤
t
▲
%
zt
∨ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴ ❴/
Σ(Coker ϑ)
Σ(Coker ϑ1 )∨
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
23
Commutativity in the derived category of the dashed square is now a consequence.
Functoriality in the other arguments is handles similarly.
The Lemma immediately yields a useful duality.
A.8 Theorem. Let M and N be R◦ -modules and assume that M has a degreewise finitely generated projective resolution. For every injective k-module E and
for every i ∈ Z there is an isomorphism of k-modules,
g i ◦ (M, N ), E) ∼
g R (M, Homk (N, E)) ,
Homk (Ext
= Tor
R
i
and it is functorial in M , N , and E.
To prove the next two results one proceeds as in the proof of Proposition A.6.
A.9 Proposition. Let X be a complex of finitely generated R-modules and let Y
and Z be as in A.3. There is a morphism of k-complexes
HomR (X, Y ) ⊗S Z −→ HomR (X, Y ⊗S Z) ,
and it is functorial in X, Y , and Z. Furthermore, it is an isomorphism under each
of the following conditions
• Z is a complex of finitely generated projective modules
• X is a complex of finitely presented modules and Z is a complex of flat modules
• X is a complex of projective modules
A.10 Proposition. Let X be a complex of R-modules, let Y be a complex of
(R, S ◦ )-bimodules, and let Z be a complex of finitely presented S-modules. There
is a morphism of k-complexes,
HomR (X, Y ) ⊗S Z −→ HomR (X, Y ⊗S Z) ,
and it is functorial in X, Y , and Z. Furthermore, it is an isomorphism if X or Z
is a complex of projective modules.
Pinched tensor products. Christensen and Jorgensen devised in [12] a pinched
tensor product, − ⊗✶
R −, to compute Tate homology. In view of Theorem 3.10 their
proof of [12, thm. 3.5] applies verbatim to yield the next result; we refer the reader
to [12] for the definition of the pinched tensor product.
A.11 Theorem. Let M be an R◦ -module that has a Tate flat resolution (T, F ),
let A be an acyclic complex of R-modules and set N = C0 (A). For every i ∈ Z,
there is an isomorphism of k-modules
✶
∼
gR
Tor
i (M, N ) = Hi (T ⊗R A) .
The next corollary is an analogue of [12, Corollary 4.10].
A.12 Corollary. Let R be commutative and let M and N be Gorenstein flat Rmodules with corresponding totally acyclic complexes of flat modules T and T ′ ,
g R (M, N ) = 0 holds for all i ∈ Z, then T ⊗✶ T ′ is an acyclic
respectively. If Tor
i
R
complex of flat R-modules, and the following statements are equivalent:
′
(i) T ⊗✶
R T is a totally acyclic complex of flat R-modules.
g R (M, N ⊗R E) = 0 holds for every injective R-module E and all i ∈ Z.
(ii) Tor
i
24
O. CELIKBAS, L. W. CHRISTENSEN, L. LIANG, AND G. PIEPMEYER
′
When these conditions hold, M ⊗R N is a Gorenstein flat R-module and T ⊗✶
RT
is a corresponding totally acyclic complex of flat R-modules.
′
Proof. It follows from the definition of pinched tensor products that T ⊗✶
RT
R
g
is a complex of flat R-modules, and if Tori (M, N ) = 0 holds for all i ∈ Z,
then the complex is acyclic by Theorem A.11. It is totally acyclic if and only
′
✶
∼
if (T ⊗✶
R T ) ⊗R E = T ⊗R (T ⊗R E) is acyclic for every injective R-module E;
g R (M, N ⊗R E) = 0 holds for every injective R-module E
that is, if and only if Tor
i
and all i ∈ Z. Finally, it follows from the definition of pinched tensor products that
′
there is an isomorphism M ⊗R N ∼
= C0 (T ⊗✶
R T ).
References
[1] Maurice Auslander and Mark Bridger, Stable module theory, Memoirs of the American Mathematical Society, No. 94, American Mathematical Society, Providence, R.I., 1969. MR0269685
[2] Luchezar L. Avramov and Hans-Bjørn Foxby, Homological dimensions of unbounded complexes, J. Pure Appl. Algebra 71 (1991), no. 2-3, 129–155. MR1117631
[3] Luchezar L. Avramov, Srikanth B. Iyengar, and Joseph Lipman, Reflexivity and rigidity for
complexes. I. Commutative rings, Algebra Number Theory 4 (2010), no. 1, 47–86. MR2592013
[4] Luchezar L. Avramov and Alex Martsinkovsky, Absolute, relative, and Tate cohomology of
modules of finite Gorenstein dimension, Proc. London Math. Soc. (3) 85 (2002), no. 2, 393–
440. MR1912056
[5] Luchezar L. Avramov and Oana Veliche, Stable cohomology over local rings, Adv. Math. 213
(2007), no. 1, 93–139. MR2331239
[6] Hyman Bass and M. Pavaman Murthy, Grothendieck groups and Picard groups of abelian
group rings, Ann. of Math. (2) 86 (1967), 16–73. MR0219592
[7] Henri Cartan and Samuel Eilenberg, Homological algebra, Princeton Landmarks in Mathematics, Princeton University Press, Princeton, NJ, 1999, With an appendix by David A.
Buchsbaum, Reprint of the 1956 original. MR1731415
[8] Olgur Celikbas, Lars Winther Christensen, Li Liang, and Greg Piepmeyer, Complete homology over associative rings, Israel J. Math, to appear. Preprint arXiv:1501.00297 [math.RA].
[9] Lars Winther Christensen, Gorenstein dimensions, Lecture Notes in Mathematics, vol. 1747,
Springer-Verlag, Berlin, 2000. MR1799866
[10] Lars Winther Christensen, Anders Frankild, and Henrik Holm, On Gorenstein projective,
injective and flat dimensions—A functorial description with applications, J. Algebra 302
(2006), no. 1, 231–279. MR2236602
[11] Lars Winther Christensen and Henrik Holm, The direct limit closure of perfect complexes, J.
Pure Appl. Algebra 219 (2015), no. 3, 449–463. MR3279365
[12] Lars Winther Christensen and David A. Jorgensen, Tate (co)homology via pinched complexes,
Trans. Amer. Math. Soc. 366 (2014), no. 2, 667–689. MR3130313
[13] Ioannis Emmanouil, On the finiteness of Gorenstein homological dimensions, J. Algebra 372
(2012), 376–396. MR2990016
[14] Edgar E. Enochs, Sergio Estrada, and Alina C. Iacob, Balance with unbounded complexes,
Bull. Lond. Math. Soc. 44 (2012), no. 3, 439–442. MR2966988
[15] Edgar E. Enochs and Overtoun M. G. Jenda, Copure injective resolutions, flat resolvents
and dimensions, Comment. Math. Univ. Carolin. 34 (1993), no. 2, 203–211. MR1241728
[16] Edgar E. Enochs and Overtoun M. G. Jenda, Relative homological algebra, de Gruyter Expositions in Mathematics, vol. 30, Walter de Gruyter & Co., Berlin, 2000. MR1753146
[17] Sergei I. Gelfand and Yuri I. Manin, Methods of homological algebra, second ed., Springer
Monographs in Mathematics, Springer-Verlag, Berlin, 2003. MR1950475
[18] François Goichot, Homologie de Tate-Vogel équivariante, J. Pure Appl. Algebra 82 (1992),
no. 1, 39–64. MR1181092
[19] Shiro Goto, Vanishing of ExtiA (M, A), J. Math. Kyoto Univ. 22 (1982/83), no. 3, 481–484.
MR0674605
[20] Laurent Gruson and Christian U. Jensen, Dimensions cohomologiques reliées aux foncteurs
lim(i) , Paul Dubreil and Marie-Paule Malliavin Algebra Seminar, 33rd Year (Paris, 1980),
←−
Lecture Notes in Math., vol. 867, Springer, Berlin, 1981, pp. 234–294. MR0633523
STABLE HOMOLOGY OVER ASSOCIATIVE RINGS
25
[21] Henrik Holm, Gorenstein homological dimensions, J. Pure Appl. Algebra 189 (2004), no. 1-3,
167–193. MR2038564
[22] Henrik Holm, Rings with finite Gorenstein injective dimension, Proc. Amer. Math. Soc. 132
(2004), no. 5, 1279–1283. MR2053331
[23] Thorsten Holm and Peter Jørgensen, Triangulated categories: definitions, properties, and
examples, Triangulated categories, London Math. Soc. Lecture Note Ser., vol. 375, Cambridge
Univ. Press, Cambridge, 2010, pp. 1–51. MR2681706
[24] Alina Iacob, Absolute, Gorenstein, and Tate torsion modules, Comm. Algebra 35 (2007),
no. 5, 1589–1606. MR2317632
[25] Christian U. Jensen, On the vanishing of lim(i) , J. Algebra 15 (1970), 151–166. MR0260839
←−
[26] David A. Jorgensen and Liana M. Şega, Independence of the total reflexivity conditions for
modules, Algebr. Represent. Theory 9 (2006), no. 2, 217–226. MR2238367
[27] Li Liang, Tate homology of modules of finite Gorenstein flat dimension, Algebr. Represent.
Theory 16 (2013), no. 6, 1541–1560. MR3127346
[28] Masayoshi Nagata, Local rings, Interscience Tracts in Pure and Applied Mathematics, No. 13,
Interscience Publishers a division of John Wiley & Sons New York-London, 1962. MR0155856
[29] Barbara L. Osofsky, Homological dimension and cardinality, Trans. Amer. Math. Soc. 151
(1970), 641–649. MR0265411
[30] Christian Peskine and Lucien Szpiro, Dimension projective finie et cohomologie locale. Applications à la démonstration de conjectures de M. Auslander, H. Bass et A. Grothendieck,
Inst. Hautes Études Sci. Publ. Math. (1973), no. 42, 47–119. MR0374130
[31] Paul Roberts, Two applications of dualizing complexes over local rings, Ann. Sci. École Norm.
Sup. (4) 9 (1976), no. 1, 103–106. MR0399075
[32] Paul Roberts, Le théorème d’intersection, C. R. Acad. Sci. Paris Sér. I Math. 304 (1987),
no. 7, 177–180. MR0880574
[33] Oana Veliche, Gorenstein projective dimension for complexes, Trans. Amer. Math. Soc. 358
(2006), no. 3, 1257–1283. MR2187653
[34] Charles A. Weibel, An introduction to homological algebra, Cambridge Studies in Advanced
Mathematics, vol. 38, Cambridge University Press, Cambridge, 1994. MR1269324
University of Missouri, Columbia, MO 65211, U.S.A.
Current address: University of Connecticut, Storrs, CT 06269, U.S.A
E-mail address: [email protected]
Texas Tech University, Lubbock, TX 79409, U.S.A.
E-mail address: [email protected]
URL: http://www.math.ttu.edu/~ lchriste
Lanzhou Jiaotong University, Lanzhou 730070, China
E-mail address: [email protected]
Columbia Basin College, Pasco, WA 99301, U.S.A.
E-mail address: [email protected]
| 0math.AC
|
Journal of Machine Learning Research 1 (2015) 1-48
Submitted 1/15; Published 1/15
Inferring and Learning from Neuronal Correspondences
Ashish Kapoor
[email protected]
Microsoft Research
One Microsoft Way
Redmond, WA 98052, USA
arXiv:1501.05973v2 [q-bio.NC] 27 Jan 2015
E. Paxon Frady
[email protected]
Department of Neurobiology
University of California
San Diego, CA 92093, USA
Stefanie Jegelka
[email protected]
Department of Electrical Engineering and Computer Science
Massachusetts Institute of Technology
Cambridge, MA 02139, USA
William B. Kristan
[email protected]
Department of Neurobiology
University of California
San Diego, CA 92093, USA
Eric Horvitz
[email protected]
Microsoft Research
One Microsoft Way
Redmond, WA 98052, USA
Editor: TBD
Abstract
We introduce and study methods for inferring and learning from correspondences among
neurons. The approach enables alignment of data from distinct multiunit studies of nervous
systems. We show that the methods for inferring correspondences combine data effectively
from cross-animal studies to make joint inferences about behavioral decision making that
are not possible with the data from a single animal. We focus on data collection, machine learning, and prediction in the representative and long-studied invertebrate nervous
system of the European medicinal leech. Acknowledging the computational intractability
of the general problem of identifying correspondences among neurons, we introduce efficient computational procedures for matching neurons across animals. The methods include
techniques that adjust for missing cells or additional cells in the different data sets that
may reflect biological or experimental variation. The methods highlight the value harnessing inference and learning in new kinds of computational microscopes for multiunit
neurobiological studies.
Keywords:
PCA
Neurobiology, Metric Learning, Correspondence Matching, Probabilistic
c 2015 Ashish Kapoor, Paxon Frady, Stefanie Jegelka, William B. Kristan and Eric Horvitz.
Kapoor, Frady, Jegelka, Kristand and Horvitz
1. Introduction
Neurobiologists have long pursued an understanding of the emergent phenomena of nervous
systems, such as the neuronal basis for choice and behavior. Much research on neuronal
systems grapples with the complex dynamics of interactions among multiple neurons. New
techniques, such as calcium imaging, voltage-sensitive dye (VSD) imaging (Cacciatore et al.,
1999; Gonzalez and Tsien, 1995) and multi-unit electrode recordings, enable larger views of
nervous systems. However, for many experimental preparations, the amount of data that
can be collected via tedious experiments is limited. As an example, data from voltagesensitive dyes are time-limited because of bleaching of the dyes and also neuronal damage
caused by phototoxicity.
We have developed methods for combining the data from multiple experiments to pool
data on neural function. The approach allows us to make inferences from data sets that
are impossible to obtain from individual preparations. Coalescing the data from multiple
experiments is an intrinsically difficult problem because of the difficulty in matching cells
and their roles across animals. Variation is observed in nervous systems of individual animals
based on developmental differences as well as artifacts introduced in the preparation and
execution of experiments. Developing a means for identifying correspondences in cells across
animals would allow data to be pooled from multiple animals supporting deeper inferences
about neuronal circuits and behaviors.
We focus specifically on experimental studies of neurons composing the ganglia of H.
verbana (Briggman et al., 2005). The leech has a stereotypical nervous system consisting
of repeating packets of about 400 neurons. About a third of these neurons have been
identified, and these neurons can be found reliably in different animals. The remaining twothirds of neurons have yet to be identified, but are believed to maintain similar properties
and functional roles across animals. The general problem of correspondence matching of
the cells of two different animals is illustrated in Figure 1. We seek to identify neurons that
are equivalent across ganglia obtained from different animals. For example, the red cell in
animal a has several candidate correspondences in animal b, but with varying degrees of
similarity (indicated by shades of red). Ideally, we will find a one-to-one match via jointly
considering multiple similarities. With only two animals, this problem can be mapped to
a bipartite graph-matching task and can be solved optimally (Munkres, 1957). However,
we want to jointly solve the matching task for larger numbers of animals. Such matching
across multiple graphs defined by the individual nervous systems is an intractable problem
in the NP-hard class (Papadimitriou and Steiglitz, 1982). In addition, the problem is even
more difficult because such matching must also take into account variations in the numbers
and properties of neurons observed in different animals. These variations can be due to
both developmental differences (e.g., some neurons may be missing or duplicated) and
experimental artifacts (e.g., some neurons may be out of the plane of focus or destroyed in
the delicate dissection). A key challenge in this endeavor is the formulation of a similarity
measure that takes into account physical parameters of cells, such as their size and location,
as well as their functional properties.
2
Inferring and Learning from Neuronal Correspondences
Figure 1: Challenge of identifying correspondence among neurons in ganglia from different
animals. Given compatibility constraints, a correspondence algorithm seeks a oneto-one mapping between neurons in two H. verbana preparations. The goal is to
find out correspondences between cells in each of the animals. The color coding
illustrates compatibility constraints: Feasible matches for the highlighted red,
green, and blue cells in animal 1 (c) are found in animal 2 (d), as indicated with
matched colors. The degree of feasibility of the matches is depicted via shading of
cells in animal b, where the most compatible cell for each source neuron in animal
a is highlighted with a white border. Although the figure shows only two animals,
such compatibility constraints occur
3 across all pairs of the 6 animals used.
Kapoor, Frady, Jegelka, Kristand and Horvitz
Table 1: The proposed algorithmic framework
Step 1:
Learn Compatibility Measure
Step 2:
Recover Correspondence Map
Step 3:
Infer Missing Data
Given training data with pairs of match and
non-match cells, estimate the parameter matrix
A that defines the compatibility function
f ab (i, j), between any ith and j th neurons for
all animal pairs.
Start with an initialized empty match set S0 .
Iteratively determine the next best match
Mt by solving equation 3
and update St = St−1 ∪ Mt .
End when all the cells are matched.
Construct the matrix Y that aggregates
data from all the animals, where each
row corresponds to cells and are permuted
according to the matching. Use Probabilistic
PCA on Y to infer the missing data.
2. Machine Learning Framework
We use a set of neuronal data collected in Briggman et al. (2005), which consists of optical
VSD recordings from populations of 123-148 neurons in a mid-body segmental ganglion from
six different leeches. Earlier research on this data identified neurons involved in decisionmaking. In particular, the study aimed at understanding the roles of neuronal populations
in decisions to swim or to crawl following stimulation. Sensory neurons (DP nerve) were
stimulated in such a way that would elicit, with equal (0.5) likelihood, swimming or crawling. This previous study considered single cell activations and joint analysis of neurons using
dimensionally reduction techniques of PCA and LDA. However, these techniques were limited to one animal at a time. In the current study, we propose a framework that analyzes
data across-animals to increase the power of the analysis.
Rather than using a handcrafted measure, we have employed a machine-learning framework that relies on supervised training data. This algorithm estimates an appropriate
similarity function between neurons in different animals based on a training set of highconfidence correspondences. These correspondences are readily identified neurons in the
nervous systems of H. verbana (Muller et al., 1981). An important capability of the algorithm is to take into account the probabilistic nature of inferred correspondences. The
algorithm begins by learning a weighting function of relevant features that maximizes the
likelihood of matches within the training set. The next step of the approach is to jointly
solve the correspondence matching problem for neurons across animals, while considering
potential missing or extra cells in each animal. The final step is to consider correspondences
with functions that fill in missing data. As we will demonstrate below, pooling neurophysiological data from multiple studies in a principled manner leads to larger effective data
with greater statistical power than the individual studies.
4
Inferring and Learning from Neuronal Correspondences
Specifically, the pipeline for the methodology includes three steps (detailed in Table 1):
(1) Determining a similarity score across pairs of cells, (2) recovering correspondences that
are consistent with the similarity measure, and (3) estimating missing data. We describe
these steps in detail below:
2.1 Learning Similarity Measure for Cells
The goal in Step 1 is to learn a similarity function f ab (i, j) indicating the feasibility of a
match between the ith cell in animal a and the j th cell in animal b. The most desirable
characteristic for such a function is a high positive value for likely matches and diminishing
values for poor matches. Such a characteristic is captured by the exponentiation of a
negative distance measure among sets of features that represent multiple properties of cells.
Formally:
f ab (i, j) = e−[φ(i,a)−φ(j,b)]
T A[φ(i,a)−φ(j,b)]
(1)
Here, φ(·) are d dimensional feature representations for the individual cells for each
animal and summarizes physical (e.g., size, location etc.) and functional (e.g., optical
recordings) properties. A is a d × d parameter matrix with positive entries that are learned
from data. Intuitively, the negative log of the similarity function is a distance function
between the feature representations: a zero distance between two feature vectors result in
highest similarity measure of 1, whereas representations at further distance away in the
feature space have a diminishing value. The matrix A parameterizes this distance measure.
Given training data consisting of several probable pairs of matched neurons, we use A to
solve an optimization problem. We describe the details below.
The following list of features were used in our work:
• Structural features: Absolute position of the cell with respect to the entire observed
frame, relative position of cell in relation to the entire ganglion, absolute size of the cell
in pixels, indicator vector specifying packet the neuron is located in (among Central,
Left Anterior, Left Posterior, Right Anterior, Right Posterior or Central Posterior
packet), and relative position coordinate of the neuron in its respective packet.
• Functional features: Coherence of electrophysiological observations with swim oscillations and single cell discrimination time (see Briggman et al. (2005)) that distinguishes from swim versus the crawl behavior.
Intuitively, the optimization problem finds the parameter A that minimizes the distance
between pairs of cells that were tagged as matches, while maximizing the distance among
other pairs. Formally, parameter A of the compatibility measure is estimated by minimizing
the objective:
X
X
X
A∗ = arg min
[−2 log f ab (i, j) + log
f ab (i, j 0 ) + log
f ab (i0 , j)]
(2)
A
i,j
j 0 ∈b,j 0 6=j
i0 ∈a,i0 6=i
subject to the constraint that all entries of A are positive. The sum is over all the labeled
training pairs (i, j) tagged as likely matches. Intuitively the first term −2 log f ab (i, j) in
the objective prefers solutions that would collapse the distance between matched pairs to
5
Kapoor, Frady, Jegelka, Kristand and Horvitz
zero, while the rest of the terms prefer a solution where the distance between the rest of the
cells are maximized. The optimization is straightforward and a simple gradient descent will
always find a locally optimal solution. Note that a more appropriate constraint is positive
semi-definite condition on A, however we suggest using a non-negativity constraint due to
simplicity in optimization with almost no reduction in performance of the pipeline.
2.2 Correspondence Matching
In a second step, we calculate the correspondence matches across all the animals. Instead
of calculating all the matches simultaneously, the framework follows an iterative procedure:
future matches are made not only by using the similarity function, but also by comparing
the geometric and structural relationship of the candidates to the past matches. Besides
considering the distances induced by the similarity function (i.e. − log f ab (i, j)), and unlike
past work on graph matching (Williams et al., 1997; Bunke, 2000), the proposed method
utilizes knowledge of landmarks by inducing constraints that impose topological and geometric invariants. This match-making algorithm considers the iteration t and denotes the
set of already determined matches by St . The algorithm then determines Mt+1 , the next
set of neurons from all the animals to be matched by solving the following optimization
task:
X
Mt+1 = arg min
− log f ab (i, j) + λDLM (i, j, St )
(3)
M
all pairs(i,j)∈M
Here λ is the trade-off parameter that balances the compatibility measure with landmark
distances DLM (·) from the matches recovered in all the prior iterations. The landmark
distance computation provides important structural and topological constraints for solving
the correspondence tasks. Given anchor points, the landmark distances attempt to capture
structural and locational relationship with respect to the available landmarks. There are
several options such as commute distance (McKay, 1981; Lovasz, 1993) on a nearest-neighbor
graph, or Euclidean distance computed by considering either the locations or the feature
representation of the neurons. In our experiments, we compute landmark distances between
neuron i in animal a and neuron j in animal b with respect to a set of anchor points S as:
X
X
DLM (i, j, S) =
log f aa (i, i0 ) −
log f bb (j, j 0 ).
(4)
i0 ∈S
j 0 ∈S
The optimization problem in the above equation is solved using off-the-shelf energy minimization procedures (Boykov et al., 2001; Minka, 2005). The set of the newly discovered
matches are then included and the process is repeated until all matches stay the same (settle). Essentially, the goal is to find a set of matched neurons across all the animals such
that objective function is minimized. We start with a reasonable initialization of solution
(for example by solving for consecutive pairs of animals). This solution is iteratively refined by considering data drawn from one study at a time and searching for a replacement
neuron which would lower the total energy. Such replacements continue until no further
minimization is observed.
Utilizing landmarks are appropriate as an informative signal for matching neurons in
the leech, because there is a typified geometric structure. Although soma positions do vary
from animal to animal, often certain somas remain arranged with particular geometrical
6
Inferring and Learning from Neuronal Correspondences
Figure 2: Cell correspondences inferred across six H. verbana. Graphics show results of the
correspondence matching procedure across six animals. Color coding indicates
the correspondences, where matched cells across different animals share the same
color. We highlight two cells (depicted as 1 and 2) and show the matches as
lines linking neurons across the animals. Several cells remain unmatched and are
depicted using the dashed lines (unfilled interior). The algorithm is capable of
handling partial matches where cells are not present in all the six animals due to
true structural differences or losses either in their preparation or in their sensing.
relationships. For instance the Nut and AE cells typically form a box-like pattern, the N
and T sensory neurons usually will be arranged in a hemi-circle along the packet edge, which
often will wrap around the AP cell. These types of arrangements are useful for identification
of cells by eye, and we extend our algorithm to utilize these relationships.
The framework is extended to handle poor matches and missing cells by considering a
sink cell in every animal. The sink cell has a fixed cost of matching, denoted as c, and acts
as a threshold such that neuron matches with costs greater than c are disallowed. The sink
cells are a soft representation of the probability that a particular neuron was not visible
during a given preparation.
2.3 Pooling Across Animals
Finally in the third step, the framework reconstructs data corresponding to cells that are
missing and remain unobserved in some animals. In particular, if we consider the elec7
Kapoor, Frady, Jegelka, Kristand and Horvitz
Figure 3: Computed canonical ganglion for H. verbana derived from the correspondence
matching algorithm. We used the results of the correspondence matching algorithm to generate an average or canonical ganglion by computing mean location
and size for each cell that was matched across at least three different animals.
The shades of neurons are colored according to the weight determined by an LDA
projection that would distinguish between swim and crawl models (brighter color
mean higher weight; the colors used were arbitrary).
trophysiological activity for unobserved cells as latent random variables, then we can infer
those latent variables by exploiting the fact that they were observed in other animals. Once
we have correspondence information across animals, we can fill in missing electrophysiological data. Formally, we invoke data completion via Probabilistic Principle Component
Analysis (PPCA) (Roweis, 1998; Tipping and Bishop, 1999). To apply PPCA, we construct
a matrix Y, where each row corresponds to a neuron and each column corresponds to the
fluorescence intensity in a short time interval. Further, since the correspondences between
all the animals are calculated, we can stack the data from all the animals in Y such that
the rows are arranged according to the discovered correspondences. (we use -1 to denote
absence of data due to missing cells in an animal). The PPCA algorithm recovers the low
dimensional structure in the data, and inserts missing data via Expectation Maximization
8
Inferring and Learning from Neuronal Correspondences
(Dempster et al., 1977). The PPCA algorithm starts with an initialized low-dimensional
projection and alternates between the E-step and the M-step. The E-step is where the
missing data is estimated by considering statistical relationships in the data. The M-step
is where the estimates of the low-dimensional projection are further refined.
Consider the matrix Y (dimensions c × n), which consists of neuronal activity recordings
of c cells from all the animals, is constructed using the methodology described in text above.
We then first scale all the values in the matrix Y between 0 and 1. Lets denote the lowdimensional representation of the data as matrix X (dimensions k × n, where k < c) and
the principal components as C (dimensions c × k). The PPCA algorithm first initializes the
matrices X and C randomly and then alternates between the following two steps:
E-step: Estimate Ŷ = CX
T
T
M-step: Refine Xnew = (C T C)−1 C T and Cnew = Ŷ Xnew
(Xnew Xnew
)−1
The algorithm converges when the maximum change in any of individual dimensions of
estimates Y is less than 0.001. PPCA is guaranteed to converge so that it produces data
completion even for neurons that are not observed in some animals.
In our implementation, optimization for Step 1 (see Table 1) is performed via Limited
Memory BFGS (Liu and Nocedal, 1989) routine and energy minimization for the above
Equation is performed via iterative variational inference (Beal, 2003). There are three
parameters that need to be specified in the framework: c the upper limit on cost of allowed
matches, the trade-off parameter between compatibility and relative locality measure, and
k the dimensionality of the low dimensional projection in PPCA. These parameters are
determined via a cross-validation methodology. The cross validation is performed out by
considering the aggregated matrix Y , randomly reducing 10% of the observed data, and
considering the reconstruction error using an L2 norm on the removed data. This process is
repeated 10 times and parameters with minimum average reconstruction error are chosen.
The search space for parameters c and λ lie in log-scale (i.e. c and λ ∈ [10−5 , 10−4 , .., 105 ]),
while for k we try in a linear range (i.e. k ∈ [1, .., 25]).
3. Experiments
Training data for learning the parameter A was collected by an experimentalist (EPF)
who hand-annotated 815 different match pairs across all the animals. Fig. 1 shows the
resulting compatibility measure for these data. Note how the physical properties (such
as size, relative location, packet membership) of the most likely matches (highlighted by
a white outline) illustrate the quality of the learned function. The matching procedure
results in a correspondence map (Fig 2.) matching neurons across the 6 different animals.
Once the correspondence map was calculated, it was used to generate a prototypic model
of animal by averaging physical as well as functional properties (Fig. 3).
Because the resulting correspondence map was computed simultaneously across all the
animals, it provides a simple way to analyze the quality of the recovered solution. In addition
to the physical properties, the functional characteristics of any two matched neurons across
different animals are similar across animals (Figure 4). We observed that a simple estimator
9
Kapoor, Frady, Jegelka, Kristand and Horvitz
Figure 4: Activity of neurons in a leech ganglion from prior study (Briggman et al., 2005),
showing how their neuronal activity can be used to identify homologous neurons
across animals. Voltage-sensitive dye traces from two different neurons that were
considered to be matches by the correspondence matching algorithm. The traces
highlight that the algorithm has the capability of recovering correspondence across
cell that are functionally similar.
based on average activity of neurons in five animals predicted the activity of the sixth one
(Figure 5). Here we estimated the entire time series of activity for a given cell in an animal
by considering the activity for the corresponding cell across the rest of the five animals. Two
different models for swim and crawl mode are computed where the prediction is performed
by computing an average across all the observed time-series.
For all six animals, lower differences between the observed electrophysiological activity and predictions made by a model learned from rest of the animals confirm that the
framework had recovered correct correspondences between the neurons across animals.
Although the matching algorithm performs quite well, it is likely that the algorithm
is far from perfect. Many of the matched cells may not be correct. Since the functional
responses of the cells are a factor for the matching, cells with little functional signal will be
harder to match than those with big signals. Cells lacking functional signal, however, are
not providing a lot of information for predicting behavioral outcome. Thus, these cells likely
have poor matches, but are also likely the cells which are non-relevant to the swim-crawl
decision circuit. It is also possible that many cells are effectively the same given this data
set, but the matches do not truly reflect homologous pairs. We expect many cells to be
functionally the same and mismatching these similar cells may not hurt our analysis.
The correspondence matching algorithm enables pooling of the data across animals,
which allows exploration that was not feasible previously. For example, Figure 6 shows
10
Inferring and Learning from Neuronal Correspondences
Figure 5: Bar graphs that highlight the results to test the recovered correspondence using a leave-one-animal-out analysis. The plots were generated by first considering a candidate test animal and then building a predictive model for each cell
(from when the animal swam or crawled) using the remaining five animals. The
bar-chart compares mean-squared error between predicted and observed electrophysiological activity when matching using the proposed framework with random
selection. The differences across all of the six leave-one-animal-out test cases are
significant.
a 3-dimensional projection recovered by applying ISOMAP (Tenenbaum et al., 2000), a
non-linear dimensionality reduction method that is an extension to linear methods such as
PCA. Because the algorithm was applied to the entire pooled data, the recovered dimensions
are consistent across all the animals, and thus can be visualized and analyzed within the
same reference frames. Previously application of such techniques (such as PCA and LDA
in (Briggman et al., 2005)) was limited to a single animal at a time resulting in dimensions
which were incomparable across animals.
The pooling of data enabled by methodology proved to be valuable in predictive models
of decision making. Figure 7 shows that pooling the data across animals enable earlier
predictions of one of the two behaviors (swimming or crawling) following stimulation than
data from a single animal. Specifically, PCA was performed on pooled data and earliest
discrimination time between swim and crawl was determined according the procedure described in (Briggman et al., 2005). In Figure 3, we highlight cells in the composed canonical
ganglion that play an important role in the behavioral decisions of the animal. Combining
data across multiple animals enables transfer and overlay of information, allowing aggregation of important statistical parameters and more robust empirical models. Figure 8 shows
ganglion maps for six animals highlighting cells that contribute most towards discrimination
amongst the swim and the crawl trials. Note that the highly discriminative cells (towards
the red spectrum) are consistent in physical properties such as location and size across the
different animals. We also note that these cells are significantly different from cell 208 that
was identified in earlier studies (Briggman et al., 2005).
11
Kapoor, Frady, Jegelka, Kristand and Horvitz
Figure 6: Using correspondences to predict behavior from neuronal activity. Identification
of corresponding neurons across animals enables larger data sets to be constructed
by pooling observations from multiple preparations, which in turn enable deeper
and more accurate data analysis to address questions of interest. This figure shows
Non-linear projections generated by applying the ISOMAP algorithm. The blue
and red dots correspond to swim and crawl mode and depict the trajectory that
the voltage-sensitive dye trajectories take for each animal. Note that ISOMAP
applied for an individual animal might result in projections that are inconsistent
across the different animals. However, using the discovered correspondences of
neurons across animals, we combine the data from all six animals, and recover
projections that are consistent for all of the animals.
4. Related Work
The work described in this paper builds upon many different sub-areas of machine learning.
In particular, the key ingredients include metric learning, correspondence matching and
probabilistic dimensionality reduction (Roweis, 1998; Tipping and Bishop, 1999). Distance
metric learning is a fairly active research area. Most of the work in distance metric learning
focus on K-Nearest Neighbor (k-NN) classification scenario (Duda et al., 2001) and often
aim to learn a Mahalanobis metric that is consistent with the training data (Frome et al.,
2007; Bar-Hillel et al.; Weinberger et al., 2006; Davis et al., 2006). The distance metric
learning method employed in this paper is closest to the work of Goldberger et al. (2005)
and Globerson and Roweis (2006), but modified to just consider the sets of similar cells
given by the user.
12
Inferring and Learning from Neuronal Correspondences
Figure 7: Bar graphs showing that pooled data allows us to discriminate between swim and
crawl significantly earlier than what was reported earlier using a PCA analysis
on data from a single animal (Briggman et al., 2005).
Correspondence problems are employed in a multitude of applications. Computer vision
is particularly closer to our scenario. Among the simplest are transformations of rigid bodies, where geometry can be exploited (Goodrich and Mitchell, 1999; McAuley et al., 2008),
while correspondences among non-rigid objects, and between non-identical objects, can
pose significant challenges. Algorithms applied to more general correspondence problems
largely combine the compatibility of points by features with the local geometric compatibility of matches. Such models can be formulated as graphical models (McAuley et al.,
2008; Torresani et al., 2008; Starck and Hilton, 2007) or as selecting nodes in an association
graph (Cho et al., 2010; Cour et al., 2006), and have been extended to higher-order criteria
(Duchenne et al., 2009; Zass and Shashua, 2008; Lee et al., 2011). Other methods consider
the Laplacian constructed from a neighborhood graph (Umeyama, 1988; Escolano et al.,
2011; Mateus et al., 2008), and some models are learned from full training examples (Torresani et al., 2008). Closest to the idea of using reference points are approaches based on
seed points (Sharma et al., 2011), landmarks (Jegelka et al., 2014), coarse-to-fine strategies
(Starck and Hilton, 2007), and on guessing points that help orient the remaining points in
a rigid body (McAuley and Caetano, 2012).
13
Kapoor, Frady, Jegelka, Kristand and Horvitz
5. Conclusion and Future Work
The proposed methodology is likely to be even more useful in combination with other
data-centric analyses. For example, the model learned from past data can be employed
to guide future experimentation. By computing correspondences between the model and
and data from an ongoing experiment in real-time, we can then use the model to guide
information extraction strategies. The methodology can also be extended to perform withinleech analysis, such as discovering bilateral pairs of neurons. In addition, this methodology
can readily be used to analyze the simultaneous activity of multiple neurons in other animals.
We foresee valuable uses of the approach in overlaying data from larger nervous systems
and, moving beyond cells, to higher-level abstractions of nervous system organization, such
as components of retina or columns in vertebrate nervous systems. Given its simplicity and
the appeal of potentially pooling large quantities of data, the correspondence methodology
may find wide use in many areas of neuroscience.
Acknowledgments
We acknowledge the assistance of Johnson Apacible, Erick Chastain and Paul Koch.
References
A. Bar-Hillel, T. Hertz, N. Shental, and D. Weinshall. Learning a mahalanobis metric from
equivalence constraints. Journal of Machine Learning Research.
M. J. Beal. Variational algorithms for approximate bayesian inference. University College
London, 2003.
Y. Boykov, O. Veksler, and R. Zabih. Fast approximate energy minimization via graph
cuts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2001.
K. L. Briggman, H. D. I. Abarbanel, and W. B. Kristan. Optical imaging of neuronal
populations during decision-making. Science, 11, 2005.
H. Bunke. Recent developments in graph matching. 2000.
T. W. Cacciatore, P. D. Brodfuehrer, J. E. Gonzalez, T. Jiang, S. R. Adams, R. Y. Tsien,
W. B. Kristan Jr., and D. Kleinfeld. Identification of neural circuits by imaging coherent
electrical activity with fret-based dyes. Neuron, 23, 1999.
Y. Cho, J. Lee, and K. M. Lee. Reweighted random walks for graph matching. In European
Conference on Computer Vision, 2010.
T. Cour, P. Srinivasan, and J. Shi. Balanced graph matching. In Advances in Neural
Information Processing Systems, 2006.
J. Davis, B. Kulis, P. Jain, S. Sra, and I. Dhillon. Information-theoretic metric learning. In
International Conference on Machine Learning, 2006.
14
Inferring and Learning from Neuronal Correspondences
A. Dempster, N. Laird, and D. Rubin. Maximum-likelihood from incomplete data via the
em algorithm. Journal of Royal Statistical Society Series B, 39, 1977.
O. Duchenne, F. Bach, I. Kweon, and J. Ponce. A tensor-based algorithm for high-order
graph matching. In Computer Vision and Pattern Recognition, 2009.
R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley and Sons,
2001.
F. Escolano, E. Hancock, and M. Lozano. Graph matching through entropic manifold
alignment. In Computer Vision and Pattern Recognition, 2011.
A. Frome, Y. Singer, and J. Malik. Image retrieval and classification using local distance
functions. In Advances in Neural Information Processing Systems, 2007.
A. Globerson and S. Roweis. Metric learning by collapsing classes. In Advances in Neural
Information Processing Systems, 2006.
J. Goldberger, S. Roweis, G. Hinton, and R. Salakhutdinov. Neighbourhood components
analysis. In Advances in Neural Information Processing Systems, 2005.
J. E. Gonzalez and R. Y. Tsien. Voltage sensing by fluorescence resonance energy transfer
in single cells. Biophysical Journal, 1995.
M.T. Goodrich and J.S.B. Mitchell. Approximate geometric pattern matching under rigid
motions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(4):371–
379, 1999.
S. Jegelka, A. Kapoor, and E. Horvitz. An interactive approach to solving correspondence
problems. International Journal of Computer Vision, 108, 2014.
J. Lee, M. Cho, and K. M. Lee. Hyper-graph matching via reweighted random walks. In
Computer Vision and Pattern Recognition, 2011.
D. C. Liu and J. Nocedal. On the limited memory method for large scale optimization.
Mathematical Programming B, 1989.
L. Lovasz. Random walks on graphs: a survey. Combinatorics: Paul Erdos is Eighty, 2,
1993.
D. Mateus, R. Horaud, D. Knossow, F. Cuzzolin, and E. Boyer. Articulated shape matching
using laplacian eigenfunctions and unsupervised point registration. In Computer Vision
and Pattern Recognition, 2008.
J. McAuley and T. Caetano. Fast matching of large point sets under occlusion. Pattern
recognition, 45, 2012.
J.J. McAuley, T.S. Caetano, and M. S. Barbosa. Graph rigidity, cyclic belief propagation and point pattern matching. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 30(11):20472054, 2008.
15
Kapoor, Frady, Jegelka, Kristand and Horvitz
B. D. McKay. Practical graph isomorphism. Congressus Numerantium, 1981.
T. P. Minka. Divergence measures and message passing. Microsoft Research Technical
Report, 2005.
J. Munkres. Algorithms for the assignment and transportation problems. Journal of the
Society for Industrial and Applied Mathematics, 5, 1957.
C. Papadimitriou and K. Steiglitz. Combinatorial optimization: Algorithms and complexity.
Prentice Hall, Englewood Cliffs, NJ, 1982.
S. Roweis. Em algorithms for pca and spca. 1998.
A. Sharma, R. P. Horaud, J. Cech, and E. Boyer. Topologically-robust 3d shape matching based on diffusion geometry and seed growing. In Computer Vision and Pattern
Recognition, 2011.
J. Starck and A. Hilton. Correspondence labelling for wide-timeframe free-form surface
matching. In International Conference on Computer Vision, 2007.
J. B. Tenenbaum, V. de Silva, and J. C. Langford. A global geometric framework for
nonlinear dimensionality reduction. Science, 2, 2000.
M. E. Tipping and C. M. Bishop. Journal of the Royal Statistical Society, 61, 1999.
L. Torresani, V. Kolmogorov, and C. Rother. Feature correspondence via graph matching:
Models and global optimization. In European Conference on Computer Vision, 2008.
S. Umeyama. An eigendecomposition approach to weighted graph matching problems. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 10(5), 1988.
K. Weinberger, J. Blitzer, , and L. Saul. Distance metric learning for large margin nearest
neighbor classification. In Advances in Neural Information Processing Systems, 2006.
M. L. Williams, R. C. Wilson, and E. R. Hancock. Multiple graph matching with bayesian
inference. Pattern Recognition Letters, 18, 1997.
R. Zass and A. Shashua. Probabilistic graph and hypergraph matching. In Computer Vision
and Pattern Recognition, 2008.
16
Inferring and Learning from Neuronal Correspondences
Figure 8: Determining influential cells using linear discriminant analysis. The ganglion
maps from 6 experiments are shown. The maps are from the same experiments
as in Fig. 4. Cells are color-coded based on the magnitude of the contribution
to the linear discriminant direction. Red and yellow represent large magnitude
contributions, blue represents small contributions. We can see that there are at
17 do not include cell 208 (marked using white
least 3 cells that are influential and
arrow).
| 5cs.CE
|
Strongly verbally closed groups
Andrey M. Mazhuga
arXiv:1707.02464v2 [math.GR] 18 Jul 2017
Faculty of Mechanics and Mathematics
Moscow State University
Moscow 119991, Leninskie gory, MSU
[email protected]
Abstract
It was recently proven that all free and many virtually free verbally closed subgroups are algebraically closed in any group. We establish sufficient conditions for a group that is an extension
of a free non-abelian group by a group satisfying a non-trivial law to be algebraically closed in any
group in which it is verbally closed. We apply these conditions to prove that the fundamental groups
of all closed surfaces, except the Klein bottle, and almost all free products of groups satisfying a
non-trivial law are algebraically closed in any group in which they are verbally closed.
1
Introduction
A subgroup H of a group G is called verbally closed (in G) [8] (see also [10], [11], [4], [7]) if any equation
of the form w(x1 , . . . , xn ) = h, where w(x1 , . . . , xn ) ∈ Fn (x1 , . . . , xn ) and h ∈ H, having a solution in G
has a solution in H too.
A subgroup H of a group G is called algebraically closed (in G) if any system of equations of the
form {w1 (x1 , . . . , xn , H) = 1, . . . , wm (x1 , . . . , xn , H) = 1}, where wi (x1 , . . . , xn , H) ∈ Fn (x1 , . . . , xn ) ∗
H, having a solution in G has a solution in H too.
A subgroup H of a group G is called retract (of G) if G is a semidirect product of a normal
subgroup N and H (i.e. G = N ⋊ H).
It is easy to see that any retract is an algebraically closed subgroup and any algebraically closed
subgroup is a verbally closed subgroup. Thus, the question naturally arises: under what conditions on
a subgroup H and a group G the revers implications hold. It is known (see, e.g., [4]) that both of these
reverse implications do not hold in general. The following was established in [8]
R1) if G is finitely presented and H is a finitely generated algebraically closed subgroup in G, then H
is a retract of G;
R2) if G is finitely generated over H † and H is a equationally Noetherian‡ algebraically closed subgroup
in G, then H is a retract of G.
The work of the author was supported by the Russian Foundation for Basic Research, project no. 15-01-05823.
A group G is finitely generated over H if G = hH, Xi for some finite subset X ⊆ G.
‡
A group H is equationally Noetherian if any system of equations with coefficients from H and finitely many unknowns
is equal to its finite subsystem.
†
1
For the class of verbally closed subgroups, no similar structural descriptions in known. However,
in finitely generated free groups (see [8]) and finitely generated free nilpotent groups (see [11]) the
situation is rather simple: verbally closed subgroups, algebraically closed subgroups and retracts are the
same things. In article [4], the following theorem was established
Theorem [4]. Let G be any group and let H be its verbally closed virtually free infinite non-dihedral†
subgroup containing no infinite abelian noncyclic subgroups. Then
1) H is algebraically closed in G;
2) if G is finitely generated over H, then H is a retract of G.
We call H a strongly verbally closed group if H is an algebraically closed subgroup in any group
containing H as a verbally closed subgroup. Notice, that the first assertion of the given above theorem
describes a certain class of strongly verbally closed groups (in particular, all non-trivial free groups
belong to this class). Abelian groups is another class of strongly verbally closed groups (see Corollary
4). In this article we establish sufficient conditions for a group H that is an extension of a free nonabelian group by a group satisfying a non-trivial law‡ to be strongly verbally closed. These conditions
can be used to establish strong verbal closedness of rather wide class of groups. For instance, is Section
2 we apply them to prove the following
C1) the fundamental groups of all closed surfaces, except the Klein bottle, are strongly verbally closed;
C2) if Hi , i = 1, . . . , n, n ≥ 2 are non-trivial groups satisfying a non-trivial law and H = H1 ∗· · ·∗Hn 6≃
D∞ , then H is strongly verbally closed.
Notice, that any fundamental group H of a closed surface is linear because it admits faithful
representation in PSL2 (R) (see, e.g., [2]), and it is well-known that PSL2 (R) is linear. Therefore, H is
equationally Noetherian [1]. Thus, from C1) and R2) we have
Theorem. If H is the fundamental group of a closed surface, which is not the Klein bottle, then H is
a retract of any finitely generated over H group containing H as a verbally closed subgroup.
Before proceeding to the formulation of the main result, we introduce some notation. The symbol
hxi∞ means an infinite cyclic group generated by x. The symbols IA (H) and IA (hxi∞ ) mean the verbal
subgroups of H and hxi∞ generated by a word IA †† , respectively. The symbol IA (H)hxi∞ ∗IA (H) means
the normal closure of IA (H) in hxi∞ ∗ IA (H). The symbols Z(H) and H ′ mean the center and the
derived subgroup of H, respectively.
Main Theorem. If a group H such that there exists
T1) a short exact sequence of groups 1 → F → H → A → 1 such that F is a free non-abelian group
and A is a group satisfying a non-trivial law IA ;
T2) a divisible subgroup Q of H such that Q ⊆ Z(H) and Q ∩ H ′ = {1};
†
for an infinite group, non-dihedral means non-isomorphic to the free product of two groups of order two.
Let I(x1 , . . . , xr ) be an element of the free group Fr (x1 , . . . , xr ) with a basis x1 , . . . , xr , then we say that a group G
satisfies the law I if I(g1 , . . . , gr ) = 1 in G for all g1 , . . . , gr ∈ G.
††
Let I(x1 , . . . , xr ) be an element of the free group Fr (x1 , . . . , xr ) with a basis x1 , . . . , xr , then the verbal subgroup of
a group G determined by I, I(G), is I(G) = hI(g1 , . . . , gr )| g1 , . . . , gr ∈ Gi .
‡
2
T3) a generation set U = U −1 of H such that for any u ∈ U there are elements Eu,k (x, IA (H)) ∈
IA (hxi∞ ) · IA (H)hxi∞ ∗IA (H) , k = 1, . . . , nu such that the system of equations
{Eu,k (x, IA (H)) = Eu,k (u, IA (H)) | k = 1, . . . , nu }
has a unique solution in H/Q,
then H is strongly verbally closed.
Hence, to use this theorem one has to choose at one’s discretion: 1) a generating set U = U −1 of
H, 2) a non-trivial law, IA , in A, 3) a central divisible subgroup Q of H such that Q ∩ H ′ = {1} (it can
be shown that in this case Q is a direct factor of H, however, we will neither need or use this fact), 4) a
set of words {Eu,k (x, IA (H)) | k = 1, . . . , nu } ⊂ IA (hxi∞ ) · IA (H)hxi∞ ∗IA (H) for each u ∈ U (note that
these sets may differ for the different elements u ∈ U ).
Now we want to draw the reader’s attention to the fact that at the definition of a strongly verbally
closed group H verbal closedness of H (in some group G) is given as the assumption. We do not consider
the question in which groups a given group H can be verbally closed, and we restrict ourselves to the
following remark. It follows from Example 2 that any free group is strongly verbally closed. On the other
hand
Observation 1. If a group G has a generating set consisting of finite-order elements, then G does not
have a verbally closed non-trivial free subgroup.
Proof. Let Fr be a verbally closed free subgroup of G of rank r ≥ 1 and let S be a generating set of G
consisting of finite-order elements. If r = 1, then Fr = hf i∞ for some f ∈ G \ {1}. Let s1 s2 · · · sm be a
decomposition of f into the elements si ∈ S. Then, there is a natural number p ≥ 2 which is relatively
prime to the orders, ordG (si ), of all the elements si of this decomposition. Consider the equation:
xp1 xp2 · · · xpm = f.
i
This equation has a solution in G (it is easy to see that xi = sm
i , where pmi ≡ 1 (mod ordG (si )) is a
solution of the equation), but it has no solution in hf i∞ (since the element f is not a proper power in
hf i∞ ). Thus, this contradicts the verbal closedness of H in G.
If r ≥ 2, then Fr = hf i∞ ∗Fr−1 , for some f ∈ G\{1} and a free subgroup Fr−1 of G. The subgroup
hf i∞ is verbally closed in Fr (since hf i∞ is a free factor of Fr , and a free factor is a retract), therefore
(by the transitivity of verbal closedness), it is verbally closed in G and we again get a contradiction.
In Section 2 we discuss examples and corollaries showing how the Main Theorem can be applied.
Section 3 contains the proof of the Main Theorem. Our argument is a refined and more sophisticated
version of the proof of the Main Theorem of [4], but it also based on the use of Lee words [5].
Let us fix the notations. If k ∈ Z, x and y are elements of a group, then xy , xky , x−y and x−ky
denote y −1 xy, y −1 xk y, y −1 x−1 y and y −1 x−k y, respectively. The derived subgroup and the center of a
group G are denoted by G′ and Z(G), respectively. The commutator, [x, y], of two elements, x and y,
of a group we define as x−1 y −1 xy. If X is a subset of a group G, then |X|, hXi and CG (X) mean the
cardinality of X, the subgroup generated by X and the centralizer of X in G, respectively. Notations
hxin , n ∈ N and hxi∞ mean a cyclic group of order n and an infinite cyclic group generated by an
element x, respectively. The index of a subgroup H of a group G is denoted by (G : H). If A and B
are subgroups of a group G, then A · B = hab|a ∈ A, b ∈ Bi. The symbol AB means the normal closure
3
of a subgroup A in a group B. The symbols A ∗ B, A ∗ B, A ⋊ B и A × B mean the free product of
C
groups A and B, a free product of groups A and B with the amalgamated subgroup C, a semidirect
product of groups A and B and the direct product of groups A and B, respectively. Fn (x1 , . . . , xn ) or
Fn is the free group of rank n (with a basis x1 , . . . , xn ). Sometimes instead of a notation x1 , . . . , xn we
use an abridged notation x rewriting, for example, w(x1 , . . . , xn ) as w(x) or Fn (x1 , . . . , xn ) as F (x).
2
Examples and Corollaries
The following example has no particular theoretical interest and was included to demonstrate how the
Main Theorem can be applied to a concrete group.
Example 1. The group H = F2 (b, c) ⋊ hai2 , where F2 (b, c) is the free group freely generated by b, c and
ba = b−1 , ca = c is strongly verbally closed.
Proof. Consider the following short exact sequence of groups
1 → F2 (b, c) → F2 (b, c) ⋊ hai2 → hai2 → 1
and choose: 1) the generating set U = U −1 = (b, a), (c, a), (c−1 , a), (1, a) , 2) the law IA = t2 and
3) the group Q = {1}. Next, we choose for each u ∈ U the words Eu,k and verify that the conditions of
the Main Theorem hold.
For (b, a) we consider the following words: E(b,a),1 = x2 , E(b,a),2 = x, (b2 , 1) , E(b,a),3 =
x, (c2 , 1) ∈ IA (hxi∞ ) · IA (H)hxi∞ ∗IA (H) and show that the system
2
x = (b, a)2 = (1, 1), x, (b2 , 1) = (b, a), (b2 , 1) = (b4 , 1), x, (c2 , 1) = (b, a), (c2 , 1) = ( b, c2 , 1)
has the unique solution, x = (b, a), in H/Q = H. It is straightforward to verify that the solutions of
the equation x2 = (1, 1) in H are either of the form x = (w1 bk w2 , a), where k ∈ Z and w1a = w2−1 or
x = (1, 1). It is clear that x = (1, 1) is not a solution of the second equation. Substituting x = (w1 bk w2 , a)
−1 −k −1
in x, (b2 , 1) = (b4 , 1) we have w1 bk w2 b2 w2−1 b−k w1−1 b2 = b4 in F2 (b, c), that is b2w2 b w1 = b2 in
F (b, c), whence w1 bk w2 = bm for some m ∈ Z. Since we can assume that the word w1 bk w2 is reduced
and that the last letter of w1 and the first letter of w2 is not b, the last equality means that w1 = w2 = 1
(and k = m). Thus, the solutions of the first two equations of the system are of the form x = (bk , a),
k ∈ Z. Substituting x = (bk , a) in the third equation we obtain the equality bk , c2 = b, c2 in F2 (b, c),
from which it follows that k = 1.
A similar verification shows that one can choose the words E(c,a),1 = E(c−1 ,a),1 = x2 , E(c,a),2 =
E(c−1 ,a),2 = x, (b2 , 1) for (c, a) and (c−1 , a), and the words E(1,a),1 = x2 , E(1,a),2 = x, (b2 , 1) ,
E(1,a),3 = x, (c2 , 1) for (1, a).
The following two examples are the special cases of assertion 1) of the stated above theorem of
[4].
Example 2. A free group is strongly verbally closed.
Proof. Let H be a free group. If H is abelian, then H is strongly verbally closed in accordance with
Corollary 4. Let H be a non-abelian free group. Consider the following short exact sequence of groups
1→H →H →1→1
and choose U = H, IA = t, Q = {1} and Eu,1 = x ∈ IA (hxi∞ ) · IA (H)hxi∞ ∗IA (H) for each u ∈ U . Now,
the verification of the conditions of the Main Theorem is trivial.
4
Example 3. Virtually free non-virtually cyclic group with the unique extraction of roots of infinite-order
elements† is strongly verbally closed.
Proof. If H is a virtually free non-virtually cyclic group, then it is clear that there exists a short exact
sequens of groups of the form
1 → F → H → A → 1,
where F is a non-abelian free group, and A is a finite group. Let U be the set of all infinite-order elements
of H, IA = t|A| , Q = {1}, and Eu,1 = x|A| ∈ IA (hxi∞ ) · IA (H)hxi∞ ∗IA (H) for each u ∈ U .
In [4], the following was established
in a virtually free group which is not virtually cyclic, any element
decomposes into a product of two infinite-order elements.
It follows immediately from this statement that U is a generating set of H. The uniqueness of a
solution of an equation x|A| = u|A| , u ∈ U in H/Q = H is assumed in the hypothesis of the example.
The verification of the remaining conditions of the Main Theorem is trivial.
Corollary 1. Let 1 → F → H → A → 1 be a short exact sequence of groups such that F is a non-abelian
free group, and A is a group satisfying a non-trivial law IA . If Z(H) is a divisible group, Z(H)∩H ′ = {1}
and there exist elements f1 , . . . , fm ∈ IA (H) such that CH (f1 , . . . , fm ) = Z (H), then H is strongly
verbally closed.
Proof. Let us put U = H, Q = Z(H) and Eu,k = [fk , x] ∈ IA (hxi∞ ) · IA (H)hxi∞ ∗IA (H) , k = 1, . . . , m
for each u ∈ U .
To verify the conditions of the Main Theorem we only need to check that for u ∈ U the system
{[fk , x] = [fk , u] | k = 1, . . . , m}
has a unique solution in H/Q. It is clear that the system has a solution, x = uQ, in H/Q. Let x
bQ
′
be a solution of this system in H/Q, then (since Q = Z(H), and Z(H) ∩ H = {1}) for all k we have
T
bu−1 ∈ CH (fk ), k = 1, . . . , m. That is, x
bu−1 ∈ m
b−1 fk x
b = fk−1 u−1 fk u in H, thus x
fk−1 x
k=1 CH (fk ) =
CH (f1 , . . . , fm ) = Q, therefore x
bQ = uQ in H/Q.
Further we will need the following technical lemma.
Lemma 1. Let 1 → F → H → A → 1 be a short exact sequence of groups such that F is a non-abelian
free group, and A is a group satisfying a non-trivial law IA . Then IA (H) is a non-abelian free subgroup
of H.
Proof. The group IA (H) is free as a subgroup of a free group F . Subgroup IA (F ) in normal (even
fully invariant) in F as a verbal subgroup and IA (F ) 6= {1} since IA is a non-trivial word. Therefore,
if IA (H) is abelian, then IA (F ) ≃ Z. It can be easily deduce from the Nielsen-Schreier Theorem that a
non-abelian free group cannot have a normal infinite cyclic subgroup. Thus, IA (H) is a non-abelian free
group.
†
that is, if h1 , h2 are the infinite-order elements and hk1 = hk2 , k ∈ N, then h1 = h2 .
5
Corollary 2. Let H = H1 ∗ · · · ∗ Hn 6≃ D∞ , n ≥ 2 be a free product of non-trivial groups Hi satisfying
a non-trivial law, then H is strongly verbally closed.
Proof. Consider the following short exact sequence of groups:
1 → F → H → A → 1,
where A = H1 × · · · × Hn and F is the kernel of the natural epimorphism H1 ∗ · · · ∗ Hn ։ A (that is, F
is the cartesian subgroup). It is well-known that the cartesian subgroup is a free group (and it is clear
that if H 6≃ D∞ and n ≥ 2, then F is non-abelian free).
Let [x1 , x2 , . . . , xn−1 , xn ] = [[[. . . [[x1 , x2 ] , x3 ] , . . . ] , xn−1 ] , xn ] be the left-normed commutator of
x1 , x2 , . . . , xn−1 , xn , and let Wi be a non-trivial law in Hi (we assume that for the different indices the
words Wi do not contain the same letters). It is easy to see that in this case A satisfies the non-trivial
law IA = [W1 , W2 , . . . , Wn−1 , Wn ]. We shall verify the conditions of Corollary 1.
The group B = IA (H) is non-abelian free by Lemma 1. Therefore, there exist elements f1 , f2 ∈ B
such that [f1 , f2 ] 6= 1. It is well-known that the cartesian subgroup trivially intersects the free factors
Hi , therefore (since B is a subgroup of the cartesian subgroup F ), the elements f1 , f2 cannot belong to
a subgroup of the form Hig , g ∈ H. If CH (f1 ) ∩ CH (f2 ) 6= {1}, then the center of the group hf1 , f2 , zi,
where z ∈ CH (f1 ) ∩ CH (f2 ) \ {1} is non-trivial, therefore the group hf1 , f2 , zi does not decompose as a
proper free product. But by the Kuros̆ Subgroup Theorem (see. e.g., [9]) this would imply that hf1 , f2 , zi
is a subgroup of Hig , g ∈ H. Thus, CH (f1 ) ∩ CH (f2 ) = CH (f1 , f2 ) = 1.
Sice the fundamental group of a connected non-closed surface is free (see, e.g., [12])), by Example
2, it is strongly verbally closed. It is well-known (see, e.g., [6]) that the fundamental group of a closed
surface does not decompose as a proper free product, nevertheless
Corollary 3. The fundamental group of a closed surface, except the Klein bottle, is strongly verbally
closed.
Further in the text of the article a surface means a closed (i.e. connected, compact, without
boundary) surface.
It is well-known that the fundamental group of an orientable surface of genus g ≥ 0 (we denote
such a surface as Sg+ ) is of the type:
π1 (Sg+ ) = hx1 , y1 , . . . , xg , yg | [x1 , y1 ] · · · [xg , yg ]i ,
and that the fundamental group of a non-orientable surface of genus g ≥ 1 (we denote such a surface as
Sg− ) is of the type:
π1 (Sg− ) = x1 , . . . , xg | x21 · · · x2g .
The following technical lemma is rather well-known.
Lemma 2. Let π1 (S) be the fundamental group of a surface S whose Euler characteristic† is χ(S).
S1) If χ(S) < 0 and elements g1 , g2 ∈ π1 (S) \{1} commute, then hg1 , g2 i ≃ Z.
S2) If χ(S) < 0, then π1 (S)′ is a non-abelian free group.
†
The Euler characteristic, χ(Sg+ ), of Sg+ is χ(Sg+ ) = 2 − 2g and the Euler characteristic, χ(Sg− ), of Sg− is χ(Sg− ) = 2 − g.
6
Proof. For a proof of S1) see, for example, [3].
It is well-known (see, e.g., [3]) that π1 (S)′ is a free group. If χ(S) < 0, then g ≥ 2 for Sg+ and
g ≥ 3 for Sg− . Therefore, the following epimorphisms are naturally defined
π1 (Sg+ ) = hx1 , y1 , . . . , xg , yg | [x1 , y1 ] · · · [xg , yg ]i ։ hx1 , y1 , x2 , y2 | [x1 , y1 ][x2 , y2 ]i ≃ F2 ∗ F2 ;
Z
π1 (Sg− )
=
x1 , . . . , xg | x21 · · · x2g
։ x1 , x2 , x3 |
x21 , x22 , x23
≃ Z2 ∗ Z2 ∗ Z2 .
Since the commutator subgroups of the images of these epimorphisms are non-abelian groups, it follows
that the commutator subgroups, π1 (Sg+ )′ , g ≥ 2 and π1 (Sg− )′ , g ≥ 3, are also non-abelian.
Proof. It is clear that the groups π1 (S0+ ), π1 (S1+ ) and π1 (S1− ) are abelian, therefore, by Corollary
4, these groups are strongly verbally closed. The group π1 (S2− ) is excluded by the hypothesis of the
corollary. Thus, it remains to consider the surfaces whose Euler characteristic is negative.
For H = π1 (S), χ(S) < 0 consider the following short exact sequence of groups
1 → H ′ → H → A → 1,
where A = H/H ′ . By assertion S2) of Lemma 2, H ′ is a non-abelian free group. Assuming that IA =
[t1 , t2 ], we verify the conditions of Corollary 1.
By Lemma 1, B = IA (H) is a non-abelian free subgroup of H and, therefore, there are
elements f1 , f2 ∈ B such that [f1 , f2 ] 6= 1. Suppose that there exists an element, say z, such that
z ∈ CH (f1 ) ∩ CH (f2 ) \ {1}. Then, by assertion S1) of Lemma 2, there are elements u1 , u2 ∈ H such
1
2
that f1 = uk11 , z = um
and f2 = uk22 , z = um
for some k1 , k2 , m1 , m2 ∈ Z \ {0}. Therefore, we
1
2
k2 m1
k1 k2 m1
k
k
m
k
m
k
k
1
2
2
1
2
have f1
= u1
= z 1 2 = u2
= f2
. Since f1 and f2 are elements of a free group,
k2 m1
k1 m2
the equality f1
= f2
means that f1 and f2 commute, but this contradicts our choice. Whence,
CH (f1 , f2 ) = CH (f1 ) ∩ CH (f2 ) = {1}.
3
Proof of the Main Theorem
The following lemma is well known.
Lemma 3. If a subgroup H of a group G is such that any finite system of equations of the form
{wi (x1 , . . . , xn ) = hi | i = 1, . . . , m} ,
(1)
where wi (x1 , . . . , xn ) ∈ Fn (x1 , . . . , xn ), hi ∈ H, having a solution in G has a solution in H too, then H
is algebraically closed.
Proof. Just denote the coefficients by new letters and interpret them as variables. For example, the
solvability of the equation xyh1 [x2018 , h2 ]y −1 = 1 is equivalent to the solvability of the system
{xyz[x2018 , t]y −1 = 1, z = h1 , t = h2 }.
Now, recall that any integer matrix can be reduced to a diagonal matrix by integer elementary
transformations. This means that any finite system of equations of the form (1) can be reduced to a
system of the form
i
{xm
i ui (x1 , . . . , xn ) = hi , uj (x1 , . . . , xn ) = hj | i = 1, . . . , l, j = l + 1, . . . , m} ,
7
(2)
where ui (x1 , . . . , xn ), uj (x1 , . . . , xn ) ∈ F (x1 , . . . , xn )′ , mi > 0, 0 ≤ l ≤ n, hi , hj ∈ H, by means of a
finite sequence of transformations of the form wi → wi wj±1 and xi → xi x±1
j . It follows from the form of
these transformations that system (1) has a solution in a group G (containing H as a subgroup) if and
only if system (2) has a solution in G. Therefore, the following lemma holds.
Lemma 4. If a subgroup H of a group G is such that any finite system of the form (2) having a solution
in G has a solution in H too, then H in algebraically closed.
Corollary 4. An abelian group is strongly verbally closed.
Proof. Let H be a verbally closed abelian subgroup in a group G. Suppose that a system of the form
(2) has a solution in G. Since H is a verbally closed subgroup in G, each equation of this system has
a solution in H. For each i = 1, . . . , l let a tuple (ai,1 , ai,2 , . . . , ai,n ) be a solution of the equation
′
i
xm
i ui (x1 , . . . , xn ) = hi in H. Since ui (x1 , . . . , xn ), uj (x1 , . . . , xn ) ∈ F (x1 , . . . , xn ) and H is abelian, we
have ui (g1 , . . . , gn ) = 1, uj (g1 , . . . , gn ) = 1 for any g1 , . . . , gn ∈ H, thus hj = 1, j = l + 1, . . . , m. Now
it is easy to see that xi = ai,i , i = 1, . . . , l and xi = 1, i = l + 1, . . . , m is a solution of (2) in H.
Recall that a Lee word in m variables for the free group of rank r is an element Lm (z1 , . . . , zm ) of
the free group of rank m such that
′ ) 6= 1 in F , then v ′ ∈ F are obtained from v ∈ F by
L1) if Lm (v1 , . . . , vm ) = Lm (v1′ , . . . , vm
r
r
i
r
i
simultaneous conjugation, i.e., there exists s ∈ Fr such that vi′ = vis for all i = 1, . . . , m;
L2) Lm (v1 , . . . , vm ) = 1 if and only if the elements v1 , . . . , vm of Fr generate a cyclic subgroup.
In [5], such words were constructed for all integers r, m ≥ 2. Actually, it is easy to see that Lee’s
result implies the existence of a universal Lee word in m variables.
Lemma [4]. For any positive integer m, there exists an element Lm (z1 , . . . , zm ) ∈ Fm such that
properties L1) and L2) hold in all free groups Fr and even in F∞ .
Proof. This assertion follows immediately from Lee’s result and the following simple fact:
F∞ embeds into F2 as a malnormal subgroup,
i.e. a subgroup S ⊂ F2 such that S f ∩ S = {1} for all f ∈ F2 \ S. This fact follows, e.g., from a result
of [13]:
in a free group, any set satisfying small-cancellation condition C(5)
freely generates a malnormal subgroup.
Thus, a Lee word for F2 is universal, i.e. it is suitable also for F∞ .
The proof of the following lemma is based on the use of universal Lee words.
Lemma 5. Let I(H) and I (F (x)) be the verbal subgroups of H and F (x) generated by a word I,
respectively. If I(H) is a non-abelian free group and H is a verbally closed subgroup in a group G, then
any finite system of equations of the form
{wi (x, I(H)) = 1| i = 1, . . . , m} ,
where wi (x, I(H)) ∈ I (F (x)) · I(H)F (x)∗I(H) , having a solution in G has a solution in H too.
Proof. Let S be a system of the form (3), having a solution in G.
8
(3)
By the hypothesis of the lemma, the left-hand sides, wi (x, I(H)), of the equations of S have
decompositions of the form
mi
Y
c
bi,ji,j ,
(4)
wi (x, I(H)) = ai
j=1
where ai ∈ I (F (x)), bi,j ∈ I(H) and ci,j ∈ F (x) ∗ I(H). We fix such a decomposition for the left-hand
side of each equation of S. Let Mi,j be the set of all the elements of I(H) which are present in the
normal form (for F (x) ∗ I(H)) of the element ci,j . We define the set M as the union of all the sets Mi,j
and all the elements bi,j . Since I(H) is a non-abelian free group, there exist elements h′ , h′′ ∈ I(H) such
that hM, h′ , h′′ i 6≃ Z. If hM i ≃ Z or M = ∅, we add the elements h′ , h′′ to the set M .
Since M ⊂ I(H), for each hβ ∈ M there is a word, tβ (y β ) ∈ I(F (y β )), such that tβ (g β ) = hβ for
some elements g β of H (we assume that for the different indices the tuples y β do not contain the same
letters). For each element hβ ∈ M we add to the system S the equation tβ (y β ) = hβ , and each occurrence
of the constant hβ in the fixed above decompositions (4) of the left-hand sides of the equations of the
system S we replace with the word tβ (y β ). We denote this new system by S 1 . Note that the left-hand
sides of the equations of S 1 do not contain constants from H; their right-hand sides are the constants
from I(H); the system S 1 has a solution in G if and only if the system S has a solution in G.
Let us denote the equations of the system S 1 by uα (x, y) = fα , where y is a concatenation of all
the tuples y β † . That is
S 1 = uα (x, y) = fα | α = 1, . . . , N ,
where N = m + |M |. Let LN (z1 , . . . , zN ) be a universal Lee word in N variables. Consider the equation
LN (u1 (x, y), . . . , uN (x, y)) = LN (f1 , . . . , fN ).
This equation has a solution in G by construction (just take the following: a solution of S as x and g β
as y β ). Thus, since H is a verbally closed subgroup in G and LN (f1 , . . . , fN ) ∈ I(H) ⊆ H, the equation
has a solution, say x̂, ŷ, in H.
By construction a word uα (x, y) is either of the form
wi (x, y) ∈ I (F (x)) · I(F (y))F (x)∗I(F (y))
(5)
tβ (y β ) ∈ I(F (y)).
(6)
or of the form
It is clear from (5) and (6) that for any values of x, y in H the value of uα (x, y) is an element of the
free group I(H).
Since I(H) is a free group, fα , uα (x̂, ŷ) ∈ I(H) for all α = 1, . . . , N and hf1 , . . . , fN i 6≃ Z
(indeed, by construction M ⊆ {f1 , . . . , fN } and hM i 6≃ Z), Property L2) of Lee words implies that
LN (f1 , . . . , fN ) 6= 1. Then, according to Property L1) of Lee words, there exists s ∈ I(H) such that
Ŝ 1 = uα (x̂, ŷ) = fαs | α = 1, . . . , N .
−1
−1
That is, x̂s , ŷ s is a solution to the system S 1 in H. Recall how the system S 1 was obtained from the
system S. Now we carry out the inverse transformation, but with the set of equalities Ŝ 1 . More precisely,
using the equalities of the form tβ (ŷ β ) = hsβ of Ŝ 1 , in the left-hand sides of all the equalities of the form
wi (x̂, ŷ) = 1s of Ŝ 1 we replace (according to the fixed above decompositions (4) of the left-hand sides of
†
if e.g., (y1 , y2 ) and (y3 , y4 ) are tuples, then we call the tuples of the form (y1 , y2 , y3 , y4 ) or (y3 , y4 , y1 , y2 ) their
concatenation.
9
the equations of S) all the elements tβ (ŷ β ) with the elements hsβ . This leads us to the following set of
equalities
Ŝ = {wi (x̂, I(H)s ) = 1| i = 1, . . . , m} .
Whence, x̂s
−1
is a solution to the system S in H as required.
Now we are ready to prove the Main Theorem.
Proof of the Main Theorem. Let H be a verbally closed subgroup in a group G, then (by Lemma
4) it suffices to show that a system S of the form (2) having a solution in G has a solution in H too.
To each equation wi (x) = hi of the system S we associate the system
n
o
−1
−1
Si = wi (x) yi,t
,
·
·
·
y
=
u
,
y
=
u
,
.
.
.
,
y
=
u
i,0
i,1
i,1
i,t
i,t
i
i
i,1
i
(7)
where hi = ui,0 ui,1 . . . ui,ti , ui,j ∈ U and yi,j are the new unknowns. Consider the system
S 1 = {Si | i = 1, . . . , m} ,
consisting of all the equations of all the systems Si . It is clear that the system S 1 has a solution in G if
and only if the system S has a solution in G.
Using the elements Eui,j ,k (x, IA (H)), k = 1, . . . , nui,j from Condition T3) of the Main Theorem,
−1
−1
= yi,j , j =
x,
·
·
·
y
and
w
y
to each equation wi,j x, y = ui,j (where wi,0 x, y = wi (x) yi,t
i,j
i,1
i
1, . . . , ti ) of the system S 1 we associate the system
Si,j = Eui,j ,k wi,j x, y , IA (H) = Eui,j ,k (ui,j , IA (H)) | k = 1, . . . , nui,j .
Next, consider the system
S 2 = {Si,j | i = 1, . . . , m, j = 0, . . . , ti } ,
consisting of all the equations of all the systems Si,j . It is clear that if the system S 1 has a solution in
G, then the system S 2 has a solution in G too.
Since IA (H) is a non-abelian free group (by Lemma 1), Eui,j ,k wi,j x, y , IA (H) ∈ IA (F (x, y)) ·
IA (H)F (x,y)∗IA (H) and Eui,j ,k (ui,j , IA (H)) ∈ IA (H), it is clear that the system S 2 is of the form (3).
Therefore, by Lemma 5, this system has a solution, say x̂, ŷ, in H. In turn, in accordance with Condition
T3) of the Main Theorem, this implies that the following equalities hold in H/Q
wi,j x̂, ŷ Q = ui,j Q| i = 1, . . . , m, j = 0, . . . , ti .
Therefore, for some elements qi,j ∈ Q we have the following set of equalities (in H)
o
n
−1
−1
·
·
·
ŷ
=
u
q
,
ŷ
=
u
q
,
.
.
.
,
ŷ
=
u
q
|
i
=
1,
.
.
.
,
m
.
Ŝ 1 = wi (x̂)ŷi,t
i,0 i,0
i,1
i,1 i,1
i,ti
i,ti i,ti
i,1
i
Whence, since Q is central, there exist the elements qi ∈ Q such that the following equalities hold
Ŝ = {wi (x̂) = ui,0 ui,1 · · · ui,ti qi = hi qi | i = 1, . . . , m} .
(8)
This means that x̂ is a solution to S in H up to some elements qi ∈ Q.
In accordance with (2), the equality wi (x̂) = hi qi from (8) can be either of the form
i
ui (x̂1 , . . . , x̂n ) = hi qi or of the form x̂m
i ui (x̂1 , . . . , x̂n ) = hi qi , mi > 0.
i
Suppose that x̂m
i ui (x̂1 , . . . , x̂n ) = hi qi , mi > 0. Then (since Q is a divisible group) there is ri ∈ Q
such that rimi = qi . If we replace the element x̂i of the tuple x̂ with the element ri−1 x̂i (and leave the
10
other elements unchanged), then (since Q is a cental subgroup of H and ui (x1 , . . . , xn ) ∈ F (x1 , . . . , xn )′ )
we have
ri−1 x̂i
mi
−1
i
ui (x̂1 , . . . , ri−1 x̂i , . . . , x̂n ) = ri−mi x̂m
i ui (x̂1 , . . . , ri x̂i , . . . , x̂n )
i
= qi−1 x̂m
i ui (x̂1 , . . . , x̂i , . . . , x̂n )
= qi−1 hi qi = hi ,
while the other equalities of Ŝ remain unchanged.
Suppose that ui (x̂1 , . . . , x̂n ) = hi qi . The equation ui (x1 , . . . , xn ) = hi has a solution in H (indeed,
this equation has a solution in G since the system S does, and H is a verbally closed subgroup in G).
′
It means (since ui (x1 , . . . , xn ) ∈ F (x1 , . . . , xn )′ ) that hi ∈ H ′ , thus qi = h−1
i ui (x̂1 , . . . , x̂n ) ∈ H . But
by Condition T2) of the Main Theorem we have Q ∩ H ′ = {1}, therefore qi = 1.
Thus, the tuple (r1−1 x̂1 , . . . , rl−1 x̂l , x̂l+1 , . . . , x̂n ) is a solution to the system S in H as required.
4
Acknowledgments
The author thanks A. A. Klyachko for many useful conversations and many useful remarks.
References
[1] G. Baumslag, A. Myasnikov, V. Remeslennikov, Algebraic geometry over groups I. Algebraic sets
and ideal theory. J. Algebra. 219 (1999), 16-79.
[2] C. M. Campbell, M. R. Quick, E. F. Robertson, G. C. Smith, Groups St Andrews 2005 Volume 1.
Cambridge University Press, 2007.
[3] W. Jaco, On certain Subgroups of the fundamental group of a closed surface. Mathematical Proceedings of the Cambridge Philosophical Society. Vol. 67, Issue 01, (1970), 17-18.
[4] A. A. Klyachko, A. M. Mazhuga, Verbally closed virtually free subgroups, Sb. Math. (to appear).
See also arXiv: 1702.07761
[5] D. Lee, On certain C-test words for free groups. J. Algebra. 247 (2002), 509-540.
[6] R. C. Lyndon, P. E. Schupp, Combinatorial Group Theory. Springer-Verlag, New York, 1977.
[7] A. M. Mazhuga, On free decompositions of verbally closed subgroups of free products of finite
groups, Journal of Group Theory (to appear). See also arXiv:1605.01766
[8] A. Myasnikov, V. Roman’kov, Verbally closed subgroups of free groups, Journal of Group Theory
17 (2014), 29-40. See also arXiv:1201.0497.
[9] D. J. S. Robinson, A Course in the Theory of Groups. Springer, 1995.
[10] V. A. Roman’kov, Equations over groups, Groups Complexity Cryptology 4:2 (2012), 191-239.
[11] V. A. Roman’kov, N. G. Khisamiev, Verbally and existentially closed subgroups of free nilpotent
groups. Algebra and Logic 52 (2013), 336-351.
11
[12] J. Stillwell, Classical Topology and Combinatorial Group Theory. Springer, 1995.
[13] D. T. Wise, The residual finiteness of positive one-relator groups, Comment. Math. Helv 76 (2001),
314-338.
12
| 4math.GR
|
An Improved Algorithm for Diameter-Optimally
Augmenting Paths in a Metric Space
Haitao Wang
Department of Computer Science, Utah State University, Logan, UT 84322, USA.
[email protected]
Abstract
arXiv:1608.04456v1 [cs.DS] 16 Aug 2016
Let P be a path graph of n vertices embedded in a metric space. We consider the problem of
adding a new edge to P such that the diameter of the resulting graph is minimized. Previously
(in ICALP 2015) the problem was solved in O(n log3 n) time. In this paper, based on new
observations and different algorithmic techniques, we present an O(n log n) time algorithm.
Keywords and phrases diameter, path graphs, augmenting paths, minimizing diameter, metric
space
1
Introduction
Let P be a path graph of n vertices embedded in a metric space. We consider the problem
of adding a new edge to P such that the diameter of the resulting graph is minimized. The
problem is formally defined as follows.
Let G be a graph and each edge has a non-negative length. The length of any path of
G is the total length of all edges of the path. For any two vertices u and v of G, we use
dG (u, v) to denote the length of the shortest path from u to v in G. The diameter of G is
defined as maxu,v∈G dG (u, v).
Let P be a path graph of n vertices v1 , v2 , . . . , vn and there is an edge e(vi−1 , vi ) connecting
vi−1 and vi for each 1 ≤ i ≤ n−1. Let V be the vertex set of P . We assume (V, |·|) is a metric
space and |vi vj | is the distance of any two vertices vi and vj of V . Specifically, the following
properties hold: (1) the triangle inequality: |vi vk | + |vk vj | ≥ |vi vj |; (2) |vi vj | = |vj vi | ≥ 0;
(3) |vi vj | = 0 if i = j. In particular, for each edge e(vi−1 , vi ) of P , its length is equal to
|vi−1 vi |. We assume that given any two vertices vi and vj of P , the distance |vi vj | can be
obtained in O(1) time.
Our goal is to find a new edge e connecting two vertices of P and add e to P , such that
the diameter of the resulting graph P ∪ {e} is minimized.
The problem has been studied before. Große et al. [10] solved the problem in O(n log3 n)
time. In this paper, we present a new algorithm that runs in O(n log n) time. Our algorithm is
based on new observations on the structures of the optimal solution and different algorithmic
techniques. Following the previous work [10], we refer to the problem as the diameteroptimally augmenting path problem, or DOAP for short.
1.1
Related Work
If the path P is in the Euclidean space Rd for a constant d, then Große et al. [10] also gave
an O(n + 1/3 ) time algorithm that can find a (1 + )-approximation solution for the problem
DOAP, for any > 0. If P is in the Euclidean plane R2 , De Carulfel et al. [4] gave a linear
time algorithm for adding a new edge to P to minimize the continuous diameter (i.e., the
diameter is defined with respect to all points of P , not only vertices).
The more general problem and many variations have also been studied before, e.g., see
[1, 3, 5, 6, 9, 12, 13, 15] and the references therein. Consider a general graph G in which edges
© Haitao Wang;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
2
Diameter-Optimally Augmenting Paths
have non-negative lengths. For an integer k, the goal of the general problem is to compute a
set F of k new edges and add them to G such that the resulting graph has the minimum
diameter. The problem is NP-hard [15] and some other variants are even W[2]-hard [6, 9].
Approximation results have been given for the general problem and many of its variations,
e.g., see [3, 6, 13]. The upper bounds and lower bounds on the values of the diameters of the
augmented graphs have also been investigated, e.g. see [1, 12].
Since diameter is an important metric of network performance, which measures the
worst-case cost between any two nodes of the network, as discussed in [3, 5], the problem of
augmenting graphs for minimizing the diameter and its variations have many practical applications, such as in data networks, telephone networks, transportation networks, scheduling
problems, etc.
As an application of our problem DOAP, consider the following scenario in transportation
networks. Suppose there is a highway that connects several cities. In order to reduce the
transportation time, we want to build a new highway connecting two cities such that the
distance between the farthest two cities using both highways is minimized. Clearly, this is a
problem instance of DOAP.
1.2
Our Approaches
To tackle the problem, Große et al. [10] first gave an O(n log n) time algorithm for the
decision version of the problem: Given any value λ, determine whether it is possible to add
a new edge e into P such that the diameter of the resulting graph is at most λ. Then, by
implementing the above decision algorithm in a parallel fashion and applying Megiddo’s
parametric search [14], they solved the original problem DOAP in O(n log3 n) time [10]. For
differentiation, we referred to the original problem DOAP as the optimization problem.
Our improvement over the previous work [10] is twofold.
First, we solve the decision problem in O(n) time. Our algorithm is based on the O(n log n)
time algorithm in the previous work [10]. However, by discovering new observations on the
problem structure and with the help of the range-minima data structure [2, 11], we avoid
certain expensive operations and eventually achieve the O(n) time complexity.
Second, comparing with the decision problem, our algorithm for the optimization problem
is completely different from the previous work [10]. Let λ∗ be the diameter of the resulting
graph in an optimal solution. Instead of using the parametric search, we identify a set S of
candidate values such that λ∗ is in S and then we search λ∗ in S using our algorithm for the
decision problem. However, computational difficulties arise for this approach due to that
the set S is too large (|S| = Ω(n2 )) and computing certain values of S is time-consuming
(e.g., for certain values of S, computing each of them takes O(n) time). To circumvent these
difficulties, our algorithm has several steps. In each step, we shrink S significantly such that
λ∗ always remains in S. More importantly, each step will obtain certain formation, based on
which the next step can further reduce S. After several steps, the size of S is reduced to O(n)
and all the remaining values of S can be computed in O(n log n) time. At this point we can
use our decision algorithm to find λ∗ from S in additional O(n log n) time. Equipped with
our linear time algorithm for the decision problem and utilizing several other algorithmic
techniques such as the sorted-matrix searching techniques [7, 8] and range-minima data
structure [2, 11], we eventually solve the optimization problem in O(n log n) time.
The rest of the paper is organized as follows. In Section 2, we introduce some notation
and observations. In Section 3, we present our algorithm for the decision problem. The
optimization problem is solved in Section 4.
H. Wang
3
vi
v1
vn
vk
vj
vl
Figure 1 Illustrating the resulting graph after a new edge e(vi , vj ) is added.
2
Preliminaries
In this section, we introduce some notation and observations, some of which are from Große
et al. [10].
For any two vertices vi and vj of P , we use e(vi , vj ) to denote the edge connecting vi
and vj in the metric space. Hence, e(vi , vj ) is in P if and only if |i − j| = 1. The length of
e(vi , vj ) is |vi vj |.
For any i and j with 1 ≤ i ≤ j ≤ n, we use G(i, j) to denote the resulting graph by
adding the edge e(vi , vj ) into P . If i = j, G(i, j) is essentially P . Let D(i, j) denote the
diameter of G(i, j).
Our goal for the optimization problem DOAP is to find a pair of indices (i, j) with
1 ≤ i ≤ j ≤ n such that D(i, j) is minimized. Let λ∗ = min1≤i≤j≤n D(i, j), i.e., λ∗ is the
diameter in an optimal solution.
Given any value λ, the decision problem is to determine whether λ ≥ λ∗ , or in other
words, determine whether there exist a pair (i, j) with 1 ≤ i ≤ j ≤ n such that D(i, j) ≤ λ.
If yes, we say that λ is a feasible value.
Recall that for any graph G, dG (u, v) refers to the length of the shortest path between
two vertices u and v in G.
Consider any pair of indices (i, j) with 1 ≤ i ≤ j ≤ n. We define α(i, j), β(i, j), γ(i, j),
and δ(i, j) as follows (refer to Fig. 1).
I Definition 1. 1. Define α(i, j) to be the largest shortest path length in G(i, j) from v1 to
all vertices vk with k ∈ [i, j], i.e., α(i, j) = maxi≤k≤j dG(i,j) (v1 , vk ).
2. Define β(i, j) to be the largest shortest path length in G(i, j) from vn to all vertices vk
with k ∈ [i, j], i.e., β(i, j) = maxi≤k≤j dG(i,j) (vk , vn ).
3. Define γ(i, j) to be the largest shortest path length in G(i, j) from vk to vl for any k and
l with i ≤ k ≤ l ≤ j, i.e., γ(i, j) = maxi≤k≤l≤j dG(i,j) (vk , vl ).
4. Define δ(i, j) to be the shortest path length in G(i, j) from v1 to vn , i.e., δ(i, j) =
dG(i,j) (v1 , vn ).
It can be verified (also shown in [10]) that the following observation holds.
I Observation 2. ([10]) D(i, j) = max{α(i, j), β(i, j), γ(i, j), δ(i, j)}.
Further, due to the triangle inequality of the metric space, the following monotonicity
properties hold.
I Observation 3. ([10])
1. For any 1 ≤ i ≤ j ≤ n − 1, α(i, j) ≤ α(i, j + 1), β(i, j) ≥ β(i, j + 1), γ(i, j) ≤ γ(i, j + 1),
and δ(i, j) ≥ δ(i, j + 1).
2. For any 1 ≤ i < j ≤ n, α(i, j) ≤ α(i + 1, j), β(i, j) ≥ β(i + 1, j), γ(i, j) ≥ γ(i + 1, j), and
δ(i, j) ≤ δ(i + 1, j).
4
Diameter-Optimally Augmenting Paths
For any pair (i, j) with 1 ≤ i ≤ j ≤ n, let P (i, j) denote the subpath of P between vi and
P
vj . Hence, dP (vi , vj ) is the length of P (i, j), i.e., dP (vi , vj ) = i≤k≤j−1 |vk vk+1 | if i < j
and dP (vi , vj ) = 0 if i = j.
In our algorithms, we will need to compute f (i, j) for each f ∈ {α, β, γ, δ}. The next
observation gives an algorithm. The result was also given by Große et al. [10] and we present
the proof here for completeness of this paper.
I Lemma 4. ([10]) With O(n) time preprocessing, given any pair (i, j) with 1 ≤ i ≤ j ≤ n,
we can compute dP (i, j) and δ(i, j) in O(1) time, and compute α(i, j) and β(i, j) in O(log n)
time.
Proof. As preprocessing, we compute the prefix sum array A[1 · · · n] such that A[k] =
P
1≤l≤k−1 |vl vl+1 | for each k ∈ [2, n] and A[1] = 0. This can be done in O(n) time. This
finishes our preprocessing.
Consider any pair (i, j) with 1 ≤ i ≤ j ≤ n. Note that dP (vi , vj ) = A[j] − A[i], which
can be computed in constant time.
For δ(i, j), it is easy to see that δ(i, j) = min{dP (1, n), dP (1, i) + |vi vj | + dP (j, n)}. Hence,
δ(i, j) can be computed in constant time.
For α(i, j), if we consider dG(i,j) (v1 , vk ) as a function of k ∈ [i, j], then dG(i,j) (v1 , vk ) is a
unimodal function. More specifically, as k changes from i to j, dG(i,j) (v1 , vk ) first increases
and then decreases. Hence, α(i, j) can be computed in O(log n) time by binary search on
the sequence vi , vi+1 , . . . , vj .
Computing β(i, j) can be also done in O(log n) time in a similar way to α(i, j). We omit
the details.
J
For computing γ(i, j), although one may be able to do so in O(n) time, it is not clear
to us how to make it in O(log n) time even with O(n log n) time preprocessing. As will be
seen later, this is the major difficulty for solving the problem DOAP efficiently. We refer to
it as the γ-computation difficulty. Our main effort will be to circumvent the difficulty by
providing alternative and efficient solutions.
For any pair (i, j) with 1 ≤ i ≤ j ≤ n, we use C(i, j) to denote the cycle P (i, j) ∪ e(vi , vj ).
Consider dG(i,j) (vk , vl ) for any k and l with i ≤ k ≤ l ≤ j. Notice that the shortest
path from vk to vl in C(i, j) is also a shortest path in G(i, j). Hence, dG(i,j) (vk , vl ) =
dC(i,j) (vk , vl ). There are two paths in C(i, j) from vk to vl : one is P (k, l) and the other
uses the edge e(vi , vj ). We use d1C(i,j) (vk , vl ) to denote the length of the above second
path, i.e., d1C(i,j) (vk , vl ) = dP (vi , vk ) + |vi vj | + dP (vl , vj ). With these notation, we have
dC(i,j) (vk , vl ) = min{dP (vk , vl ), d1C(i,j) (vk , vl )}. According to the definition of γ(i, j), we
summarize our discussion in the following observation.
I Observation 5. For any pair (i, j) with 1 ≤ i ≤ j ≤ n, γ(i, j) = maxi≤k≤l≤j dC(i,j) (vk , vl ),
with dC(i,j) (vk , vl ) = min{dP (vk , vl ), d1C(i,j) (vk , vl )} and d1C(i,j) (vk , vl ) = dP (vi , vk ) + |vi vj | +
dP (vl , vj ).
In the following, to simplify the notation, when the context is clear, we use index i to
refer to vertex vi . For example, dP (i, j) refers to dP (vi , vj ) and e(i, j) refers to e(vi , vj ).
3
The Decision Problem
In this section, we present our O(n) time algorithm for solving the decision problem. For any
value λ, our goal is to determine whether λ is feasible, i.e. whether λ ≥ λ∗ , or equivalently,
H. Wang
5
α
γ
λ
β
δ
Ii (δ) Ii (β)
Ii (α) Ii (γ)
Figure 2 Illustrating f (i, j) as j changes in [i, n] and Ii (f ) for f ∈ {α, β, γ, δ}.
whether there is a pair (i, j) with 1 ≤ i ≤ j ≤ n such that D(i, j) ≤ λ. If yes, our algorithm
can also find such a feasible edge e(i, j).
By Observation 2, D(i, j) ≤ λ holds if and only if f (i, j) ≤ λ for each f ∈ {α, β, γ, δ}.
To determine whether λ is feasible, our algorithm will determine for each i ∈ [1, n], whether
there exists j ∈ [i, n] such that f (i, j) ≤ λ for each f ∈ {α, β, γ, δ}.
For any fixed i ∈ [1, n], we consider α(i, j), β(i, j), γ(i, j), and δ(i, j) as functions of
j ∈ [i, n]. In light of Observation 3, α(i, j) and γ(i, j) are monotonically increasing and β(i, j)
and δ(i, j) are monotonically decreasing (e.g., see Fig. 2). We define four indices Ii (f ) for
f ∈ {α, β, γ, δ} as follows. Refer to Fig. 2.
I Definition 6. Define Ii (α) to be the largest index j ∈ [i, n] such that α(i, j) ≤ λ. We
define Ii (γ) similarly to Ii (α). If β(i, n) ≤ λ, then define Ii (β) to be the smallest index
j ∈ [i, n] such that β(i, j) ≤ λ; otherwise, let Ii (β) = ∞. We define Ii (δ) similarly to Ii (β).
As discussed in [10], λ is feasible if and only if [1, Ii (α)]∩[Ii (β), n]∩[1, Ii (γ)]∩[Ii (δ), n] 6= ∅
for some i ∈ [1, n]. By Observation 3, we have the following lemma.
I Lemma 7. For any i ∈ [1, n − 1], Ii (α) ≥ Ii+1 (α), Ii (β) ≥ Ii+1 (β), Ii (γ) ≤ Ii+1 (γ), and
Ii (δ) ≤ Ii+1 (δ) (e.g., see Fig. 3).
Proof. According to Observation 3, α(i, j) ≤ α(i + 1, j). This implies that Ii (α) ≥ Ii+1 (α)
by the their definitions (e.g., see Fig. 3). The other three cases for β, γ, and δ are similar. J
β(i, j)
β(i + 1, j)
γ(i, j)
γ(i + 1, j)
δ(i, j)
λ
λ
Ii+1 (β) Ii (β)
δ(i + 1, j)
α(i + 1, j)
α(i, j)
Ii+1 (α)Ii (α)
Ii (δ) Ii+1 (δ)
Ii (γ) Ii+1 (γ)
Figure 3 Illustrating f (i, j) and f (i + 1, j) as j changes and Ii (f ) and Ii+1 (f ) for f ∈ {α, β, γ, δ}.
3.1
Computing Ii (α), Ii (β), and Ii (δ) for all i ∈ [1, n]
In light of Lemma 7, for each f ∈ {α, β, δ}, we compute Ii (f ) for all i = 1, 2, . . . , n in O(n)
time, as follows.
We discuss the case for δ first. According to Lemma 4, δ(i, j) can be computed in constant
time for any pair (i, j) with 1 ≤ i ≤ n. We can compute Ii (δ) for all i ∈ [1, n] in O(n) time
by the following simple algorithm. We first compute I1 (δ), which is done by computing
δ(1, j) from j = 1 incrementally until the first time δ(1, j) ≤ λ. Then, to compute I2 (δ),
6
Diameter-Optimally Augmenting Paths
vi
v1
vn
vk−1
vk
vj
Figure 4 Illustrating the path (the dotted curve) from vn to vk−1 using the edge e(i, j).
we compute δ(2, j) from j = I1 (δ) incrementally until the first time δ(2, j) ≤ λ. Next, we
compute Ii (δ) for i = 3, 4, . . . , n in the same way. The total time is O(n). The correctness is
based on the monotonicity property of Ii (δ) in Lemma 7.
To compute Ii (α) or Ii (β) for i = 1, 2, . . . , n, using a similar approach as above, we can
only have an O(n log n) time algorithm since computing each α(i, j) or β(i, j) takes O(log n)
time by Lemma 4. In the following Lemma 8, we give another approach that only needs
O(n) time.
I Lemma 8. Ii (α) and Ii (β) for all i = 1, 2, . . . , n can be computed in O(n) time.
Proof. We only discuss the case for β since the other case for α is analogous.
The key idea is that for each pair (i, j), instead of computing the exact value of β(i, j), it
is sufficient to determine whether β(i, j) ≤ λ. In what follows, we show that with O(n) time
preprocessing, we can determine whether β(i, j) ≤ λ in O(1) time for any index pair (i, j)
with 1 ≤ i ≤ j ≤ n.
Let k be the smallest index in [1, n] such that dP (k, n) ≤ λ. This implies that dP (k−1, n) >
λ if k > 1. As preprocessing, we compute the index k, which can be easily done in O(n) time
(or even in O(log n) time by binary search).
Consider any pair (i, j) for 1 ≤ i ≤ j ≤ n. Our goal is to determine whether β(i, j) ≤ λ.
1. If k ≤ i, then it is vacuously true that β(i, j) ≤ λ.
2. If k > j, then β(i, j) > λ.
3. If i < k ≤ j, a crucial observation is that β(i, j) ≤ λ if and only if the length of the path
from vn to vk−1 using the new edge e(i, j), i.e., dP (i, k − 1) + |vi vj | + dP (j, n), is less than
or equal to λ. See Fig. 4. Clearly, the above path length can be computed in constant
time, and thus, we can determine whether β(i, j) ≤ λ in constant time.
Therefore, we can determine whether β(i, j) ≤ λ in constant time for any pair (i, j) with
1 ≤ i ≤ j ≤ n. With this result, we can use a similar algorithm as the above for computing
Ii (δ) to compute Ii (β) for all i ∈ [1, n] in O(n) time. The lemma thus follows.
J
Due to the γ-computation difficulty mentioned in Section 2, it is not clear whether it
possible to compute Ii (γ) for all i = 1, . . . , n in O(n log n) time.
Recall that λ is feasible if and only if there exists an i ∈ [1, n] such that [1, Ii (α)] ∩
[Ii (β), n] ∩ [1, Ii (γ)] ∩ [Ii (δ), n] 6= ∅. Now that Ii (f ) for all i = 1, 2, . . . , n and f ∈ {α, β, δ}
have been computed but the Ii (γ)’s are not known, in the following we will use an “indirect”
approach to determine whether the intersection of the above four intervals is empty for every
i ∈ [1, n].
3.2
Determining the Feasibility of λ
For each i ∈ [1, n], define Qi = [1, Ii (α)] ∩ [Ii (β), n] ∩ [1, Ii (γ)] ∩ [Ii (δ), n]. Our goal is to
determine whether Qi is empty for each i = 1, 2, . . . , n.
H. Wang
7
i
j
ai gj + 1 gj
Figure 5 Illustrating the graph G(i, ai ) with gj + 1 ≤ ai .
Consider any i ∈ [1, n]. Since Ii (f ) for each f ∈ {α, β, δ} is known, we can determine the
intersection [1, Ii (α)] ∩ [Ii (β), n] ∩ [Ii (δ), n] in constant time. If the intersection is empty,
then we know that Qi = ∅. In the following, we assume the intersection is not empty.
Let ai be the smallest index in the above intersection. As in [10], an easy observation is
that Qi 6= ∅ if and only if ai ∈ [1, Ii (γ)]. If ai ≤ i (note that ai ≤ i actually implies ai = i
since ai ≥ Ii (β) ≥ i), it is obviously true that ai ∈ [1, Ii (γ)] since i ≤ Ii (γ). Otherwise (i.e.,
i < ai ), according to the definition of Ii (γ), ai ∈ [1, Ii (γ)] if and only if γ(i, ai ) ≤ λ. Große
et al. [10] gave an approach that can determine whether γ(i, ai ) ≤ λ in O(log n) time after
O(n log n) time preprocessing. In the following, by new observations and with the help of the
range minima data structure [2, 11], we show that whether γ(i, ai ) ≤ λ can be determined in
constant time after O(n) time preprocessing.
For each j ∈ [1, n], define gj as the largest index k in [j, n] such that dP (j, k) ≤ λ. Observe
that g1 ≤ g2 ≤ · · · ≤ gn .
Consider any i and the corresponding ai with i < ai . Our goal is to determine
whether γ(i, ai ) ≤ λ. Since we are talking about γ(i, ai ), we are essentially considering
the graph G(i, ai ). Recall that C(i, ai ) is the cycle P (i, ai ) ∪ e(i, ai ). By Observation 5,
γ(i, ai ) = maxi≤k≤l≤ai dC(i,ai ) (k, l), and further, dC(i,j) (k, l) = min{dP (k, l), d1C(i,ai ) (k, l)}
and d1C(i,ai ) (k, l) = dP (i, k) + |vi vai | + dP (l, ai ).
For any j ∈ [i, ai − 1], if gj ≤ ai − 1, then vertex gj + 1 is in the cycle C(i, ai ). Note that
1
dC(i,ai ) (j, gj + 1) = dP (i, j) + |vi vai | + dP (gj + 1, ai ). See Fig. 5.
We have the following lemma.
I Lemma 9. γ(i, ai ) ≤ λ if and only if for each j ∈ [i, ai −1], either gj ≥ ai or d1C(i,ai ) (j, gj +
1) ≤ λ.
Proof.
Suppose γ(i, ai ) ≤ λ. Consider any j ∈ [i, ai − 1] such that gj ≤ ai − 1. Below we
prove d1C(i,ai ) (j, gj + 1) ≤ λ must hold.
By the definition of gj , it holds that dP (j, gj +1) > λ. Since γ(i, ai ) ≤ λ and dC(i,ai ) (j, gj +
1) ≤ γ(i, ai ), we obtain dC(i,ai ) (j, gj +1) ≤ λ. Note that dC(i,ai ) (j, gj +1) = min{dP (j, gj +
1), d1C(i,ai ) (j, gj + 1)}. Hence, it must hold that d1C(i,ai ) (j, gj + 1) ≤ λ.
This proves one direction of the lemma.
Suppose it is true that for each j ∈ [i, ai − 1], either gj ≥ ai or d1C(i,ai ) (j, gj + 1) ≤ λ. We
prove γ(i, ai ) ≤ λ below.
Consider any pair of indices (k, l) with i ≤ k ≤ l ≤ ai . To prove γ(i, ai ) ≤ λ, it is sufficient
to show that dC(i,ai ) (k, l) ≤ λ. If k = l, then dC(i,ai ) (k, l) = 0 and thus dC(i,ai ) (k, l) ≤ λ
obviously holds. In the following we assume k < l. This implies that k ≤ ai − 1. Hence,
either gk ≥ ai or d1C(i,ai ) (k, gk + 1) ≤ λ.
Recall that dC(i,ai ) (k, l) = min{dP (k, l), d1C(i,ai ) (k, l)}.
If gk ≥ ai , then l ≤ ai ≤ gk , and thus, dP (k, l) ≤ λ by the definition of gk . Hence, we
obtain dC(i,ai ) (k, l) ≤ λ.
8
Diameter-Optimally Augmenting Paths
Otherwise, we have d1C(i,ai ) (k, gk + 1) ≤ λ. If l ≤ gk , we again have dP (k, l) ≤ λ and thus
dC(i,ai ) (k, l) ≤ λ. If l ≥ gk + 1, then dP (l, ai ) ≤ dP (gk + 1, ai ). Hence, d1C(i,ai ) (k, l) =
dP (i, k) + |vi vai | + dP (l, ai ) ≤ dP (i, k) + |vi vai | + dP (gk + 1, ai ) = d1C(i,ai ) (k, gk + 1) ≤ λ.
Consequently, we again obtain dC(i,ai ) (k, l) ≤ λ.
This proves the other direction of the lemma.
J
Recall that g1 ≤ g2 ≤ · · · ≤ gn . For each k ∈ [1, n], define hk to be the smallest index j
in [1, k] with gj ≥ k. Observe that h1 ≤ h2 ≤ · · · ≤ hn .
Note that if i < hai , then for each j ∈ [i, hai − 1], gj < ai and gj + 1 ≤ ai . Due to the
preceding lemma, we further have the following lemma.
I Lemma 10. γ(i, ai ) ≤ λ if and only if either hai ≤ i or d1C(i,ai ) (j, gj + 1) ≤ λ holds for
each j ∈ [i, hai − 1].
Proof.
Suppose γ(i, ai ) ≤ λ. If hai ≤ i, then we do not need to prove anything. In the
following, we assume hai > i. Consider any j ∈ [i, hai − 1]. Our goal is to show that
d1C(i,ai ) (j, gj + 1) ≤ λ holds.
Indeed, since γ(i, ai ) ≤ λ, by Lemma 9, we have either gj ≥ ai or d1C(i,ai ) (j, gj + 1) ≤ λ.
Since j ∈ [i, hai − 1], gj < ai . Hence, it must be that d1C(i,ai ) (j, gj + 1) ≤ λ.
This proves one direction of the lemma.
Suppose either hai ≤ i or d1C(i,ai ) (j, gj + 1) ≤ λ holds for each j ∈ [i, hai − 1]. Our goal is
to show that γ(i, ai ) ≤ λ. Consider any k ∈ [i, ai − 1]. By Lemma 9, it is sufficient to
show that either gk ≥ ai or d1C(i,ai ) (k, gk + 1) ≤ λ.
If hai ≤ i, then since k ≥ i, we obtain gk ≥ ai by the definition of hai .
Otherwise, d1C(i,ai ) (j, gj + 1) ≤ λ holds for each j ∈ [i, hai − 1]. If k ≥ hai , then we still
have gk ≥ ai . Otherwise, k is in [i, hai − 1], and thus it holds that d1C(i,ai ) (k, gk + 1) ≤ λ.
This proves the other direction of the lemma.
J
Let |C(i, ai )| denote the total length of the cycle C(i, ai ), i.e., |C(i, ai )| = dP (i, ai ) +
|vi vai |. The following observation is crucial because it immediately leads to our algorithm in
Lemma 12.
I Observation 11. γ(i, ai ) ≤ λ if and only if either hai ≤ i or minj∈[i,hai −1] {dP (j, gj +1)} ≥
|C(i, ai )| − λ.
Proof. Suppose hai > i. Then, for each j ∈ [i, hai − 1], gj < ai and gj + 1 ≤ ai . Note that
d1C(i,ai ) (j, gj + 1) = |C(i, ai )| − dP (i, gj + 1). Hence, d1C(i,ai ) (j, gj + 1) ≤ λ is equivalent to
dP (j, gj + 1) ≥ |C(i, ai )| − λ. Therefore, d1C(i,ai ) (j, gj + 1) ≤ λ holds for each j ∈ [i, hai − 1]
if and only if minj∈[i,hai −1] {dP (j, gj + 1)} ≥ |C(i, ai )| − λ.
By Lemma 10, the observation follows.
J
I Lemma 12. With O(n) time preprocessing, given any i ∈ [1, n] and the corresponding ai
with i < ai , whether γ(i, ai ) ≤ λ can be determined in constant time.
Proof. As preprocessing, we first compute gj for all j = 1, 2, . . . , n, which can be done in
O(n) time due to the monotonicity property g1 ≤ g2 ≤ . . . ≤ gn . Then, we compute hk for
all k = 1, 2, . . . , n, which can also be done in O(n) time due to the monotonicity property
h1 ≤ h2 ≤ . . . ≤ hn . Next, we compute an array B[1, . . . , n] with B[j] = dP (j, gj + 1) for
each j ∈ [1, n] (let dP (j, gj + 1) = ∞ if gj + 1 > n). We build a range-minima data structure
H. Wang
on B [2, 11]. The range minima data structure can be built in O(n) time such that given any
pair (i, j) with 1 ≤ i ≤ j ≤ n, the minimum value of the subarray B[i · · · j] can be returned
in constant time [2, 11]. This finishes the preprocessing step, which takes O(n) time in total.
Consider any i and the corresponding ai with i < ai . Our goal is to determine whether
γ(i, ai ) ≤ λ, which can be done in O(1) time as follows.
By Observation 11, γ(i, ai ) ≤ λ if and only if either hai ≤ i or minj∈[i,hai −1] {dP (j, gj +
1)} ≥ |C(i, ai )| − λ. Since hai has been computed in the preprocessing, we check whether
hai ≤ i is true. If yes, then we are done with the assertion that γ(i, ai ) ≤ λ. Otherwise, we
need to determine whether minj∈[i,hai −1] {dP (j, gj + 1)} ≥ |C(i, ai )| − λ holds. To this end,
we first compute minj∈[i,hai −1] {dP (j, gj + 1)} in constant time by querying the range-minima
data structure on B with (i, hai − 1). Note that |C(i, ai )| can be computed in constant time.
Therefore, we can determine whether γ(i, ai ) ≤ λ in O(1) time. This proves the lemma. J
With Lemma 12, the decision problem can be solved in O(n) time. The proof of the
following theorem summarizes our algorithm.
I Theorem 13. Given any λ, we can determine whether λ is feasible in O(n) time, and
further, if λ is feasible, a feasible edge can be found in O(n) time.
Proof. First, we do the preprocessing in Lemma 4 in O(n) time. Then, for each f ∈ {α, β, δ},
we compute Ii (f ) for all i = 1, 2, . . . , n, in O(n) time. We also do the preprocessing in
Lemma 12.
Next, for each i ∈ [1, n], we do the following. Compute the intersection [1, Ii (α)] ∩
[Ii (β), n] ∩ [Ii (δ), n] in constant time. If the intersection is empty, then we are done for this
i. Otherwise, obtain the smallest index ai in the above intersection. If ai ≤ i, then we
stop the algorithm with the assertion that λ is feasible and report e(i, ai ) as a feasible edge.
Otherwise, we use Lemma 12 to determine whether γ(i, ai ) ≤ λ in constant time. If yes, we
stop the algorithm with the assertion that λ is feasible and report e(i, ai ) as a feasible edge.
Otherwise, we proceed on i + 1.
If the algorithm does not stop after we check all i ∈ [1, n], then we stop the algorithm
with the assertion that λ is not feasible. Clearly, we spend O(1) time on each i, and thus,
the total time of the algorithm is O(n).
J
4
The Optimization Problem
In this section, we present our algorithm that solves the optimization problem in O(n log n)
time, by making use of our algorithm for the decision problem given in Section 3 (we will
refer to it as the decision algorithm). It is sufficient to compute λ∗ , after which we can use
our decision algorithm to find an optimal new edge in additional O(n) time.
We start with an easy observation that λ∗ must be equal to the diameter D(i, j) of G(i, j)
for some pair (i, j) with 1 ≤ i ≤ j ≤ n. Further, by Observation 2, λ∗ is equal to f (i, j) for
some f ∈ {α, β, γ, δ} and some pair (i, j) with 1 ≤ i ≤ j ≤ n.
For each f ∈ {α, β, γ, δ}, define Sf = {f (i, j) | 1 ≤ i ≤ j ≤ n}. Let S = ∪f ∈{α,β,γ,δ} Sf .
According to our discussion above, λ∗ is in S. Further, note that λ∗ is the smallest feasible
value of S. We will not compute the entire set S since |S| = Ω(n2 ). For each f ∈ {α, β, γ, δ},
let λf be the smallest feasible value in Sf . Hence, we have λ∗ = min{λα , λβ , λγ , λδ }.
In the following, we first compute λα , λβ , λδ in O(n log n) time by using our decision
algorithm and the sorted-matrix searching techniques [7, 8].
9
10
Diameter-Optimally Augmenting Paths
4.1
Computing λα , λβ , and λδ
For convenience, we begin with computing λβ .
We define an n × n matrix M [1 · · · n; 1 · · · n]: For each 1 ≤ i ≤ n and 1 ≤ j ≤ n, define
M [i, j] = β(i, j) if j ≥ i and M [i, j] = β(i, i) otherwise. By Observation 3, the following
lemma shows that M is a sorted matrix in the sense that each row is sorted in descending
order from left to right and each column is sorted in descending order from top to bottom.
I Lemma 14. For each 1 ≤ i ≤ n, M [i, j] ≥ M [i, j + 1] for any j ∈ [1, n − 1]; for each
1 ≤ j ≤ n, M [i, j] ≥ M [i + 1, j] for any i ∈ [1, n − 1].
Proof. Consider any two adjacent matrix elements M [i, j] and M [i, j + 1] in the same
row. If j ≥ i, then M [i, j] = β(i, j) and M [i, j + 1] = β(i, j + 1). By Observation 3,
M [i, j] ≥ M [i, j + 1]. If j < i, then M [i, j] = M [i, j + 1] = β(i, i). Hence, in either case,
M [i, j] ≥ M [i, j + 1] holds.
Consider any two adjacent matrix elements M [i, j] and M [i + 1, j] in the same column.
If j ≥ i + 1, then M [i, j] = β(i, j) and M [i + 1, j] = β(i + 1, j). By Observation 3, we obtain
M [i, j] ≥ M [i, j + 1]. If j < i + 1, then M [i, j] = β(j, j) and M [i, j + 1] = β(j + 1, j + 1).
Note that β(j, j) is essentially equal to dP (j, n) and β(j + 1, j + 1) is equal to dP (j + 1, n).
Clearly, dP (j, n) ≥ dP (j + 1, n). Hence, in either case, M [i, j] ≥ M [i, j + 1].
J
Note that each element of Sβ is in M and vice versa. Since λβ is the smallest feasible
value of Sβ , λβ is also the smallest feasible value of M . We do not construct M explicitly.
Rather, given any i and j, we can “evaluate” M [i, j] in O(log n) time since β(i, j) can be
computed in O(log n) time if i ≤ j by Lemma 4. Using the sorted-matrix searching techniques
[7, 8], we can find λβ in M by calling our decision algorithm O(log n) times and evaluating
O(n) elements of M . The total time on calling the decision algorithm is O(n log n) and the
total time on evaluating matrix elements is also O(n log n). Hence, we can compute λβ in
O(n log n) time.
Computing λα and λδ can done similarly in O(n log n) time, although the corresponding
sorted matrices may be defined slightly differently. We omit the details. However, we cannot
compute λγ in O(n log n) time in the above way, and again this is due to the λ-computation
difficulty mentioned in Section 2.
Note that having λα , λβ , and λδ essentially reduces our search space for λ∗ from S to
Sγ ∪ {λα , λβ , λδ }.
We compute λ1 = min{λα , λβ , λδ }. Thus, λ∗ = min{λ1 , λγ }. Hence, if λγ ≥ λ1 , then
∗
λ = λ1 and we are done for computing λ∗ . Otherwise (i.e., λγ < λ1 ), it must be that
λ∗ = λγ and we need to compute λγ . To compute λγ , again we cannot use the similar way
as the above for computing λβ . Instead, we use the following approach. We should point out
that the success of the approach relies on the information implied by λγ < λ1 .
4.2
Computing λ∗ in the Case λγ < λ1
We assume λγ < λ1 . Hence, λ∗ = λγ . Let e(i∗ , j ∗ ) be the new edge added to P in an optimal
solution. We also call e(i∗ , j ∗ ) an optimal edge.
Since λ∗ = λγ < λ1 , we have the following observation.
I Observation 15. If λγ < λ1 and e(i∗ , j ∗ ) is an optimal edge, then λ∗ = γ(i∗ , j ∗ ).
Proof. Assume to the contrary that λ∗ =
6 γ(i∗ , j ∗ ). Then, by Observation 2, λ∗ is equal to
one of α(i∗ , j ∗ ), β(i∗ , j ∗ ), and δ(i∗ , j ∗ ). Without loss of generality, assume λ∗ = α(i∗ , j ∗ ).
H. Wang
11
γ
λ1
α
β
δ
ai ∗ − 1 ai ∗
ai∗ + 1
Figure 6 Illustrating f (i∗ , j) as j changes for f ∈ {α, β, γ, δ}. The three indices ai∗ − 1, ai∗ , and
a
i∗
+ 1 are shown.
Since α(i∗ , j ∗ ) is in Sα , λ∗ must be the smallest feasible value of Sα , i.e., λ∗ = λα . However,
this contradicts with that λ∗ = λγ < λ1 = min{λα , λβ , λγ } ≤ λα .
J
For any i ∈ [1, n], for each f ∈ {α, β, γ, δ}, with respect to λ1 , we define Ii0 (f ) in a similar
way to Ii (f ) defined in Section 3 with respect to λ except that we change “≤ λ” to “< λ1 ”.
Specifically, define Ii0 (α) to be the largest index j ∈ [i, n] such that α(i, j) < λ1 . I 0 (γ) is
defined similarly to Ii0 (α). If β(i, n) < λ1 , then define Ii0 (β) to be the smallest index j ∈ [i, n]
such that β(i, j) < λ1 ; otherwise Ii0 (β) = ∞. Ii0 (δ) is defined similarly to Ii0 (β). Note that
similar monotonicity properties for Ii0 (f ) with f ∈ {α, β, γ, δ} to those in Lemma 7 also hold.
Recall that e(i∗ , j ∗ ) is an optimal edge. An easy observation is that since λ1 is strictly
larger than λ∗ , the intersection [1, Ii0∗ (α)] ∩ [Ii0∗ (β), n] ∩ [Ii0∗ (δ), n] cannot be empty. Let ai∗
be the smallest index in the above intersection. Note that i∗ ≤ ai∗ since i∗ ≤ Ii0∗ (β) ≤ ai∗ .
The following lemma shows that e(i∗ , ai∗ ) is actually an optimal edge.
I Lemma 16. If λγ < λ1 and e(i∗ , j ∗ ) is an optimal edge, then j ∗ = ai∗ .
Proof. For any pair (i, j) with 1 ≤ i ≤ j ≤ n, let η(i, j) = max{α(i, j), β(i, j), δ(i, j)}. By
Observation 2, D(i, j) = max{γ(i, j), η(i, j)}.
We first prove the following claim: If γ(i∗ , ai∗ ) ≥ η(i∗ , ai∗ ), then j ∗ = ai∗ (e.g., see
Fig. 6).
On the one hand, consider any j ∈ [i∗ , ai∗ − 1]. By the definition of ai∗ , η(i∗ , j) ≥ λ1 .
Since λ1 > λγ = λ∗ , η(i∗ , j) > λ∗ . By Observation 2, D(i∗ , j) ≥ η(i∗ , j) > λ∗ . Hence, j
cannot be j ∗ since otherwise D(i∗ , j) would be equal to λ∗ , incurring contradiction.
On the other hand, consider any j ∈ [a∗ + 1, n]. By Observation 2, D(i∗ , j) ≥ γ(i∗ , j). By
Observation 3, γ(i∗ , j) ≥ γ(i∗ , ai∗ ). Hence, D(i∗ , j) ≥ γ(i∗ , ai∗ ). Further, since γ(i∗ , ai∗ ) ≥
η(i∗ , ai∗ ) (the claim hypothesis), we have D(i∗ , ai∗ ) = max{γ(i∗ , ai∗ ), η(i∗ , ai∗ )} = γ(i∗ , ai∗ ).
Therefore, we obtain D(i∗ , ai∗ ) ≤ D(i∗ , j). This implies that j ∗ = ai∗ . Hence, the claim
follows.
We proceed to prove the lemma. Based on the above claim, it is sufficient to show that
∗
γ(i , ai∗ ) ≥ η(i∗ , ai∗ ), as follows.
Assume to the contrary that γ(i∗ , ai∗ ) < η(i∗ , ai∗ ). Then, D(i∗ , ai∗ ) = η(i∗ , ai∗ ). According to the definition of ai∗ , η(i∗ , ai∗ ) < λ1 . Hence, D(i∗ , ai∗ ) < λ1 . Let λ0 = D(i∗ , ai∗ ).
Since λ0 = η(i∗ , ai∗ ), λ0 is a value in Sα ∪ Sβ ∪ Sδ . Since λ0 = D(i∗ , ai∗ ), λ0 is a feasible value
(i.e., λ0 ≥ λ∗ ). Recall that λ1 is the smallest feasible value of Sα ∪ Sβ ∪ Sδ . Thus, we obtain
contradiction since λ0 < λ1 .
Therefore, γ(i∗ , ai∗ ) ≥ η(i∗ , ai∗ ) holds. The lemma thus follows.
J
Lemma 16 is crucial because it immediately suggests the following algorithm.
We first compute the indices Ii0 (α), Ii0 (β), Ii0 (δ) for i = 1, . . . , n. This can be done in O(n)
time using the similar algorithms as those for computing Ii (α), Ii (β), Ii (δ) in Section 3.1. In
12
Diameter-Optimally Augmenting Paths
fact, here we can even afford O(n log n) time to compute these indices. Hence, for simplicity,
we can use the similar algorithm as that for computing Ii (δ) in Section 3.1 instead of the
one in Lemma 8. The total time is O(n log n).
Next, for each i ∈ [1, n], if [1, Ii0 (α)] ∩ [Ii0 (β), n] ∩ [Ii0 (δ), n] 6= ∅, then we compute ai , i.e.,
the smallest index in the above intersection. Let I be the set of index i such that the above
interval intersection for i is not empty. Lemma 16 leads to the following observation.
I Observation 17. If λγ < λ1 , then λ∗ is the smallest feasible value of the set {γ(i, ai ) | i ∈
I}.
Proof. By Lemma 16, one of the edges of {e(i, ai ) | i ∈ I} is an optimal edge. By
Observation 15, λ∗ is in {γ(i, ai ) | i ∈ I}. Thus, λ∗ is the smallest feasible value in
{γ(i, ai ) | i ∈ I}.
J
We can further obtain the following “stronger” result, although Observation 17 is sufficient
for our algorithm.
I Lemma 18. If λγ < λ1 , then λ∗ = mini∈I γ(i, ai ).
Proof. For any pair (i, j) with 1 ≤ i ≤ j ≤ n, let η(i, j) = max{α(i, j), β(i, j), δ(i, j)}. By
Observation 2, D(i, j) = max{γ(i, j), η(i, j)}.
We first prove the following claim: For any i ∈ I, η(i, ai ) < γ(i, ai ). Indeed, assume
to the contrary that η(i, ai ) ≥ γ(i, ai ) for some i ∈ I. Then, D(i, ai ) = η(i, ai ). By the
definition of ai , η(i, ai ) < λ1 . Hence, D(i, ai ) < λ1 . Let λ0 = D(i, ai ). Note that λ0 is a
feasible value that is in Sα ∪ Sβ ∪ Sδ . However, λ0 < λ1 contradicts with that λ1 is the
smallest feasible value in Sα ∪ Sβ ∪ Sδ .
Next, we prove the lemma by using the above claim. For each i ∈ I, by the above clam,
D(i, ai ) = γ(i, ai ), and thus, γ(i, ai ) is a feasible value. By Lemma 16, we know that λ∗ is in
{γ(i, ai ) | i ∈ I}. Therefore, λ∗ is the smallest value in {γ(i, ai ) | i ∈ I}. The lemma thus
follows.
J
Observation 17 essentially reduces the search space for λ∗ to {γ(i, ai ) | i ∈ I}, which has
at most O(n) values. It is tempting to first explicitly compute the set and then find λ∗ from
the set. However, again, due to the γ-computation difficulty, we are not able to compute the
set in O(n log n) time. Alternatively, we use the following approach to compute λ∗ .
4.3
Finding λ∗ in the Set {γ(i, ai ) | i ∈ I}
Recall that according to Observation 5, γ(i, j) = maxi≤k≤l≤j dC(i,j) (k, l), with dC(i,j) (k, l) =
min{dP (k, l), d1C(i,j) (k, l)} and d1C(i,j) (k, l) = dP (i, k) + |vi vj | + dP (l, j). Hence, γ(i, j) is
equal to dP (k, l) or d1C(i,j) (k, l) for some k ≤ l. Therefore, by Observation 17, there exists
i ∈ I such that λ∗ is equal to dP (k, l) or d1C(i,j) (k, l) for some k and l with i ≤ k ≤ l ≤ ai .
Let Sp = {dP (k, l) | 1 ≤ k ≤ l ≤ n} and Sc = {d1C(i,j) (k, l) | i ≤ k ≤ l ≤ ai , i ∈ I}. Based
on our above discussion, λ∗ is in Sp ∪ Sc . Further, λ∗ is the smallest feasible value in Sp ∪ Sc .
Let λp be the smallest feasible value of Sp and let λc be the smallest feasible value of
Sc . Hence, λ∗ = min{λp , λc }. By using the technique of searching sorted-matrices [7, 8], the
following lemma computes λp in O(n log n) time.
I Lemma 19. λp can be computed in O(n log n) time.
H. Wang
13
i
k
h′ai
gk′
ai
l
l−1
Figure 7 Illustrating the graph G(i, ai ) whose diameter is λ∗ and λ∗ = dC(i,ai ) (k, l).
Proof. We define an n × n matrix M [1 · · · n; 1 · · · n]: For each 1 ≤ i ≤ n and 1 ≤ j ≤ n,
define M [i, j] = dP (i, j) if j ≥ i and M [i, j] = 0 otherwise. It is easy to verify that each row
of M is sorted in ascending order from the left to right and each column is sorted in ascending
order from bottom to top. Consequently, by using the sorted-matrix searching technique
[7, 8], λp can be found by calling our decision algorithm O(log n) times and evaluating O(n)
elements of M . Clearly, given any i and j, we can evaluate M [i, j] in constant time. Hence,
λp can be computed in O(n log n) time.
J
Recall that λ∗ = min{λp , λc }. In the case λp ≤ λc , λ∗ = λp and we are done with
computing λ∗ . In the following, we assume λp > λc . Thus, λ∗ = λc . With the help of the
information implied by λp > λc , we will compute λ∗ in O(n log n) time. The details are given
below.
For any j ∈ [1, n], let gj0 denote the largest index k ∈ [j, n] such that the subpath length
dP (j, k) is strictly smaller than λp . Note that the definition of gj0 is similar to gj defined in
Section 4.3 except that we change “≤ λ” to “< λp ”.
For each k ∈ [1, n], let h0k denote the smallest index j ∈ [1, k] with gj0 ≥ k. Let I 0 be the
subset of i ∈ I 0 such that i ≤ h0ai − 1. Hence, for each i ∈ I 0 and each j ∈ [i, h0ai − 1], gj0 < ai
and thus gj0 + 1 ≤ ai .
For each i ∈ I 0 , define d1max (i, ai ) = maxj∈[i,h0a −1] d1C(i,j) (j, gj0 + 1). The following lemma
i
gives a way to determine λ∗ .
I Lemma 20. If λγ < λ1 and λc < λp , then λ∗ = d1max (i, ai ) for some i ∈ I 0 .
Proof. Since λγ < λ1 and λc < λp , by our above discussions, λ∗ = λc .
By Observation 17, λ∗ is the diameter of the graph G(i, ai ) for some i ∈ I and λ∗ is equal
to the length of the shortest path of two vertices vk and vl in C(i, ai ) for i ≤ k ≤ l ≤ ai ,
i.e., λ∗ = dC(i,ai ) (k, l) = min{dP (k, l), d1C(i,j) (k, l)}. See Fig. 7. Since λ∗ = λc , we further
know that λ∗ = d1C(i,j) (k, l) and d1C(i,j) (k, l) ≤ dP (k, l). In fact, d1C(i,j) (k, l) < dP (k, l), since
otherwise if d1C(i,j) (k, l) = dP (k, l), then λ∗ = dP (k, l) would be in the set Sp , contradicting
with that λp is the smallest feasible value in Sp and λ∗ < λp .
For simplicity of discussion, we assume |vl vl−1 | > 0 (since otherwise we can keep updating
l to l − 1 until we find |vl vl−1 | > 0; note that such an l will eventually be found before we
reach k since 0 ≤ λ∗ = d1C(i,j) (k, l) < dP (k, l)).
We prove the following claim: dP (k, l − 1) < λp ≤ dP (k, l).
On the one hand, since λ∗ = d1C(i,ai ) (k, l) and |vl vl−1 | > 0, we obtain that λ∗ <
d1C(i,ai ) (k, l − 1). Since λ∗ is the diameter in the graph G(i, ai ), dG(i,ai ) (k, l − 1) =
dC(i,ai ) (k, l − 1) ≤ λ∗ . Further, because dC(i,ai ) (k, l − 1) = min{dP (k, l − 1), d1C(i,ai ) (k, l −
1)} and λ∗ < d1C(i,ai ) (k, l − 1), we obtain dP (k, l − 1) ≤ λ∗ . As λ∗ = λc < λp , it follows
that dP (k, l − 1) < λp .
14
Diameter-Optimally Augmenting Paths
On the other hand, assume to the contrary that λp > dP (k, l). Then, since dP (k, l) >
d1C(i,ai ) (k, l) = λ∗ , dP (k, l) is a feasible value. Clearly, dP (k, l) is in the set Sp . However,
λp > dP (k, l) contradicts with that λp is the smallest feasible value in Sp .
This proves the claim. With the claim, we show below that λ∗ = d1max (i, ai ), which will
prove the lemma.
We first show that i is in I 0 , i.e., i ≤ h0ai − 1. Indeed, since λp ≤ dP (k, l) (by the claim),
based on the definition of gk0 , it holds that gk0 < l (e.g., see Fig. 7). Since l ≤ ai , we obtain
gk0 ≤ ai − 1. This implies that k < h0ai and thus k ≤ h0ai − 1. Since i ≤ k, i ≤ h0ai − 1.
It remains to prove λ∗ = d1max (i, ai ). Indeed, recall that λ∗ = d1C(i,ai ) (k, l). Note
that the above claim in fact implies that gk0 = l − 1, and thus, gk0 + 1 = l. Hence, we
have λ∗ = d1C(i,ai ) (k, l) = d1C(i,ai ) (k, gk0 + 1). Note that k is in [i, h0ai − 1]. Consider any
j ∈ [i, h0ai − 1]. To prove λ∗ = d1max (i, ai ), it is now sufficient to prove λ∗ ≥ d1C(i,ai ) (j, gj0 + 1),
as follows.
Recall that gj0 + 1 = l ≤ ai . Since λ∗ is the diameter of G(i, ai ), dG(i,ai ) (j, gj0 + 1) =
dC(i,ai ) (j, gj0 + 1) ≤ λ∗ . Recall that dC(i,ai ) (j, gj0 + 1) = min{dP (j, gj0 + 1), d1C(i,ai ) (j, gj0 + 1)}.
By the definition of gj0 , we know that dP (j, gj0 + 1) ≥ λp . Since λp > λ∗ , dP (j, gj0 + 1) > λ∗ .
Hence, it must be that λ∗ ≥ d1C(i,ai ) (j, gj0 + 1).
This proves that λ∗ = d1max (i, ai ). The lemma thus follows.
J
In light of Lemma 20, in the case of λc < λp , λ∗ = λc is the smallest feasible value of
for all i ∈ I 0 . Note that the number of such values d1max (i, ai ) is O(n). Hence, if
we can compute d1max (i, ai ) for all i ∈ I 0 , then λ∗ can be easily found in additional O(n log n)
time using our decision algorithm, e.g., by first sorting these values and then doing binary
search.
The next lemma gives an algorithm that can compute d1max (i, ai ) for all i ∈ I 0 in O(n)
time, with the help of the range-minima data structure [2, 11].
d1max (i, ai )
I Lemma 21. d1max (i, ai ) for all i ∈ I 0 can be computed in O(n) time.
Proof. Consider any i ∈ I 0 . For any j ∈ [i, h0ai −1 ], it is easy to see that d1C(i,ai ) (j, gj + 1) =
|C(i, ai )| − dP (j, gj + 1), where |C(i, ai )| is the length of the cycle C(i, ai ). Hence, we can
obtain the following,
d1max (i, ai ) =
max
j∈[i,h0a −1]
i
d1C(i,j) (j, gj0 + 1) =
max
j∈[i,h0a −1]
{|C(i, ai )| − dP (j, gj0 + 1)}
i
= |C(i, ai )| −
min
j∈[i,h0a −1]
dP (j, gj0 + 1).
i
minj∈[i,h0a −1] dP (j, gj0
i
Define dmin (i, ai ) =
+ 1). By the above discussions we have
= |C(i, ai )| − dmin (i, ai ). Therefore, computing d1max (i, ai ) boils down to computing dmin (i, ai ). In the following, we compute dmin (i, ai ) for all i ∈ I 0 in O(n) time, after
which d1max (i, ai ) for all i ∈ I 0 can be computed in additional O(n) time.
First of all, we compute gj0 and h0j for all j = 1, 2, . . . , n. This can be easily done in O(n)
time due to the monotonicity properties: g10 ≤ g20 ≤ · · · ≤ gn0 and h01 ≤ h02 ≤ · · · ≤ h0n . Recall
that for each i ∈ I, ai has already been computed. Then, we can compute I 0 in O(n) time
by checking whether i ≤ h0ai − 1 for each i ∈ I.
Next we compute an array B[1 · · · n] such that B[j] = dP (j, gj0 + 1) for each j ∈ [1, n].
Clearly, the array B can be computed in O(n) time. Then, we build a range-minima data
structure on B [2, 11]. The range-minima data structure can be built in O(n) time such that
given any pair (i, j) with 1 ≤ i ≤ j ≤ n, the minimum value of the subarray B[i · · · j] can be
computed in constant time.
d1max (i, ai )
H. Wang
Finally, for each i ∈ I 0 , we can compute dmin (i, ai ) in constant time by querying the
range-minima data structure on B with (i, h0ai − 1).
Therefore, we can compute dmin (i, ai ) for all i ∈ I 0 , and thus compute d1max (i, ai ) for all
i ∈ I 0 in O(n) time.
J
In summary, we can compute λ∗ in O(n log n) time in the case λγ < λ1 and λc < λp .
Our overall algorithm for computing an optimal solution is summarized in the proof of
Theorem 22.
I Theorem 22. An optimal solution for the optimization problem can be found in O(n log n)
time.
Proof. First, we compute λα , λβ , and λδ , in O(n log n) time by using our decision algorithm
and the sorted-matrix searching techniques. Then, we compute λ1 = min{λα , λβ , λδ }.
Second, by using λ1 , we compute the indices Ii0 (α), Ii0 (β), and Ii0 (δ) for all i = 1, 2, . . . , n.
This can be done in O(n) time. For each i ∈ [1, n], if [1, Ii0 (α)] ∩ [Ii0 (β), n] ∩ [Ii0 (δ), n] 6= ∅, we
compute ai (i.e., the smallest index in the above intersection) and add i to the set I (initially
I = ∅). Hence, all such ai ’s and I can be computed in O(n) time.
If I = ∅, then we return λ1 as λ∗ .
If I 6= ∅, then we compute λp in O(n log n) time by Lemma 19. We proceed to compute
d1max (i, ai ) for all i ∈ I 0 by Lemma 21, and then find the smallest feasible value λ0 in the set
{d1max (i, ai ) | i ∈ I 0 } in O(n log n) time. Finally, we return min{λ1 , λp , λ0 } as λ∗ .
The above computes λ∗ in O(n log n) time. Applying λ = λ∗ on our decision algorithm
can eventually find an optimal edge in additional O(n) time.
J
References
1
2
3
4
5
6
7
8
9
10
N. Alon, A. Gyárfás, and M. Ruszinkó. Decreasing the diameter of bounded degree graphs.
Journal of Graph Theory, 35:161–172, 2000.
M. Bender and M. Farach-Colton. The LCA problem revisited. In Proc. of the 4th Latin
American Symposium on Theoretical Informatics, pages 88–94, 2000.
D. Bilò, Luciano Gualà, and Guido Proietti. Improved approximability and nonapproximability results for graph diameter decreasing problems. Theoretical Computer
Science, 417:12–22, 2012.
J.-L. De Carufel, C. Grimm, A. Maheshwari, and M. Smid. Minimizing the continuous diameter when augmenting paths and cycles with shortcuts. In Proc. of the 15th Scandinavian
Workshop on Algorithm Theory (SWAT), pages 27:1–27:14, 2016.
E.D. Demaine and M. Zadimoghaddam. Minimizing the diameter of a network using shortcut edges. In Proc. of the 12th Scandinavian conference on Algorithm Theory (SWAT),
pages 420–431, 2010.
F. Frati, S. Gaspers, J. Gudmundsson, and L. Mathieson. Augmenting graphs to minimize
the diameter. Algorithmica, 72:995–1010, 2015.
G. Frederickson and D. Johnson. Generalized selection and ranking: Sorted matrices. SIAM
Journal on Computing, 13(1):14–30, 1984.
G.N. Frederickson and D.B. Johnson. Finding kth paths and p-centers by generating and
searching good data structures. Journal of Algorithms, 4(1):61–80, 1983.
Y. Gao, D.R. Hare, and J. Nastos. The parametric complexity of graph diameter augmentation. Discrete Applied Mathematics, 161:1626–1631, 2013.
U. Große, J. Gudmundsson, C. Knauer, M. Smid, and F. Stehn. Fast algorithms for
diameter-optimally augmenting paths. In Proc. of the 42nd International Colloquium on
Automata, Languages and Programming (ICALP), pages 678–688, 2015.
15
16
Diameter-Optimally Augmenting Paths
11
12
13
14
15
D. Harel and R.E. Tarjan. Fast algorithms for finding nearest common ancestors. SIAM
Journal on Computing, 13:338–355, 1984.
T. Ishii. Augmenting outerplanar graphs to meet diameter requirements. Journal of Graph
Theory, 74:392–416, 2013.
C.-L. Li, S.T. McCormick, and D. Simchi-Levi. On the minimum-cardinality-boundeddiameter and the bounded-cardinality-minimum-diameter edge addition problems. Operations Research Letters, 11:303–308, 1992.
N. Megiddo. Applying parallel computation algorithms in the design of serial algorithms.
Journal of the ACM, 30(4):852–865, 1983.
A.A. Schoone, H.L. Bodlaender, and J. Van Leeuwen. Diameter increase caused by edge
deletion. Journal of Graph Theory, 11:409–427, 1997.
| 8cs.DS
|
Distributed Event-Triggered Control for Asymptotic
Synchronization of Dynamical Networks ⋆
Tao Liu aMing Cao bClaudio De Persis bJulien M. Hendrickx c
arXiv:1508.04606v3 [cs.SY] 6 Jan 2018
a
Department of Electrical and Electronic Engineering
The University of Hong Kong, Hong Kong S.A.R., China
b
Faculty of Science and Engineering
University of Groningen, 9747 AG, Groningen, The Netherlands
c
ICTEAM Institute, Université catholique de Louvain, Louvain-la-Neuve, Belgium
Abstract
This paper studies synchronization of dynamical networks with event-based communication. Firstly, two estimators are
introduced into each node, one to estimate its own state, and the other to estimate the average state of its neighbours. Then,
with these two estimators, a distributed event-triggering rule (ETR) with a dwell time is designed such that the network
achieves synchronization asymptotically with no Zeno behaviours. The designed ETR only depends on the information that
each node can obtain, and thus can be implemented in a decentralized way.
Key words: distributed event-triggered control, asymptotic synchronization, dynamical networks.
1
Introduction
Synchronization of dynamical networks, and its related problem—consensus of multi-agent systems, have
attracted a lot of attention due to their extensive applications in various fields (see Arenas et al. (2008);
Olfati-Saber et al. (2007); Ren et al. (2007); Wu (2007)
for details). Motivated by the fact that connected nodes
in some real-world networks share information over
a digital platform, these problems have recently been
investigated under the circumstance that nodes communicate to their neighbours only at certain discrete-time
instants. To use the limited communication network
resources effectively, event-triggered control (ETC) (see
Heemels et al. (2012) and reference therein) introduced
in networked control systems has been extensively used
to synchronize networks. Under such a circumstance,
each node can only get limited information, and the
main issue becomes how to use these limited information to design an ETR for each node such that the
network achieves synchronization asymptotically and
⋆ The material in this paper was not presented at any conference.
Email addresses: [email protected] (Tao Liu),
[email protected] (Ming Cao), [email protected] (Claudio
De Persis), [email protected] (Julien M.
Hendrickx).
Preprint submitted to Automatica
meanwhile to prevent Zeno behaviours that are caused
by the continuous/discrete-time hybrid nature of ETC,
and undesirable in practice (Tabuada (2007)).
Early works in ETC focused on dynamical networks
with simple node dynamics such as single-integrators
and double-integrators. In Dimarogonas and Johansson
(2009), distributed ETC was used to achieve consensus.
To prevent Zeno behaviour, a decentralized ETR with
a time-varying threshold was introduced to achieve consensus in Seyboth et al. (2013). Self-triggered strategies
were proposed in De Persis and Frasca (2013) and shown
to be robust to skews of the local clocks, delays, and limited precision in the communication.
Most recently, attention has been increasingly paid to
networks with generalized linear node dynamics. Different types of ETC have been developed to achieve either
bounded or asymptotic synchronization for such networks (e.g., Demir and Lunze (2012); Zhu et al. (2014);
Liu et al. (2013); Meng and Chen (2013); Xiao et al.
(2015); Garcia et al. (2015); Yang et al. (2016); Hu et al.
(2016)). In order to achieve asymptotic synchronization
as well as to prevent Zeno behaviours, two main methods are developed in the literature. One uses bidirectional communication, i.e., at each event time, a node
sends its sampled state to its neighbours and meanwhile
asks for its neighbours’ current states to update the
9 January 2018
We consider a dynamical network described by
control signal (e.g., Meng and Chen (2013); Xiao et al.
(2015); Hu et al. (2016)). The other uses unidirectional
communication, i.e., a node only sends its sampled
information to its neighbours but does not require information from its neighbours (e.g., Liu et al. (2013);
Garcia et al. (2015); Yang et al. (2016)). However, the
latter needs di + 1 ≥ 2 estimators in each node and uses
an exponential term in the ERT in oder to prevent Zeno
behaviours.
ẋi (t) = Hxi (t) + Bui (t), ∀i ∈ V
where xi = (xi1 , xi2 , . . . , xin )⊤ ∈ Rn is the state of
node i. H ∈ Rn×n , B ∈ Rn , and ui ∈ R are the
node dynamics matrix, input matrix, and control
input, respectively. Generally, continuous communication between neighbouring nodes is assumed, i.e.,
P
ui (t) = K N
j=1 aij (xj (t) − xi (t)). This yields the following network
In this paper, we study asymptotic synchronization
of networks with generalized linear node dynamics by
using the unidirectional communication method. The
main differences from the existing results are as follows.
Firstly, a new sampling mechanism is used with which
two estimators are introduced into each node, whereas
most existing results need every node to estimate the
states of all its neighbours. Secondly, inspired by the
method proposed in Tallapragada and Chopra (2014),
we replace the exponential term extensively used in the
literature by a dwell time that was originally introduced
in switched systems (Cao and Morse (2010)), which
can simplify the implementation of the designed ETR.
Thirdly, a distributed ETR for each node is designed
based on the two estimators and dwell time, whereas
most of the existing results use decentralized ETRs
that only consist of local information of the node itself,
i.e., the state error between the node and its own estimator and the time-dependent exponential term (e.g.,
Garcia et al. (2015); Yang et al. (2016)). By introducing an estimation of the synchronization errors between
neighbours using the neighbours’ sampled information,
the proposed ETR method can reduce the number of
sampling times for each node significantly.
2
(1)
ẋi (t) = Hxi + BK
XN
j=1
aij (xj (t) − xi (t)).
(2)
In this paper, we assume that connections in (1) are realized via discrete communication, i.e., each node only obtains information from its neighbours at certain discretetime instants. We will present an event-triggered version of network (2), and study how to design an ETR
for each node to achieve asymptotic synchronization. We
suppose that the topological structure of the network is
fixed, undirected and connected.
We introduce two estimators Oi and OVi into each node
i, where Oi is used to estimate its own state, and OVi is
used to estimate the average state of its neighbours. We
adopt the following control input
ui (t) = K (x̂Vi (t) − lii x̂i (t))
(3)
where K ∈ R1×n is the control gain to be designed, x̂i ∈
Rn and x̂Vi ∈ Rn are states of Oi and OVi , respectively.
The state equations of Oi and OVi are given by
Oi :
Network Model and Preliminaries
Notation: Denote the set of real numbers, non-negative
real numbers, and non-negative integers by R, R+ , and
Z+ ; the set of n-dimensional real vectors and n × m
real matrices by Rn and Rn×m . In , 1n and 1n×m are
the n-dimensional identity matrix, n-dimensional vector and n × m matrix with all entries being 1, respectively. k · k represents the Euclidean norm for vectors
and also the induced norm for matrices. The superscript
(·)⊤ is the transpose of vectors or matrices. ⊗ is the Kronecker product of matrices. For a single ω : R+ → Rn ,
ω(t− ) = lims↑t ω(s). Let G be an undirected graph consisting of a node set V = {1, 2, . . . , N } and a link set
E = {ē1 , ē2 , . . . , ēM }. If there is a link ēk between nodes
i and j, then we say node j is a neighbour of node i and
vice versa. Let A = (aij ) ∈ RN ×N be the adjacency matrix of G, where aii = 0 and aij = aji > 0, i 6= j, if node
i and node j are neighbours, otherwise aij = aji = 0.
The Laplacian matrix L = (lij ) ∈ RN ×N is defined by
PN
lij = −aij , if j 6= i and lii = j=1 aij .
OVi :
x̂˙ i (t) = H x̂i (t), t ∈ [tki , tki +1 )
x̂i (t) = xi (t),
t = tki
x̂˙ Vi (t) = H x̂Vi (t),
−
x̂Vi (t) = x̂Vi (t ) −
P
j∈Ji
t ∈ [tk̄i , tk̄i +1 )
ej (t− ),
t = tk̄i .
(4)
(5)
The increasing time sequences {tki } and {tk̄i }, ki , k̄i ∈
Z+ represent time instants that node i sends updates
to its neighbours and that it receives updates from
one or more of its neighbours, respectively. We assume that: there is no time delay for computation and
execution, i.e., tki represents both the ki th sampling
time and the ki th time when node i broadcasts updates; and the communication network is under an
ideal circumstance, i.e., there are no time delays or
data dropouts in communication. Therefore, the set
Ji = Ji (tk̄i ) = j | tkj = tk̄i , j ∈ Vi is a subset of
Vi , from which node i receives updated information at
t = tk̄i , and Vi = {j | aij > 0, j ∈ V} is the index set of
the neighbours for node i. The vector ei (t) = x̂i (t)−xi (t)
represents the deviation between the state of estimator
Oi and its own, and which node i can easily compute.
2
In network (7), ẑi in ETR (10) contains information of
x̂j , j ∈ Vi which are not available for node i as node
i only has estimator (7b). Therefore, one estimator for
each node is insufficient to implement ETR (10) in practice. To overcome this difficulty, we introduce another
estimator (5) into each node. It is shown that network
(7) is theoretically equivalent to network (1) with the
two estimators Oi , OVi , and ETR (10) is equivalent to
ETR (6) which can be implemented in practice.
The time sequence {tki } is decided by the ETR
tki +1 = inf {t > tki | ri (t, xi , x̂i , x̂Vi ) > 0}
(6)
where ri (·, ·, ·, ·) : R+ × Rn × Rn × Rn → R is the eventtriggering function to be designed. For t > tki , if ri > 0
−
−
at t = t−
ki +1 , then node i samples xi (tki +1 ), x̂i (tki +1 ),
−
−
calculates ei (tki +1 ), sends ei (tki +1 ) to its neighbours,
and reinitialize the estimator Oi at t = tki +1 by
xi (tki +1 ). In addition, node i will reinitialize
P the estima−
)
−
tor OVi by x̂Vi (tk̄i +1 ) = x̂Vi (t−
j∈Ji ej (tk̄i +1 )
k̄i +1
each time when it receives updates from its neighbours. We further assume the network is well initialized at t = t0 , i.e., x̂i (t0 ) = 0 and each node samples
and sends ei (t0 ) to its neighbours.
Therefore, we have
P
x̂i (t0 ) = xi (t0 ), x̂Vi (t0 ) = j∈Vi xj (t0 ) and Ji (t0 ) = Vi
for all i ∈ V. Then, the problem is with the given network topology, to design a proper ETR (6) such that
network (1) achieves synchronization asymptotically
without Zeno behaviours.
Remark 1 It is shown in Liu et al. (2013) that under
the same assumptions, a network with di + 1 estimators
for each node (di is the number of neighbours of node i)
is also theoretically equivalent to network (7), and thus
equivalent to network (1) with two estimators Oi and
OVi . On the other hand, the error ei (t) = x̂i (t) − xi (t) is
extensively used in the literature to design ETR, where
each node sends its sampled state to its neighbours. By
having each node sending ei (tki ) instead of xi (tki ), it
turns out that we can reduce the number of estimators.
The implementation of this new sampling mechanism
needs no more information than that used in the literature. Further, instead of calculating di +1 ≥ 2 estimators
x̂j , our method only calculates x̂i and x̂Vi for each node i,
and hence, our method has implementation advantages,
in particular for networks with large di and limited embedded computing resources in each node. Like most of the
existing results in the literature of ETC, in our method
each node needs to send ei (t) (or xi (t)) to its neighbours
rather than the relative state information (xj (t) − xi (t))
that is extensively used in network (2) with continuously
interconnected nodes. Of course, it is important to study
network (7) by only using the relative state information
for the design purposes which should be studied in the
future.
To simplify the analysis, we will show that network (1)
with controller (3) and estimators (4), (5) is equivalent
to the following system where each node maintains an
estimator of the state of each of its neighbours.
ẋi (t) = Hxi (t) − BK
Oi :
XN
j=1
lij x̂j (t), ∀i ∈ V (7a)
x̂˙ i (t) = H x̂i (t), t ∈ [tki , tki +1 )
x̂i (t) = xi (t),
(7b)
t = tki .
P
P
˙
˙
Defining z̄i =
j∈Vi x̂j gives z̄i (t) =
j∈Vi x̂j (t) =
H z̄i (t), t ∈ [tk̄i , tk̄i +1 ), which has the same dynamics as
x̂Vi defined in (5). Moreover, at t = tk̄i , we have
z̄i (t) =
X
x̂j (t− ) +
X
j∈Vi /Ji (t)
= x̂Vi (t).
xj (t)
j∈Ji (t)
j∈Vi /Ji (t)
=
X
This paper will use model (7) and ETR (10) for the
analysis. But the obtained results can be implemented by
using controller (3) with the two estimators Oi , OVi and
ETR (6). Based on network (7), we give the definition
of asymptotic synchronisation.
x̂j (t− ) +
X
j∈Ji (t)
x̂j (t− ) − ej (t− )
⊤
⊤
⊤
∈
Definition 1 Let x(t) = x⊤
1 (t), x2 (t), . . . , xN (t)
⊤
⊤
⊤
nN
(t),
.
.
.
,
x̂
(t)
∈
R
be
(t),
x̂
RnN and x̂(t) = x̂⊤
2
1
N
a solution of network (7) with initial condition x0 =
⊤
⊤ ⊤
(x⊤
10 , x20 , . . . , xN 0 ) and xi0 = xi (t0 ). Then, the network
is said to achieve synchronization asymptotically, if for
every x0 ∈ RnN the following condition is satisfied
(8)
Thus, we have z̄i (t) = x̂Vi (t) for all t ≥ t0 . Then, controller (3) becomes
ui = K (z̄i − lii x̂i ) = K (x̂Vi − lii x̂i ) .
(9)
lim kxi (t) − xj (t)k = 0, ∀ i, j ∈ V.
t→∞
Substituting (9) into (1) gives that network (1) with (3),
(4), and (5) is equivalent to (7).
Remark 2 When the communication network is not
ideal, model (1) with (3) and Oi , OVi cannot be simplified
to (7). A more complicated model is needed to describe
the network dynamics. Time delays and packet loss will
influence the synchronization performance. However,
due to the robust property of asymptotic synchronization, bounded synchronization can be guaranteed where
P
Moreover, let ẑi = j∈Vi (x̂j − x̂i ). We have x̂Vi = z̄i =
ẑi + lii x̂i . Then, ETR (6) can be reformulated as
tki +1 = inf {t > tki | ri (t, xi , x̂i , ẑi ) > 0} .
(11)
(10)
3
It is shown in Trentelman et al. (2013) that a necessary
and sufficient condition for asymptotic synchronization
of network (2) with continuous interconnections is the
existence of positive definite matrices Pi such that
the final synchronization error may depend on the time
delay magnitude and probability of packet loss. Another
important problem for this case is under what conditions
the network can still achieve synchronization asymptotically. These issues should be studied in the future.
Hi⊤ Pi + Pi Hi = −2In , i = 2, 3, . . . , N.
3
Event-Triggered Control
This condition requires all the linear systems with system matrices Hi = H − λi BK, i = 2, . . . , N are asymptotically stable simultaneously, which is stronger than
that (H, B) is stabilizable. Another alternative is to find
a common P > 0 for all Hi , i = 2, . . . , N (e.g., Wu et al.
(2017)). From (14), network (7) with ETC can be regarded as network (2) with an external input (or a disturbance) Λ̄Φ̄⊤ e. According to input-to-state stability theory, a necessary condition for system (14) to be asymptotically stable is that the the corresponding system
(also described by (14) but without the term Λ̄Φ̄⊤ e) is
asymptotically stable. Hence, the existence of matrix solutions Pi to Lyapunov equations (16) is also a fundamental requirement for network (7) with ETC to achieve
asymptotic synchronization. In this paper, we assume
that such matrices Pi exist.
⊤
⊤
⊤
Denote e(t) = e⊤
with ei (t) =
1 (t), e2 (t), . . . , eN (t)
x̂i (t) − xi (t). Network (7a) can be rewritten by
ẋ = (IN ⊗ H − L ⊗ BK)x − (L ⊗ BK)e.
(12)
Since the topology of the network is undirected and connected, the Laplacian matrix L is irreducible, symmetric,
and has only one zero eigenvalue. Further, there exists
an orthogonal matrix Ψ = (ψ1 , ψ2 , . . . , ψN ) ∈ RN ×N
with ψi = (ψi1 , ψi2 , . . . , ψiN )⊤ and Ψ⊤ Ψ = IN such
that Ψ⊤ LΨ = Λ = diag(λ1 , λ2 , . . . , λN ) where
√0 =
λ1 < λ2 ≤ λ3 ≤ · · · ≤ λN . Choose ψ1 = 1/ N 1⊤
N
for λ1 . Due to the zero row sum property of L, we
PN
have
j=1 ψij = 0 for all i = 2, 3, . . . , N . Defining
Φ = (ψ2 , ψ3 , . . . , ψN ) ∈ RN ×(N −1) gives
Φ⊤ Φ = IN −1 ,
ΦΦ⊤ = IN −
1
1N ×N .
N
P
P
Let zi =
j∈Vi (x̂j − x̂i ),
j∈Vi (xj − xi ), ẑi =
⊤ ⊤
z = (z1⊤ , z2⊤ , . . . , zN
)
= (−L ⊗ In )x, and ẑ =
⊤ ⊤
(ẑ1⊤ , ẑ2⊤ , . . . , ẑN
) = (−L ⊗ In )x̂. Next, we give a useful
lemma which will be used to prove the main result.
(13)
Let Λ1 = Φ⊤ LΦ = diag{λ2 , λ3 , . . . , λN }, Φ̄ = Φ ⊗ In
and Λ̄ = Λ1 ⊗ BK = diag {λ2 BK, λ3 BK, . . . , λN BK}.
Defining y = Φ̄⊤ x gives
ẏ(t) =Φ̄⊤ (IN ⊗ H)x − (L ⊗ BK)(IN n − Φ̄Φ̄⊤
+Φ̄Φ̄⊤ )(x + e)
=(IN −1 ⊗ H − Λ1 ⊗ BK)y − Λ̄Φ̄⊤ e
⊤
Lemma 2 Consider network (7). The following two inequalities hold for any t ≥ t0
kẑk ≤ λN (kek + kyk)
λ2 kyk ≤ λN kek + kẑk.
(14)
(17)
(18)
PROOF. Due to k(L ⊗ In )k = λN , we have
⊤
where we use properties Φ̄ (IN ⊗ H) = (IN −1 ⊗ H)Φ̄
and (L ⊗ BK)(IN n − Φ̄Φ̄⊤ ) = 0 for any BK, which
are supported by facts L1N = 0 and (13). Denoting
H̄ = (IN −1 ⊗ H) − (Λ1 ⊗ BK) = diag {H2 , H3 , . . . , HN }
with Hi = H − λi BK, system (14) can be simplified to
ẏ = H̄y − Λ̄Φ̄⊤ e.
(16)
kẑk = k(L ⊗ In )(x + e)k ≤ kzk + λN kek
(19)
kzk = k(L ⊗ In )(x̂ − e)k ≤ kẑk + λN kek.
(20)
⊤
Let U = ΦΦ , then for any L, we have LU =
U L, i.e., L and U are diagonalizable simultaneously. Further, we have Ψ⊤ LΨ = Λ and Ψ⊤ U Ψ =
diag{λu1 , λu2 , . . . , λuN }, where λu1 = 0 and λui = 1, i =
2, 3, . . . , N are eigenvalues of U . Let λ̄i , i = 1, 2, . . . , N
be eigenvalues of the matrix (λ2N U 2 − L2 ). Then with
U 2 = U , we have λ̄1 = 0 and λ̄i = λ2N − λ2i ≥ 0,
i = 2, 3, . . . , N , which gives L2 ≤ λ2N U 2 . Thus, we have
(15)
PN
By defining x̄ = N1 i=1 xi , we have kyk2 = x⊤ Φ̄Φ̄⊤ x =
PN
2
i=1 kxi − x̄k where the last equality follows from
⊤
Φ Φ = IN −1 and (Φ̄Φ̄⊤ )2 = Φ̄Φ̄⊤ . Therefore, if
limt→∞ ky(t)k = 0, then xi (t), xj (t), and x̄(t) are
asymptotically equal when t → ∞, i.e., network (7)
achieves synchronization asymptotically. This result is
summarized in the following lemma.
kzk2 =x⊤ (L2 ⊗ In )x ≤ λ2N x⊤ (U 2 ⊗ In )x
=λ2N kΦ̄⊤ xk2 = λ2N kyk2 .
(21)
Combining (19) with (21) gives inequality (17). Similar
to (21), we have kyk2 = x⊤ (U 2 ⊗ In )x ≤ 1/λ22 x⊤ (L2 ⊗
In )x which with (20) gives (18).
✷
Lemma 1 If system (15) is asymptotically stable, i.e.,
limt→∞ ky(t)k = 0, then network (7) achieves synchronization asymptotically.
4
Let ρ =
λN
δ
,
2N (α2 +δ 2 )
√
ρ1 =
1 √
δ
λ2 ( 2(α2 +δ 2 )
Based on the comparison theory (Khalil (2002)), we
have kel (t)k/ky(t)k ≤ φ(t − tkl ), whenever kel (tkl )k
/ky(tkl )k ≤ φ(tkl ) where φ(t − tkl ) is the solution of the
ordinary differential equation
+ 1),
δ ∈ (0, 1), α = maxi=2,3,...,N {λi kPi BKk}, a =
δ
δ
kHk + kH̄k
+ λN α kBKk, b = λN kBKk(1 + α ), and
τ∗ =
1
a
ln
aρ
bρ1
+ 1 > 0. We have the following result.
φ̇ = aφ + b
Theorem 1 Network (7) achieves synchronization
asymptotically under the distributed ETR
tki +1 = inf {t ≥ tki + τ ∗ | kei k > ρkẑi k} .
with the initial condition φ(tkl ). At t = tkl , we have
kel (tkl )k/ky(tkl )k = 0. Setting φ(tkl ) = 0 gives
(22)
kel (t)k
≤ φ(t − tkl ), ∀t ∈ [tkl , t∗ ).
ky(t)k
Moreover, no Zeno behaviour occurs in the network.
kel (t)k
kel (t)k
≤ ρ1
≤ ρ1 φ(t − tkl ), ∀t ∈ [tkl , t∗ ).
kẑ(t)k
ky(t)k
(23)
Solving (29) with φ(tkl ) = 0 shows that it will take φ(t−
tkl ) a positive time constant τ ∗ to change its values from
0 to ρ/ρ1 , so does kel (t)k/ky(t)k. Therefore, it requires
at least τ ∗ to make kel (t)k move from 0 to ρkẑ(t)k.
This is true at t = t0 , as we have kei (t0 )k = 0 and
hence kei (t0 )k ≤ ρkẑ(t0 )k, ∀i ∈ V. Suppose to obtain a
contradiction that (23) does not always hold, and let t∗
be the infimum of times at which it does not hold. Due to
the finite number of nodes, there exists a node l such that
kel k > ρkẑk for times arbitrarily close t∗ from above,
i.e., ∀ǫ > 0, ∃t ∈ [t∗ , t∗ + ǫ] such that kel (t)k > ρkẑ(t)k.
It follows from ETR (22) that t∗ must be in (tkl , tkl + τ ∗ ]
for some kl ∈ Z+ . We now show that there cannot exist
a t∗ in (tkl , tkl + τ ∗ ], which will establish (23). Since
kei (t)k ≤ ρkẑ(t)k, ∀i ∈ V, ∀t < t∗ , which gives
kek2 =
N
X
i=1
kei k2 ≤
δ2
kẑk2 .
2
2λN (α2 + δ 2 )
Suppose, to obtain a contradiction, that t∗ < tkl + τ ∗ . In
that case, kel (t)k/ky(t)k ≤ φ(t − tkl ) < φ(τ ∗ ) ≤ ρ/ρ1 ,
for all t ≤ t∗ . By continuity of kel k/kyk, this implies the
existence of an ε > 0 such that kel (t)k/ky(t)k < φ(τ ∗ )
for all t ≤ t∗ + ε. Therefore, there holds then kel (t)k <
ρkẑ(t)k for all t < t∗ + ε, in contradiction with t∗ being
the infimum of times at which kel (t)k > ρkẑ(t)k.
Now, select the Lyapunov function candidate V = y ⊤ P y
with P = diag{P2 , P3 , . . . , PN }. Then, the derivative of
V along system (15) satisfies
(24)
On the other hand, inequality (17) gives
2
kẑk ≤
2λ2N (kek2
2
+ kyk ).
V̇ ≤ −2kyk2 + 2αkykkΦ̄⊤ek.
(25)
Calculating
d kel k
dt kyk
δ
ky(t)k, ∀t ∈ [t0 , t∗ ).
α
δ
kyk.
α
(32)
V̇ ≤ −2(1 − δ)kyk2 < 0, ∀kyk =
6 0.
(33)
kΦ̄⊤ ek ≤ kΦ̄⊤ kkek = kek ≤
(26)
Substituting (32) into (31) gives
for t ∈ [tkl , t∗ ) directly gives
kel k kΛ̄kkel kkek
d kel k
≤ kHk + kH̄k
+
dt kyk
kyk
kyk2
kek
+ λN kBKk
+ λN k BKk
(27)
kyk
Therefore, the equilibrium point y = 0 of system (15) is
asymptotically stable. Based on Lemma 1, the network
achieves synchronization asymptotically.
✷
where we use (17) in Lemma 2 to get (27). Substituting
(26) into (27) gives
kel k
d kel k
≤a
+ b.
dt kyk
kyk
(31)
The inequality (23) holds, so does (26). Combining (26)
with kΦ̄k = 1 yields
Substituting (25) into (24) yields
ke(t)k ≤
(30)
Further, combining (18) with (24) gives kẑk ≥ kyk/ρ1
which with (30) leads to
PROOF. Under ETR (22), the existence of τki =
tki +1 − tki > 0 is guaranteed by the dwell time τ ∗ . To
show asymptotic synchronization, we claim that the
network with (22) satisfies
kei k ≤ ρkẑk, ∀i ∈ V, ∀t ≥ t0 .
(29)
Remark 3 Like most results in synchronization of dynamical network with/without ETC (e.g., Trentelman et al.
(2013); Guinaldo et al. (2013)), one usually needs some
global parameters to guarantee asymptotic synchronization and exclude Zeno behaviours. These parameters can
(28)
5
be estimated by using methods proposed in the related
literature (e.g., Franceschelli et al. (2013)), and can be
initialized to each node at the beginning. However, how
to use local parameters rather than global ones (e.g., how
to replace N by using local parameter such as the degree
of the node di ) remains open, and deserves attention.
4
An Example
To show the effectiveness of our method, consider a network with 10 nodes that have parameters as follows
H=
0 −0.5
!
, B=
0
!
, K=
−0.5 1 .
1
0.5 0
Remark 4 Most of the existing results (e.g., Demir and Lunze
(2012); Guinaldo et al. (2013); Seyboth et al. (2013);
Zhu et al. (2014); Garcia et al. (2015); Yang et al.
We adopt the two-nearest-neighbour graph to describe
(2016) ) use decentralized ETRs which can be summathe topology, i.e., Vi = {j | j = i ± 1, i ± 2}, i =
rized in the following compact from
1, 2, . . . , 10. If j ∈ Vi and j < 0 (j > 10), then j = j + 11
(j = j − 10). Since the matrix H has two eigenvalues
on the imaginary axis of the complex plane, the network
will synchronize to a stable time-varying solution deter(34)
tki +1 = inf t | kei k > c0 + c1 e−γt
mined by the initial condition. By calculating, we get
α = 2.9061. We select δ = 0.9. Figure 1 gives the simulation results of the network with the distributed ETR
(22) (DDT), which shows the effectiveness of the prowhere c0 ≥ 0, c1 ≥ 0, γ > 0 are three design parameters.
posed method. In the figure, we only give the sampling
Obviously, ETR (34) only depends on local information
time instants in the first 2 seconds for clarity. The thefrom node i itself, and asymptotic synchronization can
oretical value of τ ∗ is 0.0013 s. The minimum and maxbe achieved only when c0 = 0. In our paper, we introduce
imum sample periods (τmin /τmax ) for each node during
kẑi k into the proposed ETR (22). The term kẑi k updated
the simulation time are given in Table 1 which shows
by xj (tkj ) estimates the synchronization errors between
that the actual sample periods are much larger than τ ∗ .
neighbours continuously, and thus provides each node
useful information for determining its sampling times.
Therefore, the proposed ETR can reduce the sampling
We also compared our method with the decentralized
times significantly, in particular for cases where kẑi k is
ETR (34) (DET) proposed in Guinaldo et al. (2013).
large (see the example in Section 4 for details). Further,
According to Remark 4, only bounded synchronization
it is shown in Liu et al. (2013) that a similar distributed
can be guaranteed with c0 6= 0 in (34) (Seyboth et al.
−γt
ETR as (22) but with an exponential term c1 e
can also
(2013)). For this case, the advantage of our method is
guarantee asymptotic synchronization. However, this paclear. So here, we only compare our method with the case
per replaces the exponential term by a dwell time which
c0 = 0 where asymptotic synchronization under (34) can
can be implemented easily in practice. Such a τ ∗ gives an
also be achieved. We select c1 = ρ and γ = 0.30579. Durupper bound for the designed ETR (22), and therefore, a
ing the simulation period (0 – 18 s), the network with
modified ETR with 0 < τi∗ ≤ τ ∗ can also synchronize the
DDT samples 3432 times in total, whereas the network
network without Zeno behaviours.
with DET samples 212 times more (3644 times in total).
Remark 5 To simplify notations, this paper only considers the case where ui is a scalar. However, the obtained
results can extend to multiple-input case directly. It is
pointed out in Heemels et al. (2013) that the joint design
of the controller and event-triggering rule is a hard problem. However, we can select any control gain K to synchronizes the continuous-time network (2), i.e., to stabilize (H, λi B), i = 2, . . . , N simultaneously. It can be
selected by solving a group of linear matrix inequalities.
Moreover, a periodic ETC method was proposed to stabilize linear systems in Heemels et al. (2013) where the
triggering condition was verified periodically. In the paper, we do not check the event-triggering condition in the
time interval [tki , tki + τ ∗ ), but check the condition continuously during the period [tki + τ ∗ , tki +1 ). It is of great
interest to study asymptotic synchronization by using periodic ETC and one-directional communications.
Fig. 1. Simulation for the network with DDT.
6
Table 1. The minimum/maximum sample period
5
Node 1
Node 2
Node 3
Node 4
τmin
0.0153
0.0114
0.016
0.0214
0.0188
τmax
0.2651
0.5292
0.6336
0.0817
0.1851
Node 6
Node 7
Node 8
Node 9
Node 10
τmin
0.0046
0.0688
0.0116
0.0117
0.0117
τmax
0.3584
0.2841
1.4677
0.5347
0.5238
Demir, O., Lunze, J., 2012. Event-based synchronziation of multi-agent systems. In: IFAC Conf. on Analysis and Design of Hybrid Systems. Eindhoven, The
Netherlands, pp. 1–6.
Dimarogonas, D. V., Johansson, K. H., 2009. Eventbased control for multi-agent systems. In: IEEE Conference on Decison and Control. Shanghai, China, pp.
7131–7136.
Franceschelli, M., Gasparri, A., Giua, A., Seatzu, C.,
2013. Decentralized estimation of Laplacian eigenvalues in multi-agent systems. Automatica 49, 1031–
1036.
Garcia, E., Cao, Y., Wang, X., Casbeer, D., 2015. Decentralized event-triggered consensus of linear multiagent systems under directed graphs. In: American
Control Conference. Chicago, USA, pp. 5764–5769.
Guinaldo, M., Dimarogonas, D., Johansson, K., Sanchez,
J., Dormido, S., 2013. Distributed event-based control strategies for interconnected linear systems. IET
Control Theory and Applications 7 (6), 877886.
Heemels, W. P. M. H., Donkers, M. C. F., Teel, A. R.,
2013. Periodic event-triggered control for linear systems. IEEE Trans. on Automatic Control 58 (4), 847–
861.
Heemels, W. P. M. H., Johansson, K. H., Tabuada,
P., 2012. An introduction to event-triggered and selftriggered control. In: IEEE Conference on Decison and
Control. Maui, USA, pp. 3270–3285.
Hu, W., Liu, L., Feng, G., 2016. Consensus of linear
multi-agent systems by distributed event-triggered
strategy. IEEE Trans. on Cybernetics 46 (1), 148–157.
Khalil, H. K., 2002. Nonlinear Systems, 3rd Edition.
Prentice Hall, New Jersey.
Liu, T., Cao, M., De Persis, C., Hendrickx, J. M., 2013.
Distributed event-triggered control for synchronization of dynamical networks with estimators. In: IFAC
Workshop on Distributed Estimation and Control in
Networked Systems. Koblenz, Germany, pp. 116–121.
Meng, X., Chen, T., 2013. Event based agreement protocols for multi-agent networks. Automatica 49, 2125–
2132.
Olfati-Saber, R., Fax, J. A., Murray, R. M., 2007. Consensus and cooperation in networked multi-agent systems. Proceedigs of the IEEE 95 (1), 215–233.
Ren, W., Beard, R. W., Atkins, E., 2007. Informaiton
consensus in multivechicle cooperative control: collective behavior through local interation. IEEE Control
System Magazine 27 (2), 71–82.
Seyboth, G. S., Dimarogonas, D. V., Johansson, K. H.,
2013. Event-based broadcasting for multi-agent average consensus. Automatica 49, 245–252.
Tabuada, P., 2007. Event-triggered real-time scheduling
of stabilizing control tasks. IEEE Trans. on Automatic
Control 52 (9), 1680–1685.
Tallapragada, P., Chopra, N., 2014. Decentralized eventtriggering for control of nonlinear systems. IEEE
Trans. on Automatic Control 59 (12), 3312–3324.
Trentelman, H. L., Takaba, K., Monshizadeh, N., 2013.
Robust synchronization of uncertain linear multi-
Node 5
Conclusion
This paper has studied asymptotic synchronization of
networks by using distributed ETC. By using the introduced estimators, a distributed ETR for each node has
been explored, which only relies on the state of the node
and states of the estimators. It has been shown that the
proposed ETC synchronizes the network asymptotically
with no Zeno behaviours. It is worth pointing out that
time-delay and data packet dropout are common phenomena which definitely affects the synchronization of
networks with event-based communication. It appears
that synchronization of such networks with imperfect
communication is an important issue to pursue further
for both theoretical interest and practical consideration.
Acknowledgements
Liu’s work was supported by The University of Hong
Kong Research Committee Research Assistant Professor Scheme and a grant from the RGC of the Hong Kong
S. A. R. under GRF through Project No. 17256516.
Cao’s work was supported in part by the European Research Council (ERC-StG-307207) and the Netherlands
Organization for Scientific Research (NWO-vidi-14134).
De Persis’s work was partially supported by the Dutch
Organization for Scientific Research (NWO) under the
auspices of the project Quantized Information Control
for formation Keeping (QUICK) and by the STW Perspectief program “Robust Design of Cyber-physical Systems” under the auspices of the project “Cooperative
Networked Systems”. Hendrickx’s work was supported
by the Belgian Network DYSCO (Dynamical Systems,
Control, and Optimization), funded by the Interuniversity Attraction Poles Program, initiated by the Belgian
Science Policy Office.
References
Arenas, A., Dı́az-Guilera, A., Kurths, J., Moreno, Y.,
Zhou, C., 2008. Synchronization in complex networks.
Physics Reports 469, 93–153.
Cao, M., Morse, A. S., 2010. Dwell-time switching.
Sysstems & Control Letters 59 (1), 57–65.
De Persis, C., Frasca, P., 2013. Robust self-triggered coordination with ternary controllers. IEEE Trans. on
Automatic Control 58 (12), 3024–3038.
7
agent systems. IEEE Trans. on Automatic Control
58 (6), 1511–1523.
Wu, C. W., 2007. Synchronization in complex networks
of nonlinear dynamical systems. World Scientific, Singapore.
Wu, Y., Meng, X., Xie, L., Lu, R., Su, H., Wu, Z.,
2017. An input-based triggering approach to leaderfollowing problems. Automatica 75, 221–228.
Xiao, F., Meng, X., Chen, T., 2015. Sampled-data consensus in switching networks of integrators based on
edge events. International Journal of Control 88 (2),
391–402.
Yang, D., Ren, W., Liu, X., Chen, W., 2016. Decentralized event-triggered consensus for linear multi-agent
systems under general directed graphs. Automatica
69, 242–249.
Zhu, W., Jiang, Z.-P., Feng, G., 2014. Event-based consensus of multi-agent systems with general linear models. Automatica 50, 552–558.
8
| 3cs.SY
|
Analyses of Aggregate Fluctuations of Firm Network Based on the
Self-Organized Criticality Model∗
Hiroyasu Inoue†
Graduate School of Simulation Studies, University of Hyogo
Abstract
arXiv:1512.05066v2 [q-fin.GN] 4 Apr 2016
This study examine the difference in the size of avalanches among industries triggered by demand shocks, which can be rephrased by control of the economy or fiscal policy, and by using
the production-inventory model and observed data. We obtain the following results. (1) The size of
avalanches follows power law. (2) The mean sizes of avalanches for industries are diverse but their
standard deviations highly overlap. (3) We compare the simulation with an input-output table and
with the actual policies. They are compatible.
Keywords: Aggregate Fluctuation, Demand, Network, Firm, Production, Inventory, Control Theory
∗ This study is conducted as a part of the project “Price Network and Dynamics of Small and Medium Enterprises”
undertaken at the Research Institute of Economy, Trade and Industry (RIETI). The authors thank the institute for various
means of support. We thank Hiroshi Yoshikawa, Hideaki Aoyama, Hiroshi Iyetomi, Yuichi Ikeda, Yoshi Fujiwara, Wataru
Soma, Yoshiyuki Arata, and members who attended the internal seminar of RIETI for their helpful comments. We gratefully
acknowledge financial support from the Japan Society for the Promotion of Science (No. 15K01217).
† Graduate School of Simulation Studies, University of Hyogo, 7-1-28 Minatojima-minamimachi, Chuo-ku, Kobe, Hyogo
650-0047, Japan
1
1
Introduction
Giving stimulus to firms and prompting the spillover effect is a way for a government to affect its
economy, which includes purchasing goods and services, giving grants to firms, and fine-tuning taxes.
Governments consider fiscal policy as an important determinant of growth [1, 2]. Currently, analysis
of input-output table is considered a strong tool to predict the spillover effect [3]. It enables us to
obtain a single predicted value of the spillover effect caused by the stimulus. However, it is obvious
that we cannot obtain the same result as that of the prediction, even if we use exactly the same
volume as that used in the calculation. Nevertherless, we normally expect that the result should be
substantially around the prediction. This concern, that, whether the expectation is correct is the
main topic of this study.
If the size of the spillover effect is around the average of the prediction, it should be true that
the propagation is never amplified or reduced through firm networks. However, Gabaix showed that
if the firm size distribution is fat-tailed, the hypothesis breaks down [4]. In addition, Acemoglu et
al. pointed out that microeconomic shocks may lead to aggregate fluctuations in the presence of
intersectoral input-output linkages [5]. These studies suggest that the stimulus and spillover effect
do not result in the proximity of the prediction. In other words, normal distribution is usually
assumed, but this assumption is doubtful.
This study reveals how demand shocks from outside cause the spillover effect. We use a micro
model invented by Bak et al. [6] and employ observed data. We clear up the following points: (1)
the diversity of the spillover effect. (it must depend on industries in which shocks are given.) and
(2) the extent of getting involved in the spillover effect. (it must also depend on industries.)
The remainder of this paper is organized as follows. In Section 2, we introduce the dataset.
Section 3 describes the methodologies that we utilize in analyses. Section 4 presents the results.
Finally, Section 5 concludes.
2
Data
We use datasets, TSR Company Information Database and TSR Company Linkage Database, collected by Tokyo Shoko Research (TSR), one of the major corporate research companies in Japan.
The datasets are provided by the Research Institute of Economy, Trade and Industry (RIETI). In
particular, we use the dataset collected in 2012. The TSR data contain a wide range of firm informa-
2
tion. As necessary information for our study, we use identification, capital, industry type, suppliers
and clients. We construct an entire network of firms based on suppliers and clients. Note that there
are up to 24 suppliers and up to 24 clients for each firm in the data. It may be considered that
the constraint limits the number of links for each node. However, a node can be suppliers of other
nodes without limitation, as long as those clients designate the node as a supplier, and vice versa.
Therefore, the numbers of suppliers or clients are not limited to 24. The number of firms, that is,
nodes, is 1,109,549. The number of supplier-client ties, that is, links, is 5,106,081. This network has
direction and the direction is important in our study.
We split firms based on industries. The industries are classified by the Japan Standard Industrial
Classification [7]. We mainly use the division levels that have 20 classifications. However, we make
alterations to the classification. Since the classifications “S: Government, except elsewhere classified”
and “T: Industries unable to classify” are less important in our study, we omit them. In addition, we
separate “I: Whole sale and retail trade” into wholesale and retail. The difference of the divisions is
not negligible in our study because shocks from outside, such as fiscal policies, often occur in retail.
Therefore, the division level in our study after alterations shows 19 industries.
Moreover, we use three industries at the group level to compare the effects of some Japanese
fiscal policies. The groups are “5911: New motor vehicle stores,” “5931: Electrical appliance stores,
except secondhand goods,” and “6821: Real estate agents and brokers.” The difference between the
division and group levels is clear in the later industries and there is no concern about confusing the
two levels.
We use an input-output table to compare the prediction of the spillover-effect size between a
micro model and the table. As the closest table in time, we use the 2011 updated input-output table
[8].
Figure 1 shows the degree distribution of the observed network. The red plots are the distribution
of the observed network. An important point is that the distribution is fat-tailed, which means the
distribution does not decay super-linearly. It seems that we can fit plots to a line P ∝ k −λ , where P
is the cumulative probability, k is the degree, and λ is a positive constant. If the degree distribution
is the normal distribution, the plot is shaped as the blue plots in Figure 1. Since normal distribution
exponentially decays, we can observe the blue plots decrease super-linearly on the log-log plot. How
to create the random network is explained in Section 4. The reason we should compare the observed
network with the random network is that the random network creates aggregate fluctuation that
3
decays exponentially, as we show in Section 4. In other words, it tells us that if the random network is
the network in the real economy, there is no fat-tailed aggregate fluctuation. However, the observed
network is not the random network as we see here.
[Figure 1 here]
If a probability distribution or a cumulative probability distribution can be fitted to a line, it is
said that the distribution is a power-law distribution. A network with a power-law distribution is
often called a scale-free network. It has been pointed out that the power-law or scale-free nature of
networks is a determinant of fat-tailed aggregate fluctuations [4]. Since the observed network is the
scale-free network, we expect that the aggregate fluctuations of the network are fat-tailed.
3
Methodology
We use a modified model [9] based on a production model [6]. The modified model enables us to
conduct micro-level simulations and investigate the characteristics of aggregate fluctuations. The
model of production and inventory was originally invented by Bak et al. [6]. The model assumes that
firms connect with each other on supply chains. Each firm has some amount of inventory. When a
firm receives orders from clients, it supplies intermediate goods or services to clients. If the firm does
not have enough inventory, it sends orders to suppliers. Therefore, cascades of orders and production
sometimes occur. The size of the cascades can be defined by the total extent of production due to
activated firms. Bak et al. showed that the distribution of the cascade size follows the power law.
This result underlies recent network-based studies related to aggregate fluctuations [4, 5]. That is,
the cascade reaction can be understood as aggregate fluctuations. Here, for brevity, we call the
cascade reaction an avalanche.
The result obtained by Bak et al. strongly depends on the regularity of supply chain network.
A node has two suppliers and two clients in the regular network, except the nodes in the top and
bottom layers. As we have already shown that the real supply-chain network is not a regular network
but a scale-free network, the assumption is too strong to apply the model to the real supply-chain
network. To mitigate the limitation of the regular network, Iino and Iyetomi generalized the model
so that a node has arbitrary numbers of in-degree or out-degree and analyzed the nature of the
generalized production model [9]. We employ their generalized model with a minor modification.
Here, we describe the model used in our analyses. For every time step t and every firm i, a new
4
amount of inventory is decided based on the following equation.
zi (t + 1) = zi (t) − si (t) + yi (t),
where zi (t) is the amount of inventory of firm i at time t, si (t) is the amount of orders received by
firm i at time t, and yi (t) is the amount of production conducted by firm i at time t. The equation
renews the inventory and is depicted in Figure 2(a). We assume that (1) the firm equally sends
orders to its suppliers, (2) each firm produces one unit of production from one unit of material
that it obtains, and (3) a firm produces the minimum goods necessary to meet requests from its
consumers. Assumptions (1) and (2) simply result in a production feature in which yi (t) is a multiple
of ni , where ni is the number of suppliers for the firm i. In addition, the assumption (3) results in
zi (t) ≤ ni . Based on the inventory renewal equation and the assumptions, the amount of production
yi (t) is given by
0
ni
yi (t) =
..
.
a (t) · n
i
i
(si (t) ≤ zi (t))
(zi (t) < si (t) ≤ zi (t) + ni )
..
.
(1)
(zi (t) + (ai (t) − 1)ni < si (t) ≤ zi (t) + ai (t)ni ),
where ai (t) is the number of orders that firm i places with each supplier. ai (t) is calculated by a
ceiling function
ai (t) =
&
'
si (t) − zi (t)
.
ni
Since the quantity of the received orders si (t) is the sum of orders to firm i,
si (t) =
X
aj (t),
j
where j is one of clients of firm i. If a firm does not have a client and the firm needs to produce, the
firm regarded as belonging to primary industry and is assumed able to produce an arbitrary amount
of production.
[Figure 2 here]
The first orders are placed from outside. Depending on the analyses, a firm is selected from all
firms or firms in a specific industry to place an order. The selection is uniformly random.
Two firms may mutually supply or a long step of supply chains may form a cycle. It is possible
5
that firms on a loop indefinitely produce goods or services, although this never occurs in the real
economy. Iino and Iyetomi assume that firms have randomly assigned potential values, which is an
analogy of electrostatics. A firm with more potential than another can supply, which is similar to
water flow. Although this assumption helps avoid a loop and is useful for analyzing the nature of
randomly created networks, it is not particularly clear how to assign a value to each firm. Here,
we make a simple assumption. A firm that has already supplied products, that is, a firm that is
already in a propagation process, is ignored as a supplier. Figure 2(b) shows an example of a loop
with three firms. Firm 3 ignores firm 1 when it needs production. More precisely, the supply link
from firm 1 to firm 3 is tentatively ignored.
Since the observed data include all industries, it may be considered that it is unnatural to
contemplate inventory for service industries. This is because it is not understandable to consider
inventory for intangible products, such as insurance or healthcare. However, service industries do
have inventory. That is, if any service is ready to be used, it should be considered as inventory. For
example, a vacant hotel room ready for use incurs cost. Therefore, we can discuss all industries that
deal with tangible or intangible products in the same network.
4
Results
In this section, we show the fat-tailed nature of avalanches and their diversity over industries. We
start with the results of avalanches comparing a random network and the observed network. In a
random network, every pair of nodes is connected
according
to constant probability p. The expected
N
number of links of the random network with p is
p , where N is the number of nodes. The
2
random network is created so that the network has the same number of nodes and the same number
of links as the observed network. The observed network has 1,109,549 nodes and 5,106,081 links.
Therefore, we set p as approximately 8.30 × 10−6 . As a result, we obtain a random network with
1,109,549 nodes and 5,366,223 links. This accordance is necessary to compare the two networks.
The experiments proceed as follows for both networks. (1) At time t, a firm is randomly chosen
from all the firms. (2) A chosen firm sells a unit of product. (3) An avalanche is calculated. (4)
Repeat (1)–(3) for 1 billion times (t proceeds from 1 to 1 billion.) For an avalanche size, that is,
aggregated production, we obtain
Y (t) =
X
i
6
yi (t),
for every time step t.
Figure 3 shows the avalanche sizes for the two networks. The red plots are for the observed
network and the blue plots are for the random network. The random network obviously decays
fast and it seems that the distribution cannot be fitted to a line. On the other hand, the observed
network has a part that can be fitted to a line. This result, that a scale-free network has a fat-tailed
avalanche size, has already been shown analytically [9] and has been proved partly under some
constraints [10].
[Figure 3 here]
The results here tell us that uniformly random stimuli cause scale-free avalanche on the real
network. That is, the average avalanche size is not a representative value of it. The input-output
table analysis results in a single value of prediction, which means some representative value is used
to predict the spillover effect. However, the power-law distribution does not have a typical scale. It
seems that careful analyses are required by the input-output table.
From the viewpoint of fiscal policy, it is important to know how stimulus received by different
industries causes differences. We conduct the experiments with a few changes to the previous experiments. A firm is randomly chosen from an industry. The industry is fixed through an experiment.
In every experiment, 1 billion instances of demand are given. The experiments are conducted for the
19 industry divisions. Figure 4 shows the distributions of avalanche sizes. While we had expected
that the distributions of avalanche sizes would have different shapes, this turns out to be untrue.
As can be observed, there is no apparent difference in shapes.
[Figure 4 here]
Figure 5 shows the mean sizes of avalanches. However, as mentioned in the last paragraph, the
size distribution is fat-tailed and therefore, the average is not a representative of the avalanche size.
The error bars in Figure 5 show standard deviations. Since we already observed that the avalanche
size has the power law distribution in Figures 3 and 4, we know that standard deviations or variances
are large. We do not conduct the statistical test for the difference of the average because the 1 billion
samples cause small standard errors and always show significance of the difference. Therefore, the
test is pointless. Instead, it is important that standard deviations are overlapping. It seems that
we cannot strongly expect that the spillover effect started from a specific industry that is certainly
superior or inferior to other industries.
7
[Figure 5 here]
In Figure 5, the order of the industries that are aligned in the horizontal axis roughly shows the
advancement from the primary industries. Since an advanced industry, such as manufacturing or
services, is in the downstream of supply chains and has long chains from the primary industries, it
may be considered that the distance to the primary industries has a positive correlation coefficient
with the avalanche size of an industry. However, we do not observe such a relationship, which can
be attributed to the network structure. Scale-free network is a small-world network [11]. A firm has
a short path to a hub. Since the hub yields a large yi (t), the magnitude of avalanches seems to be
dominated by this mechanism.
It may be argued that the production model of this study is only a model and it is doubtful how
much it can explain the actual economy. Therefore, we simply compare the size of avalanches and
the inverse matrix of the input-output table [8]. The Pearson’s correlation coefficient is 0.28. If we
consider that the production model does not include the data of trade volume at all, it can be said
that the coefficient is surprisingly large. It should be noted that the production model can simulate
the variances that cannot be obtained from the input-output table.
We conduct further experiments that start from the group level of industries so that we can compare the effect of actual policies. The results of Figure 6 are based on the division level industries.
Those industries are “5911: New motor vehicle stores” (CarSale), “5931: Electrical appliance stores,
except secondhand goods” (ElectronicsSale), and “6821: Real estate agents and brokers” (HouseSale), which are at the group level. They correspond to the target industries of past Japanese fiscal
policies: eco-vehicle tax breaks, eco-point system for housing system, and eco-point system for home
electronics. The setup of experiments is the same as that for the previous experiments.
[Figure 6 here]
Since the government publishes the actual size of budgets and some institutes publish the estimated economic results, we can validate the predictability of the model. The government lost tax
revenue corresponding to 241.0 billion Japanese yen (1.98 billion US dollars at an assumed exchange
rate of 122 Japanese yen to 1 US dollar) in 2009 for the eco-vehicle tax breaks [12]. The economic
result was estimated about 2,937.53 billion Japanese yen (24.08 billion US dollars) [13, 14]. However,
this estimation does not include indirect effects of other industries. The leverage of the return to
the investment is 12.18. The government spent 494.8 billion Japanese yen (4.06 billion US dollars)
8
for the eco-point system for housing in total. The economic result was estimated at about 414.0
billion Japanese yen (3.39 billion US dollars) [15]. The leverage is 0.84. The government spent
692.9 billion Japanese yen (5.68 billion US dollars) for the eco-point system for home electronics in
total. The economic result was estimated at about 5 trillion Japanese yen (40.98 billion US dollars)
[16]. The leverage is 72.16. Although the economic result and leverage are just estimations in the
abovementioned studies, the eco-point system for housing is apparently small, as we find in our
experiments. Three samples are small but we find no apparent contradiction in the results.
We observe the distributions of avalanche sizes that are stated from specific industries in Figure 4
and the shapes of the tails are similar. This can be interpreted as a certain supply chain always being
used in those large avalanches and the supply chain may lie on particular industries. To examine
the hypothesis, we obtain a different measure from the first experiments. The measure is how often
firms in an industry become involved in avalanches. Note that we examine the avalanche sizes that
start from specific industries thus far and a firm is randomly selected from all firms.
Surprisingly, as is observed in Figure 7, the extent to which a firm becomes involved in avalanches
is sharply different. The sharpness is totally different from that observed in Figure 5. Wholesale
and manufacturing are distinctly large and construction can be included in the largest group. This
result means that firms in those industries are apparently always involved in large avalanches that
start from any industry.
[Figure 7 here]
5
Conclusion
This paper analyzed the cause of the diversity of the spillover effect. We used observed data of
transactions in Japan. For the data, we employed the production model. As a result, we confirmed
the size of the spillover effect triggered by demand follows the power law. Therefore, the normal
distribution, which is usually expected in analyses of input-output tables, cannot be a reliable
assumption. Although we did not use the volume of trade, the results of the simulations show
significant correlation coefficients. Moreover, the simulated avalanche sizes for the policies actually
conducted correspond to the estimation given by the ex-post evaluations of the policies. In addition,
industries have diverse potential to become involved in avalanches.
9
References
[1] W. Easterly and S. Rebelo. Fiscal policy and economic growth. Journal of Monetary Economics,
32:417–458, 1993.
[2] W. Romp and J. de Haan. Public Capital and Economic Growth: A Critical Survey. Perspektiven der Wirtschaftspolitik, 8(S1):6–52, 2007.
[3] W.W. Leontief. Quantitative Input and Output Relations in the Economic Systems of the
United States. The Review of Economics and Statistics, 18(3):105–125, 1936.
[4] X. Gabaix. The Granular Origins of Aggregate Fluctuations. Econometrica, 79(3):733–772,
2011.
[5] D. Acemoglu, V.M. Carvalho, A. Ozdaglar, and T. Alireza. The network origins of aggregate
fluctuations. Econometrica, 80(5):1977–2016, 2012.
[6] P. Bak, K. Chen, J. Scheinkman, and M. Woodford. Aggregate fluctuations from independent
sectoral shocks: self-organized criticality in a model of production and inventory dynamics.
Ricerche Economiche, 47(1):3–30, 1993.
[7] Ministry of Internal Affairs and Communications. The Japan standard industrial classification
(JSIC) summary of development of the JSIC and its eleventh revision, 2013.
[8] Trade Ministry of Economy and Japan Industry. The 2011 updated Input-output table, 2011.
[9] T. Iino and H. Iyetomi. Modeling of Relation between Transaction Network and Production
Activity for Firms. Progress of Theoretical Physics Supplement, 179:134–144, 2009.
[10] N. Zachariou, P. Expert, M. Takayasu, and K. Christensen. Generalised Sandpile Dynamics on
Artificial and Real-World Directed Networks. Plos One, 10(11):e0142685, 2015.
[11] D.J. Watts. Small Worlds: The Dynamics of Networks Between Order and Randomness. Princeton University Press, 1999.
[12] Ministry of Internal Affairs and Communications. Policy evaluation pertaining to special taxation measures (In Japanese), 2015.
[13] D. Shirai.
The eco-vehicle tax breaks cut 0.1% of the total domestic CO2 emissions (In
Japanese). Japan Center for Economic Resarch, 2010.
10
[14] Monthly report of new vehicle registration (In Japanese). http://www.jada.or.jp/.
[15] Mizuho Research Institute. Evaluation of the eco-point system for housing (In Japanese), 2012.
[16] Board of Audit Japan. Report of the eco-point system for home electronics (In Japanese), 2012.
11
Figure 1: Comparison of degree distributions between the random network and observed network: The
horizontal axis shows degree and the vertical axis shows cumulative probability. The blue plots indicate
the random network. The red ones indicate the observed network.
12
Figure 2: Generalized production and inventory model: (a) Scheme of the inventory renewal. Arrows
show the flows of products. Therefore, orders of supplies are opposite. (b) Example of loop avoidance
assumption. There are three firms 1, 2, and 3. Firm 1 asks firm 2 to supply products (given firm 1’s
inventory is not enough for demand.) Firm 2 asks firm 3 to supply products. However, firm 3 does not
take supply from firm 1 and the link is ignored.
13
Figure 3: Comparison of avalanche size distribution between the random network and the observed
network: The horizontal axis shows sizes of avalanches caused by demand. The vertical axis shows the
cumulative probability. The blue plots indicate the random network. The red ones indicate the observed
network. Although there are size zero avalanches, we ignore them because the main aim of the figure is
to show the shapes of the tails and zero cannot be included into log plots.
14
Figure 4: Comparison of avalanche size distribution between industries: The horizontal axis shows sizes
of avalanches caused by demand. There is repeated demand for a firm that is chosen randomly from the
industry. The vertical axis shows the cumulative probability. Although there are size zero avalanches, we
omit them because the aim of the figure is to show the shapes of the tails and zero cannot be included in
log plots.
15
Figure 5: Comparison of average avalanche size starting from specific industry: The horizontal axis lists
the industries. There is repeated demand for a firm that is chosen randomly from the industry. The
vertical axis shows the average of avalanches. The error bars show standard deviations.
16
Figure 6: Simulation for industries to compare fiscal policies: The horizontal axis lists the industries.
They correspond to the target industries of past Japanese fiscal policies: eco-vehicle tax breaks, eco-point
system for housing, and eco-point system for home electronics. The vertical axis shows the average of
avalanches. There is demand for a firm in each industry.
17
Figure 7: Difference of expectations for being involved in avalanches: The horizontal axis lists the
industries. The vertical axis shows the expectation to be involved in avalanches per instance of demand.
The firm with demand is chosen randomly from all the firms. The expectation in each industry is averaged
over the firms in each industry.
18
| 3cs.SY
|
Dynamic graph connectivity with improved worst case update time
and sublinear space
arXiv:1509.06464v1 [cs.DS] 22 Sep 2015
David Gibb
∗
Bruce Kapron
Valerie King
∗
Nolan Thorn
∗
Abstract
This paper considers fully dynamic graph algorithms with both faster worst case update
time and sublinear space. The fully dynamic graph connectivity problem is the following: given
a graph on a fixed set of n nodes, process an online sequence of edge insertions, edge deletions,
and queries of the form “Is there a path between nodes a and b?” In 2013, the first data structure
was presented with worst case time per operation which was polylogarithmic in n. In this paper,
we shave off a factor of log n from that time, to O(log 4 n) per update. For sequences which are
polynomial in length, our algorithm answers queries in O(log n/ log log n) time correctly with
high probability and using O(n log2 n) words (of size log n). This matches the amount of space
used by the most space-efficient graph connectivity streaming algorithm. We also show that
2-edge connectivity can be maintained using O(n log2 n) words with an amortized update time
of O(log6 n).
∗
Department of Computer Science, University of Victoria, BC, Canada; [email protected], [email protected],
[email protected], [email protected]. This research was supported by NSERC and a Google Faculty Research Grant
1
Introduction
In the dynamic connectivity problem, we are given an undirected graph G = (V, E), where V is a
fixed set of n nodes, and an online sequence of updates and queries of the following form:
• Delete(e): Delete edge e from E.
• Insert(e): Insert edge e into E.
• Query(x, y): Is there a path between nodes x and y?
The goal is to process graph updates and queries as efficiently as possible. Our data structure
solves the dynamic connectivity problem using O(log4 ) worst case time per update operation and
O(log n/ log log n) worst case time per query. In addition, if the sequence of updates is polynomial
in length, our data structure needs only a sublinear amount of storage, namely O(n log2 n) words
(of size O(log n)), to ensure that with high probability, all queries are answered correctly. By way
of comparison, the dynamic graph stream problem is to process an online sequence of polynomial
number of edge insertions and deletions so that when the stream ends, a spanning forest is output,
where the goal is to minimize space. Our algorithm is the first sublinear space dynamic graph
algorithm: it matches the space needed by the most space efficient dynamic graph stream algorithm
by Ahn, Guha, and McGregor from 2012 [1], and it shaves a factor of O(log n) off the worst case
time per update of the best previously known dynamic connectivity algorithm by Kapron, King,
and Mountjoy from 2013 [6].
As an application, we show the first sublinear space dynamic 2-edge connectivity algorithm,
which uses O(log6 n) amortized update time, O(log n/ log log n) query time, and O(n log2 n) words.
This compares to the O(m + n log n) word algorithm with amortized update time O(log4 n) and
O(log n/ log log n) query time by Holm, de Lichtenberg, and Thorup from 1998 [5].
The sublinear space algorithm has 2-sided error of 1/nc for any constant c. If we keep a list
of edges with additional space O(m) then there is 1-sided error: if the answer to a query is “yes,”
then the answer is always correct; if it is “no,” then it is correct with probability 1 − 1/nc for c any
constant. In both cases, we assume the adversary knows the edges in the graph, but constructs
the sequence of updates and queries with no knowledge of the random bits used by the algorithm,
other than whether the queries have been answered correctly or not. The edges of the spanning
forest used by the data structure are not revealed to the adversary.
Several dynamic graph problems can easily be reduced to the problem of maintaining connectivity. See [4]. Thus our algorithm yields a dynamic graph algorithm for determining the weight
of a minimum spanning tree in a graph with k different edge weights, where the cost of updates is
increased by a factor of k. It also can be used to maintain the property of bipartiteness with only
constant factor increase in cost. See [1],[4].
An Overview of Techniques: Nearly every dynamic connectivity data structure maintains a spanning
forest F . Dealing with edge insertions is relatively easy. The challenge is to find a replacement
edge when a tree edge is deleted, splitting a tree into two subtrees. A replacement edge is an edge
reconnecting the two subtrees, or, in other words, in the cutset of the cut (T, V \ T ) where T is one
of the subtrees. To simplify notation, where there is no confusion, we allow T to denote the set of
nodes in the tree T . Otherwise we will use the notation v(T ).
We compare the techniques of our paper to the 2013 dynamic graph connectivity [6] and the
2012 dynamic graph stream paper [1]. All three papers use random sampling to identify the
1
name of a edge which crosses a cut. The 2013 dynamic graph paper introduced the cutset data
structure. It maintains a dynamic forest of disjoint trees which is a subgraph of a graph G and, for
each tree, a “leaving” edge, i.e., an edge with exactly one endpoint in the tree. That paper uses
O(m log n + n log2 n) words to implement the cutset data structure and ensures that edges leaving
are identified with high probability. For each edge inserted, it flips fresh coins to determine which
samples it appears in and records these outcomes in a table.
The graph stream paper implicitly uses a restricted cutset data structure which accommodates a
continuous stream of nontree edge insertions and deletions, with one final batch change of nontree
edges to tree edges, at which point the leaving edge from a tree is determined with constant
probability. The streaming paper uses O(n log n) words to implement this restrictive cutset data
structure and O(n log2 n) words overall. It uses hash functions rather than independent random
bits to determine whether an edge is in a sample, eliminating the need for a table to remember the
sampling outcomes. It uses a well-studied technique known as L0 sampling and 1-sparse recovery
to sample, find an edge with constant probability, and then verify with high probability it is an
edge.
Our cutset data structure also determines an edge leaving with constant probability using
O(n log n) words. It can be seen as using L0 sampling and a novel, possibly more practical, method
of 1-sparse recovery.
All three papers use O(log n) tiers of cutset data structures to build a spanning forest, similar
to the Borůvka’s parallel minimum spanning tree algorithm. Each tier contains a forest, and each
forest is a subforest of the tier above it. In the cutset data structure on tier i, the leaving edges
discovered there become tree edges in the forest on tier i + 1, and so on. The streaming paper
needs to form these forests only once at the end of the stream. In worst case, a single insertion or
deletion can cause a superpolylogarithmic number of changes to edges leaving on the various tiers.
Thus maintaining the spanning forest for each update that the streaming algorithm would have
produced if the stream had ended at that update, is too costly in terms of time. By using high
probability to find edges leaving, the 2013 dynamic graph paper avoids this problem by finding an
edge leaving every non-maximal tree in a tier at an extra cost of log n factor in time and space.
Our algorithm avoids the cost of using high probability; it does not necessarily keep the same tree
edges as the streaming algorithm, but it ensures that about the same number of trees are merged
on the next tier. This induces a subtle change to the invariants, which are less restrictive.
Related work: For other related work on dynamic connectivity, see [6]
Organization: In Section 2 we define the cutset data structure. In Section 3, we show how it can
be used to solve dynamic connectivity. In Section 4, we discuss the implementation of cutset data
structures. In Section 5, we give a space efficient dynamic 2-edge connectivity algorithm.
2
Cutset data structure definition and results
Consider a forest F of disjoint trees which are subtrees of a graph G. G and F are dynamic, that
is, they are undergoing updates in the form of edge insertions and deletions. In addition, an edge
in G which links two trees may be updated to become a tree edge which would then cause the
two trees in F to be joined as one. Similarly a tree edge may be changed to a non-tree edge. The
problem is to design a data structure with the goal of maintaining, for each tree T in F which is
non-maximal, an edge in the cutset of (T, V \ T ). A successful Search(T ) returns this edge if the
2
cutset is non-empty and null otherwise. We assume that the graph updates are oblivious to the
(randomized) outcomes of Search. I.e., we may consider the sequence of updates to be fixed in
advance, but updates are revealed to the algorithm one at a time. The operations are:
• Insert edge(x, y): Adds edge {x, y} to E.
• Make tree edge(x, y): For {x, y} ∈ E but not in F , adds {x, y} to F .
• Make nontree edge(x, y): Removes {x, y} from F .
• Delete edge(x, y): Removes {x, y} from E. If {x, y} is an edge of F , removes it from F .
• Search(T ): For T ∈ F , returns null if there are no edges of E in the cutset of (T, V \ T );
returns an edge in th cutset of (T, V \ T ) otherwise.
Lemma 2.1. Let c > 0 be any constant. There is a cutset data structure for graph G with forest F
with worst case update time O(log2 n) and space O(n log n) words, such that for any T ∈ F , where
C is the cutset of (T, V \ T ), if C 6= ∅, then with constant probability p ≥ 1/8 − 1/nc , Search(T )
returns an edge in the cutset; with probability less than 1/nc it returns a non-edge or edge not in
C; otherwise it returns null. If C = ∅, it always returns null.
With an auxilliary list of O(m) words to keep track of the current edge set in the graph, Search(T )
returns an edge in C with probability 1/8 and always returns an edge in C, if it returns an edge.
The implementation of the cutset data structure is discussed in Section 4.
3
Fully dynamic connectivity algorithm
Theorem 3.1. For any constant c and any sequence of polynomially many updates, there is a
dynamic connectivity algorithm which answers all queries correctly with high probability and uses
O(log4 n) time per update and O(n log2 n) words. The error is 2-sided, i.e., ”yes” may be output
when the correct answer is ”no” and vice versa, with probability less than 1/nc .
If a list of current edges in the graph is maintained (with an additional O(m) words), any
sequence of any length of updates may be processed so that with high probability any single query
is correct. In this case, the error is 1-sided; if an answer to a query is “yes” the answer is always
correct and if it is “no”, it may be incorrect with probability 1/nc .
3.1
Tiers and Invariants
Recall Borůvka’s parallel algorithm for building a minimum spanning tree [2]. We may view that
algorithm as constructing a sequence of forests we call tiers. On tier 0 are the set of nodes in the
graph; on tier ℓ + 1 is a subforest of the minimum spanning forest formed when each ℓ − 1 tree
picks a minimum cost edge linking it to another ℓ − 1 tree,
Our algorithm has similar structure. For each tier ℓ = 0, 1, ..., top, where top = O(log n), we
maintain a cutset data structure CDℓ for G with a forest Fℓ and refer to the collection as CD. Each
cutset data structure in CD is generated using independent randomness. Because tier top is not
used to form a forest on a higher tier, the data structure only consists of a forest Ftop , i.e., there is
no randomness array on this tier.
For each tier ℓ < top, Search(T, ℓ) refers to the Search(T ) operation in the cutset data structure
CD ℓ where T ∈ Fℓ ; we similarly extend the definitions of other operations of a cutset data structure
3
for each tier ℓ. Since there is no randomness array on tier top, for Delete edge(e, top) we remove e
from Ftop if it appears there but do nothing else.
As the graph is updated, the cutset data structure for tier ℓ is used to find edges leaving nonmaximal trees (“fragments”) in that tier. Given an edge e, let ℓ be the minimum tier in which e
first appears in Fℓ . We then refer to it as a tier ℓ edge. We maintain the following invariants of
CD. Let G be the current graph.
1. F0 = V , the vertices of G;
2. For every tier ℓ, 0 ≤ ℓ ≤ top − 1, Fℓ ⊆ Fℓ+1 ⊆ G.
3. Let T be a node in tier ℓ < top. If Search(T, ℓ) is successful then v(T ) is properly contained
in some v(T ′ ) for T ′ a node on tier ℓ + 1.
Discussion: Note that these invariants do not necessarily preserve the property that every edge in
the spanning tree Ftop is currently an edge leaving some fragment in some CD. Rather, Invariant
3 ensures that any fragment T in tier ℓ for which Search(T ) is successful has some edge leaving it
in Fℓ+1 . Also, the invariants do not preclude there being a tree edge in Fℓ+1 which is incident to a
fragment T in Fℓ when Search(T ) is not successful. The edges in the forests may be thought of as
“historical” Search edges, edges which were once returned by Search. This gives us the flexibility
to perform updates quickly.
Markov’s inequality implies that with constant probability, a constant fraction of executions of
Search on fragments are successful:
Lemma 3.2. Let p be defined as in Lemma 2.1. In a cutset data structure with fnum fragments,
with probability at least α = (1 − p)/(1 − p/2) the number of fragments T such that Search(T )
succeeds is at least (p/2)fnum .
Proof. If there are fnum fragments in the cutset data structure, then the expected number of
fragments for which search fails to succeed is no greater than µ = (1−p)f since each has probability
at least p of finding a tree edge. By Markov’s Inequality, the probability that (1 − p/2)fnum fail to
find a tree edge is no more than µ/(1 − p/2)fnum = (1 − p)/(1 − p/2).
As any edge can leave at most two node-disjoint fragments, it is easy to see the following:
Observation 1. Let F be a forest containing f fragments and E ′ be a set of edges leaving at least
cf fragments T then the spanning forest of F ∪ E ′ has no more than (1 − c/2)f fragments.
Lemma 3.3. For any graph G and any constant c, let a = ⌈log4/(4−p) n⌉ and top = max{2a/α, 8c ln n/α},
where α = (1 − p)/(1 − p/2). If the invariants hold then with probability 1 − 1/nc , Ftop is a spanning
forest of G.
Proof. Let p be the constant in Lemma 2.1. We call a tier with fnum fragments successful if the
number of fragments T where Search(T ) is successful is at least (p/2)fnum . Let Xℓ = 1 if tier ℓ
is successful and 0 otherwise. Then by Lemma 3.2, P r(Xℓ = 1) ≥ α. For any tier ℓ, if there are
(p/2)fnum fragments with successful searches in a tier ℓ then by Invariant 3, each of these fragments
is properly contained in Fℓ+1 , i.e., the set of tier ℓ + 1 edges includes an edge leaving for each of
these. Then by Observation 1, the number of fragments in Fℓ+1 ≤ (1 − p/4)fnum .
Since (1 − p/4)a n ≤ 1, a successful tiers suffice to bring the number of fragments to 0 (as there
can’t be only 1). If we flip max{2a/α, 8c ln n/α} coins with probability of heads α, the expected
4
number of heads at µ = max{2a, 8c ln n}. The probability that the number of heads is less than
2
2
a = µ/2 is is less than e−(1/2) µ/2 = e−((1/2) (8c ln n)/2) = 1/nc , by a Chernoff bound.
We note that the results for a fixed graph can be extended to hold for poly(n) arbitrary subgraphs, including those generated by a sequence of polynomially many updates:
Corollary 3.4. Let Gu be the graph G after u updates, let c′ , d be any constants, and assume
u ≤ nd . If top = max{2a/α, 8(c′ + d) ln n/α}, then in the cutset data structures CD u for each
′
u is a spanning forest of Gu . This holds even if there is
Gu , with probability 1 − 1/nc every Ftop
dependence in the randomness between any pair of CD u ’s.
Proof. Consider a collection of nd graphs Gi . Set c = c′ + d in Lemma 3.3. By a union bound, the
u fail to be a spanning forest of Gi is no greater than nd (1/nc ) = 1/nc′ .
probability that some Ftop
3.2
Performing updates
We say a fragment in a tier ℓ is isolated if it is equal to a fragment in tier ℓ+1. To efficiently maintain
the invariants as edges are updated, we only check a cutset induced by a tree if the invariants could
have become violated. The invariants are maintained from the bottom up. Changes to a tree may
be needed because (1) a fragment becomes isolated when a tree edge is deleted; (2) for an isolated
fragment T , Search(T ) may become successful, when before it wasn’t or (3) a fragment which is
newly formed is isolated.
Let T (x) denote the tree containing node x in a forest F . If edge {x, y} is inserted, we start at
the lowest tier, checking on each tier ℓ as follows: if T (x) 6= T (y) this may affect Search(T (x), ℓ)
and Search(T (y), ℓ). If T (x) is isolated and Search(T (x), ℓ) is successful, the edge returned by T (x)
is added to Fℓ+1 . Once a tree edge is added it must be added in all higher tiers. If this forms a
cycle, a highest tier edge in the cycle is made into a non-tree edge. This may in turn cause an
isolated fragment T ′ on a higher tier ℓ′ , for which Search(T ′ , ℓ′ ) must be checked. If edge {x, y} is
deleted, it does not affect the results of a search if both x and y are in the same fragment. If they
are not, then calls to Search for both x and y are potentially affected. At this point we proceed as
we do in the case of insertions.
We now give the algorithms for handling updates and queries. The following procedure, which
restores the invariants after an update involving an edge {x, y}, as described above, is used for both
insertions and deletions.
• Initialize: For each tier ℓ = 0, 1, 2, ..., top − 1, initialize a cutset data structure CDℓ where Fℓ
contains n trees consisting of the single vertices in G.
• Insert(e): For ℓ < top, call Insert edge(e, ℓ) to insert e into CD ℓ , and call Refresh(e).
• Delete(e): For ℓ ≤ top, call Delete edge(e, ℓ) to delete e from CD ℓ , and call Refresh(e).
• Query(x, y): If T (x) = T (y) in Ftop , return “yes”, else return “no”.
A proof of the following is given in the Appendix.
Lemma 3.5. Invariants (1), (2) and (3) are maintained by Initialize, Insert, and Delete.
5
Algorithm 1 Refresh({x, y})
1: for ℓ = 0, . . . , top − 1 do
2:
for u ∈ {x, y} do
3:
if T (u) is isolated and Search(T (u), ℓ) returns an edge eu = {a, b} then
4:
if a path between a and b exists on some tier > ℓ then
5:
Let j be the lowest tier on which there is a path
6:
Let e′ be an edge of maximum tier in the path on tier j
7:
Call Insert edge(e′ , k) for each k ≥ j
8:
Call Make tree edge(eu , ℓ′ ) for each ℓ′ , top ≥ ℓ′ > ℓ
9:
end if
10:
end if
11:
end for
12: end for
3.3
Implementation details, running time and space analysis
We briefly review basic data structures used. ET-trees are based on the observation that when
a tree is linked with another tree, the sequence given by the Euler tour of the new tree can be
obtained by a constant number of links and cuts to the sequences of the old trees. The sequence is
stored in a balanced ordered binary tree.
ET-tree properties: If each node in the graph has a value stored in x words of length O(log n) then
the following hold.
1. A tree edge can be deleted or inserted in O(x log n) time.
2. The sum of the values stored in the nodes of a given tree can be returned in O(x) time.
3. A value at a node can be updated in O(x log n) time per word.
4. Given a node v, the name of the tree T (v) containing v can be returned in O(log n) time.
5. The name of the tree can be obtained more quickly if one keeps an ET-tree of degree O(log n)
solely for this purpose. That is, it can be obtained in O(log n/ log log n) if one is willing to
spend O(log2 n) time doing an insertion or deletion of a tree edge.
ST-trees (aka link-cut trees) provide a means of finding the maximum weight edge on a path
between two nodes in a dynamic forest, in O(log n) time per tree edge link, cut, and the find
operation. ST-trees and ET-trees require only O(n ∗ |value|) space where |value| is the size of the
vector or value stored at each node or edge.
Implementation details: Ftop is stored as a degree-lg n ET-tree where no values are associated with
a node, so that queries can be answered in time O(log n/ log log n). Ftop is also stored in an ST-tree
where each tier ℓ tree edges has weight ℓ. An ST tree data structure allows us to find the cycle
edge in O(log n) time.
In the worstPcase, a new tree edge is found in every tier twice and these changes are filtered up,
for a total of 2 ℓ top − ℓ tree edge insertions or deletions:
Lemma 3.6. Let t(n) be the update time for a cutset data structure and s(n) be the space. Then
the running time and space for the fully dynamic connectivity algorithm is O((top2 ∗t(n)+top∗logn)
and the space needed is O(top ∗ s(n) + n).
6
Since top = O(log n), Lemma 3.6 and Lemma 2.1 imply Theorem 3.1.
4
Implementation of the Cutset Data Structure
To illustrate the main idea of our data structure, we will first explain a simplified deterministic
version which works when there is exactly one edge in each cutset.
Let G = (V, E) be the current graph. Each edge {x, y} ∈ G, 0 ≤ x < y < n, has the name
xb · yb where xb and yb are the ⌈lg n⌉ bit vectors containing the binary representations of nodes x
and y respectively and xb · yb denotes the 2⌈lg n⌉ bit vector xb followed by yb . We say that edge
{x, y} is given by xb · yb . Each node will maintain a vector equal to the bitwise XOR (i.e., sum mod
2) of the vectors assigned to each of its incident edges and we use summation notation to denote
bitwise XOR of two or more vectors. For each tree T ∈ F , the ET-tree for T maintains the bitwise
XOR of the vectors stored at T ’s nodes. We use “addition” or “subtraction” of vectors to refer to
the bitwise XOR of the vectors. The operations are implemented as follows:
• Insert edge(x, y): Add its name xb · yb to the vectors at nodes x and y.
• Make tree edge(x, y): Insert {x, y} into F (by linking the ET-trees representing T (x) and
T (y)).
• Make nontree edge(x, y): Delete {x, y} from F by removing it from the ET-tree that contains
it.
• Delete edge(x, y): Subtract vector xb · yb from the vectors at the nodes x and y. If the edge
is in F , remove it from F by calling Make nontree edge(x, y).
• Search(T ): . Let z = z1 z2 ...z2⌈lg n⌉ be the sum of the vectors of nodes in T . If z 6= ~0 then
return z.
Lemma 4.1. For any tree T ∈ F , if there is exactly one edge in the cutset C of the cut (T, V \ T ),
then Search(T ) is always successful. The amount of space used is O(n log n) bits.
Proof. Every edge with two endpoints in a tree T is added to the vectors of two nodes in T and
contributes twice to z, thus it appears not all in the sum z, while every edge with exactly one
endpoint contributes exactly once to z. If the cutset has exactly one edge, z is the name of the
edge.
4.1
Extension to cutsets with more than one edge
We extend our data structure to handle the case of when the cutset of (T, V \ T ) has more than
one edge. We may do this in sublinear space, but then we require the assumption that no edge is
deleted unless it is currently in the graph and the number of updates is bounded by nd for some
constant d. Alternatively, we can keep a list of edges currently in the graph.
We maintain for each node x ∈ V an array indexed by levels i where 0 ≤ i < ⌈2 lg n⌉ =
levelN um. Let si (x) denote the vector on level i for node x. When edge {x, y} is inserted into G,
for each i = 0, 1, ..., levelN um − 1 with probability 1/2i we add xb · yb to both si (x) and si (y). The
idea is that when the size of the cutset is ≈ 2i , with constant probability there is exactly one edge
from P
the cutset and a unique x in T , such that the name of the edge is added to si (x). In this case,
zi = x∈T si (x) is the name of that one edge and Search(T ) is successful.
7
To decide which edges to included in the sums on the various levels without recording each
coinflip, we use a 2-wise independent hash function h mapping edges into [2levelN um ], where for
k ≥ 1, [k] denotes {1, . . . , k}. An edge e is sampled in level i, i = 0, ..., levelN um of the cutset data
structure iff h(e) ≤ 2i .
The following lemma is straightforward and has been observed previously. A proof is in the
Appendix.
Lemma 4.2. Let W be the cutset of (T, V \ T ). With probability 1/8, there is an integer j such
that exactly one e ∈ W hashes to a value in [2j ].
To verify with high probability that zi is actually the name of one edge in the cut rather than
the sum of several names, we make use of an odd hash function 1
We say that a random hash function f : [1, m] → {0, 1} is ǫ-odd, if for any given non-empty set
S ⊆ [1, m], there are an odd number of elements in S which hash to 1 with probability ǫ, that is,
#
"
X
Pr
f (x) = 1 mod 2 ≥ ǫ.
(1)
f
x∈S
We use the construction of [7] to create a (1/8)-odd hash function. Let m ≤ 2w , where w =
2⌈lg n⌉. We pick uniformly at random an odd multiplier k from [1, 2w ] and a threshold t ∈ [1, 2w ].
From these two components, we define f : [1, 2w ] → {0, 1} as follows:
f (x) = 1 if (kx mod 2w ) ≤ t and x 6= 0
= 0 otherwise.
The idea of the verification is as follows: we use a 2-wise independent function to randomly
partition a set into two and test each part using using an odd hash function to determine if both
parts have at least one element. Repeating this O(log n), if all tests fails, then with high probability
there cannot be more than one element
in a set.
n
In particular, letting S = [ 2 ], the algorithm is initialized by choosing the following functions
independently at random.
• hj : S → {0, 1}, j = 1, 2, ..., c lg n. Each hj is a 2-wise independent hash function.
• fj,b : S → {0, 1} for j = 1, 2, ..., c lg n, and b = 0, 1. Each fj,b is an independently random
(1/8)-odd hash function.
The fj,b ’s are used to determine a 2c lg n-bit tag of an edge at a particular level. For each node
x, we maintain an auxilliary vector s′i,j,b (x). The tag of an edge is added (by bitwise exclusive-or)
1
We may view this verification problem as an instance of 1-sparse recovery and it is possible to apply known
techniques for this problem in our setting. In particular [3] proposes an approach which uses polynomial evaluation
modulo a prime p to verify uniqueness forPrecovered values (from a set of size t,) with success probability (1 − t/p).
This method maintains sums of the form i∈S ai z i , where S ⊆ [t], ai is the weight of item i, and z ∈ Zp is randomly
chosen. Using this approach for our purposes, in order for Search to maintain correct values with probability (1−1/nc )
over a sequence of nd edge updates,
we must work over p = ω(nc+d+2 ). Adding an edge would require adding (mod p)
n
i
a term ±z , where 1 ≤ i ≤ 2 . To the best of our knowledge, this requires performing roughly 2 lg n multiplications
of (c + d + 2) lg n-bit numbers, with reduction modulo p, or alternatively, c + d independent parallel repetitions of
the technique with (c + d) primes of size > 4 lg n bits. In contrast, our approach may be done with (c + d) lg n
multiplications of 2 lg n-bit random numbers, with reduction modulo a power of 2.
8
to the vector s′i,j,b (x) when the edge name is added to si (x). For each level i, the sums over the
nodes in
Pa tree of the auxillary vectors are maintained as well as the sums of the vectors. Let
′
= x∈T s′i,j,b (x). The specifics are described below.
zi,j,b
The operations Make tree edge(x, y) and Make nontree edge(x, y) are implemented as in the case
where cutsets have size one. The remaining operations are implemented as follows:
• Insert edge(x, y):
For i = 0, 1, 2, . . . , levelN um − 1
If h(e) ≤ 2i then
–{Add e to sample } add its name xb · yb to value si (x) and si (y) and
For j = 1, ..., c′ lg n and b = 0, 1:
–{Add tag for e:} If hj (e) = b and fj,b (e) = 1 then add 1 to s′i,j,b (x) and s′i,j,b(y).
• Delete edge(x, y):
For i = 0, 1, 2, .., levelN um − 1:
If h(e) ≤ 2i then
–subtract its name xb · yb to value si (x) and si (y) and
–For j = 1, ..., c′ lg n and b = 0, 1:
If hj (e) = b and fj,b (e) = 1 then subtract 1 from s′i,j,b(x) and s′i,j,b (y).
If the edge is in F , call Make nontree edge(x, y) to remove it from the ET-tree containing it.
• Search(T ):
For i = 0, 1, 2, ..., levelN um − 1:
While an edge in the cutset (T, V \ T ) has not been found, test zi as follows:
Let {z 1 , z 2 } be the edge given by the minimum i such that zi 6= 0.
′
′
If there is no pair j such that zi,j,0
= zi,j,1
= 1, then it is an edge in the cutset and returned.
Else null is returned.
Let zi be the sum of the level i vectors for a tree T , and i be the minimum level such that zi 6= 0̄.
Then with constant probability p, zi is the name of an edge in the cut (T, V \ T ). If zi 6= 0 then if
′
′
if zi is not the name of an edge in G, then with high probability, for some pair j, zi,j,0
= zi,j,1
= 1.
If zi 6= 0 there is must be at least one edge in the cut, since otherwise h(e) = 0̄ for all edges
e. If there is more than one edge in the cut, let e and e′ be two such edges. For b = 0, 1, let
Sb = {e | hj (e) = b}. Since each hj is 2-wise independent then for any j, hj (e) 6= hj (e′ ) with
probability 1/2, and therefore with probability 1/2, |S0 | > 0 and |S1 | > 0. Recall that hj , fj,0 and
fj,1 are independently chosen random hash functions. Therefore, given that |S0 | > 0 and |S1 | > 0,
′
′
zi,j,0
= zi,j,1
= 1 with probability (1/2)(1/8)2 = 1/128. The probability that this equality fails
for all j = 1, 2, ..., c(128) lg n pairs is (1 − 1/128)c(128 lg n) which is less than 1/nc for sufficiently
large n. Therefore w.h.p., when Search returns an edge it is an edge in the cut. The probability of
correctness as stated in Lemma 2.1 then follows.
Running time for the cutset data structure
Each hash function requires a constant time to perform the hash. O(log n) hashes are performed
to determine if an edge should be sampled for level i and for computing its tag. Each node in a tree
in the forest of a cutset data structure has a vector of O(log n) words, since there are O(log n) levels
for each level there are two O(log n) bit vectors, si (x) and s′i (x). Maintaining the sum of vectors of in
9
an ET-tree of O(log n) vectors of size O(log n) requires O(log2 n) time. Similarly, Make nontree edge
or Make tree edge requires O(log2 n) time. Search(T ) involves finding the maximum level i with
non-zero zi and checking its tag or determining if zi is on the current edge list and determining the
name of the trees containing these endpoints all take no more than O(log n) time each. Hence we
can conclude that the update time for edge insertion or deletion is O(log2 n) time while the search
time is O(log n). This concludes the proof of Lemma 2.1.
5
2-edge-connectivity
In this section, we treat the dynamic 2-edge connectivity algorithm [5] and two instances of the space
efficient dynamic connectivity data structure presented previously as black boxes, which we denote
by 2-EDGE, CONN1, and CONN2 respectively. All are generated using independent randomness.
Using these black boxes, we will present an algorithm to solve dynamic 2-edge connectivity in
O(n log2 n) words with the same query time and an update time which is within a O(log 2 n) factor
of [5], i.e., O(log6 n) amortized update time and O(log n/ log log n) query time. The idea is to
maintain a certificate for 2-edge connectivity using the dynamic connectivity method shown in the
previous sections, and use that certificate as the input into 2-EDGE. In the algorithm below, We
give some definitions:
• G = (V, E): the dynamic graph
• F1 : A maximal forest of G maintained by CONN1
• G1 = (V, E1 ): The graph with the forest edges removed. E1 = E \ F1
• F2 : A maximal forest of G1 maintained by CONN2
• Gc = (V, Ec ) The graph which is the input to 2-EDGE.
• F : The 2-edge connected forest of Gc maintained by 2-EDGE
We now present algorithms for handling updates and queries:
Insert(e):
1. Insert e into E and E1
{Process updates to F1 }
2. For each edge e′ added to F1 , delete e′ from E1
3. For each edge e′ removed from F1 insert e′ into E1
{Process updates to F1 ∪ F2 }
4. For each edge e′ added to F1 ∪ F2 , insert e′ into Ec
5. For each edge e′ removed from F1 ∪ F2 , remove e′ from Ec
Delete edge e = (u, v) is the same as Insert(e) except for the first step which is replaced by “Delete
e from E and E1 ”.
Query(u, v): Query 2EDGE to see if u and v are 2-edge connected.
Theorem 5.1. W.h.p., fully dynamic 2-edge connectivity can be maintained in O(log6 n) amortized
time per update and O(log n/ log log n) per query using space O(n log2 n) words, over a polynomial
length sequence of updates.
A proof is given in the Appendix.
10
References
[1] Kook Jin Ahn, Sudipto Guha, and Andrew McGregor. Analyzing graph structure via linear
measurements. In Yuval Rabani, editor, SODA, pages 459–467. SIAM, 2012.
[2] Otakar Boru̇vka. O jistem problemu minimalnim (about a certain minimal problem) (in czech,
german summary). Prace mor. pnrodoved., v(3):37–58, 1926.
[3] Graham Cormode and Donatella Firmani. A unifying framework for ℓ 0-sampling algorithms.
Distributed and Parallel Databases, 32(3):315–335, 2014.
[4] Monika R. Henzinger and Valerie King. Randomized fully dynamic graph algorithms with
polylogarithmic time per operation. J. ACM, 46(4):502–516, July 1999.
[5] Jacob Holm, Kristian de Lichtenberg, and Mikkel Thorup. Poly-logarithmic deterministic fullydynamic algorithms for connectivity, minimum spanning tree, 2-edge, and biconnectivity. J.
ACM, 48(4):723–760, 2001.
[6] Bruce M. Kapron, Valerie King, and Ben Mountjoy. Dynamic graph connectivity in polylogarithmic worst case time. In Sanjeev Khanna, editor, Proceedings of the Twenty-Fourth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2013, New Orleans, Louisiana, USA,
January 6-8, 2013, pages 1131–1142. SIAM, 2013.
[7] Mikkel Thorup. Sample(x)=(a*x<=t) is a distinguisher with probability 1/8.
abs/1411.4982, 2014.
CoRR,
Appendix
Additional proofs
Lemma 3.5. Invariants (1), (2) and (3) are maintained by Initialize, Insert, and Delete.
Proof. It is clear that all invariants hold after a call to Initialize. We now consider update operations.
Maintenance of Invariant (1) holds trivially because Refresh never makes an edge into a tree
edge on tier 0.
For Invariant (2), first note that Refresh checks for a cycle and removes an edge from a potential
cycle before inserting an edge which causes a cycle in a tree, and so each Fℓ is a forest. For the
inclusion property Fℓ ⊆ Fℓ+1 , first note that when handling deletions, the same edge is deleted
from every tier, so this does not violate the invariant. For both insertions and deletions, line 8 of
Refresh makes eu a tree edge on all tiers > ℓ, so inclusion is immediately maintained when no cycles
are created. Now suppose edge eu would form a cycle if inserted. Let e′ be an edge of maximal
tier j in this cycle. We first note that j > ℓ + 1, due to the fact that T (u) is isolated. So by the
minimality of j and the maximality of the tier of e′ , it must be the case that e′ is not an edge in
the forest on any tier < j. Therefore, e′ is removed from all forests containing it, and the inclusion
property of Invariant 2 is maintained.
Invariant (3) fails only if there is an isolated fragment T on some tier ℓ for which Search(T, ℓ)
is successful. We call this a bad T . Initially, if we start with an empty graph, there is no bad T .
We assume there is a first update during which a bad T is created and show a contradiction. A
11
bad T is created either if (1) T is newly formed by the update or (2) T pre-exists the update and
Search(T, ℓ) is successful in the current graph G. In (2) there are two subcases: T wasn’t isolated
before the update or T was isolated before the update but Search(T, ℓ) became successful after the
update. Let {x, y} be the updated edge. We first show:
Claim: Before the call to Refresh, the only bad fragment introduced on any tier has the form T (x)
or T (y), and this remains the case during the execution of Refresh.
Proof of Claim: Before the call to Refresh, there are two ways in which this update could create a
bad fragment. The first, which may occur for both insertions and deletions, happens when there is
a fragment T on a tier ℓ for which Search(T, ℓ) did not return an edge before the update, but now
does. This can only be the case if the cut induced by T has changed, and this is only when exactly
one of x or y is a node in T . The second can happen before a Refresh and only for deletions, in
particular when {x, y} is removed. In this case, two new fragments T (x) and T (y) are created and
no other cuts are changed. Hence the claim is proved for the bad fragments created before the call
to Refresh.
Now, we consider what happens during Refresh. During the execution of Refresh, a bad fragment may result when eu is made into a tree edge on some tier j > ℓ on Line 8, causing two
fragments T and T ′ to be joined into one new fragment which may be bad. Since eu is the result of
Search(T (x), ℓ) or Search(T (y), ℓ) for some ℓ, the new fragment contains x or y, and by Invariant 2,
so does every tree on higher tiers created by inserting eu . Note that replacing one edge by another
edge in a cycle does not affect any cut, so the replacement of e′ by eu does not cause the creation
of a new bad fragment. This concludes the proof of the claim.
Given the claim, we now show that after calling Refresh, Invariant (3) holds. The proof is by
induction on the value of ℓ in the for-loop of Refresh. Assume ℓ < top. Suppose that before iteration
ℓ, Invariant 3 holds on all tiers < ℓ or ℓ = 0. This will still be the case after iteration ℓ, since this
iteration only makes changes on tiers > ℓ. Suppose that T (x) ∈ Fℓ is bad. This means that T (x)
is isolated, and Search(T (x), ℓ) returns an edge eu . Since eu is inserted into Fℓ+1 , T (x) in tier ℓ
is no longer bad, and Invariant (3) holds on all tiers ≤ ℓ. The same argument applies to T (y).
Maintenance of Invariant (3) now holds by induction for all ℓ < top.
Lemma 4.2 follows immediately from the following.
Lemma 5.2. With probability 1/8, there is an integer j such that exactly one w ∈ W hashes to a
value in [2j ].
Proof. Let ℓ = n2 . We prove the statement of the lemma for j = ℓ − ⌈lg |W | − 2⌉. Then
12
1/(4|W |) < 2j /2ℓ < 1/(2|W |).
Pr[∃! w ∈ W : h(w) ∈ [2j ]]
h
X
=
Pr[h(w) ∈ [2j ] ∧ ∀w′ ∈ W \ {w} : h(w′ ) 6∈ [2j ]]
w∈W
h
We evaluate the inside of the summation:
= Pr[h(w) ∈ [2j ]] Pr[∀w′ ∈ W \ {w} : h(w′ ) 6∈ [2j ] | h(w) ∈ [2j ]]
h
h
X
≥ Pr[h(w) ∈ [2j ]](1 −
Pr[h(w′ ) ∈ [2j ] | h(w) ∈ [2j ]])
h
h
w ′ ∈W \{w}
By 2-wise independence:
= Pr[h(w) ∈ [2j ]](1 −
h
X
Pr[h(w′ ) ∈ [2j ]])
w ′ ∈W \{w}
h
= (2j /2ℓ · (1 − [(|W | − 1)2j /2ℓ ])
> (1/(4|W |)(1 − [(|W | − 1)/(2|W |)] > 1/(4|W |)(1/2)] = 1/(8|W |)
Then Pr[∃! w ∈ W : h(w) ∈ [2j ]] > 1/8.
h
Theorem 5.1. W.h.p., fully dynamic 2-edge connectivity can be maintained in O(log6 n) amortized
time per update and O(log n/ log log n) per query using space O(n log2 n) words, over a polynomial
length sequence of updates.
Proof. First we prove correctness: Let F be a spanning forest of G and F ′ be a spanning forest
of G \ F . Then F ∪ F ′ is a certificate of 2-edge connectivity, i.e., if G is 2-edge connected then
so is F ∪ F ′ . The correctness then depends on whether F1 and F2 are correctly maintained as the
spanning forests of G and G \ F1 respectively. By the proof of Theorem 3.1, F1 = Ftop maintained
by CONN1 is the spanning forest of G w.h.p. Since CONN2 uses independent randomness from
CONN1, the updates to F1 and therefore the updates of the input graph G \ F1 to CONN2 are
independent of the randomness in CONN2, similarly F2 = Ftop maintained by CONN2 is a spanning
forest of G \ F1 w.h.p. By a union bound, both forests are maintained correctly w.h.p. We note
that 2EDGE is always correct.
We observe from the proof of Theorem 3.1 that a single update to E results in O(log n) updates
to Ftop in CONN1, which in turn result in O(log2 n) updates to EC . Each update to EC has
amortized time O(log4 n) in 2EDGE. Hence the total time per update is O(log4 n) worst case time
in CONN1 plus O(log5 ) worst case time in CONN2 plus O(log6 ) amortized time in 2EDGE, for a
total amortized update time of O(log6 n). As the space used by 2EDGE is O(m + n log n) words
and |Ec | < 2n, and the space used by CONN1 and CONN2 is O(n log2 n) words, the total space
used is O(n log2 n) words.
13
| 8cs.DS
|
arXiv:1711.05900v1 [cs.AI] 16 Nov 2017
Using Noisy Extractions to
Discover Causal Knowledge
Dhanya Sridhar
University of California Santa Cruz
[email protected]
Jay Pujara
Information Sciences Institute
[email protected]
Lise Getoor
University of California Santa Cruz
[email protected]
1 Introduction
Knowledge bases (KB) constructed through information extraction from text play an important role
in query answering and reasoning. Since automatic extraction methods yield results of varying quality, typically, only high-confidence extractions are retained, which ensures precision at the expense
of recall. Consequently, the noisy extractions are prone to propagating false negatives when used for
further reasoning. However, in many problems, empirical observations of entities, or observational
data, are readily available, potentially recovering information when fused with noisy extractions.
For reasoning tasks where both empirical observations and extractions can be obtained, an open and
critical problem is designing methods that exploit both modes of identification.
In this work, we study a particular reasoning task, the problem of discovering causal relationships
between entities, known as causal discovery. There are two contrasting types of approaches to discovering causal knowledge. One approach attempts to identify causal relationships from text using
automatic extraction techniques, while the other approach infers causation from observational data.
For example, prior extraction-based approaches have mined causal links such as regulatory relationships among genes directly from scientific text [9, 12]. However, the extracted links often miss
complex and longer-range patterns that require observational data. On the other hand, given observations alone, extensive work has studied the problem of inferring a network of cause-and-effect
relationships among variables [13, 2]. Observational data such as gene expression measurements
are used to infer causal relationships such as gene regulation [7]. Prior approaches use constraints
to find valid causal orientations from observational data [5, 6, 7, 13, 4]. Although the constraints
offer attractive soundness guarantees, the need for observed measurements of variables remains
costly and prohibitive when experimental data is unpublished. Extractions such as interactions between genes mined directly from text provide a coarse approximation of unseen observational data.
Combining extractions mined from KBs with observed measurements where available to for causal
discovery can alleviate the cost of obtaining experiment-based data.
We propose an approach for fusing noisy extractions with observational data to discover causal
knowledge. We introduce C AUS F USE, a probabilistic model over causal relationships that combines
commonly used constraints over observational data with extractions obtained from a KB. C AUS F USE uses the probabilistic soft logic (PSL) modeling framework to express causal constraints in a
natural logical syntax that flexibly incorporates both observational and KB modes of evidence. As
our main contributions:
1. We introduce the novel problem of combining noisy extractions from a KB with observational data.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
2. We propose a principled approach that uses well-studied causal discovery constraints to
recover long-range patterns and consistent predictions, while cheaply acquired extractions
provide a proxy for unseen observations.
3. We apply our method gene regulatory networks and show the promise of exploiting KB
signals in causal discovery, suggesting a critical, new area of research.
We compare C AUS F USE with a conventional logic-based approach that uses only observational data
to perform causal discovery. We evaluate both methods on transcriptional regulatory networks of
yeast. Our results validate two strengths of our approach: 1) C AUS F USE achieves comparable performance with the well-studied conventional method, suggesting that noisy extractions are useful
approximations for unseen empirical evidence; and 2) global logical constraints over observational
data enforce consistency across predictions and bolster C AUS F USE to perform on par with the competing method. The results suggest promising new directions for integrating knowledge bases in
causal reasoning, potentially mitigating the need for expensive observational data.
2 Background on Logical Causal Discovery
The inputs to traditional causal discovery methods are m independent observations of n variables
V. The problem of causal discovery is to infer a directed acyclic graph (DAG) G ∗ = (V, E) such
that each edge Eij ∈ E corresponds to Vi being a direct cause of Vj , where changing the value of
Vi always changes the value of Vj .
Since graphical model G ∗ encodes conditional independences among V, causal discovery algorithms exploit the mapping between observed independences in the data and paths in G ∗ to specify
constraints on the output. The PC algorithm [13] is a canonical such method that performs independence tests on the observations to rule out invalid causal edges. Constraints over causal graph
structure can also be encoded with logic [5, 6, 7]. In a logical causal discovery system, independence
relations are represented as logical atoms. Logical atoms consist of a predicate symbol pi (·) with i
variable or constant arguments and take boolean or continuous truth values. To avoid confusion with
logical variables, for the remainder of this paper, we refer to V ∈ V as vertices. As inputs to logical
causal discovery, we require the following predicates to represent the outcomes of independence
tests among V:
– D EP(A, B), I NDEP(A, B) refers to statistical (in)dependence between vertices VA and VB
as measured by the independence test VA ⊥
⊥ VB . The conditioning set is the empty set.
– C OND D EP(A, B, S), C OND I NDEP(A, B, S) corresponds to statistical (in)dependence between vertices VA and VB when conditioned on set S ⊂ V \ {VA , VB }. The independence
test VA ⊥
⊥ VB |S is performed.
The outputs are of a logical causal discovery system are represented by the following target predicates:
– C AUSES (A, B) refers to the absence or presence of a causal edge between VA and VB , and
is substituted with all pairs of vertices A, B ∈ V. Finding truth value assignments to these
atoms is the goal of causal discovery.
– A NCESTOR(A, B) corresponds to the absence or presence of an ancestral edge between all
vertices VA and VB , where VA is an ancestor of VB if there is a directed causal path from
VA to VB . We may additionally infer the truth values of ancestral atoms jointly with causal
atoms.
Given the independence tests over V as input, the goal of logical causal discovery is to find consistent assignments to the causal and ancestral output atoms.
3 Using Extractions in Causal Discovery
In the problem of fusing noisy extractions with causal discovery, in addition to the observations, we
are given a set of variables K = {K11 . . . Knn } of evidence from knowledge base (KB) K, where
Kij is an affinity score of the interaction between Vi and Vj based on text extraction.
2
Extending previous logical causal discovery methods, we additionally represent K in the predicate
set with T EXTA DJ(A, B). T EXTA DJ(A, B) corresponds to KAB and denotes the absence or presence of an undirected edge, or adjacency, between VA and VB as extracted from text. Evidence
of adjacencies is critical to inference of C AUSES (A, B). However, adjacencies in standard causal
discovery are inferred from statistical tests alone. In our approach, we replace statistical adjacencies with T EXTA DJ(A, B). The goal of fusing KB evidence in logical causal discovery is to find
maximally satisfying assignments to the unknown causal atoms based on constraints over both independence and text-based signals. In section 4, we present a probabilistic logic approach defining
constraints using statistical and KB evidence.
4 A Probabilistic Approach to Inferring Causal Knowledge
Our approach uses probabilistic soft logic (PSL) [1] to encode constraints for causal discovery. A
key advantage of PSL is exact and efficient MAP inference for finding most probable assignments.
We first review PSL and then present our novel encoding constraints that combine statistical and KB
information.
4.1 Probabilistic Soft Logic
PSL is a probabilistic programming framework where random variables are represented as logical
atoms and dependencies between them are encoded via rules in first-order logic. Logical atoms in
PSL take continuous values and logical satisfaction of the rule is computed using the Lukasiewicz
relaxation of Boolean logic. This relaxation into continuous space allows MAP inference to be
formulated as a convex optimization problem that can be solved efficiently.
Given continuous evidence variables X and unobserved variables Y, PSL defines the following
Markov network, called a hinge-loss Markov random field (HL-MRF), over continuous assignments
to Y:
M
X
1
P (Y = y|X = x) = exp −
wr φr (y, x) ,
(1)
Z
r=1
where Z is a normalization constant, and φr (y, x) = (max{lr (y, x), 0}) is an efficient-to-optimize
hinge-loss feature function that scores configurations of assignments to X and Y as a linear function
lr of the variable assignments.
An HL-MRF is defined by PSL model M = {(R1 , w1 ) . . . (Rm , wm )}, a set of m weighted disjunctions, or rules, where wi is the weight of i-th rule. Rules consist of logical atoms and are called
ground rules if only constants appear in the atoms. To obtain the HL-MRF, we first substitute logical
variables appearing in M with constants from observations, producing M ground rules. We observe
truth values ∈ [0, 1] for a subset of the ground atoms, X and infer values for the remaining unobserved ground atoms, Y. The ground rules and their corresponding weights map to φr and wr . To
derive φr (y, x), the Lukasiewicz relaxation is applied to each ground rule to derive a hinge penalty
function over y for violating the rule. Thus, MAP inference minimizes the weighted rule penalties
to find the minimally violating joint assignment for all the unobserved variables:
m
X
arg min n
wr max{lr (y, x), 0}
y∈[0,1]
r=1
PSL uses the consensus based ADMM algorithm to perform exact MAP inference.
4.2
C AUS F USE
C AUS F USE extends constraints introduced by the PC algorithm [13]. Whereas PC infers adjacencies
from conditional independence tests, C AUS F USE uses text-based adjacency evidence in all causal
constraints. The text-based adjacency evidence bridges domain-knowledge contained in KBs with
statistical tests that propagate causal information.
Figure 1 shows all the rules used in C AUS F USE. The first set of rules follow directly from the
three constraints introduced by PC. We additionally introduce joint rules to induce dependencies
between ancestral and causal structures to propagate consistent predictions. We describe below how
C AUS F USE rules upgrade PC to combine KB and statistical signals for causal discovery.
3
Figure 1: PSL rules for combining statistical tests and KB evidence in causal discovery.
Rule Type
PC-inspired Rules
Joint Rules
Rules
C1) ¬T EXTA DJ (A, B) → ¬C AUSES (A, B)
C2) C AUSES (A, B) → ¬C AUSES (B, A)
C3) T EXTA DJ (A, B) ∧ T EXTA DJ (C, B) ∧ ¬T EXTA DJ (A, C) ∧ C OND D EP (A, C, S) ∧ IN SET (B, S) →
C AUSES (A, B)
C4) T EXTA DJ (A, B) ∧ T EXTA DJ (C, B) ∧ ¬T EXTA DJ (A, C) ∧ C OND D EP (A, C, S) ∧ IN SET (B, S) →
C AUSES (C, B)
C5) C AUSES (A, B) ∧ D EP (A, C) ∧ C OND INDEP (A, C, S) ∧ IN SET (B, S) ∧ T EXTA DJ (B, C) →
C AUSES (B, C)
C6) C AUSES (A, B) ∧ C AUSES (B, C) ∧ T EXTA DJ (A, C) → C AUSES (A, C)
J1) C AUSES (A, B) → A NC (A, B)
J2) ¬A NC (A, B) → ¬C AUSES (A, B)
J3) A NC (A, B) ∧ A NC (B, C) → A NC (A, C)
J4) A NC (A, B) ∧ T EXTA DJ (A, B) → C AUSES (A, B)
J5) T EXTA DJ (A, B) ∧ T EXTA DJ (B, C) ∧ D EP (A, C) ∧ C OND INDEP (A, C, S) ∧ IN SET (B, S) ∧
C AUSES (B, A) ∧ ¬A NC (C, A) → C AUSES (B, C)
PC-inspired Rules PC uses conditional (in)dependence and adjacency to rule out violating causal
orientations. However, in C AUS F USE, all adjacencies are directly mined from a KB. Rule C1 discourages causal edges between vertices that are not adjacent based on evidence in text. Rule C2
penalizes simple cycles between two vertices. Rules C3 and C4 capture the first PC rule and orient
chain Vi − Vj − Vk as Vi → Vj ← Vk , a v-structure, based on independence criteria. Rule C5 orients
path Vi → Vj − Vk as Vi → Vj → Vk to avoid orienting additional v-structures. Rule C6 maps to
the third PC rule, and if Vi → Vj → Vk and Vi − Vk , orients Vi → Vk to avoid a cycle. PC applies
these rules iteratively to fix edges whereas in C AUS F USE, the rules induce dependencies between
causal edges to encourage parsimonious joint inferences.
Joint Rules Joint rules encourage consistency across ancestral and causal predictions through constraints such as transitivity that follow from basic definitions. Rule J1 encodes that causal edges are
also ancestral by definition and rule J2 is the contrapositive that penalizes causal edges to nondescendants. Rule J3 encodes transitivity of ancestral edges, encouraging consistency across predictions. Rule J4 infers causal edges between probable ancestral edges that are adjacent based on
textual evidence. Rule J5 orients chain Vi − Vj − Vk as a diverging path Vi ← Vj → Vk when Vk
is not likely an ancestor of Vi . Joint rules give preference to predicted structures that respect both
ancestral and causal graphs.
In our evaluation, we investigate the implications of using a noisy extraction-based proxy for adjacency and the benefits of joint modeling.
5 Experimental Evaluation
Our experiments investigate the two main claims of our approach:
1. We study whether the noisy extractions are a suitable proxy for latent adjacencies and give
similar performance to a conventional logic-based approach that impute adjacency values
using only observations.
2. We understand the role of joint ancestral and causal rules over observational data in mitigating noise from the extraction-based evidence.
We evaluate C AUS F USE on real-world gene regulatory networks in yeast. We compare against C AU S O BS , the PSL model variant that performs prototypical causal discovery using only observational
data. C AUS O BS replaces T EXTA DJ with S TANDARDA DJ , adjacencies computed from conditional
independence tests.
5.1 Data
Our dataset for evaluation consists of a transcriptional regulatory network across 300 genes in yeast
with simulated gene expression from the DREAM4 challenge [8, 10]. We snowball sample 10
smaller subnetworks of sizes 20 with low Jaccard overlap to perform cross validation. The data contains 210 gene expression measurements simulated from differential equation models of the system.
4
We perform independence tests on the real-valued measurements which are known to contribute
numerous spurious correlations. In addition to the gene expression data, we model domain knowledge based on undirected protein-protein interaction (PPI) edges extracted from the Yeast Genome
Database:
A NC(A, B) ∧ L OCAL PPI (A, B) → C AUSES(A, B)
We obtain text-based affinity scores of interaction between pairs of yeast genes from the STRING
database. STRING finds mentions of gene or protein names across millions of scientific articles and
computes the co-occurrence of mentions between genes. As an additional step, STRING extracts
relations between genes and increases the affinity score if genes are connected by salient terms such
as “binds to” or “phosphorylates.”
5.2 Results
Model Variant
F1
C AUS F USE
C AUS O BS
0.19 ± 0.08
0.20 ± 0.05
C AUS F USES -PC
C AUS O BS -PC
0.17 ± 0.07
0.19 ± 0.05
Table 1: C AUS F USE achieves comparable performance with C AUS O BS, suggesting that noisy extractions can approximate unseen adjacencies. Without joint rules, C AUS F USES -PC shows worse
performance, pointing to the benefit of sophisticated joint modeling in mitigating noisy extractions.
We evaluate C AUS F USE and C AUS O BS using 10-fold cross validation on DREAM4 networks. C AU S O BS uses the same rules as our approach but computes S TANDARDA DJ as ground D EP (A, B)
atoms that never appear in groundings of C OND I NDEP(A, B, S), based on definition.
To evaluate the additional benefit of joint rules, we compare sub-models of C AUS F USE and C AU S O BS run with causal orientation rules only, denoted C AUS F USES -PC and C AUS O BS -PC respectively. Table 1 shows average F1 scores of all model variants for the regulatory network prediction
task on DREAM4.
Noisy Extractions Maintain Performance First, we see comparable performance between C AUS F USE and C AUS O BS, answering our first experimental question on how closely noisy extractions
approximate adjacencies. In table 1, there is no statistically significant difference between the F1
scores of C AUS F USE and C AUS O BS. The comparable performance between C AUS F USE and C AU S O BS suggests that the noisy extractions can substitute observational data computations without significantly degrading performance.
Joint Rules Overcome Noise Our investigation into model variants sheds light on the second
experimental question around how logical rules overcome the noise from extractions. When comparing PC-only variants of each method, C AUS O BS -PC gains over C AUS F USES -PC, suggesting
that sophisticated joint rules are needed to mitigate the noise from KB extractions. The consistency
across predictions encouraged by the joint rules bolsters the extraction-based adjacency signal.
Extractions Yield Higher Precision, Lower Recall To further investigate the extraction evidence
mined from STRING, we compare both S TANDARDA DJ and T EXTA DJ against gold-standard adjacencies, which we obtain from undirected regulatory links. Table 2 shows the average precision and
Adjacency
Precision
Recall
T EXTA DJ
S TANDARDA DJ
0.32
0.27
0.11
0.3
Table 2: Extraction-based adjacencies achieve higher-precision but lower recall, further substantiating need for joint rules in recovering missing causal orientations.
5
recall of each adjacency evidence type across the DREAM4 subnetworks. Interestingly, T EXTA DJ
achieves higher precision than its statistical counterpart. However, S TANDARDA DJ gains over T EX TA DJ in recall. The result further substantiates the benefit of joint modeling in recovering additional
orientations under low-recall inputs. Nonetheless, the comparison points to the need for a deeper
understanding of the role KBs play in causal reasoning.
5.3 Experiment Details
To obtain marginal and conditional (in)dependence tests, we use linear and partial correlations with
Fisher’s Z transformation. We condition on all sets up to size two. We set rule weights for both
PSL models to 5.0 except for rule C2 which is set to 10.0, since it encodes a strong acyclicity
constraint. Both models use an α threshold on the p-value to categorize independence tests as
C OND D EP,D EP or C OND I NDEP,I NDEP. We select α with 10-fold cross validation. We hold out
each subnetwork in turn and use the best average F1 score across the other subnetworks to pick
α ∈ {0.1, 0.05} raised to powers {1, 2, 3, 4, 5}. C AUS O BS selects two different α values for binning
independence tests and computing adjacencies, and C AUS F USE requires a single α for tests only. We
also select rounding thresholds for both PSL models within the same cross-validation
framework.
√
Since α is typically small, we rescale truth values p for C OND I NDEP,I NDEP by 3 p to reduce rightskewness of values. We rescale all STRING affinity scores to be between 0 and 1.
6 Related Work
Our work extends constraint-based methods to causal discovery, most notably the PC algorithm
[13], which first infers adjacencies and maximally orients them using deterministic rules based on
conditional independence. PC only supports external evidence in the form of fixed edges or nonedges. Our work is motivated by recent approaches that cast causal discovery as a SAT instance
over conditional independence statements [6, 7, 5]. SAT-based approaches are based on logical
representations that more readily admit additional constraints and relations from domain knowledge.
However, so far, logical causal discovery methods use external evidence to identify probable edges.
In a separate vein, prior work has extended text-mining to identify regulatory networks and genetic
interactions only from scientific literature [11, 12, 9]. In contrast, our goal is to propose techniques
that leverage both statistical test signals and text evidence. The work most similar to ours combines
gene expression data with evidence mined from knowledge bases to infer gene regulatory networks
[3]. However, the regulatory network inference orients edges using hard-coded knowledge of transcription factors instead of reasoning about causality. In our approach, we propose a principled
causal discovery formulation as the basis of incorporating KB evidence.
7 Discussion and Future Work
In this work, we present an initial approach for reasoning with noisy extraction-based evidence
directly in a logical causal discovery system. We benefit from a flexible logical formulation that supports replacing conventional adjacencies computed from observational data with cheaply obtained
extractions. Our evaluation suggests that the noisy KB-based proxy signal achieves comparable performance to conventional methods. The promising result points to future research in exploiting KBs
for causal reasoning, greatly mitigating the need for costly observational data. We see many directions of future work, including better extraction strategies for mining scientific literature and finding
text-based proxies for additional statistical test signals. KBs could provide ontological constraints
or semantic information useful for causal reasoning. We additionally plan to study knowledge-based
constraints for causal discovery.
References
[1] Stephen H. Bach, Matthias Broecheler, Bert Huang, and Lise Getoor. Hinge-loss markov
random fields and probabilistic soft logic. Journal of Machine Learning Research (JMLR),
2017. To appear.
[2] David Maxwell Chickering. Learning bayesian networks is np-complete. In Learning from
data, pages 121–130. Springer, 1996.
6
[3] Panagiotis Chouvardas, George Kollias, and Christoforos Nikolaou. Inferring active regulatory
networks from gene expression data using a combination of prior knowledge and enrichment
analysis. BMC bioinformatics, 17(5):181, 2016.
[4] Tom Claassen and Tom Heskes. A logical characterization of constraint-based causal discovery.
In UAI, 2011.
[5] Antti Hyttinen, Patrik O Hoyer, Frederick Eberhardt, and Matti Jarvisalo. Discovering cyclic
causal models with latent variables: A general sat-based procedure. In UAI, 2013.
[6] Antti Hyttinen, Frederick Eberhardt, and Matti Järvisalo. Constraint-based causal discovery:
Conflict resolution with answer set programming. In UAI, 2014.
[7] Sara Magliacane, Tom Claassen, and Joris M Mooij. Ancestral causal inference. In NIPS,
2016.
[8] Daniel Marbach, Robert J Prill, Thomas Schaffter, Claudio Mattiussi, Dario Floreano, and Gustavo Stolovitzky. Revealing strengths and weaknesses of methods for gene network inference.
Proceedings of the national academy of sciences, 107:6286–6291, 2010.
[9] Hoifung Poon, Chris Quirk, Charlie DeZiel, and David Heckerman. Literome: Pubmed-scale
genomic knowledge base in the cloud. Bioinformatics, 30(19):2840–2842, 2014.
[10] Robert J Prill, Daniel Marbach, Julio Saez-Rodriguez, Peter K Sorger, Leonidas G Alexopoulos, Xiaowei Xue, Neil D Clarke, Gregoire Altan-Bonnet, and Gustavo Stolovitzky. Towards
a rigorous assessment of systems biology models: the dream3 challenges. PloS one, 5:e9202,
2010.
[11] Carlos Rodríguez-Penagos, Heladia Salgado, Irma Martínez-Flores, and Julio Collado-Vides.
Automatic reconstruction of a bacterial regulatory network using natural language processing.
BMC bioinformatics, 8(1):293, 2007.
[12] Yong-Ling Song and Su-Shing Chen. Text mining biomedical literature for constructing gene
regulatory networks. Interdisciplinary Sciences: Computational Life Sciences, 1(3):179–186,
2009.
[13] Peter Spirtes and Clark Glymour. An algorithm for fast recovery of sparse causal graphs. Social
science computer review, 9:62–72, 1991.
[14] Damian Szklarczyk, John H Morris, Helen Cook, Michael Kuhn, Stefan Wyder, Milan Simonovic, Alberto Santos, Nadezhda T Doncheva, Alexander Roth, Peer Bork, et al. The string
database in 2017: quality-controlled protein–protein association networks, made broadly accessible. Nucleic acids research, 45(D1):D362–D368, 2017.
7
| 2cs.AI
|
Beyond SIFT using Binary Features in Loop Closure Detection
arXiv:1709.05833v1 [cs.CV] 18 Sep 2017
Lei Han1 , Guyue Zhou2 , Lan Xu1 , Lu Fang3
Abstract— In this paper a binary feature based Loop Closure
Detection (LCD) method is proposed, which for the first time
achieves higher precision-recall (PR) performance compared
with state-of-the-art SIFT feature based approaches. The proposed system originates from our previous work Multi-Index
hashing for Loop closure Detection (MILD), which employs
Multi-Index Hashing (MIH) [1] for Approximate Nearest Neighbor (ANN) search of binary features. As the accuracy of MILD
is limited by repeating textures and inaccurate image similarity
measurement, burstiness handling is introduced to solve this
problem and achieves considerable accuracy improvement.
Additionally, a comprehensive theoretical analysis on MIH
used in MILD is conducted to further explore the potentials
of hashing methods for ANN search of binary features from
probabilistic perspective. This analysis provides more freedom
on best parameter choosing in MIH for different application
scenarios. Experiments on popular public datasets show that
the proposed approach achieved the highest accuracy compared
with state-of-the-art while running at 30Hz for databases
containing thousands of images.
I. INTRODUCTION
Loop closure detection serves as a key component for
globally consistent visual SLAM systems [2]. Various approaches have been proposed to address this problem, but
either suffer from low accuracy [3] and perform unstably
under different scenarios, or from low efficiency [4] and
take too many computational resources to find a loop closure
candidate.
Binary feature (ORB [5] or BRISK [6]) based methods [3],
[7] benefit from their low computational complexity and
efficient memory storage requirements, but also suffer from
low precision and instability, showing a wide gap in accuracy
compared with real-valued feature (SIFT [8] or SURF [9])
based approaches [4], [10]. Our previously work MILD [11]
is the first binary feature based method that achieves comparable accuracy performance by employing MIH in ANN
search for binary features instead of conventional methods
that rely on a Bag-Of-Words (BOW) scheme. However, the
accuracy of MILD is still inferior to state-of-the-art realvalued feature based approaches [4], [10].
It is worthwhile to doubt whether the ceiling of binary
feature based LCD approaches has been reached and
whether real-valued algorithms perform better than binary
This work is supported in part by Natural Science Foundation of China
(NSFC) under contract No. 61331015.
1 Lei Han and Lan Xu are with the Robotics Institute, Hong Kong
University of Science and Technology, Clear Water Bay, Hong Kong
{lhanaf,lxuan}@connect.ust.hk
2 Guyue
Zhou
is
with
the
DJI-Innovations,
Shenzhen,
China
[email protected]
3 Lu
Fang is with the Dept. of Automation, Tsinghua University, Beijing,
China [email protected]
methods in terms of accuracy. [12] studied this problem in
the context of 3D reconstruction, and argue that most of
the popular binary features (ORB or BRISK) are inferior
to SIFT based method in terms of reconstruction accuracy
and completeness without a significant better computational
performance.
We would argue that binary or real-valued features have
different characteristics and should be treated differently. To
exploit the potential of binary features in LCD approaches,
the procedure of MILD is re-examined. A binary feature
(ORB) based LCD approach that achieves better accuracy
than SIFT based methods [4], [10] is presented to support
our idea.
One big challenge of MILD in achieving high accuracy is
the inaccurate image similarity measurement. E.g., the two
images shown in Fig. 1 are taken at different places while
sharing a high similarity score. The same elements (windows,
wall tiles) existing in different places tend to cause confusion
and decrease the accuracy. Tracing to its source, we find
that binary features are less discriminative compared with
real-valued features since binary features carry much less
information. The burstiness phenomenon [13] is more severe
for binary features, which leads to a high similarity score for
images taken at different places. Inspired by [13], we propose
burstiness handling procedure to overcome the misleading
cases. No extra computational burden is introduced in this
procedure by employing the specific features of MIH.
While MIH is an important tool for ANN search of binary
features in MILD, the analysis of MIH is incomplete as the
search radius (r) in the substring hashing procedure is limited
to 0 for the lack of immediate probabilistic conclusions. As
shown in [15], multi-probe techniques may lead to better
performance (considering both accuracy and efficiency) for
Locality Sensitive Hashing (LSH) of SIFT features. However,
no theoretical analysis on the performance gain has been
provided for the case of binary features. We propose a more
comprehensive analysis of MIH by allowing arbitrary positive integers r using the inclusion-exclusion-principle [16].
Based on this extension, the effect of Multi-Probe LSH for
binary features is presented clearly as shown in Fig. 3.
Even though MIH is very efficient in filtering out dissimilar features, a considerable part of the candidates selected
using MIH would still be outliers for large datasets, a robust
early termination strategy would reduce the frequency of
memory access as well as Hamming distance calculation
without loss of accuracy.
In summary, contributions of this paper include:
• A more comprehensive theoretical analysis of MIH in
ANN search of binary features is presented by enlarging
(a) The 199th image taken by left (b) The 199th image taken by right
camera
camera
(c) The 300th image taken by left (d) The 300th image taken by right
camera
camera
Fig. 1. Image [14] 199 and 300 are captured at different places, but shared
a high similarity score.
•
•
the search radius r in the substring hashing procedure.
While in [11], r is limited to zero.
The original MILD [11] is improved in both accuracy
and efficiency. Higher accuracy is achieved due to the
dealing with burstiness of similar features, while the
higher efficiency is due to the use of early termination
and omit features selectively, based on probabilistic
analysis. And, more remarkable, the designed burstiness
handling technique is coped with the MIH procedure so
that no extra computational burden is introduced in this
procedure.
Sparse match is proposed for frame matching using
binary features, which is 16 times faster than traverse
search at an average precision of 97%, as shown in
Sec. IV-B.
II. RELATED WORKS
Various approaches have been proposed for loop closure
detection, either using global image signatures [17] or local
image features [18], [14], [3], [7]. While global signatures
are relatively compact and can measure the similarity of two
images efficiently, it is computationally expensive to extract
global signatures, thus unpractical for on-line visual SLAM
applications. Recently, deep-learnt features [19], [20] have
also shown potentials in LCD for the ability of extracting
semantic information from images. In this paper, we will
focus on the local feature based approaches aiming to provide
an efficient and robust loop closure detector for visual SLAM
applications, as local features such as ORB [5] or SIFT [8]
are widely used in visual SLAM systems for frame registration [2], [21] and they have been proven to be accurate and
robust for various scenarios.
Local feature based LCD approaches mainly explore on
two sub-problems: feature classification and image similarity
measurement. Feature classification tries to cluster different
features indicating the same place, while image similarity
measurement reveals the possibility of two frames indicating
the same place based on previous observations.
Most of the previous methods rely on the Bag-of-words
(BOW) scheme for efficiency considerations. In BOW, features are clustered into different centroids (visual words)
using offline/online trained dictionaries. Frames are represented by the histogram of visual words, and the loop
closure likelihood is calculated based on the difference of
visual word histograms. However, at least two well-known
problems exist in such methods: 1. Perceptual Aliasing as
features clustered into the same visual word may indicate
different locations, and 2. They have high complexity for
real-valued features while low accuracy for binary features.
Among these methods, RTAB-MAP [4] achieves the highest
accuracy performance in terms of recall at 100% precision,
but takes 700 ms per detection. [3] first tries binary features
in the BOW scheme called bag of binary words, which is
able to handle 20,000 images in 50 ms with a much lower
accuracy.
Both MILD [11] and [22] argue the inefficiency of BOW
for binary features. Unlike MILD using MIH [1], [23]
to replace BOW, [22] simply project binary features into
real-valued space which is feasible for the BOW scheme.
However, the transformation from Hamming distance to Euclidean distance is highly nonlinear and a training procedure
is required for choosing projection functions, which limits
the range of applications for their proposed system.
Various methods have been proposed to improve the
image similarity measurement including weak-geometriccheck (WGC) [24] and burstiness [13]. [24] improves image retrieval quality by employing the rotation and scale
information of local features. [13] handles the burstiness
phenomenon of visual elements by assigning weights for
features based on their similarity with others. However,
such procedure requires a complex preprocessing to remove
repeating features in each image. In this paper, we will show
that combining the characteristics of MIH, burstiness can be
handled with barely any computational burden.
III. LOOP CLOSURE DETECTION
As noted in [11], our previous work MILD is proposed for
LCD by employing MIH in ANN search of binary features,
which achieves significantly better PR performance compared with other binary feature based LCD approaches [3],
[7]. However, the accuracy of MILD is still limited to repeating textures which is a common phenomenon in artificial or
natural sceneries. The exploration of MIH for ANN search
of binary features in [11] is inadequate, as search radius r
in substring hashing procedure is limited to 0.
For completeness, we will provide a brief review of
MILD firstly. Then, based on the framework of MILD,
the theoretical analysis of MIH for ANN search of binary
features is further explored and extended in Sec. III-B by
long binary feature descriptor fin is divided into m disjoint
substrings as shown in Fig. 2. Only features that fall into
the same hash entry in at least 1 hash table are considered
as nearest neighbor candidates of fin . Let Ωni denote the
collection of nearest neighbor candidates of fin . Thus, image
similarity measurement is approximated as:
P|Fn | P
φ(fin , fjk )
i=1
fjk ∈Ωn
i
0
Φ (In , Ik ) =
.
(3)
|Fn ||Fk |
Fig. 2. Framework of MIH. Binary feature fin is divided into m disjoint
substrings. k-th substring is the hash index of the kth hash table. Image
index n and feature index i are stored in corresponding m entries as
reference for feature fin .
considering multi-probe MIH. Next, a more precise model
for image similarity measurement is provided in Sec. IIIC to improve the accuracy of LCD. Finally, techniques on
early termination are discussed in Sec. III-D to improve the
efficiency of MILD. An assumption throughout this paper is
that the Hamming errors of two feature descriptors are evenly
distributed, which will be used for the efficiency-accuracy
analysis for MIH.
A. Review of MILD
In MILD, LCD is divided into two stages: image similarity
measurement and Bayesian inference. Image similarity is
used to compute the likelihood of two frames as a loop closure while Bayesian inference employs temporal coherency
to get a final probability of loop closure based on image
similarity. In this paper, we focus on improving the image
similarity measurement, which is calculated by binary features directly and hence is more accurate than conventional
methods using BOW representation of images. For a query
p
image Ip , firstly binary features Fp = f1p , f2p , · · · , f|F
are
p|
extracted. The image similarity measurement between Ip and
Iq is denoted as Φ(Ip , Iq ):
P|Fp | P|Fq |
p
q
i=1
j=1 φ(fi , fj )
Φ(Ip , Iq ) =
,
(1)
|Fp ||Fq |
where φ(fip , fjq ) refers to the binary feature similarity, i.e.,
(
exp(−d2 /σ 2 ), d ≤ d0 ,
p
q
φ(fi , fj ) =
(2)
0, d > d0 .
here d denotes the Hamming distance between binary features fip and fjq , σ is the weighting parameter, and d0 is the
pre-defined Hamming distance threshold. Φ(Ip , Iq ) evaluates
the similarity of two images using a voting approach [24].
The intuition behind Eqn. (1) is that if two images can be
registered using binary features in visual SLAM, they tend
to have feature pairs with small Hamming distance which
results in a high similarity score.
Image similarities between Ip and all candidate images
stored in the database are computed using MIH, where a
[11] evaluates the performance of MIH from two aspects:
Complexity E and Accuracy R. Complexity indicates the
ratio between features counted in the candidate set and the
total dataset. Accuracy is measured by the probability of
two features representing the same location encountered in
MIH. Lower E leads to a more efficient image similarity
computation while higher R indicates the approximation is
more accurate.
The probability that a feature pair with Hamming distance
d encountered in MIH is denoted as Precall , which is a
function of d, m and r. m is the number of hash tables and
r is the search radius in the substring hashing procedure.
In [11] only the situations of r = 0 are analyzed.
m!Θ(d, m)
.
(4)
md
Prior statistics of Hamming distance distribution for ORB
features [5] are employed in computing E and R. For
the ORB descriptor, the Hamming distance distribution for
features of the same location (inliers) and different locations
(outliers) are approximated as normal distribution Pi (d) ∈
N (32, 10) and Po (d) ∈ N (128, 20), respectively. Thus, the
complexity and accuracy of MIH can be computed as:
Precall (m, d) = 1 −
R(m)
E(m)
=
L
P
=
d=0
L
P
[Precall (m, d) × Pi (d)],
(5)
[Precall (m, d) × Po (d)].
d=0
Best parameter of MIH (m) is chosen by considering both
R and E. Only features with the same substring of fin are
selected into Ωni in [11]. The performance of MIH when
features with similar substrings fall into the candidate set
will be discussed in Sec. III-B.
B. Multi-Probe MIH
So far, only a few approaches have addressed the problem
of fast search for binary features using hashing technique,
including MIH [23] for exact nearest neighbor search and
LSH [25] for ANN search. While in [11] and this paper, the
statistical information of MIH is employed for approximate
nearest neighbor search.
In LSH, multiple substrings are extracted from the original
binary descriptor independently (e.g., binary elements are
randomly selected from the original feature descriptor). To
improve the search precision and maintain low complexity,
multiple hash table and multi-probe hashing [15] strategies
are employed. Although extensive experiments conducted
in [25] show that the Hierarchical Clustering Tree (HCI)
performs slightly better than LSH, HCI is not considered
in our work. Because as a randomized algorithm, the performance of HCI may be unstable even on the same dataset.
Beside this, for the problems of LCD, random selection of
cluster centers is not possible as features are streamed into
the database, it would be impractical to reorganize the tree
structure every time a new feature enters.
MIH resembles LSH methods for it divides long binary
codes into m short substrings. Unlike LSH [25] tries to make
each substring independent from each other, MIH takes the
inherent correlations in the substrings into consideration. For
MIH, feature pairs with a Hamming distance less than m will
be discovered for sure.
Multi-probe Hashing [15] has shown its ability to improve
the search performance of LSH for real-valued features and
is adopted in [25] for fast match of binary features by
enlarging the substring search radius r. However, in [11],
only the situations where r = 0 are analyzed as no off-theshelf mathematical tools can be found for r > 0. In this
paper, we complete the analysis of MIH in ANN search of
binary features by computing Precall (r, m, d) for arbitrary
nonnegative integer r. The derivation is implemented in
an iterative way, where the probability Precall (r, m, d) is
computed from Precall (r − 1, m, d) following the inclusionexclusion-principle [16].
Precall (r, m, d) is equal to the probability that d independent balls are thrown into m bins randomly, at least one
bin has at most r balls. The situations of r = 0 have been
analyzed in [11]. Without loss of generality, r = 1 equals to
the union of two independent events:
1) At least one bin has no ball.
2) S bins have one ball in each bin, for the rest of the
bins, each bin has at least one ball.
The probability of event 1 (P1 ) equals to Precall (0, m, d)
computed in [11]. Event 2 is the union of Ak , k =
1, 2, · · · , m, where Ak indicates that the kth bin has one
ball, and the rest of the bins have at least one ball. Based on
combinatorial analysis, we have
d(m − 1)d−1 (1 − precall (0, m − 1, d − 1))
. (6)
md
Following the inclusion-exclusion-principle, the probability of event 2 can be computed as:
P (Ak ) =
1
2
3
P2 =CN
P (A1 ) − CN
P (A1 A2 ) + CN
P (A1 A2 A3 )−
N
· · · ± CN
P (A1 A2 · · · AN ).
(7)
Then, Precall (1, m, d) is given by P1 + P2 . Similarly,
Precall (r, m, d) can be computed based on Precall (r −
1, m, d) when r is larger than 1.
Given Precall (r, m, d), we can calculate the complexity and accuracy under different parameter configurations,
R(r, m) and E(r, m), as shown in Fig. 3. It can be concluded
that:
• A larger m will lead to higher accuracy as well as higher
complexity when r is fixed;
• To achieve the same accuracy, it would be more efficient
to use a smaller m and a larger r.
Fig. 3. Complexity and Accuracy under different MIH parameters including
search radius r and hashing table number m. Complexity and Accuracy
grows monotonously with m when r is fixed. m starts with 4 for different
r as shown in this figure.
Additionally, there is a fixed overhead for MIH: the
construction of the hash tables, which is denoted by Ef ixed .
L
Suppose there are m hash tables in MIH, each with 2 m
hash entries (L indicates the feature descriptor length, e.g.,
L
L = 256 for ORB features). Ef ixed = m2 m .
From this analysis, given a required search precision, we
can find the best parameter configurations for MIH to minimize the computational cost including both complexity E
and overhead Ef ixed . E.g., in LCD problems, the minimum
accuracy required is expected to be larger than 0.8, three
parameter candidates for (m, r) include: (15, 0), (10, 1),
(8, 2). For on-line loop closure detection problems, (15, 0)
is more attractive as all the hash entries can be stored in
memory. For applications when the hash tables can be built
off-line, (8, 2) would be a better choice as the complexity of
MIH is minimized and there is no need to consider Ef ixed .
In our implementations for LCD, (16, 0) is chosen as all
the substrings can be represented as a short data structure,
which is more efficient for CPU operations.
C. Burstiness Handling
As discussed in [13], visual elements may appear more
times in an image than a statistically independent model
would predict, which is described as the burstiness phenomenon. Burstiness corrupts the visual similarity measure
in the context of image search. Two types of burstiness exist:
intra-image, when repeating texture exists in the scenery;
and inter-image, when similar objects exist in many places.
Various weighting strategies were proposed to handle the
burstiness phenomenon in [13], while in this paper we
combining the framework of MIH and handling burstiness
without an increase of computational burden.
For intra-image burstiness, typically it may take a timeconsuming preprocessing procedure to detect intra-image
burstiness by a traversal of each feature extracted from the
image [13]. Following the formulations in Sec. III-A, the
probability of independent features falling into the same
hash entry is the complexity E(r, m), which is approximated
to 0. While for repeating features, the probability is the
accuracy R(r, m), close to 100% under our configurations
(r = 0, m = 16). By limiting the number of features falling
into the same hash entry for each image, we can accomplish
the detection of intra-image repeating features without extra
preprocessing procedure.
For inter-image burstiness, we weight the feature similarity
based on the total similarity score of the feature with all
candidates provided by MIH. Given a query feature fin , the
feature similarity measurement is modified as
φ(f n , fj )
,
φ0 (fin , fj ) = log(N/Ni ) P i
n
fk ∈Ω φfi fk
(8)
N
) is the inverse document frequency term, N
where log( N
i
represents the total number of candidate frames, and Ni is
the number of frames that have a similar feature with fin .
Repeating features may cause inefficiency as these features may have little contribution to the image similarity
measurement, but may exist in the candidate set frequently.
Such inefficiency can be solved by limiting the maximum
number of buckets Nbuckets in each hash entry. The average
number of features falling into each hash entry would be:
N
, where Nf eatures is the maximum
Af eatures = f eatures
2ls
number of features stored in the dataset and ls is the substring
length. Nbuckets is set to be 50 × Af eatures by experiment.
Entries that have more buckets than Nbuckets are discarded.
D. Early Termination
The computational complexity of MILD grows linearly
with the number of images stored in the database. For large
datasets, the most time-consuming part of MILD is the
memory access of feature descriptors and Hamming distance
calculation. Given the sparsity of repeating locations, we
have found that a majority of features in the candidate set
are outliers with large Hamming distances. Based on this
observation, we adopt early termination to avoid unnecessary memory access and computational cost for outliers by
exploiting partial information of the feature descriptor. For
the feature fi and its nearest neighbor candidate fj , we
only load the first 64 bits of fj to calculate the partial
Hamming distance hp between fi and fj . If hp is larger
than a threshold, fj is regarded as outliers directly instead
of loading the rest bits of fj . Experiments show that early
termination succeeds to reject around 50% of the outliers
(Hamming distance is larger than 50).
IV. EXPERIMENTS
To evaluate the performance of our proposed approach, we
conduct extensive experiments1 on different datasets2 . Two
individual experiments are implemented: LCD and frame
match. Experiments in LCD reveal the superior performance
of our method on loop closure detection. Experiments in
frame match verify the precision of our proposed fast match
algorithm: sparse match.
1 Experiments are implemented on an Intel-core i7 @ 2.3 GHz processor
with 8 GB RAM. Only one core is used to compare the computational
efficiency of the proposed approach with other algorithms. For loop closure
detection, up to 800 ORB features are extracted for each image using
OpenCV. Feature descriptor length L is set to be 256.
2 NewCollege [14] contains 1073 images of size 1280 × 480. CityCentre [14] contains 1237 images of size 1280 × 480. Lip6Indoor [18] has
388 images of size 192 × 240. Lip6Outdoor [18] has 1063 images of size
192 × 240. BovisaOutdoor [26] contains 2277 images of size 320 × 240.
Fig. 4. Computational complexity of the proposed method, in terms of
running time (ms) for each frames for NewCollege dataset.
Fig. 5.
Precision-recall curve for each dataset.
A. Loop Closure Detection
Accuracy Evaluation: The precision-recall curves of all
the datasets are provided in Fig. 5. The proposed approach
achieves high accuracy on all the datasets provided, including
indoor, outdoor, natural and artificial sceneries. In particular,
we present the results on the NewCollege dataset to demonstrate the reliability of our proposed approach. As shown in
Fig. 6. nearly all the ground truth closures are detected by
the proposed method.
Efficiency Evaluation: The runtime of LCD is composed
by two main parts: tf for feature detection and extraction and
tq for image similarity measurement. The other parts such
as Bayesian inference can be completed efficiently within
1ms, thus can be omitted. tf is fixed given image resolution
and the number of features selected from each image, while
tq grows linearly with the number of frames stored in the
database. For the NewCollege dataset with 1073 frames, the
average runtime of each procedure is tf = 22.5 ms and
tq = 13 ms for the original version of MILD. Based on
the proposed early termination technique, tq can be reduced
to 10 ms without influencing of accuracy. The speed up
factor would be even greater for memory IO inefficient
systems, such as embedded chips or FPGA implementations.
The time cost of LCD for each frame is also presented in
Fig. 4. Although the running time grows linearly with the
number of candidate images, it can efficiently handle loop
closures for datasets containing thousands of key frames as
shown in the experiments, which is more than enough for
visual SLAM designed for VR/AR applications or indoor
navigation systems.
Fig. 7. Accuracy distribution of approximate nearest neighbour search using
LSH and SparseMatch. Three nearest neighbour search methods are used in
image matching: SparseMatch, LSH and brute-force method. Accuracy is
evaluated by the percentage of correctly matched feature pairs in all valid
feature pairs (Hamming distance less than 50).
Fig. 6.
Detected closure VS ground truth closure. The coordinate of
each pixel (i, j) represents the index for candidate image and query image
respectively. Red indicates the detected closure by our approach, Yellow
indicates the ground truth closure. Brown indicates the intersection of the
detected closure and ground truth closure. All the red points are neighbors
of the brown points and cannot be counted as false positives.
RTABMAP [4]
BOWP [10]
DBOW [3]
CNN Feature [20]
IBuILD [7]
To further verify the performance of proposed scheme,
we compare our work with state of the art approaches,
RTABMAP [4] and BOWP [10], which are based on
SIFT/SURF features, as well as DBOW [3], IBuILD [7] and
MILD [11] that use binary features, are used as references
for comparison with the proposed system. The quantitative
comparisons regarding accuracy (recall rate at 100% precision) and runtime of the whole system (including feature
extraction and loop closure detection) are shown in Table IVA. Examining the results presented, the proposed approach
achieves the highest recall rate on nearly all the datasets.
The real-valued (SIFT or SURF) feature based approach
RTABMAP [4] ranks as the second place, while being 20
times slower than our proposed approach. Note that for
the BovisaOutdoor dataset, the proposed approach achieves
much higher accuracy compared with the other approaches,
because similar features (overexposure by the sun glare) exist
in almost every frame, causing confusion in conventional
LCD methods. Benefitting from the handling of burstiness,
the repeating texture has negligible influence on the image
similarity measurement.
B. Frame Matching
Benefit from our analysis on the efficiency-accuracy performance of MIH in ANN search of binary features, sparse
match is proposed to find the corresponding features in
images I1 and I2 . For frame matching applications, overhead
of the database construction must be small and high precision
of the matching results is required. Conventional methods
based on KD-trees [25] are not suitable for such applications
as they require a time-consuming overhead for initialization
which can only work repeating usage situations. Sparse
match is treated as a light version of LCD, where only I1
MILD [11]
Proposed
City
Centre
81%
700ms
86%
441ms
30.6%
20ms
84.8%
155ms
38%
83%
36ms
89.6%
33ms
New
College
89%
700ms
77%
393ms
55.9%
20ms
82.4%
155ms
87.3%
35ms
91.8%
32ms
Lip6
Indoor
98%
100ms
92%
69ms
41.9%
94.5%
8ms
97.8%
8ms
Lip6
Outdoor
95%
400ms
94%
120ms
25.5%
93.4%
9ms
96.3%
9ms
Bovisa
Outdoor
52%
700ms
40%
209ms
6%
20ms
50%
30ms
85.6%
25ms
Table 1. Comparisons with state-of-the-art algorithms in terms of accuracy
(recall rate at 100% precision) and efficiency (average running time perframe).
is stored in the database, and I2 is the query image. To
minimize the overhead Ef ixed and maintain high precision
R, the parameter of MIH used in sparse match is chosen as:
m = 24, r = 0.
Experiments on 1000 image pairs (consecutive images in
the NewCollege dataset, 1200 ORB features are extracted
from each image) are implemented to verify the performance
of MILD in terms of both accuracy and efficiency. In the
experiments, three image match methods are implemented:
sparse match, LSH implemented in FLANN and bruteforce match. In LSH, we choose the same parameters as
SparseMatch (24 hash tables, 10 bits for each key, multiprobe level equal to 0). As shown in Fig. 7, SparseMatch
achieved higher accuracy than LSH. The average processing
time of each image pair for sparse match is 3.3 ms including
all overheads required, while 54.8 ms for brute-force search
and 14.88 ms for LSH. The experiments are implemented on
a lap-top computer and the number of available CPU cores is
limited to 1 to compare the efficiency of different algorithms.
V. CONCLUSIONS
A binary feature based LCD approach is presented in this
paper, which achieves the highest accuracy compared with
state-of-the-art as shown in the experiments while running at
30Hz on a laptop. Higher accuracy is achieved based on the
handling of burstiness, which reduces the confusion caused
by the repeating features. The complexity is also reduced
by filtering outliers based on the partial Hamming distance.
The main bottleneck for efficiency of the proposed system
lies in the extraction of binary features, which takes 70% of
the processing time. Fortunately, such computation can be
shared with binary feature based SLAM systems.
We will keep on improving the performance of the
proposed LCD system and maintain the open-source implementation for the community3 . Currently, the proposed
approach is only suitable for datasets containing thousands
of keyframes. For larger datasets, the superiority of effiency
may decrease as the complexity of MILD increases linearly
with the number of candidates in the dataset. Memory
management schemes such as [4] can be further combined to
enable the proposed algorithm running at constant time for
large scale problems. A better data structure that is suitable
for large hash tables may also be adopted to further improve
the efficiency of the proposed LCD system.
R EFERENCES
[1] Dan Greene, Michal Parnas, and Frances Yao, “Multi-index hashing
for information retrieval,” in Foundations of Computer Science, 1994
Proceedings., 35th Annual Symposium on. IEEE, 1994, pp. 722–731.
[2] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos,
“Orb-slam: a versatile and accurate monocular slam system,” IEEE
Transactions on Robotics, vol. 31, no. 5, pp. 1147–1163, 2015.
[3] Dorian Gálvez-López and Juan D Tardos, “Bags of binary words for
fast place recognition in image sequences,” IEEE Transactions on
Robotics, vol. 28, no. 5, pp. 1188–1197, 2012.
[4] Mathieu Labbe and Francois Michaud, “Appearance-based loop
closure detection for online large-scale and long-term operation,” IEEE
Transactions on Robotics, vol. 29, no. 3, pp. 734–745, 2013.
[5] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski,
“Orb: An efficient alternative to sift or surf,” in 2011 International
conference on computer vision. IEEE, 2011, pp. 2564–2571.
[6] Stefan Leutenegger, Margarita Chli, and Roland Y Siegwart, “Brisk:
Binary robust invariant scalable keypoints,” in 2011 International
conference on computer vision. IEEE, 2011, pp. 2548–2555.
[7] Sheraz Khan and Dirk Wollherr, “Ibuild: Incremental bag of binary
words for appearance based loop closure detection,” in 2015 IEEE
International Conference on Robotics and Automation (ICRA). IEEE,
2015, pp. 5441–5447.
[8] David G Lowe, “Distinctive image features from scale-invariant
keypoints,” International journal of computer vision, vol. 60, no. 2,
pp. 91–110, 2004.
[9] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool, “Surf: Speeded up
robust features,” in European conference on computer vision. Springer,
2006, pp. 404–417.
[10] Nishant Kejriwal, Swagat Kumar, and Tomohiro Shibata, “High
performance loop closure detection using bag of word pairs,” Robotics
and Autonomous Systems, vol. 77, pp. 55–65, 2016.
[11] Lei Han and Lu Fang, “Mild: Multi-index hashing for loop closure
detection,” arXiv preprint arXiv:1702.08780, accepted in Multimedia
and Expo (ICME), 2017 IEEE International Conference on, 2017.
[12] Bin Fan, Qingqun Kong, Wei Sui, Zhiheng Wang, Xinchao Wang,
Shiming Xiang, Chunhong Pan, and Pascal Fua, “Do we need
binary features for 3d reconstruction?,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition Workshops,
2016, pp. 53–62.
[13] Hervé Jégou, Matthijs Douze, and Cordelia Schmid, “On the burstiness
of visual elements,” in Computer Vision and Pattern Recognition,
2009. CVPR 2009. IEEE Conference on. IEEE, 2009, pp. 1169–1176.
[14] Mark Cummins and Paul Newman, “Fab-map: Probabilistic localization and mapping in the space of appearance,” The International
Journal of Robotics Research, vol. 27, no. 6, pp. 647–665, 2008.
3 Code
avaliable from: https://github.com/lhanaf/MILD
[15] Qin Lv, William Josephson, Zhe Wang, Moses Charikar, and Kai Li,
“Multi-probe lsh: efficient indexing for high-dimensional similarity
search,” in Proceedings of the 33rd international conference on Very
large data bases. VLDB Endowment, 2007, pp. 950–961.
[16] Kiyosi Itô, An Introduction to Probability Theory, Cambridge
University Press, 1984.
[17] Relja Arandjelovic and Andrew Zisserman, “All about vlad,” in
Proceedings of the IEEE conference on Computer Vision and Pattern
Recognition, 2013, pp. 1578–1585.
[18] Adrien Angeli, David Filliat, Stéphane Doncieux, and Jean-Arcady
Meyer, “Fast and incremental method for loop-closure detection using
bags of visual words,” IEEE Transactions on Robotics, vol. 24, no. 5,
pp. 1027–1037, 2008.
[19] Silvia Cascianelli, Gabriele Costante, Enrico Bellocchio, Paolo Valigi,
Mario L Fravolini, and Thomas A Ciarfuglia, “Robust visual semisemantic loop closure detection by a covisibility graph and cnn
features,” Robotics and Autonomous Systems, vol. 92, pp. 53–65, 2017.
[20] Yi Hou, Hong Zhang, and Shilin Zhou, “Convolutional neural
network-based image representation for visual loop closure detection,”
in Information and Automation, 2015 IEEE International Conference
on. IEEE, 2015, pp. 2238–2245.
[21] Hauke Strasdat, Local accuracy and global consistency for efficient
visual slam, Ph.D. thesis, Citeseer, 2012.
[22] Simon Lynen, Michael Bosse, Paul Furgale, and Roland Siegwart,
“Placeless place-recognition,” in 2014 2nd International Conference
on 3D Vision. IEEE, 2014, vol. 1, pp. 303–310.
[23] Mohammad Norouzi, Ali Punjani, and David J Fleet, “Fast exact
search in hamming space with multi-index hashing,” IEEE transactions on pattern analysis and machine intelligence, vol. 36, no. 6, pp.
1107–1119, 2014.
[24] Herve Jegou, Matthijs Douze, and Cordelia Schmid, “Hamming
embedding and weak geometric consistency for large scale image
search,” in European conference on computer vision. Springer, 2008,
pp. 304–317.
[25] Marius Muja and David G Lowe, “Fast matching of binary features,”
in Computer and Robot Vision (CRV), 2012 Ninth Conference on.
IEEE, 2012, pp. 404–410.
[26] Simone Ceriani, Giulio Fontana, Alessandro Giusti, Daniele Marzorati,
Matteo Matteucci, Davide Migliore, Davide Rizzi, Domenico G Sorrenti, and Pierluigi Taddei, “Rawseeds ground truth collection systems
for indoor self-localization and mapping,” Autonomous Robots, vol.
27, no. 4, pp. 353, 2009.
| 1cs.CV
|
Parameterized and approximation complexity of the
detection pair problem in graphs
Florent Foucaud∗
Ralf Klasing†
arXiv:1601.05003v2 [cs.DS] 30 Jan 2018
January 31, 2018
Abstract
We study the complexity of the problem Detection Pair. A detection pair of a graph
G is a pair (W, L) of sets of detectors with W ⊆ V (G), the watchers, and L ⊆ V (G), the
listeners, such that for every pair u, v of vertices that are not dominated by a watcher of W ,
there is a listener of L whose distances to u and to v are different. The goal is to minimize
|W | + |L|. This problem generalizes the two classic problems Dominating Set and Metric
Dimension, that correspond to the restrictions L = ∅ and W = ∅, respectively. Detection
Pair was recently introduced by Finbow, Hartnell and Young [A. S. Finbow, B. L. Hartnell
and J. R. Young. The complexity of monitoring a network with both watchers and listeners.
Networks, accepted], who proved it to be NP-complete on trees, a surprising result given that
both Dominating Set and Metric Dimension are known to be linear-time solvable on trees.
It follows from an existing reduction by Hartung and Nichterlein for Metric Dimension that
even on bipartite subcubic graphs of arbitrarily large girth, Detection Pair is NP-hard to
approximate within a sub-logarithmic factor and W[2]-hard (when parameterized by solution
size). We show, using a reduction to Set Cover, that Detection Pair is approximable
within a factor logarithmic in the number of vertices of the input graph. Our two main results
are a linear-time 2-approximation algorithm and an FPT algorithm for Detection Pair on
trees.
Keywords: Graph theory, Detection pair, Metric dimension, Dominating set, Approximation algorithm, Parameterized complexity
1
Introduction
In order to monitor faults in a network, one can place detectors on its nodes. One possibility is to
use “local” devices that are able to detect the location of a fault within distance one (we call them
watchers). Another kind of detectors, that are more far-reaching, are able to determine the exact
distance of a fault to the device, but not its precise location (we call such detectors listeners). If
we wish to monitor a network by using only watchers, we have the classic problem Dominating
Set [14, GT2] (see the books [19, 20] for a survey of results on this problem). If, on the other
hand, we want to use only listeners, we have the problem Metric Dimension [14, GT61] (see
for example the papers [2, 4, 15, 17, 22, 24] and references therein). However, it can be useful to
use both kinds of detectors: for example, one watcher is enough to monitor a complete graph of
order n, while we would need n − 1 listeners to do so. On the other hand, one listener would suffice
to monitor a path of order n, but for the same task we would need ⌈n/3⌉ watchers. Therefore,
Finbow, Hartnell and Young [12] recently proposed the concept of a detection pair in a graph G,
that is, a pair (W, L) where W ⊆ V (G) is a set of watchers and L ⊆ V (G) is a set of listeners that,
together, monitor the graph G.
More formally, we say that a vertex w dominates all the vertices in the closed neighbourhood
N [w] of w (that is, N [w] is the set of neighbours of w together with w itself). Moreover, a vertex
∗ LIMOS - CNRS UMR 6158 - Université Blaise Pascal, Clermont-Ferrand (France).
E-mail: [email protected]
† LaBRI - CNRS UMR 5800 - Université de Bordeaux, F-33400 Talence (France). E-mail: [email protected]
1
l separates two vertices u and v if the distance d(u, l) is different from the distance d(v, l). Given
a pair (W, L) of sets with W ⊆ V (G) (the watchers) and L ⊆ V (G) (the listeners), we say that
a vertex u is distinguished by (W, L) either if u is dominated by a watcher of W , or if for every
other vertex v, either v is dominated by a watcher of W , or u and v are separated by a listener of
L. Then, (W, L) is a detection pair of G if every vertex of G is distinguished by (W, L). The size
of (W, L) is the sum |W | + |L| and is denoted by ||(W, L)||. Note that we may have W ∩ L 6= ∅ if
we choose to place both a listener and a watcher at the same position. We denote by DP (G) the
minimum size of a detection pair of G. Any vertex of (W, L) is called a detector.
If we have a detection pair (∅, L) in a graph G, then L is called a resolving set of G, and
the smallest size of a resolving set of G is called its metric dimension, denoted M D(G). On
the other hand, if we have a detection pair (W, ∅), then W is a dominating set of G, and the
smallest size of a dominating set of G is its domination number, denoted γ(G). Clearly, we have
DP (G) ≤ min{γ(G), M D(G)}, and the inequality can be strict [12]. It follows that for any graph
G without isolated vertices, we have DP (G) ≤ γ(G) ≤ |V (G)|/2 [19].
The goal of this paper is to study the decision and optimization problems that are naturally
associated to the notion of a detection pair (other computational problems that are mentioned in
this paper are defined formally in Section 2).
Detection Pair
Instance: A graph G, a positive integer k.
Question: Is there a detection pair (W, L) of G with ||(W, L)|| ≤ k?
Opt-Detection Pair
Instance: A graph G.
Task: Compute an optimal detection pair of G.
A c-approximation algorithm for a given optimization problem ΠO is an algorithm that returns
a solution whose size is always at most c times the optimum. We refer to the book [1] for more
details. For a decision problem Π and for some parameter p of the instance, an algorithm for
Π is said to be fixed parameter tractable (FPT for short) if it runs in time f (p)nc , where f is a
computable function, n is the input size, and c is a constant. In this paper we will always consider
the solution size k as the parameter. We refer to the books [8, 23] for more details.
Finbow, Hartnell and Young proved that Detection Pair is NP-complete on trees [12], while
Opt-Detection Pair is linear-time solvable on trees containing no pair of leaves with a common
neighbour. This hardness result is quite surprising since the related problems Opt-Dominating
Set and Opt-Metric Dimension, while being NP-hard in general [14, GT2,GT61], can be solved
in linear time on trees (see respectively [5] and [15, 24]). We note that Dominating Set is among
the most classic and well-studied graph problems (see the books [19, 20]), and that Metric
Dimension has enjoyed a lot of interest in the recent years: see the papers [3, 6, 9, 10, 11, 13, 17, 18].
In this paper, we continue the study of the complexity of Detection Pair and Opt-Detection
Pair initiated in [12]. In Section 3, we describe a reduction to Opt-Set Cover that shows that
Opt-Detection Pair can be solved within a factor logarithmic in the size of the input graph. On
the other hand, we observe that a reduction of Hartung and Nichterlein [16, 17] for Opt-Metric
Dimension can also be applied to Opt-Detection Pair. This implies that Opt-Detection
Pair is NP-hard to approximate within a factor that is sublogarithmic in the input graph’s order,
and that Detection Pair is W[2]-hard when parameterized by the solution size k. (These hardness results hold even for graphs that are bipartite, subcubic, and have arbitrarily large girth.) In
Section 4, we prove that Opt-Detection Pair is 2-approximable in linear time on trees. In Section 5, we show that there is an algorithm for Detection Pair running in FPT time 2O(k log k) n2
on trees.1 We start the paper with some preliminary considerations in Section 2, and conclude in
Section 6.
1 In
this paper, “log” denotes the natural logarithm function.
2
2
Preliminaries
We start with some preliminary considerations.
2.1
Definitions of used computational problems
We now formally define a few auxiliary computational problems that are used or mentioned in this
paper.
Dominating Set [14, GT2]
Instance: A graph G, a positive integer k.
Question: Do we have γ(G) ≤ k, that is, is there a dominating set D of G with |D| ≤ k?
Opt-Dominating Set
Instance: A graph G.
Task: Compute an optimal dominating set of G.
Metric Dimension [14, GT61]
Instance: A graph G, a positive integer k.
Question: Do we have M D(G) ≤ k, that is, is there a resolving set R of G with |R| ≤ k?
Opt-Metric Dimension
Instance: A graph G.
Task: Compute an optimal resolving set of G.
Set Cover [14, SP5]
Instance: A hypergraph H = (X, S), a positive integer k.
Question: Is there a set cover C of H with |C| ≤ k (that is, C ⊆ S and each vertex of X belongs
to some set of C)?
Opt-Set Cover
Instance: A hypergraph H.
Task: Compute an optimal set cover of H.
2.2
Specific terminology
In a graph G, a vertex of degree 1 is called a leaf. A vertex is called a branching point if its degree
is at least 3. A branching point v is special if there is a path L starting at v and ending at a vertex
of degree 1, and whose inner-vertices all have degree 2. Path L − v is called a leg of v, and we say
that L is attached to v (note that vertex v does not belong to the leg). A special branching point
with t leaves as neighbours is called a t-stem.2
Given a special branching point x of a tree T , we define the subtree L(x) of T as the tree
containing x and all legs of T attached to x.
2.3
The classic algorithm for Opt-Metric Dimension on trees
We will use the following results of Slater [24] about Opt-Metric Dimension on trees, that are
classic in the literature about metric dimension. See also [15, 22] for similar considerations.
Proposition 1 (Slater [24]). Let T be a tree and R be the set of vertices containing, for each
special branching point x of T that has ℓ legs attached, the leaves of the ℓ − 1 longest of these legs.
Then, R is an optimal resolving set of T .
We have the following consequence of Proposition 1.
Theorem 2 (Slater [24]). Opt-Metric Dimension can be solved optimally in linear time on
trees.
2 The
previous terminology is from [12] (branching points, stems) and [22] (legs).
3
2.4
Observations and lemmas about detection pairs
The following easy observations and lemmas will be useful.
Observation 3. Let G be a graph and B2+ the set of its special branching points with at least two
legs attached. Then, we have DP (G) ≥ |B2+ |.
Proof. Let (W, L) be a detection pair of G. We show that for each special branching point x of
B2+ , there is at least one detector of (W, L) among x and the vertices belonging to a leg attached
to x. Indeed, if not, then any two neighbours of x belonging to this set would neither be dominated
by a watcher, nor separated by a listener, a contradiction.
Lemma 4. Let x be a special branching point of a graph G with a set Sx of t leaf neighbours and
a set Lx of ℓ legs of length at least 2 attached to x. Let (W, L) be a detection pair of G. If there
is a watcher at x, at least ℓ − 1 of the legs in Lx contain a detector of (W, L). Otherwise, at least
t + ℓ − 1 legs attached to x contain a detector.
Proof. If there is a watcher at x, assume for a contradiction that two legs L1 and L2 of Lx contain
no detector. Then, neither the listeners of (W, L) nor the watcher at x can distinguish the two
vertices of L1 and L2 that are at distance 2 from x, a contradiction. A similar argument holds for
vertices at distance 1 of x when x has no watcher.
Lemma 5. Let G be a graph with a special branching point x having a leg attached, whose leaf is
y. If G has a detection pair (W, L) with x ∈ L, then (W, L \ {x} ∪ {y}) is also a detection pair of
G.
Proof. Let Vx be the set of vertices containing x and the vertices of the leg attached to x whose
leaf is y. Clearly, a pair u, v of vertices in V (G) \ Vx are separated by x if and only if they are
separated by y. Moreover, all the vertices in the leg containing y are clearly distinguished, because
each such vertex is the unique one with its distance to y. This completes the proof.
3
General approximability and non-approximability
In this section, we discuss the general approximation complexity of Opt-Detection Pair.
Theorem 6. Opt-Detection Pair can be approximated within a factor 2 log n + 1 on graphs of
order n.
Proof. Given a graph G, we build an instance (X, S) of Opt-Set Cover as follows. We let X be
the set of vertex pairs of G. For each vertex v of G, we have a set Wv of S that contains an element
{x, y} of X if v dominates at least one of x and y. We also have a set Lv that contains an element
{x, y} of X if we have d(v, x) 6= d(v, y). Therefore we have |X| = |V (G)|
and |S| = 2|V (G)|. Now,
2
we claim that we have a one-to-one correspondance between the detection pairs of G and the set
covers of (X, S). Indeed, for every vertex v of G, the set Wv of S corresponds to a watcher placed
at vertex v, and the set Lv corresponds to a listener placed at vertex v. Moreover set Uv of S (with
U ∈ {W, L}) covers exactly the elements of X that correspond to a pair that is separated by the
according detector placed on v in G. Since Opt-Set Cover is approximable within a factor of
log(|X|) + 1 in polynomial time [21], the result follows.
Hartung and Nichterlein [17] provided a reduction from Dominating Set to Metric Dimension. This reduction was improved by Hartung in his PhD thesis [16] to get a reduction mapping
any instance (G, k) to an instance (G′ , k + 4), where G′ is a bipartite graph of maximum degree 3
and has girth 4|V (G)| + 6. In fact, it is not difficult to see that for any detection pair P of G′ of
size k ′ , there is a detection pair P ′ of G′ of size at most k ′ containing no watchers. In other words,
we have DP (G′ ) = M D(G′ ), and therefore the aforementioned reduction is also a reduction to
Detection Pair. Hence we have the following.3
3 Hartung and Nichterlein only state the hardness of approxmation for factors in o(log n), but a recent result
on the inapproximability of Opt-Set Cover [7], that transfers to Opt-Dominating Set via standard reductions,
implies our stronger statement.
4
Theorem 7 (Hartung and Nichterlein [16, 17]). Opt-Detection Pair is NP-hard to approximate
within a factor of (1 − ǫ) log n (for any ǫ > 0) for instances of order n, and Detection Pair is
W[2]-hard (parameterized by the solution size k). Let g ≥ 4 be an arbitrary integer. These results
hold even for instances that are bipartite, have maximum degree 3, and girth at least g.
4
A 2-approximation algorithm on trees
In this section, we prove the following approximability result.
Theorem 8. There is a linear-time 2-approximation algorithm for Opt-Detection Pair on
trees.
Proof. Given an input tree T , we denote by S3+ , the set of vertices of T that are t-stems for some
t ≥ 3. Moreover, we define T ′ as the tree obtained from T as follows. For each vertex v of S3+
(that is, v is a t-stem for some t ≥ 3), assuming that there are ℓ legs of length at least 2 attached
to v in T , we remove from T : (i) t − 2 leaf neighbours of v if ℓ = 0, (ii) t − 1 leaf neighbours of
v if ℓ = 1, and (iii) all t leaf neighbours of v if ℓ ≥ 2. Therefore, in T ′ , there are at least two legs
attached to each vertex of S3+ .
Let us describe our algorithm A.
1. Compute the set S3+ and build T ′ from T .
2. Compute an optimal resolving set R′ of T ′ .
3. Output A(T ) = (S3+ , R′ ).
We first prove that A(T ) is a detection pair of T . Notice that the distances from listeners in
R′ and vertices of V (T ′ ) are the same in T and T ′ . Therefore, all pairs of vertices of V (T ′ ) are
distinguished in T by the listeners of A(T ). On the other hand, all vertices of V (T ) − V (T ′ ) are
dominated by the watchers of S3+ . Therefore, A(T ) is a detection pair of T .
Moreover, it is clear that S3+ and T ′ can be computed in linear time, and R′ can be computed
in linear time using Slater’s algorithm (Theorem 2). Therefore A is a linear-time algorithm.
It remains to prove that ||A(T )|| = |S3+ | + M D(T ′ ) ≤ 2 · DP (T ) for any tree T . Notice that
A(T ) contains only vertices that are part of some subtree L(x), where x is a special branching
point of T ; indeed, any vertex of S3+ is a special branching point of T , and any special branching
point of T ′ is also a special branching point of T (by Proposition 1, any optimal resolving set of T ′
only contains vertices that belong to some subtree L(x) of T ′ ). Therefore, it suffices to prove that
for any detection pair D = (W, L) of T , for any optimal resolving set R′ of T ′ and for every special
branching point x of T , we have |S3+ ∩V (L(x))|+|R′ ∩V (L(x))| ≤ 2|W ∩V (L(x))|+2|L∩V (L(x))|.
To this end, let us assume that x is a t-stem for some t ≥ 0, and that there are ℓ legs of length at
least 2 attached to x. We distinguish the following cases.
Case 1: ℓ ≤ 1. Then, L(x) contains at most two detectors of A(T ). Indeed, if t ≤ 2, it contains
one or two listeners; if t ≥ 3, it contains one watcher and one listener. By Lemma 4, there is at
least one detector in L(x), which completes the proof of this case.
Case 2: ℓ ≥ 2. Then, by Lemma 4, any detection pair (W, L) of T has a detector in at
least ℓ − 1 legs of length at least 2 of L(x). If t ≥ 3 (that is, x ∈ S3+ ), A(T ) has exactly
ℓ − 1 listeners and one watcher in L(x). Therefore, for any detection pair (W, L) of T , we have
2|W ∩ V (L(x))| + 2|L ∩ V (L(x))| ≥ 2ℓ − 2 ≥ ℓ = |S3+ ∩ V (L(x))| + |R′ ∩ V (L(x))| (because ℓ ≥ 2),
and we are done.
Hence, assume that t ≤ 2. Then, L(x) contains only listeners of A(T ), that is, precisely t + ℓ − 1
leaves of L(x) that belong to R′ . If t ≤ 1, we have ℓ−1 ≥ t and then, for any detection pair (W, L) of
T , we have 2|W ∩V (L(x))|+2|L∩V (L(x))| ≥ 2ℓ−2 ≥ t+ℓ−1 = |S3+ ∩V (L(x))|+|R′ ∩V (L(x))|, and
we are done. If however, t = 2, by Lemma 4, there are at least ℓ detectors of (W, L) in L(x). Hence,
we have 2|W ∩V (L(x))|+ 2|L ∩V (L(x))| ≥ 2ℓ ≥ t+ ℓ > t+ ℓ − 1 = |S3+ ∩V (L(x))|+ |R′ ∩V (L(x))|.
This completes the proof.
5
There are two infinite families of trees that show that our algorithm has no better approximation
ratio than 2.
The first family consists of trees Tℓ1 (ℓ ≥ 1) that are built from ℓ disjoint stars with at least three
leaves each, all whose centers are adjacent to an additional single vertex. We have ||A(Tℓ1 )|| = 2ℓ
(this solution set contains one watcher and one listener for every star), while DP (Tℓ1 ) ≤ ℓ (simply
let every center of a star be a watcher, and select no listener). Hence the approximation ratio of
A on Tℓ1 is 2.
A second family consists of trees Tℓ2 (ℓ ≥ 1) built from a path P of ℓ+2 vertices. For each of the
two end-vertices of P , add two leaves that are adjacent to this end-vertex. For each degree 2-vertex
v of P , build a star Sv with three leaves and subdivide one of its edges once. Moreover, the center
of Sv is made adjacent to v. We have ||A(Tℓ2 )|| = 2ℓ + 2 (each star contains two listeners, and
two additional listeners are selected among the four degree 1-vertices adjacent to end-vertices of
P ), while DP (Tℓ2 ) ≤ ℓ + 2 (put a watcher on each center of a star, and add one neighbour of each
2
.
endpoint of P as a listener). The approximation ratio of A on Tℓ2 is 2 − ℓ+2
5
A fixed parameter tractable algorithm on trees
In this section, we provide an exact algorithm for Detection Pair that is FPT for the natural
parameter k, the solution size.
The idea of the algorithm is as follows. We first search for a solution with |L| ≤ 1. After
this step, we may assume that |L| ≥ 2, which is a technical condition required for the subsequent
steps of the algorithm. Then we proceed in three phases. In the first phase, we handle the
solution around the special branching points of the tree. As we will see, there is only a fixed set of
possibilities to try for each such branching point. Then, in the second phase, we determine the set
of remaining listeners. Here, we are able to compute a set (whose size is bounded in terms of k)
of possible vertices that may contain a listener in an optimal solution. Finally, in the third phase,
we determine the set of remaining watchers, and again we are able to compute a set of vertices
(whose size is bounded in terms of k) that may be used as a watcher. The algorithm then checks
the validity of each possible choice of these placements.
Theorem 9. Detection Pair can be solved in time 2O(k log k) n2 on trees of order n.
5.1
Preliminary lemmas
Before describing the algorithm, we prove a series of lemmas that will be essential.
The first lemma is useful to handle special branching points of the tree, by reducing the problem
around these vertices to a fixed number of cases.
Lemma 10. Let G be a graph with a special branching point x with at least two legs and (W, L)
an optimal detection pair of G. Let Sx be the set of leaves adjacent to x, and let Lx be the set of
legs of length at least 2 attached to x. Denote by V (Lx ) the vertices belonging to a leg in Lx , and
let ℓ = |Lx | and t = |Sx | (so, x is a t-stem). Then, we can obtain from (W, L) another optimal
detection pair of G by replacing all the detectors in Vx = {x} ∪ Sx ∪ V (Lx ) by one of the following
sets of detectors:
1. If ℓ ≤ 1, we may use:
(a) a single watcher at x;
(b) a watcher at x, and a listener at the leaf of a longest leg attached to x;
(c) if t + ℓ = 2, a single listener at the leaf of a longest leg attached to x.
2. If ℓ ≥ 2, we may use:
(a) a watcher at x and ℓ listeners at all leaves of the legs of Lx ;
(b) a watcher at x and ℓ − 1 listeners at all leaves of the legs of Lx but the shortest one;
(c) if t ≤ 1, ℓ + t listeners at all leaves of the legs attached to x;
6
(d) if t ≤ 1, ℓ + t − 1 listeners at all leaves of the legs attached to x, except for a shortest
such leg.
Proof. Clearly, if there are any watchers among {x} ∪ Sx, we may replace them by a single watcher
at x and obtain a valid detection pair. We now distinguish two cases.
Case 1: ℓ ≤ 1. Then, we have t ≥ 1. If Vx contained a unique detector of (W, L) that is a
watcher, then it must have dominated a leaf neighbour of x (since t ≥ 1, such a neighbour exists).
Then as observed at the beginning of the proof, we may replace this detector by a watcher at x,
and we are in Case 1(a). If Vx had a unique detector of (W, L) that is a listener, then by Lemma 4
we must have had t + ℓ = 1. Note that any listener in Vx seperates the same set of pairs of vertices
from V (G) \ Vx . Hence we may replace the existing listener by a listener at the leaf of a longest
leg attached to x, which separates at least as many vertices as any other listener in Vx . Then we
are in Case 1(c) and we are done. Hence, we assume that Vx contained at least two detectors of
(W, L). Then, using similar arguments one can check that the solution of Case 1(b) yields a valid
detection pair of G that is not larger than (W, L).
Case 2: ℓ ≥ 2. As observed before, we can replace all watchers in {x} ∪ Sx by a single watcher
+
at x. If x has a watcher, we let L+
x = Lx ; otherwise, we let Lx be the set of all legs attached to x.
+
+
Let ℓ = |Lx |.
By Lemma 4, at least ℓ+ − 1 ≥ 1 of the legs in L+
x contain a detector of (W, L). Assume
first that (W, L) contained at least ℓ+ detectors of (W, L) on the legs of L+
x . Then, we replace
all such detectors by a listener at the leaf of each leg of L+
.
Then,
any
vertex
of a leg Li of L+
x
x
is distinguished by its distance to the listener placed at the leaf of Li . Since any listener in Vx
separates the same set of pairs in V (G) \ Vx , we have obtained a valid detection pair that is not
larger than (W, L) and we are in Case 2(a) or 2(c).
Suppose now that there are exactly ℓ+ − 1 detectors of (W, L) on the legs of L+
x . We replace
these detectors by a listener at the leaf of each of the ℓ+ −1 longest legs of L+
.
By
similar
arguments
x
as above, all vertices in Vx that do not belong to the unique leg L1 of L+
without
a
detector
are
x
distinguished. Again, any listener in Vx separates the same set of pairs of V (G) \ Vx . Hence, if
some vertex is not distinguished by the new detection pair, it must be a vertex u of L1 that is not
separated from a vertex v of V (G) \ Vx . This implies that d(u, x) = d(v, x) (otherwise u and v
would be separated by any of the listeners placed in Vx ). We know that u and v were separated
by some detector of (W, L), which must have been a detector in the leg L1 (no detector in another
leg of L+
x could possibly separate u and v since d(u, x) = d(v, x)). But we know that there was
+
one leg L2 of L+
x that contained no detector of (W, L). Since L1 is a shortest leg of Lx , there is
′
′
′
a vertex u in L2 with d(x, u) = d(x, u ). But then, u and v were not distinguished by (W, L), a
contradiction. Hence we have a valid detection pair whose size is not greater than ||(W, L)|| and
we are in Case 2(b) or 2(d), and we are done.
The next two lemmas are used in our algorithm to choose the placement of the listeners.
Lemma 11. Let G be a graph with two vertices u and v connected by a unique path P . Let V (P + )
be the set of vertices of P together with the vertices of each leg attached to an inner-vertex of
degree 3 of P . Then, by placing listeners on u and on v, all vertices of V (P + ) are distinguished
(by u and v).
Proof. For every vertex x of G, we let δu,v (x) = d(u, x) − d(v, x). Note that if for two vertices x
and y, δu,v (x) 6= δu,v (y), then x and y are separated by u and v.
Clearly, u and v are distinguished. Let w be a vertex of V (P + ) \ {u, v}, and assume for
a contradiction that there is some vertex z with d(u, w) = d(u, z) and d(v, w) = d(v, z). In
particular, δu,v (w) = δu,v (z).
Note that for every two vertices p, q of P , we have δu,v (p) 6= δu,v (q), therefore at most one of
the vertices w and z can belong to P . We assume w.l.o.g. that z does not belong to P .
The shortest paths from z to both u and v must go through an inner-vertex of P , say s. Then,
we have δu,v (z) = δu,v (s) = δu,v (w), which implies that all shortest paths from w to u and v also
go through s. But then, we have d(w, s) = d(z, s), a contradiction because by the structure of
V (P + ) this implies that w = z.
7
Lemma 12. Let G be a graph with two vertices u and v connected by a unique path P , such that
each inner-vertex of P has either degree 2 or a single leg attached. Moreover let V (P + ) be the set
of inner-vertices of P together with the vertices of each leg attached to an inner-vertex of degree 3
of P . Then, for every detection pair (W, L) of G, one can obtain a valid detection pair of G by
replacing all detectors of (W ∪ L) ∩ V (P + ) by two listeners at u and v.
Proof. By Lemma 11, all vertices in V (P + ) are distinguished by u and v. Note that no watcher
of V (P + ) could possibly distinguish any vertex outside of V (P + ) ∪ {u, v} (and u, v are clearly
distinguished). Furthermore, since P is the unique path connecting u and v, for any listener l ∈ L
in V (P + ), the set of pairs in V (G) \ V (P + ) separated by l is the same as the set of pairs separated
by u (and by v). This completes the proof.
The next lemma will be used in the algorithm to determine the placement of watchers in parts
of the tree where there is no listener.
Lemma 13. Let T be a tree with a vertex x and an optimal detection pair (W, L) of T with |L| ≥ 1.
Let Tx be a subtree of T containing x and assume that (i) Tx contains no t-stems for any t ≥ 2,
(ii) for each listener l of L and each vertex v of Tx , the shortest path from l to v goes through x (in
particular Tx − x contains no listener) and (iii) Tx is maximal with respect to (i) and (ii). Then,
there is a set X of size at most 63|W |2 with W ∩ V (Tx ) ⊆ X. Moreover, X can be determined in
time linear in |V (Tx )|.
Proof. We may view Tx as a rooted tree whose root is x. Consider the partition P of V (Tx )
into parts P0 , P1 , . . . , Pd where Pi contains the vertices of V (Tx ) that have distance i to x. Since
|L| ≥ 1 and by (ii) all shortest paths from any listener of L to a vertex of Tx go through x, the
set L of listeners separates a pair u, v of V (Tx ) if and only if u and v belong to different parts of
P. Therefore, since (W, L) is a detection pair of G, for any two distinct vertices u and v from the
same part Pi , at least one of u, v is dominated by a watcher of W .
Let S2+ be the set of parts of P of size at least 2 (that is, those parts where at least one vertex
fx be the forest induced by the set of vertices of Tx that
must be dominated by a watcher). Let F
fx ), and we shall
belong to some part of S2+ , together with their neighbours in Tx . We let X = V (F
prove that X satisfies the claim. First, we show that (W ∩ V (Tx )) ⊆ X. Assume by contradiction
that this is not the case, and let w be a watcher of W that belongs to V (Tx ) \ X. By the optimality
of (W, L), (W \ {w}, L) is not a detection pair of T . Therefore, there must be some vertex u
dominated by w that is not separated by L from some other vertex v (and v is not dominated by
a watcher of W \ {w}). In other words, for every listener l ∈ L, we have dT (u, l) = dT (v, l). Thus,
since T is a tree, all shortest paths from l to both u and v have a common part (say from l until
some vertex y) and then are disjoint. By (iii), Tx is maximal with respect to (i) and (ii), which
implies that y = x and both u and v belong to Tx . However, observe that u belongs to a part
of S2+ (otherwise w would belong to X). Therefore, u is uniquely determined (within Tx ) by its
distance to x. Since |L| ≥ 1 and by (ii), this implies that u and v are in fact distinguished by L, a
contradiction.
Thus, we have shown that W ∩ V (Tx ) ⊆ X. It is moreover clear that X can be computed in
time linear in |V (Tx )| by using the distances of the vertices of Tx to x. Therefore, it remains only
to bound the size of X.
fx contains vertices of at most 9|W | distinct parts of P. To each part Pi
First, we show that F
fx but that may
of S2+ , we can associate the two parts Pi−1 and Pi+1 , that may have vertices in F
f
not be in S2+ . Hence, the total number of parts with a vertex in Fx is at most 3|S2+ |. Since a
watcher may dominate vertices from at most three parts of S2+ and each part in S2+ has at least
fx contains vertices of at most
one vertex dominated by a watcher, we have |S2+ | ≤ 3|W |, hence F
9|W | distinct parts of P.
fx has at most |W | + 2|S2+ | ≤ 7|W | leaves. Note that a leaf of F
fx is
Second, we prove that F
either an actual leaf in T , or it is a vertex of a part Pi with |Pi | = 1 (or both). We claim that
fx can contain at most |W | leaves of T belonging to a part of S2+ . Indeed, each such leaf must
F
be dominated by a watcher, but it can be dominated only by itself or by its parent in Tx . But
since by (i) there are no t-stems for t ≥ 2 in Tx , no parent can dominate more than one leaf, which
fx contains vertices from at most 2|S2+ | parts of P of size 1,
proves the bound. Furthermore, F
8
fx that are not in a part of |S2+ | is at most 2|S2+ | ≤ 6|W | since
hence the number of leaves of F
fx .
|S2+ | ≤ 3|W |. Hence in total, we have at most 7|W | leaves in F
But any rooted forest with at most 9|W | levels and with at most 7|W | leaves may have at most
fx ) has size at most 63|W |2 . This completes the proof.
63|W |2 vertices, hence X = V (F
5.2
The algorithm (proof of Theorem 9)
We now describe the claimed FPT algorithm for Detection Pair on trees and prove its correctness
and running time.
Proof of Theorem 9. Let us describe the algorithm, that seeks to build a detection pair (W, L)
with at most k detectors of an input tree T of order n.
Preliminary phase: Searching for a solution with |L| ≤ 1. A first step of the algorithm is to
check whether there is a solution with |L| ≤ 1. If we set L = ∅, the problem is equivalent to finding
a dominating set of T of size at most k. This can be done in linear time [5]. If we set |L| = 1, one
may try the n possibilities for the placement of the unique listener. For each possibility (say we
have placed the listener at vertex v), we create a subtree Tv of T rooted at v as follows. First of all,
let Tv = T . Then, for each t-stem x of T with t ≥ 2 (except if v is a leaf attached to x and x is a
2-stem), we may assume that x belongs to W (by a similar argument as in the proof of Lemma 10,
Cases 1a and 2a.). Thus, we add x to W and we we remove from Tv all leaves but one that are
attached to x (if v is one such leaf, we keep v in Tv ). Now Tv has no t-stems for t ≥ 2 and we can
apply Lemma 13. Thus, there is a set X of at most 63k 2 vertices that may contain a watcher (X
is computable in linear time). Then, it suffices to try all the possibilities of selecting k − |W | − 1
2
additional watchers from this set. There are at most 63k
∈ O 2O(k log k) such possibilities. For
k
each of them, we check in linear time whether we have a valid detection pair. In total this phase
takes 2O(k log k) n2 time. If we find a valid solution in this phase, we return YES.
Next, we assume that |L| ≥ 2 and proceed in three phases.
Phase 1: Handling the special branching points. We now do a preprocessing step using
Lemma 10. First compute (in linear time) the set of special branching points of T , and let B2+
be the set of those special branching points that have at least two legs attached. Note that by
Observation 3, if |B2+ | > k we can return NO. Hence, assume that |B2+ | ≤ k. Now, for each special
branching point x of B2+ , by Lemma 10 we can assume that there are at most four different choices
for the set of detectors on x and the vertices belonging to a leg attached to x. Therefore, we may
go through each of the possible combinations of these choices; there are at most 4k of them. Of
course we discard the choices for which there are more than k detectors. For a combination Ci ,
Let (Wi , Li ) be the partial detection pair corresponding to Ci . In the remainder, we show how to
decide whether there is a detection pair (W, L) of T with Wi ⊆ W , Li ⊆ L and ||(W, L)|| ≤ k.
Phase 2: determining the set of listeners. Let Vi∗ be the set of vertices of T containing, for
each pair u, v of listeners of Li , all the vertices of the path Puv connecting u and v, as well as all
the vertices of each leg of T attached to a degree 3 inner-vertex of Puv . By Lemma 11, all the
vertices in Vi∗ are distinguished by (Wi , Li ). Clearly, the subgraph Ti∗ = T [Vi∗ ] of T is a connected
subtree. Let X = {x1 , . . . , xp } be the set of vertices in Vi∗ that have a neighbour in T − Ti∗ . For
each vertex xj of X, we let Txj be the subtree of T formed by xj together with all trees of T − Ti∗
containing a neighbour of xj . Let F be the forest consisting of all trees Txj (1 ≤ j ≤ p) such that
at least one vertex of Txj is not distinguished by (Wi , Li ). Next, we upper-bound the size of F .
Claim 9.A. If there is a detection pair (W, L) of T with Wi ⊆ W , Li ⊆ L and ||(W, L)|| ≤ k,
then |F | ≤ k.
Proof of claim. We prove that each tree of F must contain a detector of (W, L). Assume for
contradiction that there is a tree Txa of F such that no detector of (W, L) belongs to Txa . By
definition, Txa contains a vertex w not distinguished by (Wi , Li ), hence there is a vertex w′ of
T − Ti∗ (say, w′ ∈ Txb ) not separated by any listener of Li and none of w, w′ is dominated by a
watcher of Wi . Then, w′ must belong to Txa (hence a = b), indeed by definition of Ti∗ there are
two listeners u and v of Li whose path contains xa and xb . But then, if a 6= b, clearly u and v
9
would separate w and w′ , a contradiction. But since w and w′ both belong to Txa , no listener
outside of Txa can separate them, and clearly no watcher outside of Txa can dominate them, which
proves the claim.
By Claim 9.A, if we have |F | > k, we can discard the combination Ci ; hence, we assume that
|F | ≤ k.
d
Now, from each tree Txj of F , we build the subtree T
xj of Txj as follows: (i) remove all legs in
Txj whose closest special branching point has only one leg attached; (ii) remove each leaf of Txj
d
that is adjacent to a watcher of Wi . We also denote by Fb the forest containing each tree T
xa with
Txa ∈ F . We next claim the following.
Claim 9.B. If there is a detection pair (W, L) of T with Wi ⊆ W , Li ⊆ L and ka detectors on
d
Txa , then T
xa has at most 2ka leaves.
d
Proof of claim. Let us consider the legs of T
xa containing no detector of (Wi , Li ). By (i) of the
d
construction of Txa , such a leg must be attached to a special branching point having at least two
legs attached. (Note that operation (i) did not create any new special branching point.) Moreover,
d
by (ii), if this leg has length 1 in T
xa , its unique vertex will not be dominated by a watcher, since
the set of detectors on this leg and its special branching point are determined by the combination
Ci . Note that in (ii), if we create a new leg, this leg contains a detector (a watcher).
d
d
Let v be a special branching point of T
xa and let Lv be the set of legs in Txa attached to v that
do not contain any detector of (Wi , Li ). By the previous discussion, we have |Lv | ≥ 2. If some leg
of Lv has length 1, as we observed there will be no watcher on v. Hence, in order to separate all
the neighbours of v on the legs of Lv , we need a detector in at least |Lv | − 1 ≥ |Lv |/2 of these legs.
Otherwise, we still need that many detectors in order to separate all the vertices at distance 2 of v
d
on the legs of Lv . Hence, at most half of the leaves of T
xa belong to a leg with no detector, which
proves the claim.
By Claim 9.B, d
Txj has at most 2kj leaves (where kj is the set of detectors placed on Txj ).
Therefore, if the total number of leaves in the forest Fb is more than 2k, we can discard the current
combination Ci . Hence, we can assume that the total number of leaves in Fb is at most 2k.
Therefore, the total number of vertices of degree at least 3 in Fb is at most 2k. Let us consider
the sets of maximal threads in Fb (that is, paths with inner-vertices of degree 2). The endpoints of
such threads are vertices of degree 1 or at least 3 in Fb, and therefore the number of such threads
is at most the number of these vertices, which is upper-bounded by 4k.
b
d
Each such thread Puv between two vertices u and v of a tree T
xj of F, corresponds (in Txj )
to a path whose inner-vertices either have a single leg attached, or contain a watcher of Wi with
a set of leaves attached. Hence, we may assume that Puv contains at most two listeners: if it
contained more than two, by Lemma 12, we could replace them with the two endpoints u and v
of Puv . Moreover, if Puv contained exactly one listener, similarly we may assume it is placed on
one of the endpoints of Puv . Indeed, since |L| ≥ 2, there will be another listener in the solution
set. If this listener is closer to u, similarly as in Lemma 11, a listener on v would distinguish all
+
+
) contains all vertices of Puv and the vertices of legs attached to
vertices on V (Puv
) (where V (Puv
+
) separated by u and
inner-vertices of Puv ). Moreover, the set of pairs of vertices outside of V (Puv
by v are the same. Note that by Lemma 5, if we choose to place a listener on an endpoint u of
d
a path Puv that is a thread in T
xj and this endpoint has a leg attached, instead of placing the
listener on u we can instead place the listener on the leaf of this leg.
d
Therefore, for each thread Puv of T
xa , there are four possibilities for the placement of the
listeners: a listener on both u and v (or the leaf of a corresponding leg attached to u or v); a single
listener at u; a single listener at v; no listener. There are therefore at most 44k possibilities to guess
the placement of listeners on tree Txa . For each such placement Pj , we obtain a set Lj of listeners
(of course, if ||(W, Li ∪ Lj )|| > k we do not consider this placement). Now, we can assume that
L = Li ∪ Lj , that is, that there will be no more listeners in the sought solution (W, L). Hence, it
remains to check whether we can obtain a valid solution of size k by only adding watchers.
∗
Phase 3: Determining the set of watchers. We compute a tree Ti,j
similarly as Ti∗ but using
′
∗
the new set Li ∪ Lj of listeners. Similarly as before, we define the set X of vertices of Ti,j
having
10
∗
a neighbour in T − Ti,j
, and the set of trees Tx′ j . We also let F ′ be the forest containing those
′
trees Txj that have at least one vertex not distinguished by (Wi , Li ∪ Lj ). We know that for each
tree Tx′ j of F ′ , we will use only watchers to distinguish the vertices not yet distinguished. For each
′
′
g
tree Tx′ j in F ′ , let T
xj be the tree obtained from Txj by removing all leaves already dominated by
a watcher of Wi . If this process has created a t-stem with t ≥ 2, then at least t − 1 leaf neighbours
′
g
of this t-stem have a watcher of Wi ; simply remove them. Then, the obtained tree T
xj has no
′ to satisfy
g
t-stems with t ≥ 2. Hence, we can apply Lemma 13 to g
T ′ (if necessary, we augment T
xj
xj
the maximality condition (iii) in Lemma 13). Thus, there is a set Xj of at most 63k 2 vertices of
′
g
g
Tx′ j that may possibly contain a watcher, and Xj can be computed in time linear in |V (T
xj )|. As
′
before, there are at most k trees in F and hence in total the set X = ∪j Xj of possible placements
for the remaining watchers has size at most 63k 3 (and is computable in linear time). Therefore, we
3
can check all the 63k
∈ O 2O(k log k) possibilities of placing the remaining watchers. If we find
k
a valid detection pair (W, L) with Wi ⊆ W , we return YES; otherwise, we discard the placement
Pj and move on to the next possibility.
It is clear, by the lemmas used in the description in the algorithm, that this algorithm is correct.
The running time of the preliminary phase is 2O(k log k) n2 . The running time of Phases 1 to 3 is at
most 4k 44k 2O(k log k) n which is 2O(k log k) n. This completes the proof.
6
Conclusion
We have obtained a linear-time 2-approximation algorithm for Opt-Detection Pair on trees, but
perhaps this algorithmic upper bound could be improved to a PTAS? It seems that the reduction
of [12] does not show the inapproximability of Opt-Detection Pair for any constant greater
than one. Therefore, it remains to settle the exact approximation complexity of Opt-Detection
Pair on trees.
As a second question, can the factor 2O(k log k) in our FPT algorithm for Detection Pair be
improved to single-exponential FPT time 2O(k) (or even sub-exponential time 2o(k) )? For this,
the bottleneck is our use of Lemma 13 in the preliminary phase and in Phase 3 of the algorithm,
during which we search through all possible subsets of size k of some sets of size O(k 2 ) and O(k 3 ),
respectively. Moreover, it is perhaps possible to obtain a linear running-time (in terms of n) by
reducing the running-time of the preliminary phase.
As an extension of our results, it would be of interest to determine whether, on further graph
classes, Opt-Detection Pair is constant-factor approximable and whether Detection Pair is
FPT. One natural research direction is to consider the class of planar graphs or its subclasses,
the outerplanar graphs and the series-parallel graphs (that is, graphs of treewidth 2). These
three classes contain all trees. Although both questions are settled in the affirmative for OptDominating Set and Dominating Set (see [20] and [8, 23]), as far as we know they also remain
open for Opt-Metric Dimension and Metric Dimension (though, a polynomial-time algorithm
exists for Opt-Metric Dimension on outerplanar graphs [6]).
Acknowledgements. We thank the anonymous referees for their careful reading and valuable
comments. We wish to thank Bert L. Hartnell for sharing the manuscript [12] with us. Moreover
we acknowledge the financial support of the programme ”IdEx Bordeaux – CPU (ANR-10-IDEX03-02)”.
References
[1] G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela and M. Protasi.
Complexity and approximation. Springer, 1999.
[2] R. F. Bailey and P. J. Cameron. Base size, metric dimension and other invariants of groups
and graphs. Bulletin of the London Mathematical Society 43:209–242, 2011.
11
[3] R. Belmonte, F. V. Fomin, P. A. Golovach and M. S. Ramanujan. Metric dimension of bounded
width graphs. Proceedings of the 40th International Symposium on Mathematical Foundations
of Computer Science, MFCS 2015, LNCS 9235:115–126, 2015.
[4] G. Chartrand, L. Eroh, M. Johnson and O. Oellermann. Resolvability in graphs and the metric
dimension of a graph. Discrete Applied Mathematics 105(1-3):99–113, 2000.
[5] E. J. Cockayne, S. Goodman and S. T. Hedetniemi. A linear algorithm for the domination
number of a tree. Information Processing Letters 4:41–44, 1975.
[6] J. Diaz, O. Pottonen, M. Serna and E. J. van Leeuwen. On the complexity of metric dimension.
Proceedings of the 20th European Symposium on Algorithms, ESA 2012, LNCS 7501:419–430,
2012.
[7] I. Dinur and D. Steurer. Analytical approach to parallel repetition. Proceedings of the fortysixth annual ACM symposium on Theory of computing, STOC 2014 :624–633, 2014.
[8] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Springer, 2013.
[9] D. Eppstein. Metric dimension parameterized by max leaf number. Journal of Graph Algorithms and Applications 19(1):313–323, 2015.
[10] L. Epstein, A. Levin and G. J. Woeginger. The (weighted) metric dimension of graphs: hard
and easy cases. Algorithmica 72(4):1130–1171, 2015.
[11] H. Fernau, P. Heggernes, P. van’t Hof, D. Meister and R. Saei. Computing the metric dimension for chain graphs. Information Processing Letters 115:671–676, 2015.
[12] A. S. Finbow, B. L. Hartnell and J. R. Young. The complexity of monitoring a network with
both watchers and listeners. Networks, accepted.
[13] F. Foucaud, G. B. Mertzios, R. Naserasr, A. Parreau and P. Valicov. Identification, locationdomination and metric dimension on interval and permutation graphs. II. Algorithms and
complexity. Algorithmica, to appear.
[14] M. R. Garey and D. S. Johnson. Computers and intractability: a guide to the theory of NPcompleteness, W. H. Freeman, 1979.
[15] F. Harary and R. A. Melter. On the metric dimension of a graph. Ars Combinatoria 2:191–195,
1976.
[16] S. Hartung. Exploring parameter spaces in coping with computational intractability.
PhD
Thesis,
Technische
Universität
Berlin,
Germany,
2014.
https://opus4.kobv.de/opus4-tuberlin/frontdoor/index/index/docId/4644
[17] S. Hartung and A. Nichterlein. On the parameterized and approximation hardness of metric dimension. Proceedings of the IEEE Conference on Computational Complexity, CCC 2013 :266–
276, 2013.
[18] M. Hauptmann, R. Schmied and C. Viehmann. Approximation complexity of metric dimension
problem. Journal of Discrete Algorithms 14:214–222, 2012.
[19] T. W. Haynes, S. T. Hedetniemi and P. J. Slater. Fundamentals of domination in graphs.
Marcel Dekker, 1998.
[20] T. W. Haynes, S. T. Hedetniemi and P. J. Slater, editors. Domination in graphs: advanced
topics. Marcel Dekker, 1998.
[21] D. S. Johnson, Approximation algorithms for combinatorial problems. Journal of Computer
and System Sciences 9:256–278, 1974.
[22] S. Khuller, B. Raghavachari and A. Rosenfeld. Landmarks in graphs. Discrete Applied Mathematics 70:217–229, 1996.
12
[23] R. Niedermeier. Invitation to fixed-parameter algorithms. Oxford University Press, 2006.
[24] P. J. Slater. Leaves of trees. Congressus Numerantium 14:549–559, 1975.
13
| 8cs.DS
|
COHEN-MACAULAY CIRCULANT GRAPHS
arXiv:1210.8351v1 [math.AC] 31 Oct 2012
KEVIN N. VANDER MEULEN, ADAM VAN TUYL, AND CATRIONA WATT
Abstract. Let G be the circulant graph Cn (S) with S ⊆ {1, 2, . . . , ⌊ n2 ⌋}, and let I(G)
denote its the edge ideal in the ring R = k[x1 , . . . , xn ]. We consider the problem of
determining when G is Cohen-Macaulay, i.e, R/I(G) is a Cohen-Macaulay ring. Because
a Cohen-Macaulay graph G must be well-covered, we focus on known families of wellcovered circulant graphs of the form Cn (1, 2, . . . , d). We also characterize which cubic
circulant graphs are Cohen-Macaulay. We end with the observation that even though
the well-covered property is preserved under lexicographical products of graphs, this is
not true of the Cohen-Macaulay property.
1. Introduction
Let G = (VG , EG ) denote a finite simple graph on the vertex set VG = {x1 , . . . , xn } with
edge set EG . By identifying the vertices of G with the variables of the polynomial ring
R = k[x1 , . . . , xn ] (here, k is any field), we can associate to G the quadratic square-free
monomial ideal
I(G) = hxi xj | {xi , xj } ∈ EG i ⊆ R
called the edge ideal of G. Edge ideals were first introduced by Villarreal [17]. During
the last couple of years, there has been an interest in determining which graphs G are
Cohen-Macaulay, that is, determining when the ring R/I(G) is a Cohen-Macaulay ring
solely from the properties of the graphs. Although this problem is probably intractable
for arbitrary graphs, results are known for some families of graphs, e.g., chordal graphs
[8] and bipartite graphs [7]. Readers may also be interested in the recent survey of Morey
and Villarreal [11] and the textbook of Herzog and Hibi [6], especially Chapter 9.
Our goal is to identify families of circulant graphs that are Cohen-Macaulay. Given an
integer n ≥ 1 and a subset S ⊆ {1, 2, . . . , ⌊ n2 ⌋}, the circulant graph Cn (S) is the graph on n
vertices {x1 , . . . , xn } such that {xi , xj } is an edge of Cn (S) if and only if min{|i−j|, n−|i−
j|} ∈ S. See, for example, the graph C12 (1, 3, 4) in Figure 1. For convenience of notation,
we suppress the set brackets for the set S = {1, 3, 4} in C12 (1, 3, 4). Circulant graphs
belong to the family of Cayley graphs and are sometimes viewed as generalized cycles since
Cn = Cn (1). The complete graph is also a circulant graph because Kn = Cn (1, 2, . . . , ⌊ n2 ⌋).
In the literature, circulant graphs have appeared in a number of applications related to
2010 Mathematics Subject Classification. 13F55, 13H10, 05C75, 05E45.
Key words and phrases. Cohen-Macaulay, circulant graph, well-covered graph, vertex decomposable,
shellable.
Last updated: January 7, 2018.
Research of first two authors supported in part by NSERC Discovery Grants. Research of the third
author supported by an NSERC USRA.
1
COHEN-MACAULAY CIRCULANT GRAPHS
2
Figure 1. The circulant graph C12 (1, 3, 4).
networks [1], error-correcting codes [14], and even music [3], in part, because of their
regular structure (see [4]).
To classify families of Cohen-Macaulay circulant graphs we will use the fact that all
Cohen-Macaulay graphs must be well-covered. A graph G is well-covered if all the maximal independent sets of G have the same cardinality, equivalently, every maximal independent set is a maximum independent set (see the survey of Plummer [12]). From an
algebraic point-of-view, when a graph G is well-covered, the edge ideal is I(G) is unmixed,
that is, all of its associated primes have the same height. Some families of well-covered
circulant graphs were recently classified by Brown and Hoshino [4]. Our main results
(see Theorems 3.4 and 5.2) refine the work of Brown and Hoshino by determining which
of these well-covered circulant graphs are also Cohen-Macaulay. In particular we show
in Theorem 3.4 that for n ≥ 2d ≥ 2, the circulant Cn (1, 2, . . . , d) is Cohen-Macaulay
if and only if n ≤ 3d + 2 and n 6= 2d + 2. We also show that the Cohen-Macaulay
graphs Cn (1, 2, . . . , d) are in fact vertex decomposable and shellable. Although the wellcovered circulant graphs C2d+2 (1, 2, . . . , d) and C4d+3 (1, 2, . . . , d) are not Cohen-Macaulay,
we prove that these graphs are Buchsbaum (see Theorem 3.7). We also classify which
cubic circulant graphs are Cohen-Macaulay (see Theorem 5.5).
Our paper is structured as follows. In Section 2 we recall the relevant background
regarding graph theory and simplicial complexes. In Section 3 we classify the CohenMacaulay graphs of the form Cn (1, 2, . . . , d) with n ≥ 2d. Section 4 contains the proof
of a lemma needed to prove the main result of Section 3. In Section 5, we look at cubic
circulant graphs, and classify those that are Cohen-Macaulay. Section 6 contains some
concluding comments and open questions related to the lexicographical product of graphs.
2. Background Definitions and Results
A simplicial complex ∆ on a vertex set V = {x1 , . . . , xn } is a set of subsets of V that
satisfies: (i) if F ∈ ∆ and G ⊆ F , then G ∈ ∆, and (ii) for each i = 1, . . . , n, {xi } ∈ ∆.
Note that condition (i) implies that ∅ ∈ ∆. The elements of ∆ are called its faces. The
maximal elements of ∆, with respect to inclusion, are the facets of ∆.
The dimension of a face F ∈ ∆ is given by dim F = |F | − 1; the dimension of a
simplicial complex, denoted dim ∆, is the maximum dimension of all its faces. We call ∆
a pure simplicial complex if all its facets have the same dimension. Let fi be the number
of faces of ∆ of dimension i, with the convention that f−1 = 1. If dim ∆ = D, then the
f -vector of ∆ is the (D + 2)-tuple f (∆) = (f−1 , f0 , f1 , . . . , fD ). The h-vector of ∆ is the
COHEN-MACAULAY CIRCULANT GRAPHS
3
(D + 2)-tuple h(∆) = (h0 , h1 , . . . , hD+1 ) with (see [18, Theorem 5.4.6])
k
X
k−i D + 1 − i
fi−1 .
hk =
(−1)
k
−
i
i=0
Given any simplicial complex ∆ on V , we can associate to ∆ a monomial ideal I∆ in
the polynomial ring R = k[x1 , . . . , xn ] (with k a field) as follows:
I∆ = h{xj1 xj2 · · · xjr | {xj1 , . . . , xjr } 6∈ ∆}i.
The ideal I∆ is commonly called the Stanley-Reisner ideal of ∆, and the quotient ring
R/I∆ is the Stanley-Reisner ring of ∆.
We say that ∆ is Cohen-Macaulay (over k) if its Stanley-Reisner ring R/I∆ is a CohenMacaulay ring, that is, K-dim(R/I∆ ) = depth(R/I∆ ). Here K-dim(R/I∆ ), the Krull
dimension, is the length of the longest chain of prime ideals in R/I∆ with strict inclusions,
and depth(R/I∆ ), the depth, is length of the longest sequence f1 , . . . , fj in hx1 , . . . , xn i
that forms a regular sequence on R/I∆ .
We review the required background on reduced homology; see [10] for complete details.
To any simplicial complex ∆ with f (∆) = (f−1 , f0 , . . . , fD ) we can associate a reduced
chain complex over k:
∂
∂
∂
∂
0
1
2
D
C̃· (∆; k) : 0 ←− k f−1 ←−
k f0 ←−
k f1 ←−
· · · ←−
k fD ← 0.
Here k fi is the vector space with basis elements ej0 ,j1 ,...,ji corresponding to the i-dimensional
faces of ∆. We assume j0 < j1 < · · · < ji . The boundary maps ∂i are given by
∂i (ej0 ,j1 ,...,ji ) = eĵ0 ,j1 ,...,ji − ej0 ,ĵ1 ,...,ji + ej0 ,j1,ĵ2 ...,ji + · · · + (−1)i ej0 ,j1,...,ĵi
where ˆ denotes an omitted term. The ith reduced simplicial homology of ∆ with coefficients in k is the k-vector space
H̃i (∆; k) = ker(∂i )/ im(∂i+1 ).
The dimensions of H̃i (∆; k) are related to f (∆) via the reduced Euler characteristic:
(2.1)
D
X
(−1)i dimk H̃i (∆; k) =
i=−1
D
X
(−1)i fi .
i=−1
We will find it convenient to use Reisner’s Criterion. Given a face F ∈ ∆, the link of
F in ∆ is the subcomplex
link∆ (F ) = {G ∈ ∆ | F ∩ G = ∅, F ∪ G ∈ ∆}.
Theorem 2.1 (Reisner’s Criterion). Let ∆ be a simplicial complex on V . Then R/I∆
is Cohen-Macaulay over k if and only if for all F ∈ ∆, H̃i (link∆ (F ); k) = 0 for all
i < dim link∆ (F ).
For any vertex x ∈ V , the deletion of x in ∆ is the subcomplex
del∆ ({x}) = {F ∈ ∆ | x 6∈ F }.
The following combinatorial topology property was introduced by Provan and Billera [13].
COHEN-MACAULAY CIRCULANT GRAPHS
4
Definition 2.2. Let ∆ be a pure simplicial complex. Then ∆ is vertex decomposable if
(i) ∆ is a simplex, i.e {x1 , . . . , xn } is the unique maximal facet, or
(ii) there exists an x ∈ V such that link∆ ({x}) and del∆ ({x}) are vertex decomposable.
We will also refer to the following family of simplicial complexes.
Definition 2.3. Let ∆ be a pure simplicial complex with facets {F1 , . . . , Ft }. Then ∆ is
shellable if there exists an ordering of F1 , . . . , Ft such that such that for all 1 ≤ j < i ≤ t,
there is some x ∈ Fi \ Fj and some k ∈ {1, . . . , j − 1} such that {x} = Fi \ Fk .
The following theorem summarizes a number of necessary and sufficient conditions of
Cohen-Macaulay simplicial complexes that we require.
Theorem 2.4. Let ∆ be a simplicial complex on the vertex set V = {x1 , . . . , xn }.
(i) If ∆ is Cohen-Macaulay, then it is pure.
(ii) If ∆ is Cohen-Macaulay, then h(∆) has only non-negative entries.
(iii) If n−pdim(R/I∆ ) = K-dim(R/I∆ ), then ∆ is Cohen-Macaulay (here, pdim(R/I∆ )
denotes the projective dimension of R/I∆ , the length of a minimal free resolution
of R/I∆ ).
(iv) If ∆ is vertex decomposable, then ∆ is Cohen-Macaulay.
(v) If dim ∆ = 0, then ∆ is vertex decomposable/shellable/Cohen-Macaulay.
(vi) If dim ∆ = 1, then ∆ is vertex decomposable/shellable/Cohen-Macaulay if and
only if ∆ is connected.
Proof. Many of these results are standard. For (i) see [18, Theorem 5.3.12]; for (ii) see [18,
Theorem 5.4.8]; (iii) follows from the Auslander-Buchsbaum Theorem; for (iv) see [13,
Corollary 2.9] and the fact that shellable implies Cohen-Macaulay [18, Theorem 5.3.18];
(v) is [13, Proposition 3.1.1]; and (vi) is [13, Theorem 3.1.2].
In this paper, we will be interested in independence complexes of finite simple graphs
G = (VG , EG ). We say that a set of vertices W ⊆ VG is an independent set if for all
e ∈ EG , e 6⊆ W . The independence complex of G is the set of all independent sets:
Ind(G) = {W | W is an independent set of VG }.
The set Ind(G) is a simplicial complex. Following convention, G is Cohen-Macaulay
(resp. shellable, vertex decomposable) if Ind(G) is Cohen-Macaulay (resp. shellable,
vertex decomposable).
The facets of Ind(G) correspond to the maximal independent sets of vertices of G. It
is common to let α(G) denote the cardinality of a maximum independent set of vertices
in G. A graph G is well-covered if every maximal independent set has cardinality α(G).
Moreover, a direct translation of the definitions gives us:
Lemma 2.5. If G is Cohen-Macaulay, then G is well-covered.
COHEN-MACAULAY CIRCULANT GRAPHS
5
3. Characterization of Circulant graphs Cn (1, 2, . . . , d).
In this section, we classify which circulant graphs of the form Cn (1, 2, . . . , d) are CohenMacaulay. Brown and Hoshino recently classified the well-covered graphs in this family:
Theorem 3.1. ([4, Theorem 4.1]). Let n and d be integers with n ≥ 2d ≥ 2. Then
Cn (1, 2, . . . , d) is well-covered if and only if n ≤ 3d + 2 or n = 4d + 3.
Brown and Hoshino’s result is a key ingredient for our main result. We also need one
additional result of [4] on the independence polynomial of Cn (1, 2, . . . , d), but translated
into a statement about f -vectors. The independence polynomial of a graph G is given
Pα(G)
by I(G, x) = k=0 ik xk where ik is the number of independent sets of cardinality k (we
take i0 = 1). Note that if ∆ = Ind(G) and f (∆) = (f−1 , f0 , . . . , fD ), then ik = fk−1 for
each k. If we translate [4, Theorem 3.1] into the language of f -vectors and independence
complexes, we get the following statement.
Lemma 3.2. Let n and d be
integers with n ≥ 2d ≥ 2, G = Cn (1, 2, . . . , d), and D =
n
− 1 and f (∆) = (f−1 , f0 , . . . , fD ) where
dim Ind(G). Then D = d+1
n
n − dk
fk−1 =
for k = 0, . . . , (D + 1).
k
n − dk
By Lemma 2.5, to characterize the Cohen-Macaulay circulant graphs of the form
Cn (1, 2, . . . , d), it suffices to determine which of the well-covered graphs of Theorem 3.1 are
also Cohen-Macaulay. Interestingly, proving that Cn (1, 2, . . . , d) is not Cohen-Macaulay
when n = 4d + 3 is the most subtle part of this proof. To carry out this part of the proof,
we need the following lemma, whose proof we postpone until the next section.
Lemma 3.3. Fix an integer d ≥ 3, and let G = C4d+3 (1, 2, . . . , d). If ∆ = Ind(G), then
(4d + 3) d − 1
.
dimk H̃2 (∆; k) ≥
2
3
Assuming, for the moment, that Lemma 3.3 holds, we arrive at our main result:
Theorem 3.4. Let n and d be integers with n ≥ 2d ≥ 2 and let G = Cn (1, 2, . . . , d).
Then the following are equivalent:
(i)
(ii)
(iii)
(iv)
G is Cohen-Macaulay.
G is shellable.
G is vertex decomposable.
n ≤ 3d + 2 and n 6= 2d + 2.
Proof. We always have (iii) ⇒ (ii) ⇒ (i). We now prove that (iv) ⇒ (iii).
By Lemma 3.2, when n = 2d or n = 2d + 1, dim Ind(G) = 0. Now apply Theorem 2.4
(v).
n
When 2d + 2 ≤ n ≤ 3d + 2, dim Ind(G) = ⌊ d+1
⌋ − 1 = 1. Let V = {x1 , . . . , xn }. If
n = 2d+2, then Ind(G) is not connected, because the only edges of Ind(G) are {xi , xd+1+i }
for i = 1, . . . , d + 1. On the other hand, when 2d + 3 ≤ n ≤ 3d + 2, Ind(G) is connected.
COHEN-MACAULAY CIRCULANT GRAPHS
6
To see this, let n = 2d + c for 3 ≤ c ≤ d + 2. For each i = 1, 2, . . . , n, {xi , xi+d+2 }
and {xi+1 , xi+d+2 } ∈ Ind(G), with subscript addition adjusted modulo n. Thus, for any
xi , xj ∈ V with i < j, we can make the path xi , xi+d+2 , xi+1 , xi+d+3 , xi+2 , . . . , xj . So,
Ind(G) is connected. Applying Theorem 2.4 (vi) then shows that (iv) ⇒ (iii).
To complete the proof, we will show that if n ≥ 2d with n = 2d + 2 or n > 3d + 2,
then G is not Cohen-Macaulay. In the proof that (iv) ⇒ (iii), we already showed that if
n = 2d + 2, then Ind(G) is not connected and dim Ind(G) = 1. Again by Theorem 2.4
(vi) this implies G is not Cohen-Macaulay.
If n > 3d + 2 and n 6= 4d + 3, then by Theorem 3.1, G is not well-covered, and
consequently, by Lemma 2.5, G is not Cohen-Macaulay. It therefore remains to show that
if n = 4d + 3, then G is not Cohen-Macaulay for all d ≥ 1. The remainder of this proof
is dedicated to this case.
By Lemma 3.2, dim Ind(G) = 2 and the f -vector of Ind(G) is given by
4d3 + 15d2 + 17d + 6
2
f (Ind(G)) = 1, 4d + 3, 4d + 7d + 3,
.
6
When d = 1, then f (Ind(G)) = (1, 7, 14, 7) and hence h(Ind(G)) = (1, 4, 3, −1). When
d = 2, then f (Ind(G)) = (1, 11, 33, 21) and hence h(Ind(G)) = (1, 8, 13, −1). In these two
cases, Theorem 2.4 (ii) implies G is not Cohen-Macaulay.
We can therefore assume that d ≥ 3. To show that Ind(G) is not Cohen-Macaulay,
we will show that H̃1 (Ind(G); k) 6= 0. This suffices because Ind(G) = linkInd(G) (∅), so
Reisner’s Criterion (Theorem 2.1) would imply that Ind(G) is not Cohen-Macaulay.
Using the fact that dim Ind(G) = 2, the f -vector given above, and the reduced Euler
characteristic (2.1) we know
2
X
4d3 + 15d2 + 17d + 6
(−1)i dimk H̃i (Ind(G); k).
=
−1 + (4d + 3) − (4d + 7d + 3) +
6
i=−1
2
Because Ind(G) is a non-empty connected simplicial complex, we have dimk H̃i (Ind(G); k) =
0, for i = −1, and 0. Simplifying both sides of the above equation and rearranging gives:
d(4d2 − 9d − 1)
dimk H̃1 (Ind(G); k) = dimk H̃2 (Ind(G); k) −
.
6
By Lemma 3.3
4d + 3 d − 1
d(4d2 − 9d − 1)
dimk H̃1 (Ind(G); k) ≥
−
= 1.
2
3
6
So, H̃1 (Ind(G); k) 6= 0 as desired.
When we specialize the above theorem to the case d = 1, we recover the known classification of the Cohen-Macaulay cycles [18, Corollary 6.3.6]. Note that C2 (1) = K2 is also
Cohen-Macaulay, but it is not a cycle.
Corollary 3.5. Let n ≥ 3. Then Cn = Cn (1) is Cohen-Macaulay if and only if n = 3 or
5.
COHEN-MACAULAY CIRCULANT GRAPHS
7
Even though C2d+2 (1, 2, . . . , d) and C4d+3 (1, 2, . . . , d) are not Cohen-Macaulay, they still
have an interesting algebraic structure, as noted in Theorem 3.7 below.
Definition 3.6. A pure simplicial complex ∆ is called Buchsbaum over a field k if for
every non-empty face F ∈ ∆, H̃i (link∆ (F ); k) = 0 for all i < dim link∆ (F ). We say a
graph G is Buchsbaum if the independence complex of G is Buchsbaum.
Note that by Reisner’s Criterion (Theorem 2.1), if G is Cohen-Macaulay, then G is
Buchsbaum. We can now classify all circulant graphs of the form Cn (1, . . . , d) which are
Buchsbaum, but not Cohen-Macaulay.
Theorem 3.7. Let n and d be integers with n ≥ 2d and d ≥ 1. Let G = Cn (1, 2, . . . , d).
Then G is Buchsbaum, but not Cohen-Macaulay if and only if n = 2d + 2 or n = 4d + 3.
Proof. (⇒) For G to be Buchsbaum, Ind(G) must be pure, that is G is well-covered. By
Theorem 3.1, 2d ≤ n ≤ 3d + 2 or n = 4d + 3. Because G is not Cohen-Macaulay, Theorem
3.4 implies n = 2d + 2 or n = 4d + 3.
(⇐) We first show that if n = 4d + 3, then G is Buchsbaum. Let ∆ = Ind(G). Since
dim ∆ = 2, given any F ∈ ∆, |F | ∈ {0, 1, 2, 3}. We wish to show that if |F | = 1, 2, or 3,
then H̃i (link∆ (F ); k) = 0 for all i < dim link∆ (F ).
If |F | = 3, then link∆ (F ) = {∅}, and hence H̃i (link∆ (F ); k) = 0 for all i < dim link∆ (F ) =
−1. When |F | = 2, then dim link∆ (F) = 0, and again, we have H̃i (link∆ (F ); k) = 0 for
all i < dim link∆ (F ) = 0.
It therefore suffices to show that when |F | = 1, then H̃i (link∆ (F ); k) = 0 for all
i < dim link∆ (F ). Because of the symmetry of G, we can assume without a loss of
generality that F = {x1 }. Because G is well-covered, any independent set containing
x1 can be extended to a maximal independent set, and furthermore, this independent
set has cardinality three. This, in turn, implies that dim link∆ (F ) = 1. For any i <
0, H̃i (link∆ (F ); k) = 0, so it suffices to prove that H̃0 (link∆ (F ); k) = 0. Proving this
condition is equivalent to proving that link∆ (F ) is connected.
We first note that none of the vertices x2 , x3 , . . . , xd+1 , x3d+4 , x3d+5 , . . . , x4d+3 appear in
link∆ ({x1 }) because these vertices are all adjacent to x1 in G. On the other hand, the
following elements are facets of ∆:
{x1 , xd+2 , x2d+3 }, {x1 , xd+2 , x2d+4 }, . . . , {x1 , xd+2 , x3d+3 },
{x1 , xd+3 , x3d+3 }, {x1 , xd+4 , x3d+3 }, . . . , {x1 , x2d+2 , x3d+3 }.
Consequently the following edges are in link∆ ({x1 }):
{xd+2 , x2d+3 }, {xd+2 , x2d+4 }, . . . , {xd+2 , x3d+3 }, {xd+3 , x3d+3 }, {xd+4 , x3d+3 }, . . . , {x2d+2 , x3d+3 }.
Thus link∆ ({x1 }) is connected, as desired.
Now suppose n = 2d + 2. As shown in the proof of Theorem 3.4, Ind(G) consists of
the disjoint edges {xi , xd+1+i } for i = 1, . . . , d + 1. If F ∈ Ind(G) and |F | = 2, then
link∆ (F ) = {∅}. If F ∈ Ind(G) and |F | = 1, then link∆ (F ) = {{x}} for some variable x.
Therefore G is Buchsbaum.
COHEN-MACAULAY CIRCULANT GRAPHS
8
4. Proof of Lemma 3.3
The purpose of this section is to prove Lemma 3.3. We will be interested in finding
induced octahedrons in our independence complex.
Lemma 4.1. Fix an integer d ≥ 3. Let G = C4d+3 (1, 2, . . . , d) and let ∆ = Ind(G) be
the associated independence complex. Let W = {i1 , i2 , j1 , j2 , k1, k2 } ⊆ VG be six distinct
vertices. Then the induced simplicial complex ∆|W = {F ∈ ∆ | F ⊆ W } is isomorphic
to the labeled octahedron in Figure 2 if and only if the induced graph GW is the graph of
three disjoint edges {i1 , i2 }, {j1 , j2 }, and {k1 , k2}.
i1
k1
j2
j1
k2
i2
Figure 2. A labeled octahedron
Proof. Suppose that ∆|W is isomorphic to the octahedron in Figure 2. It follows that
{i1 , i2 }, {j1 , j2 } and {k1 , k2 }, which are not edges of the octahedron, are also not edges
of ∆. Because ∆ is an independence complex, these means that each set is not an
independent set, or in other words, e1 = {i1 , i2 }, e2 = {j1 , j2 }, and e3 = {k1 , k2 } are all
edges of G. It suffices to show that GW consists only of these edges. If not, there is a
vertex x ∈ ei and a vertex y ∈ ej with i 6= j, such that {x, y} is an edge of G. However,
for any x ∈ ei and y ∈ ej , {x, y} is an edge of ∆|W , and consequently, {x, y} cannot be
an edge of G, a contradiction.
For the converse, we reverse the argument. If GW is the three disjoint edges {i1 , i2 },
{j1 , j2 } and {k1 , k2}, then it follows that {i1 , j1 , k1 }, {i1 , j1 , k2}, {i1 , j2 , k1 }, {i1 , j2 , k2 },
{i2 , j1 , k1 }, {i2 , j1 , k2 }, {i2 , j2 , k1 }, {i2 , j2 , k2 } are all independent sets of G, and thus belong
to ∆, and consequently, ∆|W . Because {i1 , i2 }, {j1 , j2 }, and {k1 , k2 } are not faces of ∆,
the facets of the complex ∆|W are these eight faces, whence ∆|W is an octahedron.
We now come to our desired proof.
Proof. (of Lemma 3.3) We begin by first recalling some facts about ∆ = Ind(G). By
Theorem 3.1 and Lemma 3.2, the simplicial complex ∆ is pure and two dimensional with
f (∆) = (f−1 , f0 , f1 , f2 ). Therefore, the reduced chain complex of ∆ over k has the form
∂
∂
∂
0
1
2
0 ←− k f−1 ←−
k f0 ←−
k f1 ←−
k f2 ←− 0.
It follows from this chain complex that dimk H̃2 (∆; k) = dimk ker ∂2 .
COHEN-MACAULAY CIRCULANT GRAPHS
9
linearly independent elements in ker ∂2 .
Our strategy, therefore, is to identify (4d+3)
3
2
Note that if W ⊆ V is a subset of the vertices such that the induced complex ∆|W is
isomorphic to an octahedron, then this octahedron corresponds to an element of ker ∂2 .
We make this more precise. Suppose that W = {i1 , i2 , j1 , j2 , k1 , k2 } ⊆ V and ∆|W is an
octahedron, i.e, the simplicial complex with facets
d−1
∆|W = h{i1 , j1 , k1 }, {i1 , j1 , k2 }, {i1 , j2 , k1 }, {i1 , j2 , k2 },
{i2 , j1 , k1 }, {i2 , j1 , k2 }, {i2 , j2 , k1 }, {i2 , j2 , k2}i.
Note that each {ia , jb , kc } is a 2-dimensional face of ∆; we associate to ∆|W the following
element of k f2 :
OW = ei1 j1 k1 − ei1 j1 k2 − ei1 j2 k1 − ei2 j1 k1 + ei1 j2 k2 + ei2 j1 k2 + ei2 j2 k1 − ei2 j2 k2 .
Here, we have assumed that the indices of each basis element have been written in increasing order. The boundary map ∂2 evaluated at OW gives ∂2 (OW ) = 0, i.e., OW ∈ ker ∂2 .
To compute the lower bound on dimk H̃2 (∆; k), we will build a list L of octahedrons
in ∆ and then order the elements of L using the lexicographical ordering so that each
octahedron in the list L contains a face that has not appeared in any previous octahedron
in L with respect to the ordering. By associating each octahedron to the corresponding
element of k f2 , each octahedron will belong to ker ∂2 . Moreover, the fact that each octahedron in L has a face that has not appeared previously implies that the octahedron can
not be written as a linear combination of our previous elements in ker ∂2 , thus giving us
the required number of linearly independent elements.
By Lemma 4.1, there is a one-to-one correspondence between the induced octahedrons
of ∆ and the induced subgraphs of G consisting of three pairwise disjoint edges. So we
can represent an octahedron by a tuple (i1 , i2 ; j1 , j2 ; k1, k2 ) where {i1 , i2 }, {j1 , j2 }, and
{k1 , k2 } correspond to these edges.
We begin by considering the octahedrons described by the following list:
(1, 2;
(1, 2;
d + 3, d + 4;
d + 3, d + 4;
2d + 5, 3d + 3)
2d + 6, 3d + 3)
..
.
(1, 2;
(1, 2;
(1, 2;
d + 3, d + 4;
d + 3, d + 5;
d + 3, d + 5;
3d + 2, 3d + 3)
2d + 6, 3d + 3)
2d + 7, 3d + 3)
..
.
(1, 2;
(1, 2;
d + 3, d + 5;
d + 3, d + 6;
3d + 2, 3d + 3)
2d + 7, 3d + 3)
..
.
(4.1)
(1, 2; d + 3, 2d;
3d + 1, 3d + 3)
(1, 2; d + 3, 2d;
3d + 2, 3d + 3)
(1, 2; d + 3, 2d + 1; 3d + 2, 3d + 3)
If we take our list of octahedrons in (4.1) and add one to each index, we will get a new list
of octahedrons. In terms of the graph G = C4d+3 (1, 2, . . . , d), we are “rotating” our disjoint
COHEN-MACAULAY CIRCULANT GRAPHS
10
edges to the right. We “rotate” these disjoint edges, or equivalently, we add one to each
index, until k1 = 4d + 3. So, for example, the disjoint edges (1, 2; d + 3, d + 4; 2d + 5, 3d + 3)
can be rotated to the right 2d − 2 times to give us 2d − 1 octahedrons
(1, 2;
(2, 3;
(3, 4;
..
.
(2d − 1, 2d;
d + 3, d + 4;
d + 4, d + 5;
d + 5, d + 6;
..
.
3d + 1, 3d + 2;
2d + 5, 3d + 3)
2d + 6, 3d + 4)
2d + 7, 3d + 5)
..
.
4d + 3, d − 2).
On the other hand, the disjoint edges (1, 2; d + 3, 2d + 1; 3d + 2, 3d + 3) are only rotated
d + 1 times to create d + 2 octahedrons
(1, 2;
(2, 3;
(3, 4;
..
.
d + 3, 2d + 1;
d + 4, 2d + 2;
d + 5, 2d + 3;
..
.
3d + 2, 3d + 3)
3d + 3, 3d + 4)
3d + 3, 3d + 5)
..
.
(d + 2, d + 3; 2d + 4, 3d + 2; 4d + 3, 1)
.
If we carry out this procedure, we end up with an expanded list L of octahedrons with
d−2
d−2
d−2
X
X
X
4d + 3 d − 1
2
.
|L| =
k(2d − k) = 2d
k−
k =
2
3
k=1
k=1
k=1
To see why, there is only one collection of disjoint edges with k1 = 2d + 5 which is rotated
2d − 1 times, there are two tuples of disjoint edges with k1 = 2d + 6 which are rotated
2d − 2 times, and so on, until we arrive at the d − 2 tuples which are constructed from all
the tuples with k1 = 3d + 2 rotated d + 2 times.
It now suffices to show that the corresponding elements of ker ∂2 are linearly independent. In (4.2), we have arranged the list L in lexicographical order from smallest to
largest:
(1, 2;
d + 3, d + 4;
..
.
d + 3, 2d + 1;
d + 4, d + 5;
..
.
d + 4, 2d + 2;
..
.
2d + 5, 3d + 3)
..
.
3d + 2, 3d + 3)
2d + 6, 3d + 4)
..
.
3d + 3, 3d + 4)
(2d − 2, 2d − 1;
3d, 3d + 1;
(2d − 2, 2d − 1;
3d, 3d + 2;
(2d − 1, 2d;
3d + 1, 3d + 2;
4d + 3, d − 3)
4d + 3, d − 3)
4d + 3, d − 2)
(1, 2;
(2, 3;
(4.2)
(2, 3;
For each (i1 , i2 ; j1 , j2 ; k1, k2 ) in L, we consider the two-dimensional face {i2 , j2 , k1 } of the
associated octahedron. We claim that as we progress down the list in (4.2), each face
{i2 , j2 , k1 } has not appeared in a previous octahedron.
COHEN-MACAULAY CIRCULANT GRAPHS
11
In particular, suppose that (i1 , i2 ; j1 , j2 ; k1 , k2 ) is the ℓ-th item in (4.2). We wish to
show that the face {i2 , j2 , k1 } has not appeared in any of the first ℓ − 1 octahedrons in
the lexicographically ordered list (4.2). Suppose, that (a1 , a2 ; b1 , b2 , c1 , c2 ) appears earlier
in the list and contains the face {i2 , j2 , k1 }. For this face to appear, {a1 , a2 } must contain
exactly one of i2 , j2 , k1 , {b1 , b2 } must contain exactly one of the remaining two vertices,
and {c1 , c2 }, must contain the remaining vertex of the face.
By the way we listed and constructed our octahedrons, i2 < j2 < k1 , i1 = i2 − 1, and
j1 = i2 +d+1. Further, k2 = i2 +3d+1 if i2 +3d+1 ≤ (4d−3), and k2 = i2 +3d+1−(4d−3)
if i2 + 3d + 1 > (4d − 3). Note that a1 6= i2 , j2 or k2 otherwise a1 > i1 , contradicting
the lexicographical ordering. Also, if a2 = j2 , then a1 = j2 − 1 ≥ i2 > i1 , again a
contradiction. The same problem arises if a2 = k1 . Thus {a1 , a2 } = {i2 − 1, i2 } = {i1 , i2 },
i.e., (a1 , a2 ; b1 , b2 ; c1 , c2 ) = (i1 , i2 , j1 , b2 , c1 , k2 ).
Since j2 and k1 must also appear in this tuple, there are only two possibilities:
(a1 , a2 ; b1 , b2 ; c1 , c2 ) = (i1 , i2 , j1 , k1 , j2 , k2 ) or (i1 , i2 , j1 , j2 , k1 , k2 ).
But neither of these tuples appear strictly before (i1 , i2 , j1 , j2 , k1 , k2 ) with respect to our
ordering, thus completing the proof.
5. Cohen-Macaulay Circulant Cubic Graphs
Brown and Hoshino [4] classified which circulant cubic graphs are well-covered. Recall
that a cubic graph is a graph in which each vertex has degree 3. Thus, if G is a circulant
cubic graph, then G = C2n (a, n) for some 1 ≤ a < n.
There are only a finite number of connected well-covered circulant cubic graphs:
Theorem 5.1 ([4, Theorem 4.3]). Let G be a connected circulant cubic graph. Then G
is well-covered if and only if it is isomorphic to one of the following graphs: C4 (1, 2),
C6 (1, 3), C6 (2, 3), C8 (1, 4) or C10 (2, 5).
Using a computer algebra system like Macaulay2 [9], one can simply check which of
these graphs, displayed in Figure 3, are Cohen-Macaulay.
Theorem 5.2. Let G be a connected circulant cubic graph. Then is Cohen-Macaulay if
and only if it is isomorphic to C4 (1, 2) or C6 (2, 3).
Proof. By Theorem 5.1, it suffices to check which of the graphs C4 (1, 2), C6 (1, 3), C6 (2, 3),
C8 (1, 4) or C10 (2, 5), are also Cohen-Macaulay.
Figure 3. The well-covered connected cubic circulant graphs.
COHEN-MACAULAY CIRCULANT GRAPHS
12
For any graph G, dim R/I(G) = α(G). So, by Theorem 2.4 (iii), we simply need to
check if α(G) = n − pdim(R/I(G)). We can compute α(G) for each of the graphs G in
Figure 3 by inspection; on the other hand, we compute the projective dimension using a
computer algebra system. The following table summarizes these calculations:
G
C4 (1, 2) C6 (1, 3) C6 (2, 3) C8 (1, 4) C10 (2, 5)
n − pdim(R/I(G))
1
1
2
2
2
α(G)
1
3
2
3
4
The conclusion now follows from the values in the table.
As in Brown and Hoshino [4], we will use the following result to extend Theorem 5.2
to all circulant cubic graphs. The following classification is due to Davis and Domke [5].
Theorem 5.3. Let G = C2n (a, n) with 1 ≤ a < n, and let t = gcd(a, 2n).
(i) If
(ii) If
2n
t
2n
t
is even, then G is isomorphic to t copies of C 2n (1, nt ).
t
is odd, then G is isomorphic to 2t copies of C 4n (2, 2n
).
t
t
We also use the following lemma in the next proof.
Lemma 5.4 ([18, Proposition 6.2.8]). Suppose that the graph G = H ∪ K where H and
K are disjoint components of G. Then G is Cohen-Macaulay if and only if H and K are
Cohen-Macaulay.
Theorem 5.5. Let G = C2n (a, n) with 1 ≤ a < n, that is, G is a cubic circulant graph.
Let t = gcd(a, 2n). Then G is Cohen-Macaulay if and only if 2n
= 3 or 4.
t
Proof. Suppose that 2n
6= 3 or 4. If 2n
is even, then C 2n (1, nt ) is not Cohen-Macaulay by
t
t
t
2n
4n (2,
is
odd,
then
C
)
is
also
not
Cohen-Macaulay by Theorem 5.2.
Theorem 5.2 and if 2n
t
t
t
Thus by Lemma 5.4 G is not Cohen-Macaulay. Conversely, if 2n
= 4, then by Theorem
t
2n
5.3, G is isomorphic to t copies of C4 (1, 2) and if t = 3, then G is isomorphic to 2t copies
of C6 (2, 3). In both cases, Theorem 5.2 and Lemma 5.4 imply G is Cohen-Macaulay.
6. Concluding Comments and Open Questions
The question of classifying all Cohen-Macaulay circulant graphs Cn (S) is probably an
intractable problem. Even the weaker question of determining whether or not a circulant
graph Gn (S) is well-covered (equivalently, Ind(Cn (S)) is a pure simplicial complex) was
shown by Brown and Hoshino to be co-NP-complete [4, Theorem 2.5]. At present, the
best we can probably expect is to identify families of Cohen-Macaulay circulant graphs.
Brown and Hoshino observed that circulant graphs behave well with respect to the
lexicographical product. Recall this construction:
Definition 6.1. Given two graphs G and H, the lexicographical product, denoted G[H], is
graph with vertex set V (G) × V (H), where any two vertices (u, v) and (x, y) are adjacent
in G[H] if and only if either {u, x} ∈ G or u = x and {v, y} ∈ H.
COHEN-MACAULAY CIRCULANT GRAPHS
13
When G and H are both circulant graphs, then the lexicographical product G[H] is
also circulant (see [4, Theorem 4.6]). The well-covered property is also preserved with
respect to the lexicographical product (see [16]).
Theorem 6.2. Let G and H be two non-empty graphs. Then G[H] is well-covered if and
only if the graphs G and H are well-covered.
As a consequence, the families of well-covered circulant graphs discovered in [4] can
be combined into new well-covered circulant graphs using the lexicographical product. It
is therefore natural to ask if the lexicographical product allows us to build new CohenMacaulay circulant graphs from known Cohen-Macaulay circulant graphs. In other words,
can we replace “well-covered” in Theorem 6.2 by “Cohen-Macaulay”. This turns out not
to always be the case, as the following example shows.
Example 6.3. Let G and H be the Cohen-Macaulay circulant graphs G = C2 (1) and
H = C5 (1). Then G[H] = C10 (1, 4, 5) and H[G] = C10 (1, 2, 3, 5) as seen in Figure 4.
We can compute α(G) for the graphs in Figure 4 by inspection; on the other hand,
Figure 4. Lexicographical products C2 [C5 ] and C5 [C2 ]
we compute the projective dimension using Macaulay 2 [9]. We find that α(G[H]) =
dim(R/I(G[H])) = 2 = n − pdim(R/I(G[H])), so G[H] is Cohen-Macaulay. However,
α(H[G]) = dim(R/I(H[G])) = 2 > n − pdim(R/I(H[G])) = 1, so H[G] is not CohenMacaulay.
In light of the above example, we can ask what conditions on G and H allow us to
conclude that the lexicographical product G[H] is Cohen-Macaulay.
We end with a question concerning Lemma 3.3. Using Macaulay 2 [9], we found that
d−1
dimk H̃2 (∆; k) = 4d+3
for d = 1, . . . , 14. This suggests that the inequality of Lemma
3
2
3.3 is actually an equality. We wonder if this is indeed true.
Acknowledgements. We thank Brydon Eastman for writing the LATEX code to produce
circulant graphs, and Russ Woodroofe and Jennifer Biermann for useful discussions.
References
[1] J.-C. Bermond, F. Comellas, D.F. Hsu, Distributed loop computer networks: a survey. J. Parallel
Distrib. Comput. 24 (1995), 2–10.
[2] F. Boesch, R. Tindell, Circulants and their connectivities. J. Graph Theory 8 (1984), 487–499.
[3] J. Brown, R. Hoshino, Independence polynomials of circulants with an application to music. Discrete
Math. 309 (2009), 2292–2304.
[4] J. Brown, R. Hoshino, Well-covered circulant graphs. Discrete Math. 311 (2011), 244–251.
[5] G. Davis, G. Domke, 3-Circulant Graphs. J. Combin. Math. Combin. Comput. 40 (2002), 133–142.
COHEN-MACAULAY CIRCULANT GRAPHS
14
[6] J. Herzog, T. Hibi, Monomial ideals. Graduate Texts in Mathematics 260, Springer-Verlag, 2011.
[7] J. Herzog, T. Hibi, Distributive lattices, bipartite graphs and Alexander duality. J. Algebraic Combin.
22 (2005), 289–302.
[8] J. Herzog, T. Hibi, X. Zheng, Cohen-Macaulay chordal graphs. J. Combin. Theory Ser. A 113
(2006), 911–916.
[9] D. R. Grayson, M. E. Stillman, Macaulay 2, a software system for research in algebraic geometry.
http://www.math.uiuc.edu/Macaulay2/.
[10] E. Miller, B. Sturmfels, Combinatorial Commutative Algebra. Graduate Texts in Mathematics, 227,
Springer, 2005.
[11] S. Morey, R. Villarreal, Edge ideals: algebraic and combinatorial properties. In Progress in Commutative Algebra 1: Combinatorics and Homology, edited by Christopher Francisco, Lee C. Klingler,
Sean Sather-Wagstaff, Janet C. Vassilev, (2012) de Gruyter, Berlin/Boston, 85–126.
[12] M. Plummer, Well-covered graphs: a survey. Quaestiones Math. 16 (1993), 253–287.
[13] J. Provan, L. Billera, Decompositions of simplicial complexes related to the diameters of convex
polyhedra. Math. Oper. Res. 5 (1980), no. 4, 576–594.
[14] V.N. Sachkov, V.E. Tarakanov, Combinatorics of Nonnegative Matrices. In: Translations of Mathematical Monographs, 213, American Mathematical Society, Providence, 2002.
[15] R. Stanley, Combinatorics and Commutative Algebra, Second Edition. Birkhäuser, Boston (1983).
[16] J. Topp, L. Volkmann, On the well-coveredness of products of graphs. Ars Combin. 33 (1992),
199–215.
[17] R. H. Villarreal, Cohen-Macaulay graphs. Manuscripta Math. 66 (1990), 277–293.
[18] R. H. Villarreal, Monomial algebras. Monographs and Textbooks in Pure and Applied Mathematics,
238. Marcel Dekker, Inc., New York, 2001.
Department of Mathematics, Redeemer University College, Ancaster, ON, L9K 1J4,
Canada
E-mail address: [email protected]
Department of Mathematical Sciences, Lakehead University, Thunder Bay, ON, P7B
5E1, Canada
E-mail address: [email protected]
Department of Mathematics, Redeemer University College, Ancaster, ON, L9K 1J4,
Canada
E-mail address: [email protected]
| 0math.AC
|
arXiv:1802.02474v1 [cs.MS] 12 Jan 2018
High-level python abstractions for optimal checkpointing in
inversion problems
Navjot Kukreja∗
Jan Hückelheim
Michael Lange
Imperial College London
London, UK
Imperial College London
London, UK
Imperial College London
London, UK
Mathias Louboutin
Andrea Walther
Simon W. Funke
The University of British Columbia
Vancouver, BC, Canada
Universität Paderborn
Paderborn, Germany
Simula Research Laboratory
Lysaker, Norway
Gerard Gorman
Imperial College London
London, UK
ABSTRACT
1
Inversion and PDE-constrained optimization problems often rely
on solving the adjoint problem to calculate the gradient of the objective function. This requires storing large amounts of intermediate
data, setting a limit to the largest problem that might be solved with
a given amount of memory available. Checkpointing is an approach
that can reduce the amount of memory required by redoing parts
of the computation instead of storing intermediate results. The
Revolve checkpointing algorithm offers an optimal schedule that
trades computational cost for smaller memory footprints. Integrating Revolve into a modern python HPC code and combining it with
code generation is not straightforward. We present an API that
makes checkpointing accessible from a DSL-based code generation
environment along with some initial performance figures with a
focus on seismic applications.
Seismic inversion is a computationally intensive technique that
uses data from seismic wave propagation experiments to estimate
physical parameters of the earth’s subsurface. A seismic inversion
problem based on a wave equation can be viewed as an optimization problem and numerically solved using a gradient-based optimization [16]. Since the gradient is usually calculated using the
adjoint-state method, the method requires that the forward and
adjoint field are known for each time step of the simulation [9]. We
discuss this in section 2.
Previous work on similar inverse problems led to the Revolve
algorithm [4] and the associated C++ tool which provides an optimal schedule at which to store checkpoints, i.e. states from which
the forward simulation can be restored. A study of optimal checkpointing for seismic inversion was done in [14] but this was not
accompanied by a high-level abstraction that made integration
of other software easier with Revolve. The algorithm is further
discussed in section 3.
The Revolve tool and algorithm, however, only provide the schedule to be used for checkpointing. Although this eases some of the
complexity of the application code using the algorithm, the glue
code required to manage the forward and adjoint runs is still quite
complex. This acts as a deterrent to the more widespread use of the
algorithm in the community.
In this paper, we describe how the Revolve algorithm can be
combined with code generation to make checkpointing much more
accessible. The software that can enable this is described in section
4. Although we use particular examples from seismic imaging, the
abstraction and software proposed here are quite general in nature
and can be used in any problem that requires checkpointing in
combination with a variety of computational methods.
In section 5 we provide some initial performance figures on
which we judged the correctness and performance of the implementation.
CCS CONCEPTS
• Software and its engineering → Software design engineering;
KEYWORDS
HPC, Code generation, API, Checkpointing, Adjoint, Inverse Problems
ACM Reference format:
Navjot Kukreja, Jan Hückelheim, Michael Lange, Mathias Louboutin, Andrea
Walther, Simon W. Funke, and Gerard Gorman. 2017. High-level python
abstractions for optimal checkpointing in inversion problems. In Proceedings
of SuperComputing, Denver, Colorado, USA, November 2017 (SC17), 8 pages.
https://doi.org/10.475/123_4
∗ Corresponding
Author. Email: [email protected]
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).
SC17, November 2017, Denver, Colorado, USA
© 2017 Copyright held by the owner/author(s).
ACM ISBN 123-4567-24-567/08/06. . . $15.00
https://doi.org/10.475/123_4
2
INTRODUCTION
SEISMIC IMAGING AND DEVITO
Seismic imaging techniques exploit the principle that a traveling wave carries information about the physical properties of the
medium it travels through. While different techniques focus on
SC17, November 2017, Denver, Colorado, USA
Kukreja, N. et al
where δ d = dsyn − dobs is the data residual, J is the Jacobian of
the forward operator, u is the forward wavefield and vt t is the
second-order time derivative of the adjoint wavefield.
It can be seen that the evaluation of the gradient first requires the
simulation of the forward and adjoint wavefields. This is achieved
by modeling the wave equation using a discretization, usually finite
difference. Various forms of the wave equation exist, e.g. acoustic
isotropic, anisotropic - VTI/TTI and elastic. Each of these models the
physics to different levels, with corresponding levels of complexity.
Here, we focus on the acoustic equation although the analysis
applies to all the forms of the equation mentioned.
In the discrete form, the acoustic wave equation from [7] can be
written as the following linear system:
A(m)u = PTs q
Figure 1: Graphical demonstration of a seismic experiment
that produces the data used as input in a seismic imaging
workflow (Source: Open University [1])
different kinds of information and objectives, we focus here on
reverse-time migration (RTM, [3, 15]), an imaging method that relies on a good estimate of the velocity model to obtain an image
of the reflectors in the subsurface. The algorithm relies on a datafitting procedure where synthetic data dsyn is computed with the
current estimate of the physical model via a wave-equation solve
and compared to the field measured data dobs . An example of a
field data recording is illustrated in Figure 1. This problem is a leastsquare minimization. We introduce here the formulation of the
problem solved and justify the implementation of optimal checkpointing. We previously introduced Devito [7], a finite-difference
domain specific language (DSL) for time-dependent PDE solvers.
Devito provides symbolic abstractions to define the forward and
adjoint wavefields. We will not go through its implementation here
but concentrate on the computation of the image of the subsurface.
Seismic imaging, in our case reverse-time migration (RTM), provides an image of the subsurface reflectors from field recorded data
and a cinematically correct smooth background velocity model. In
practice, the recording is repeated with different source/receivers
pair (called experiments) over the same physical region. An estimate of the physical parameters m is obtained from the recorded
data with different methods such as full-waveform inversion (FWI,
iterative RTM for low frequencies). Once m is estimated, RTM provides an image of the subsurface to be interpreted. As just stated,
RTM is a single gradient of the FWI objective that can be written
as [6, 9, 16]:
minimize Φs (m) =
m
1
2
dsyn − dobs 2
2
(1)
The square slowness model m is a physical property of the
medium through which the wave is propagating. The gradient
of the objective function Φs (m) with respect to the square slowness
m is given by:
∇Φs (m) =
nt
Õ
t =1
u[t]vt t [t] = JT δ d
(2)
(3)
where A is the discretized wave-equation and Ps is the sourcerestriction operator. The wavefield u is then given by:
u = A−1 (m)PTs q
(4)
Although equation 4 can provide the value of u for the entire domain at every time step, explicitly formulating the entire matrix u is
prohibitively expensive in terms of computer memory required and
is avoided wherever possible. When doing forward-only simulations, of interest is the value of u at certain predetermined locations
in the simulated domain that we call receivers. We record the progression of u at these locations through time. This is represented
mathematically by applying the restriction operator Pr at the required receiver locations. The result of applying Pr to u, which we
call the simulated data, is given by:
dsyn = Pr A−1 (m)PTs q
(5)
We can now rewrite the objective function from equation 1 as:
2
1
Pr A−1 (m)PTs q − dobs
(6)
2
2
In equation 2 we can define the Jacobian of the forward operator
minimize Φs (m) =
m
as:
dPr A−1 (m)PTs
(7)
dm
While the u term in this equation can be calculated from equation 4,
the v term can be calculated from its adjoint equation given as:
J=
AT (m)v = PTr δ d.
2.1
(8)
Implementation
As the forward wavefield is obtained as a time-marching procedure
forward in time, the adjoint wavefield is then obtained similarly
with a backward in time time-marching procedure. The procedure
to derive an image with RTM can be summarized as:
(1) Compute the synthetic data dsyn with a forward solve with
equation (5).
(2) Compute the adjoint wavefield from the data residual with
equation (8).
(3) Compute the gradient as the correlation of the forward and
adjoint wavefield with equation (2).
For the first step, we know from [8] that equation 4 can be
modeled in Devito using the Operator defined in figure 2
High-level python abstractions for optimal checkpointing in inversion problems
def forward(model, m, eta, src, rec, order=2):
# Create the wavefield function
u = TimeData(name='u', shape=model.shape,
time_order=2, space_order=order)
# Derive stencil from symbolic equation
eqn = m * u.dt2 - u.laplace + eta * u.dt
stencil = solve(eqn, u.forward)[0]
update_u = [Eq(u.forward, stencil)]
# Add source injection and receiver interpolation
src_term = src.inject(field=u, expr=src * dt**2
/ m)
rec_term = rec.interpolate(expr=u)
# Create operator with source and receiver terms
return Operator(update_u + src_term + rec_term,
subs={s: dt, h: model.spacing})
SC17, November 2017, Denver, Colorado, USA
def gradient(model, m, eta, src, rec, order=2):
# Create the adjoint wavefield function
v = TimeData(name='v', shape=model.shape,
time_order=2, space_order=order)
gradient_update = Eq(grad, grad - u.dt2 * v)
# The adjoint equation
eqn = m * v.dt2 - v.laplace - eta * v.dt
stencil = solve(eqn, v.backward)[0]
eqn = Eq(v.backward, stencil)
# Add expression for receiver injection
ti = v.indices[0]
receivers = rec.inject(field=v, expr=rec * dt *
dt / m)
return Operator([eqn] + [gradient_update] +
receivers, subs={s:dt, h:
model.get_spacing()}, time_axis=Backward)
Figure 2: Devito code required for a forward operator
The Operator thus created can be used to model a forward
where the src_term contains the source to be injected into the
field and rec_term will be extracting the receiver information from
the simulation.
Clearly, the third step requires the intermediate data from both
the previous steps. Storing both the forward and adjoint wavefields
in memory would be naive since two complete wavefields need to
be stored. The first obvious optimization is to merge steps 2 and 3
as a single pass where step 3, the gradient calculation, can use the
output from step 2 (the adjoint wavefield) as it is calculated, hence
saving the need for storing the adjoint wavefield in memory. The
Operator for the combined steps 2 and 3 can be created in Devito
using the code given in figure 3.
This still leaves the requirement of having the result of step 1
available. The most efficient, from a computational point of view,
would be to store the full history of the forward wavefield during
the first step. However, for realistically sized models, it would require TeraBytes of direct access memory. One solution would be
to store the field on disk but would lead to slow access memory
usage making it inefficient. This memory limit leads to checkpointing, storing only a subset of the time history, then recomputing it
during the adjoint propagation. Revolve provides an optimal schedule for checkpointing to store for a given model size, number of
time steps and available memory. The next section discusses how
checkpointing is implemented.
3
REVOLVE
As we have seen, the usage of adjoint methods allows the computation of gradient information within a time that is only a very small
multiple of the time needed to evaluate the underlying function
itself. However, for nonlinear processes like the one we saw in the
previous section, the memory requirement to compute the adjoint
information is in principle proportional to the operation count of
Figure 3: Devito code required for an operator that calculates
the adjoint and gradient in a single pass
the underlying function, see, e.g., [5, Sec. 4.6]. In Chap. 12 of the
same book, several checkpointing alternatives to reduce this high
memory complexity are discussed. Checkpointing strategies use a
small number of memory units (checkpoints) to store the system
state at distinct times. Subsequently, the recomputation of information that is needed for the adjoint computation but not available is
performed using these checkpoints in an appropriate way. Several
checkpointing techniques have been developed, all of which seek
an acceptable compromise between memory requirement and runtime increase. Here, the obvious question is where to place these
checkpoints during the forward integration to minimize the overall
amount of required recomputations.
To develop corresponding optimal checkpointing strategies, one
has to take into account the specific setting of the application. A
fixed number of time steps to perform and a constant computational
cost of all time steps to calculate is the simplest situation. It was
shown in [4] that, for this case, a checkpointing scheme based on
binomial coefficients yields, for a given number of checkpoints,
the minimal number of time steps to be recomputed. An obvious
extension of this approach would be to include flexibility with respect to the computational cost of the time steps. For example, if
one uses an implicit time stepping method based on the solution of
a nonlinear system, the number of iterations needed to solve the
nonlinear system may vary from time step to time step, yielding
non-uniform time step costs. In this situation, it is no longer possible to derive an optimal checkpointing strategy beforehand. Some
heuristics were developed to tackle this situation [11]. However,
extensive testing showed that, even in the case of nonuniform step
costs, binomial checkpointing is quite competitive. Another very
important extension is the coverage of adaptive time stepping. In
SC17, November 2017, Denver, Colorado, USA
r=new Revolve(steps,snaps)
do
whatodo = r->revolve()
switch(whatodo)
case advance: for r->oldcapo < i <= r->capo
forward(x,u)
case firsturn: eval_J(x,u)
init(bu,bx)
adjoint(bx,bu,x,u)
case youturn: adjoint(bx,bu,x,u)
case takeshot: store(x,xstore, r->check)
case restore: restore(x,xstore, r->check)
while(whatodo <> terminate)
Figure 4: revolve algorithm with calls to the application interface
this case, the number of time steps to be performed is not known beforehand. Therefore, so-called online checkpointing strategies were
developed, see, e.g., [13, 18]. Finally, one has to take into account
where the checkpoints are stored. Checkpoints stored in memory
can be lost on failure. For the sake of resilience or because future
supercomputers may be memory constrained, checkpoints may
have to necessarily be stored to disk. Therefore, the access time
to read or write a checkpoint is not negligible in contrast to the
assumption frequently made for the development of checkpointing
approaches. There are a few contributions to extend the available
checkpointing techniques to a hierarchical checkpointing, see, e.g.,
[2, 10, 12].
The software revolve implements binomial checkpointing, online
checkpointing as described in [13], and hierarchical, also called
multi-stage, checkpointing derived in [12]. For this purpose, it
provides a data structure r to steer the checkpointing process and
the storage of all information required for the several checkpointing
strategies.
To illustrate the principle structure of an adjoint computation
using checkpointing, Fig. 4 illustrates the kernel of revolve used
for the binomial checkpointing. The two remaining checkpointing
strategies are implemented in a similar fashion only taking the
additional extensions into account. The forward integration as well
as the corresponding adjoint computation is performed within a
do-while-loop of the structure in Fig. 4, where steps and snaps
denote the number nt of time steps of the forward simulation and
the number c of checkpoints available for the adjoint computation,
respectively.
Hence, the routine revolve determines the next action to be
performed which must by supported by the application being differentiated. These actions are
• advance: Here, the user is supposed to perform a part of
the forward integration based on the routine forward(x,u),
where x represents the state of the system and u the control.
The variable r–>oldcapo contains the current number of the
state of the forward integration. That is, before starting the
for-loop x holds the state at time t r–>oldcapo . The variable
Kukreja, N. et al
r–>oldcapo determines the targeted number of the state of
the forward integration. Therefore, r–>capo − r–>oldcapo
time steps have to be perform to propagate the state x from
the time t r–>oldcapo to the time t r–>capo
• firstrun: This action signals the start of the adjoint computation. Therefore, first the target function is evaluated. Then,
the user has the possibility to initialize the adjoint variable
bu and bx. Subsequently, the first adjoint step is performed.
• youturn: The next adjoint step has to be performed.
• takeshot: Here, the user is supposed to store the current
state x in the checkpoint with the number r–>check. The
array of checkpoints is here denoted by xstore but the specific organisation of the checkpoints is completely up to
the user. During the adjoint computation r–>check selects
the checkpoint number appropriately such that all states
needed for the adjoint computation are available. Once the
adjoint computation has started, states that were stored in
the checkpoints are also overwritten to reuse memory.
• restore: The content of the checkpoint with the number
r–>check has to be copied into the state x to recompute the
forward integration starting from this state.
It is important to note that this checkpointing approach is completely independent from the method that is actually used to provide
the adjoint information. As can be seen, once an adjoint computation is available the implementation can incorporate binomial
checkpointing to reduce the memory requirement.
We also have to stress that revolve provides a so-called serial
checkpointing which means that only one forward time step or
one adjoint step is performed at each stage of the adjoint computation. Nevertheless, the computation of the forward time step and/or
the adjoint step may be performed heavily in parallel, i.e., may be
evaluated on a large scale computer system. This is in contrast to
so-called parallel checkpointing techniques where several forward
time steps might be performed in parallel even in conjunction with
one adjoint step. Corresponding optimal parallel checkpointing
schedules were developed in [17]. However, so far no implementation to steer such a parallel checkpointing process is available.
The revolve software also includes an adjust procedure that computes, for a given number of time steps, the number of checkpoints
such that the increase in spatial complexity equals approximately
the increase in temporal complexity. Using the computed number
as the number of checkpoints minimises cost when assuming that
the user pays computational resources per node and per time, e.g.
the cost is proportional to the available memory and the runtime
of the computation.
4
ABSTRACTIONS FOR CHECKPOINTING
In this section, we will discuss the package pyRevolve , which has
been developed during the course of this work to encapsulate Revolve checkpointing in a user-friendly, high level python library.
This library is available online, along with its source 1 . We will first
provide details of its implementation in Section 4.1. Afterwards, we
will discuss the interplay of pyRevolve with the C++ checkpointing implementation that was previously discussed in Section 3.
Finally, we will discuss the usage of pyRevolve in an application in
1 https://github.com/opesci/pyrevolve
High-level python abstractions for optimal checkpointing in inversion problems
Section 4.2 to RTM, as described in section 2, implemented in the
Devito domain specific language.
Although section 4.2 discusses the special case of devito, the
interface of the pyRevolve library was designed to allow an easy
integration into other python codes as well.
SC17, November 2017, Denver, Colorado, USA
Revolve
Original C++
CRevolve
4.1
API, pyRevolve side
The pyRevolve interface was designed as part of providing checkpointing to users of Devito with an accessible API. The design had
the following goals:
• Making checkpointing available to users of Devito without
forcing them to get involved in implementation details like
loops, callbacks, data storage mechanisms. The user should
choose whether to use checkpointing in one place, but not
be forced to do anything beyond this.
• All knowledge of checkpointing, different strategies (online/offline checkpointing, multistage) shall be contained
within one module of the python framework, while still benefiting all operations in the code.
• The checkpointing itself should be contained in a separate
library that allows others to use it easily, even if they are not
interested in using Devito. This matured into pyRevolve .
• Since the data movement requires intricate knowledge of
the data structures used and their organization in memory,
this is handled by the application code.
To achieve these goals, pyRevolve was designed for the following
overall workflow, which will be explained in more detail in the following sections. The term application here refers to the application
using pyRevolve as a library (in this case Devito). To begin with, the
application creates objects with an apply method to perform the
actual forward and reverse computations, which are both instances
of a concrete implementation of the abstract base classOperator.
The application also creates an instance of a concrete implementation of the abstract base class Checkpoint that can deep-copy all
time-dependent working data that the operators require into a specified memory location. Next, the application instantiates pyRevolve
’s Revolver object and passes the forward and reverse operators,
and the checkpoint object. When required, the application starts
the Revolver’s forward sweep, which will complete the forward
computation and store checkpoints as necessary. After the forward
sweep completes, the application can finalize any computation
that is based on the forward data, such as evaluation of objective
functions, or store the final result as necessary. This may be accompanied/followed by the initialization of the adjoint data structures.
After this, the application calls the Revolver’s reverse sweep. This
will compute the adjoint, possibly by performing partial forward
sweeps and loading checkpoint data.
The pyRevolve package contains crevolve, which is a thin C
wrapper around a previously published C++ implementation2 . The
C++ files in this package are slightly modified for compatibility
with Python, but the original is available from the link in the footnote. The crevolve wrapper around the C++ library is taken from
libadjoint3 .
2 http://www2.math.uni-paderborn.de/index.php?id=12067&L=1
3 https://bitbucket.org/dolfin-adjoint/libadjoint
C wrapper
pyRevolve
Python wrapper
Devito
Application
Figure 5: Packages overview. Devito and an example application that uses checkpointing are the subject of Section 2. Revolve has been described in Section 3. The packages pyRevolve and cRevolve and how they are used to create
a high-level abstraction of checkpointing are explained in
Section 4.
One key design aspect is that pyRevolve is not responsible for
performing the data copies, and therefore does not need to know
about the properties or structure of the data that needs to be stored.
For this purpose, pyRevolve provides the Checkpoint abstract base
class that has a size attribute, and a load(ptr) and save(ptr)
method. The user must provide a concrete implementation of such
an object. The size attribute must contain the size of a single
checkpoint in memory, this information is used by pyRevolve to
allocate the correct amount of memory. The save method must
deep-copy all working data to the memory region starting at the
provided pointer, and the load method must restore the working
data from the memory region starting at the pointer ptr, either
by performing a deep-copy, or by pointing the computation to the
existing data inside the checkpoint storage.
The pyRevolve library provides to the user the class Revolver
that must be instantiated with the following arguments:
• Checkpoint object: This has to be an implementation of
the abstract base class Checkpoint.
• Forward operator: An object that provides a function apply()
as specified in Section 4.2 that performs the forward computation.
• Reverse operator: Similarly, an object that provides a function apply() that performs the reverse computation.
• Number of checkpoints: This is optional, and specifies the
number of checkpoints that can be stored in memory. If it
is not given, a default value is computed using the adjust
method explained in Section 3.
SC17, November 2017, Denver, Colorado, USA
Kukreja, N. et al
≪enum≫
Action
Crevolve
advance
takeshot
restore
firstrun
youturn
error
+ capo : uint
+ oldcapo : uint
+ check : uint
+revolve() : Action
Figure 6: Crevolve classes.
# Example time-varying field that needs checkpointing
u = TimeData(...)
# Some expression that generates the values for u
fw = Operator(...)
# Some expression that uses the values of u
rev = Operator(...)
cp = DevitoCheckpoint([u])
revolver = Revolver(cp, fw, rev, nt)
# Forward sweep that will pause to take checkpoints
revolver.apply_forward()
# Could perform some additional steps here
# Reverse sweep that uses the checkpoints
revolver.apply_reverse()
Checkpoint
# size : int
Figure 8: Devito code to utilize checkpointing based on pyRevolve
# save(ptr : pointer) : void
# load(ptr: pointer) : void
4.2
1
Storage
- data : numpy_array
+ init(n_ckp : int, size_ckp : int)
+ get_item(i : uint) : pointer
1
Revolver
- op_forward : Operator
- op_reverse : Operator
- store : Storage
- cr : CRevolve :: Checkpointer
- ckp : Checkpoint
+ init(ckp, op_forward, op_reverse, n_ckp, n_time)
+ apply_forward() : void
+ apply_reverse() : void
Operator
# apply() : void
Figure 7: pyRevolve classes. The abstract classes Checkpoint
and Operator are implemented by the client application
• Number of time steps: This is also optional. If it is not
given, an online checkpointing algorithm is used.
Based on either the given or computed number of checkpoints,
the constructor instantiates a storage object that allocates the necessary amount of memory (number of checkpoints × checkpointsize),
and makes that memory accessible to the Revolver.
API, Application side
To introduce checkpointing using the pyRevolve library, an application must implement a particular interface. We use devito here
as an example application, however, everything discussed here is
fairly general and any application may implement checkpointing
using pyRevolve using the approach discussed in this section.
To begin with, a concrete implementation of pyRevolve ’s abstract
base class Checkpoint, called DevitoCheckpoint was created. This
class has three methods:
• save(ptr): Save the contents of the working memory into
the location ptr.
• restore(ptr): Restore a previously stored checkpoint from
location ptr into working memory.
• size: Report the amount of memory required by a single
checkpoint. This is used to decide the total amount of memory to be allocated and to calculate offsets.
Along with this, we set up two Operator, one a ForwardOperator
to carry out the forward computation and a GradientOperator
that computes the image, as explained in section 2. These can be
used to initialize a Revolver object as shown in figure 8.
On initialization of the Revolver, object, the DevitoCheckpoint
object is queried for the size of one checkpoint and pyRevolve allocates n_checkpoints*checkpoint.size bytes of memory for
the storage of checkpoints. Calling revolver.apply_forward()
carries out a forward run but broken down into chunks as specified
by the checkpointing schedule provided by Revolve, each chunk
being executed by calling fwd.apply() with arguments t_start
and t_end corresponding to the timesteps to run the simulation for.
Between these chunks, cp.save() is automatically called to save
the state to a checkpoint.
On calling revolver.apply_reverse(), the Revolver calls
rev.apply() with the relevant t_start/t_end arguments for the
sections where the result from the forward pass is available in a
checkpoint. This will be loaded by a call to cp.load. For others,
it will automatically call fwd.apply() to recompute and store in
memory the results from a part of the forward operator so the
reverse operator can be applied for that part.
High-level python abstractions for optimal checkpointing in inversion problems
SC17, November 2017, Denver, Colorado, USA
DevitoCheckpoint
+ size : int
+ save(ptr : pointer) : void
+ load(ptr: pointer) : void
ForwardOperator
ReverseOperator
- data
- data
+ apply() : void
+ apply() : void
Figure 10: Velocity model used for gradient test
Figure 9: Devito operators, and the implementation of a
Checkpoint class.
Following this work, users of Devito can easily add optimal
checkpointing to their adjoint computations by following the steps
described above.
5
EXPERIMENT
There are two possible ways of testing the numerical accuracy of
an implementation - solving a problem whose solution has certain
known mathematical properties and verifying these properties numerically, and comparing the results to a reference solution. Here
we do both - we use the gradient test as described in [7] and also
verify that the numerical results match those from a reference implementation. The test uses the Taylor property of the gradient to test
whether the calculated gradient follows the expected convergence
for small perturbations. The test can be written mathematically as:
Figure 11: Timings for gradient test for different amounts of
peak memory consumption
ϵ0 =Φs (m0 + hdm) − Φs (mt )
ϵ1 =Φs (m0 + hdm) − Φs (m0 ) − h⟨∇Φs (m0 ), dm⟩.
(9)
where Φs is defined in equation 1. This test is carried out for
a certain m, which we call here m0 . This is the smoothed version
of a two layer model mt i.e. the true model has two horizontal
sections, each with a different value of squared slowness. Figure ??
shows the true velocity model mt . The measured data required by
the objective function is modeled on the true two-layer model and
dm = m0 − mt . The constant h then varies between 10−1 and 10−4
to verify that ϵ0 = O(h) is a first order error and ϵ1 = O(h 2 ) is
a second order error. The code used for this test can be found in
the repository for devito 4 . The tests were carried out on a Intel(R)
Xeon(R) CPU E5-2640 v3 @ 2.60GHz (Haswell) with 128GB RAM.
We used a grid with 230 × 230 × 230 points. With the simulation
running for 1615 timesteps, this required about 80 GB to store
the full forward wavefield in memory. The first run was made
with the regular gradient example that stores the entire forward
wavefield in memory. This was then repeated with checkpointing,
with varying number of checkpoints. It was verified that the results
4 https://github.com/opesci/devito
from all versions matched each other exactly and also passed the
gradient test mentioned previously. The peak memory usage was
tracked for each such run, as well as the total time to solution. The
memory consumption was measured using the memory_profiler
python module and the time to solution by using the time python
command before and after the function to be profiled. To eliminate
variation in the results, the timings are the minimum value from
three runs.
As can be seen in figure 11, the reduction in runtime as more
memory is available is not in line with the theoretical predictions
from Griewank and Walther [4]. This is expected because the theoretical numbers do not take into account the cost of deep copies
(implemented here using numpy) as well as the cost of repeatedly
calling a C function from python vs doing the repetition inside the
C function. Since the adjoint computation (as well as the associated
forward computation) is carried out one time-step at a time, this
significantly reduces the amount of work available inside a single
Operator call. This might cause inefficiencies in load-balancing
across cores using OpenMP. This effect is seen most clearly when
comparing the reference implementation that stores the forward
SC17, November 2017, Denver, Colorado, USA
field in a contiguous block of memory with the checkpointed implementation that stores a checkpoint at every time step. In this case,
although the memory consumption of the two implementations is
the same, the checkpointed implementation runs slower because
of the overheads previously mentioned.
6
CONCLUSIONS
Through this work, we have shown that with high-level abstractions it is possible to greatly simplify the complexity of client code.
We have also verified the correctness of our implementation using
mathematical tests. This already enables the users of Devito to
utilize Revolve based checkpointing in their applications to solve
much bigger problems than previously possible. However, through
the experiment in section 5, we have seen that the overhead introduced by checkpointing is non-trivial. For this reason, there is
much more work to be done to implement more features that widen
the applicability of pyRevolve .
7
FUTURE WORK
This work carried out so far was a proof of concept of integration
with Revolve using high level abstractions and, as such, is still a
work-in-progress in terms of use for practical applications. The most
important limitation in the current implementation is that it implements “serial checkpointing”, i.e. during the reverse computation,
only one timestep can be advanced at a time and this severely limits
the parallelizability of this code. The high-level interface would
need to be extended to be able to manage parallelization strategies.
Another important feature that might be required in pyRevolve before it is adopted in the community is multi-stage checkpointing.
Here, some checkpoints may be transparently swapped to disk,
further increasing the amount of memory available to applications
without any change in the application code. For problems implemented with adaptive time-stepping, the number of time-steps is
not known a-priori and that would require pyRevolve to implement
online checkpointing, something that even Devito would require
in future versions.
ACKNOWLEDGMENTS
The authors are very grateful to Fabio Luporini and Nicolas Barral.
This work was carried out as part of the Intel Parallel Computing
Centre at Imperial College, London.
REFERENCES
[1] 2017. (August 2017). http://www.open.edu/openlearn/science-maths-technology/
science/environmental-science/earths-physical-resources-petroleum/
content-section-3.2.1
[2] G. Aupy, J. Herrmann, P. Hovland, and Y. Robert. 2016. Optimal Multi-stage
algorithm for Adjoint Computation. SIAM Journal on Scientific Computing 38, 3
(2016), C232–C255.
[3] Edip Baysal, Dan D. Kosloff, and John W. C. Sherwood. 1983. Reverse time
migration. Geophysics 48, 11 (1983), 1514–1524. https://doi.org/10.1190/1.1441434
arXiv:http://geophysics.geoscienceworld.org/content/48/11/1514.full.pdf
[4] Andreas Griewank and Andrea Walther. 2000. Algorithm 799: revolve: an implementation of checkpointing for the reverse or adjoint mode of computational
differentiation. ACM Transactions on Mathematical Software (TOMS) 26, 1 (2000),
19–45.
[5] A. Griewank and A. Walther. 2008. Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation. SIAM.
[6] Eldad Haber, Matthias Chung, and Felix J. Herrmann. 2012. An effective method
for parameter estimation with PDE constraints with multiple right hand sides.
SIAM Journal on Optimization 22, 3 (7 2012). http://dx.doi.org/10.1137/11081126X
Kukreja, N. et al
[7] Navjot Kukreja, Mathias Louboutin, Felippe Vieira, Fabio Luporini, Michael Lange,
and Gerard Gorman. 2016. Devito: automated fast finite difference computation.
In Domain-Specific Languages and High-Level Frameworks for High Performance
Computing (WOLFHPC), 2016 Sixth International Workshop on. IEEE, 11–19.
[8] Michael Lange, Navjot Kukreja, Fabio Luporini, Mathias Louboutin, Charles
Yount, Jan Hückelheim, and Gerard J Gorman. 2017. Optimised finite difference
computation from symbolic equations. arXiv preprint arXiv:1707.03776 (2017).
[9] R-E Plessix. 2006. A review of the adjoint-state method for computing the gradient
of a functional with geophysical applications. Geophysical Journal International
167, 2 (2006), 495–503.
[10] M. Schanen, O. Marin, and H. Zhang amd M. Anitescu. 2015. Asynchronous TwoLevel Checkpointing Scheme for Large-Scale Adjoints in the Spectral-Element Solver
Nek5000. Technical Report ANL/MCS-P5422-1015. Argonne National Laboratory.
[11] Julia Sternberg and Michael Hinze. 2010. A memory-reduced implementation of
the Newton-CG method in optimal control of nonlinear time-dependent PDEs.
Optimization Methods & Software 25, 4 (2010), 553–571.
[12] P. Stumm and A. Walther. 2009. Multi-stage Approaches for Optimal Offline
Checkpointing. SIAM Journal of Scientific Computing 31, 3 (2009), 1946–1967.
[13] Philipp Stumm and Andrea Walther. 2010. New algorithms for optimal online
checkpointing. SIAM Journal on Scientific Computing 32, 2 (2010), 836–854.
[14] William W. Symes. 2007. Reverse time migration with optimal checkpointing.
GEOPHYSICS 72, 5 (2007), SM213–SM221. https://doi.org/10.1190/1.2742686
arXiv:https://doi.org/10.1190/1.2742686
[15] Albert Tarantola. 1984. Inversion of seismic reflection data in the acoustic
approximation. GEOPHYSICS 49, 8 (1984), 1259–1266. https://doi.org/10.1190/1.
1441754 arXiv:https://doi.org/10.1190/1.1441754
[16] Jean Virieux and Stéphane Operto. 2009. An overview of full-waveform inversion
in exploration geophysics. Geophysics 74, 6 (2009), WCC1–WCC26.
[17] Andrea Walther. 2004. Bounding the number of processes and checkpoints
needed in time-minimal parallel reversal schedules. Computing 731, 2 (2004),
135–154.
[18] Qiqi Wang, Parviz Moin, and Gianluca Iaccarino. 2009. Minimal Repetition
Dynamic Checkpointing Algorithm for Unsteady Adjoint Calculation. SIAM
Journal on Scientific Computing 31, 4 (2015/05/03 2009), 2549–2567. https://doi.
org/10.1137/080727890
| 5cs.CE
|
A Fast Parameterized Algorithm for Co-Path Set
Blair D. Sullivan, Andrew van der Poel
Department of Computer Science, North Carolina State University,
Raleigh, NC 27695
arXiv:1603.04376v3 [cs.DS] 27 Jul 2016
{blair sullivan,ajvande4}@ncsu.edu
January 4, 2018
Abstract
The k-Co-Path Set problem asks, given a graph G and a positive
integer k, whether one can delete k edges from G so that the remainder
is a collection of disjoint paths. We give a linear-time fpt algorithm
with complexity O∗ (1.588k ) for deciding k-Co-Path Set, significantly
improving the previously best known O∗ (2.17k ) of Feng, Zhou, and Wang
(2015). Our main tool is a new O∗ (4tw(G) ) algorithm for Co-Path Set
using the Cut&Count framework, where tw(G) denotes treewidth. In
general graphs, we combine this with a branching algorithm which refines
a 6k-kernel into reduced instances, which we prove have bounded treewidth.
1
Introduction
We study parameterized versions of Co-Path Set [3, 16], an NP-complete
problem asking for the minimum number of edges whose deletion from a graph
results in a collection of disjoint paths (the deleted edges being a co-path set –
see Figure 1). Specifically, we are concerned with k-Co-Path Set, which uses
the natural parameter of the number of edges deleted.
k-Co-Path Set
Input: A graph G = (V, E) and a non-negative integer k.
Parameter: k
Problem: Does there exist F ⊆ E of size exactly k such that
G[E \ F ] is a set of disjoint paths?
These problems are naturally motivated by determining the ordering of
genetic markers in DNA using fragment data created by breaking chromosomes
with gamma radiation (a technique known as radiation hybrid mapping) [4,13,15].
Unfortunately, human error in distinguishing markers often means the constraints
implied by markers’ co-occurrence on fragments are incompatible with all possible
linear orderings, necessitating an algorithm to find the “best” ordering (that
violates the fewest constraints). Co-Path Set solves the special case where
each DNA fragment contains exactly two genetic markers (corresponding to an
edge in the graph); any linear ordering of the markers must correspond to some
1
Figure 1: Three co-path sets (dashed edges), including one of minimum size
(rightmost).
set of paths, and we minimize the number of unsatisfied constraints (edges in
the co-path set).
Recent algorithmic results related to Co-Path Set include a (10/7)-approximation
algorithm [2], and two parameterized algorithms deciding k-Co-Path Set [6, 7],
the faster of which [7] has time complexity1 O∗ (2.17k ). However, as written,
both parameterized results [6,7] contain a flaw in their analysis which invalidates
their probability of a correct solution in the given time2 . The best known bound
prior to [6] is an O∗ (2.45k ) algorithm [16]. In this paper, we prove:
Theorem 1. k-Co-Path Set is decidable in O∗ (1.588k ) linear-fpt time with
probability at least 2/3.
We note that standard amplification arguments apply, and Theorem 1 holds
for any success probability less than 1. Further, if f is an increasing function
with limn→∞ f (n) = 1, we can solve k-Co-Path Set with success probability
at least f (n) in O(1.588k npolylog(n)).
The remainder of this paper is organized as follows: after essential definitions
and notation in Section 2, we start in Section 3 by giving a new O∗ (4tw(G) )
algorithm tw-copath for solving Co-Path Set parameterized by treewidth (tw)
using the Cut&Count framework [5]. Finally, Section 4 describes the linear-fpt
algorithm referenced in Theorem 1, which solves k-Co-Path Set on general
graphs in O∗ (1.588k ) by applying tw-copath to a set of “reduced instances”
generated via kernelization and a branching procedure3 deg-branch.
2
Preliminaries
Let G(V, E) be the graph with vertex set V and edge set E. Unless otherwise
noted, we assume |V | = n; we let N (v) denote the set of neighbors of a vertex v,
and let deg(v) = |N (v)|. Given a graph G(V, E) and F ⊆ E, we write G[F ] for
the graph G(V, F ).
Our tw-copath algorithm in Section 3 uses dynamic programming over a
tree decomposition, and its running time depends on the related measure of
treewidth [14], which we denote tw(G). To simplify the dynamic programming,
1 Throughout this paper, we use the notation O ∗ (f (k)) for the fpt (fixed-parameter tractable)
complexity O(f (k)nO(1) ); we say an algorithm is linear-fpt if the complexity is O(f (k)n).
2 Step 2.11 in both versions of Algorithm R-MCP checks if a candidate co-path set F has
size ≤ k1 (as they are sweeping over all possible sizes of candidates and want to restrict the size
accordingly). If F is too large, the algorithm discards it and continues to the next iteration.
However, in order for their analysis to hold, the probability that the candidate is contained
in a co-path set must be ≥ (1/2.17)k1 (or (1/2.29)k1 in [6]) for every iteration. Candidates
which are too large may have significantly smaller probability of containment, yet are counted
in the exponent of the analysis.
3 The properties of our reduced instances guarantee we can find a tree decomposition in
poly(k) time.
2
we will use a variant of nice tree decompositions [5, 10] where each node in the
tree has one of five specific types: leaf, introduce vertex, introduce edge, forget
vertex, or join. The “introduce edge” nodes are labelled with an edge uv and
have one child (with an identical bag); we require each edge in E is introduced
exactly once. Additionally, we enforce that the root node is of type “forget
vertex” (and thus has an empty bag). A tree decomposition can be transformed
into a nice decomposition of the same width in time linear in the size of the
input graph [5].
When describing the dynamic programming portion of the algorithm we use
Iverson’s bracket notation: if p is a predicate we let JpK be 1 if p is true and 0
otherwise. We also use the shorthand f [x → y] to denote updating a function f
so that f (x) = y and all other values are unchanged.
Finally, we use fast subset convolution [1] to reduce the complexity of handling
join nodes in the nice tree decomposition (Section 3). This technique maps
functions of the vertices in a join bag to values in Zp = Z/pZ (where p is chosen
based on the application). The key complexity result we rely on uses the Zp
product, which is defined below. We write ZB
p for the set of all vectors t of
length |B| assigning a value t(b) ∈ Zp to each element of b ∈ B.
Definition 1 (Zp product). Let p ≥ 2 be a fixed integer and let B be a finite set.
For t1 , t2 , t ∈ ZB
p we say that t1 + t2 = t if t1 (b) + t2 (b) = t(b) (in Zp ) for all
p
b ∈ B. For a ring R and functions f, g : ZB
p → R, define the Zp product, ∗x as
P
(f ∗px g)(t) =
f (t1 )g(t2 ).
t1 +t2 =t
Fast subset convolution guarantees that certain Zp products can be computed
quickly.
Lemma 1 (Cygan et al. [5]). Let R = Z or R = Zq for some constant q. The Z4
product of functions f, g : ZB
4 → R can be computed in time and ring operations
4|B| |B|O(1) .
3
An O∗ (4tw(G) ) Algorithm via Cut&Count
We start by giving an fpt algorithm for Co-Path Set parameterized by treewidth.
Our primary tool is the the Cut&Count framework, which enables ctw nO(1) onesided Monte Carlo algorithms for connectivity-type problems with constant
probability of a false negative. Cut&Count has previously been used to improve
the best-known bounds for several well-studied problems, including Connected
Vertex Cover, Hamiltonian Cycle, and Feedback Vertex Set [5].
Pilipczuk showed that an O∗ (ctw ) algorithm for some constant c can be designed
with the Cut&Count approach for Co-Path Set because the problem can be
expressed in the specialized graph logic known as ECML+C [12]. However, since
our end goal is to improve on existing algorithms for k-Co-Path Set in general
graphs using a bounded treewidth kernel, we need to develop a specialized
dynamic programming algorithm with a small value of c. We show:
Theorem 2. There exists a one-sided fpt Monte Carlo algorithm tw-copath
deciding k-Co-Path Set for all k in a graph G in O∗ (4tw(G) ) time with failure
probability ≤ 1/3.
3
The Cut&Count technique has two main ingredients: an algebraic approach
to counting which uses arithmetic in Z2 (enabling faster algorithms) alongside
a guarantee that undesirable objects are seen an even number of times (so a
non-zero result implies a desired solution has been seen), and the idea of defining
the problem’s connectivity requirement through consistent cuts. In this context,
a consistent cut is a partitioning (V1 , V2 ) of the vertices of a graph into two sets
such that no edge uv has u ∈ V1 and v ∈ V2 and all vertices of degree 0 are in V1 .
Since each connected component must lie completely on one side of any consistent
cut, we see that a graph G has exactly 2cc(G)−nI (G) such cuts, where cc(G) is the
number of connected components and nI (G) is the number of isolated vertices. In
order to utilize parity with the number of consistent cuts, we introduce markers,
which create even numbers of consistent cuts for graphs that are not collections of
disjoint paths. Our counting algorithm tw-copath, which computes the parity of
the size of the collection of subgraphs with consistent cuts which adhere to specific
properties pertaining to Co-Path Set, employs dynamic programming over a
nice tree decomposition. We further use weights and the Isolation Lemma to
bound the probability of a false negative arising from multiple valid markings of a
solution. We use fast subset convolution [1] to reduce the complexity required for
handling join bags in the dynamic programming. In the remainder of this section,
we present the specifics for applying these techniques to solve Co-Path Set.
3.1
Cutting
We first provide formal definitions of markers and marked consistent cuts, which
we use to ensure that sets of disjoint paths are counted exactly once in some
entry of our dynamic programming table.
Definition 2. A triple (V1 , V2 , M ) is a marked consistent cut of a graph G
if (V1 , V2 ) is a consistent cut and M ⊆ E(G[V1 ]). We refer to the edges in M
as the markers. A marker set is proper if it contains at least one edge in each
non-isolate connected component of G.
Note that if a marker set is proper, all vertices are on the V1 side of the cut.
This is because by the definition of a consistent cut, all isolates are on the V1
side, and if every connected component contains a marker then all connected
components must fall entirely on the V1 side as well. Therefore for any proper
marker set there exists exactly one consistent cut, while all marker sets which
are not proper will be paired with an even number of consistent cuts because
unmarked components may lie in V1 or V2 . We use proper marker sets to
distinguish desired subgraphs by assigning markers in such a way that when we
prune the dynamic programming table for solutions (as described later in the
section), the only subgraphs we consider which may have a proper marker set
are collections of disjoint paths. We know because the marker set is proper that
the subgraph has a unique consistent cut, and thus these collections of disjoint
paths will only be counted once in some entry of the dynamic programming
table, while all other subgraphs will be counted an even number of times. Note
that we are not claiming that all collections of disjoint paths will have proper
marker sets.
We refer to the complement of a co-path set (the edges in the disjoint paths) as
a cc-solution, and call it a marked-cc-solution when paired with a proper marker
set of size exactly equal to its number of non-isolate connected components. While
4
cc-solutions can be viewed as solutions due to their complementary nature, being
marked is crucial in our counting algorithm and thus subgraphs which are markedcc-solutions are what correspond to solutions in the dynamic programming table.
We now describe our use of the Isolation Lemma, which P
guarantees we are
able to use parity to distinguish solutions. Let f (X) denote x∈X f (x).
Isolation Lemma ([11]). Let F ⊆ 2U be a non-empty set family over universe
U . A function ω : U → Z is said to isolate F if there is a unique S ∈ F with
ω(S) = minF ∈F ω(F ). Assign weights ω : U → {1, 2, ..., N } uniformly at random.
Then the probability that ω isolates F is at least 1 − |U |/N .
Intuitively, if F is the set of solutions (or complements of solutions) to an
instance of Co-Path Set and |F| is even, then tw-copath would return a false
negative. This is because while each solution is counted an odd number of times
in tw-copath, because there are an even number of solutions the total count
of solutions is even, making the combined count of solutions and non-solutions
even and the algorithm would incorrectly determine a solution does not exist
(a false negative). The Isolation Lemma allows us to partition F based on the
weight of each solution (as assigned by ω), and guarantees at least one of the
partition’s blocks has odd size with constant probability. We let U contain
two copies of every edge e ∈ E: one representing e as a marker and one as an
edge in the cc-solution. Then 2U denotes all pairs of edge subsets (potential
marked-cc-solutions), and we set N = 3|U | = 6E (selected to achieve success
probability in Theorem 1). Each copy of an edge is assigned a weight in [1, N ]
uniformly at random by ω and the probability of finding an isolating ω is thus
2/3. We denote the values assigned by ω to the set of marker copies by ωM , and
likewise to the set of edge-in-cc-solution copies by ωE .
3.2
Counting
A marked-cc-solution C of a graph G corresponds to a co-path set of size k when
the number of edges and markers in C match specific values which depend on
k and |E(G)|. These values are easily deduced because we know the deletion
of a co-path set solution of size k will leave |E(G)| − k edges in a cc-solution.
Furthermore, because a forest has n − m connected components, the number of
markers in C needs to be at most |V (C)| − |E(G)| + k. All isolates from a forest
can be removed and the resulting graph is still a forest, and thus the actual
number of markers necessary in C is |V (C)| − nI (C) − |E(G)| + k.
We now describe a dynamic programming (DP) algorithm over a nice tree
decomposition which returns mod 2 the number of appropriately sized markedcc-solutions in the root’s subtree (for a fixed k). Since no-instances have no
appropriately sized marked-cc-solutions, and yes-instances have at least one, odd
parity for the number of marked-cc-solutions of size corresponding to k implies
a solution to the k-Co-Path Set instance must exist.
During the DP algorithm we actually count (for all values (m, e)) the number
of cc-candidates, which are subgraphs G0 ⊆ G with maximum degree 2, exactly
e edges, and a marked consistent cut with m markers. The following lemma
justifies counting cc-candidates in place of marked-cc-solutions.
Lemma 2. The parity of the number of marked-cc-solutions in G with e edges
and weight w is the same as the parity of the number of cc-candidates G0 ⊆ G
with e edges, |V (G0 )| − e − nI (G0 ) markers, and weight w.
5
Proof. Consider a subgraph G0 ⊆ G with maximum degree 2 and e edges. Let
M 0 be a marking of G0 such that ωE (E(G0 )) + ωM (M 0 ) = w. Assume first
that G0 is a collection of paths. We know that G0 has |V (G0 )| − e − nI (G0 )
non-isolate connected components. If M 0 is a proper marker set of G0 , then
|M 0 | = |V (G0 )|−e−nI (G0 ) and (G0 , M 0 ) has exactly one consistent cut. Therefore
(G0 , M 0 ) contributes one to both the number of marked-cc-solutions and the
number of cc-candidates, respectively.
If otherwise M 0 is not a proper marker set, then (G0 , M 0 ) contains an unmarked connected component and has an even number of consistent cuts, and
therefore contributes an even number to the count of cc-candidates and zero
to the number of marked-cc-solutions. Finally, if G0 contains at least one cycle
then cc(G0 ) > |V (G0 )| − e − nI (G0 ). Therefore at least one connected component
does not contain a marker, and the number of consistent cuts is even, so the
contribution to the count of cc-candidates is again even and the contribution
to the count of marked-cc-solutions is zero. We conclude that the parity of the
number of marked-cc-solutions and the parity of the number of cc-candidates is
the same.
Our dynamic programming algorithm is a bottom-up approach over a nice
tree decomposition. We build cc-candidates for all values of m and e (encoding
the option to add/not add edges and select/not select edges as markers), and
keep track of various parameters ensuring that when pruning the DP table we
only consider cc-candidates which could be valid solutions to the k-Co-Path
Set instance. We use the number of edges to ensure our solution is of the
correct size, and the number of markers and non-isolate vertices to determine
when a subgraph is acyclic. The weight parameter allows us to distinguish
between solutions and decreases the likelihood of a false negative occurring via
the Isolation Lemma. Finally, we need a parameter that encodes the degree
information required to properly combine cc-candidates as we iterate up the tree.
We call this parameter a degree-function and define on the vertices V of a bag as
Variable
a
e
m
w
Parameter
Maximum value
# of non-isolated vertices
# of edges
# of markers
weight of edges and markers
n
n2
n2
4n4
Table 1: Dynamic programming table parameters and upper bounds.
f : V → Σ = {0, 11 , 12 , 2}, where f (v) corresponds to v’s degree in the associated
cc-candidates of the table entry — for vertices of degree 1, their value 1j denotes
which side of the partition (V1 , V2 ) they are on. Vertices with degree 0 are on the
V1 side of the cut by definition so we need not keep track of their side of the cut.
Similarly degree 2 vertices cannot have additional incident edges, thus the side of
the cut they fall on will not matter for selecting markers. In summary, we have
table entries Ax (a, e, m, w, s) counting the number of cc-candidates at bag x
with a non-isolated vertices, e edges, m markers, weight w, and degree-function
s.
In the following description of the dynamic programming algorithm over
a nice tree decomposition T , we let z1 , z2 denote the children of a join node;
6
otherwise, the unique child is denoted y.
Leaf :
Ax (0, 0, 0, 0, ∅) = 1; Ax (a, e, m, w, s) = 0 for all other inputs.
Introduce vertex v:
Ax (a, e, m, w, s) = Js(v) = 0KAy (a, e, m, w, s).
Introduce edge uv:
Ax (a, e, m, w, s) = Ay (a, e, m, w, s) +
X
Jφ2 (αu , αv )KAy (a0 , e − 1, m, w0 , s0 )
αt ∈subs(s(t))
t∈{u,v}
+
X
Jφ1 (αu , αv )K Ay (a0 , e − 1, m, w0 , s0 ) + Ay (a0 , e − 1, m − 1, w00 , s0 ) ,
αt ∈subs(s(t))
t∈{u,v}
where φj (αu , αv ) = (αu = 1j ∨ s(u) = 1j ) ∧ (αv = 1j ∨ s(v) = 1j ), a0 =
a − (|{11 , 12 } ∩ {s(u), s(v)}|), w0 = w − ωE (uv), w00 = w − ωE (uv) − ωM (uv),
s0 = s[u → αu , v → αv ], and the subs function returns all the values the degreefunction in child node y could have assigned to vertices u and v based on current
degree-function s (summarized below).
s(v)
0
11
12
2
subs(s(v)) ∅
0
0
{11 , 12 }
We now argue this formula’s correctness. The term Ay (a, e, m, w, s) handles
the case when uv is excluded from the cc-solution. We handle the case when
uv is added to the cc-solution by iterating over all possible subs values for each
endpoint, only considering counts in child y’s entries where u and v have the
appropriate subs values (preventing us from ever having a vertex with degree
greater than 2). Note that we use the φj function to guarantee that if s labels
u or v as an isolate, we do not use the introduced edge. We have a summation
for both possible j values in order to consider uv falling on either side of the cut.
The formulation of a0 assures that each endpoint of degree 1 is now included in
the count of non-isolates (i.e. when u and/or v had degree 0 in y). We utilize
the marker weight of uv to distinguish when we choose it as a marker (only if
on V1 side of cut), and increment m accordingly. In either case, we update w
appropriately (with w0 if no marker, w00 if marker introduced).
Forget vertex h:
X
Ax (a, e, m, w, s) =
Ay (a, e, m, w, s[h → α]).
α∈{0,11 ,12 ,2}
As a forgotten vertex can have degree 0, 1 or 2 in a cc-candidate, we must
consider all possible values that s assigns to h in child bag y. Note that cccandidates in which h is both not an isolate and not a member of a connected
component that contains a marker will cancel mod 2, as h can be on either side
of the cut and all parameters will be identical.
7
Join:
We compute Ax from Az1 and Az2 via fast subset convolution [1] taking care to
only combine table entries whose degree-functions are compatible.
Definition 3. At a join node x with children z1 and z2 , the degree-functions s1
from Az1 , s2 from Az2 , and s from Ax are compatible if one of the following
holds for every vertex v in x: (i) si (v) = 0 and sl (v) = s(v), i 6= l or (ii)
s1 (v) = s2 (v) = 1j and s(v) = 2 for i, j, l ∈ [1, 2].
In order to apply Lemma 1, we let B be the bag at x, and transform the values
assigned by the degree function s to values in Z4 . Let φ : {0, 11 , 12 , 2} → Z4 and
ρ : {0, 11 , 12 , 2} → Z be defined as in the table below, extending to vectors by
component-wise application.
φ
ρ
0
11
12
2
0
0
1
1
3
1
2
2
We use φ to apply Lemma 1, while the function ρ (which corresponds to a
vertex’s degree) is used in tandem to ensure the compatibility requirements are
met: if φ(s1 ) + φ(s2 ) = φ(s), then necessarily ρ(s1 ) + ρ(s2 ) ≥ ρ(s). From the
above table it is easy to verify that φ(s1 ) + φ(s2 ) = φ(s) and ρ(s1 ) + ρ(s2 ) = ρ(s)
together imply that s1 , s2 and s are compatible. We sum over both functions
when computing values for join nodes, to make sure that solutions from the
children are combined only when there is compatibility.
Assign P
t1 = φ(s1 ), t2 = φ(s2 ), and t = φ(s) in accordance with Lemma 1.
Let ρ(s) = v∈B ρ(s(v)); that is ρ(s) is the sum of the degrees of all the vertices
in the join node, as assigned by degree-function s. By defining functions f and
g as follows:
f hd,a,e,m,wi (φ(s)) = Jρ(s) = dKAz1 (a, e, m, w, s),
g hd,a,e,m,wi (φ(s)) = Jρ(s) = dKAz2 (a, e, m, w, s),
and writing r~i for the vector hdi , ai , ei , mi , wi i, we can now compute
X
Ax (a, e, m, w, s) =
(f r~1 ∗4x g r~2 )(φ(s))
r~1 +r~2 =hρ(s),a0 ,e,m,wi
−1
where a0 = a + |s−1
1 {11 , 12 } ∩ s2 {11 , 12 }|. We point out that
X
(f r~1 ∗4x g r~2 )(φ(s)) = 1
r~1 +r~2 =hρ(s),a0 ,e,m,wi
only if both φ(s1 ) + φ(s2 ) = φ(s) and ρ(s1 ) + ρ(s2 ) = ρ(s); that is, exactly when
s1 , s2 and s are compatible.
We conclude this section by describing how we search the DP table for
marked-cc-solutions at the root node r. By Lemma 2, the parity of the number
of marked-cc-solutions with |E| − k edges and weight w is the same as the parity
of the number of cc-candidates G0 with |E| − k edges, |V (G0 )| − (|E| − k) − nI (G0 )
markers and weight w. These candidates are recorded in the table entries
8
Ar (a, |E| − k, a − |E| + k, w, ∅), where a is the number of non-isolates. Therefore,
if there exists some a and w so that Ar (a, |E| − k, a − |E| + k, w, ∅) = 1, then we
have a yes-instance of k-Co-Path Set. Note that the degree-function is ∅ in
this entry because there are no vertices contained in the root node by definition.
By Lemma 1, the time complexity of tw-copath for a join node B is O∗ (4|B| ),
which is O∗ (4tw ). Note that for the other four types of bags, as we only consider
one instance of s per table entry, the complexity for each is O∗ (4tw ). We point
out that the size of the table is polynomial in n because there are a linear number
of bags and a polynomial number of entries (combinations of parameters) for
each bag. Since the nice tree decomposition has size linear in n, the bottomup dynamic programming runs in total time O∗ (4tw ). This complexity bound
combined with the correctness of tw-copath discussed above proves Theorem 2.
4
Achieving O∗ (1.588k ) in General Graphs
In order to use tw-copath to solve k-CoPath Set in graphs with unbounded
treewidth, we combine kernelization and a branching procedure to generate a
set of reduced instances – bounded treewidth subgraphs of the input graph G.
Specifically, we begin by constructing a kernel of size at most 6k as described
in [7]. Our reduced instances are bounded degree subgraphs of the kernel given
by a branching technique. We prove that (1) at least one reduced instance is an
equivalent instance; (2) we can bound the number of reduced instances; and (3)
each reduced instance has bounded treewidth. Finally, we analyze the overall
computational complexity of this process.
4.1
Kernelization and Branching
We start by describing our branching procedure deg-branch (Algorithm 1),
which uses a degree-bounding technique similar to that of Zhang et al. [16]. Our
implementation takes an instance (G, k) of Co-Path Set and two non-negative
integers ` and D, and returns a set of reduced instances {(Gi , k − `)} so that (1)
each Gi is a subgraph of G with exactly |E| − ` edges and maximum degree at
most D; and (2) at least one (Gi , k − `) is an equivalent instance to (G, k). The
size of the output (and hence the running time) of deg-branch depends on both
input parameters ` and D. We will select D to achieve the desired complexity in
copath in Section 4.3. We also make use of a budget parameter b, which keeps
track of how many more edges can be removed per the constraints of ` (b is
initially set to `).
Our branching procedure leverages the observation that if a co-path set S
exists, then every vertex has at most two incident edges not in S. Specifically, for
every vertex of degree greater than D, we branch on pairs of incident edges which
could remain after removing a valid co-path set (calling each pair a candidate),
creating a search tree of subgraphs.
Algorithm 1 returns the set of reduced instances, the size of which is at most
the number of leaves in the search tree of the branching process (inequality
can result from the algorithm discarding branches in which the number of
edits necessary to branch on a vertex exceeds the number of allowed deletions
remaining). We now give an upper bound on the size of this set.
9
Algorithm 1: Generating reduced instances
1
2
3
4
5
6
7
8
Algorithm deg-branch(G, k, `, D, b)
Let v be a vertex of maximum degree in G
if deg(v) ≥ D + 1 and b ≥ D − 1 then
Select vertices u1 , . . . uD+1 uniformly at random from N (v)
R = ∅, Ev = {{v, ui }}
for e1 , e2 ∈ Ev , e1 6= e2 do
Ev0 = Ev \ {e1 , e2 }
R = R ∪ deg-branch(G \ Ev0 , k, `, D, b − (D − 1))
9
10
11
return R
else if b = 0 and deg(v) ≤ D then return {(G, k − `)}
else return ∅
// Discard G
Lemma 3. Let T be a search tree formed by deg-branch(G, `, D, k, b). The
`/(D−1)
number of leaves of T is at most D+1
.
2
Proof of Lemma 3. The number of children of each interior node of T is D+1
,
2
D+1 depth(T )
resulting in at most
leaves. The depth of T is limited by the
2
second condition of the if on line 3 of Algorithm 1. For each recursive call, b is
decremented by (D − 1), until b ≤ D − 1. As b is initally set to `, this implies
depth(T ) ≤ `/(D − 1), proving the claim.
Finally, we argue that at least one member of the set of reduced instances
returned by deg-branch is equivalent to the original. Consider a solution F
to k-Co-Path Set in the original instance (G, k). Every vertex has at most
two incident edges in G[E\F ], and since all candidates are considered at every
high-degree vertex, at least one branch correctly keeps all of these edges.
4.2
Treewidth of Reduced Instances
Our algorithm deg-branch produces reduced instances with bounded degree;
in order to bound their treewidth, we make use of the following result, which
originated from Lemma 1 in [8] and was extended in [9].
Lemma 4. For > 0, there exists n ∈ Z+ s.t. for every graph G with n > n
vertices,
!
17
X
tw(G) ≤
ci ni + n≥18 + n,
i=3
where ni is the number of vertices of degree i in G for i ∈ {3, . . . , 17}, n≥18 is
the number of vertices of degree at least 18, and ci is given in Table 2. Moreover,
a tree decomposition of the corresponding width can be constructed in polynomial
time in n.
Since the structure of k-Co-Path Set naturally provides some constraints
on the degree sequence of yes-instances, we are able to apply Lemma 4 to our
reduced instances to effectively bound treewidth. We first find an upper bound
on the number of degree-3 vertices in any yes-instance of k-Co-Path Set.
10
d
cd
3
4
5
6
7
8
9
10
0.1667 0.3334 0.4334 0.5112 0.5699 0.6163 0.6538 0.6847
d
cd
11
12
13
14
15
16
17
0.7105 0.7325 0.7514 0.7678 0.7822 0.7949 0.8062
Table 2: Numerically obtained constants cd , 3 ≤ d ≤ 17, used in Lemma 4;
originally given in Table 6.1 of [9].
Lemma 5. Let ni be the number of vertices of degree i in a graph G for any
i ∈ Z+ , and ∆ be the maximum degree of G. If (G, k) is a yes-instance of
P∆
k-Co-Path Set, then n3 ≤ 2k − ( i=4 (i − 2)ni ).
Proof. Since (G, k) is a yes-instance, removing some set of at most k edges results
in a graph of maximum degree 2. For a vertex of degree j ≥ 3, at least j − 2
incident edges must be removed. Thus, n3 + 2n4 + 3n5 + . . . + (∆ − 2)n∆ ≤ 2k
(each removed edge counts twice – once for each endpoint).
Lemma 6. Let (G, k) be an instance of k-Co-Path Set such that G has n
vertices and max degree at most ∆ ∈ {3, . . . , 17}. Then the treewidth of G is
bounded by k/3 + n + c, for some constant c and any > 0. A tree decomposition
of the corresponding width can be constructed in polynomial time in n.
Proof. Let n be defined as in Lemma 4. Let G0 be the graph formed by adding
N = n isolates to G. By Lemma 4, because G0 has maximum degree at most
∆, tw(G0 ) ≤ (1/6)n3 + (1/3)n4 + . . . + c∆ n∆ + (N + n). We can substitute the
bound for n3 from Lemma 5, which yields:
P∆
2k − ( i=4 (i − 2)ni ) n4
+
+ . . . + c∆ n∆ + (N + n)
tw(G ) ≤
6
3
k
≤ + (n + N ).
3
0
Note
P∆ that the inequality holds because we can pair the negative terms of
( i=4 (i − 2)ni )/6 with the corresponding terms of n4 /3 + . . . + c∆ n∆ and
the value of cj nj − (j − 2)(nj )/6 is non-positive for all j ∈ [4, 17]. Since N = n
is a constant, we have tw(G0 ) ≤ k/3 + n + c. Since G ⊆ G0 and treewidth is
monotone under subgraph inclusion, this proves the claim.
We point out that when applying Lemma 6 to reduced instances, computing
the desired tree decomposition is polynomial in k (since they are subgraphs of a
6k-kernel).
4.3
The Algorithm copath
This section describes how we combine the above techniques to prove Theorem 1.
As shown in Algorithm 2, we start by applying 6k-kernel [7] to find G0 , a kernel
of size at most 6k; this process deletes k − k 0 edges. We then guess the number
of edges k1 ∈ [0, k 0 ] to remove during branching, and use deg-branch to create
a set of reduced instances Qk1 , each of which have k 0 − k1 edges. Note that
deg-branch considers all possible reduced instances, and thus if a (cc-)solution
11
exists, it is contained in at least one reduced instance. To ensure the complexity
of finding the reduced instances does not dominate the running time, we set the
degree bound D of the reduced instances to be 10 (any choice of 10 ≤ D ≤ 17 is
valid). By considering all possible values of k1 , we are assured that if (G, k) is a
yes-instance, some Qk1 contains a yes-instance. Each reduced instance is then
passed to tw-copath, which correctly decides the problem with probability 2/3.
Algorithm 2: Deciding k-Co-Path Set
1
2
3
4
5
6
7
Algorithm copath (G,k)
(G0 , k 0 ) = 6k-kernel(G, k)
for k1 ← 0 to k 0 do
Qk1 = deg-branch(G0 , k 0 , k1 , 10, k1 )
foreach (Gi , k2 ) ∈ Qk1 do
if tw-copath(Gi , k2 ) then return true
return false
Proof of Theorem 1. We now analyze the running time of copath, as given in
Algorithm 2. By Lemma 3, the size of each Qk1 is O(1.561k1 ). For each reduced
instance (Gi , k2 ) in Qk1 , we have tw(Gi ) ≤ k2 /3 + (6k) + c by Lemma 6.
Applying Theorem 2, tw-copath runs in time O∗ (4k2 /3+6k ) for each reduced
instance (Gi , k2 ) in Qk1 (with success probability at least 2/3). Each iteration
of the outer for loop can then be completed in time
O∗ (1.561k1 4k2 /3+6k ) = O∗ (4k/3+6k ) = O∗ (1.588k ),
where we use that k1 + k2 = k 0 ≤ k, and choose < 10−5 . Since this loop runs at
most k + 1 times, this is also a bound on the overall computational complexity
of copath. Additionally copath is linear-fpt, as the kernelization of [7] is O(n),
and the kernel has size O(k), avoiding any additional poly(n) complexity from
the tw-copath subroutine. Note that by Lemma 4 the tree decomposition can
be found in polynomial time in the size of the reduced instance. Since reduced
instances are subsets of 6k-kernels, the linearity is unaffected because the graph
has size polynomial in k.
5
Conclusion
This paper gives an O∗ (4tw ) fpt algorithm for Co-Path Set. By coupling this
with kernelization and branching, we derive an O∗ (1.588k ) linear-fpt algorithm
for deciding k-Co-Path, significantly improving the previous best-known result
of O∗ (2.17k ). We believe that the idea of combining a branching algorithm
which guarantees equivalent instances with bounds on the degree sequence from
the problem’s constraints can be applied to other problems in order to obtain
a bound on the treewidth (allowing treewidth-parameterized approaches to be
extended to general graphs).
One natural question is whether similar techniques extend to the generalization of Co-Path Set to k-uniform hypergraphs (as treated in Zhang et al. [16]).
It is also open whether the combined parameterization asking for a co-path set
of size k resulting in ` disjoint paths is solvable in sub-exponential fpt time.
12
Acknowledgements
This work supported in part by the Gordon & Betty Moore Foundation under DDD Investigator
Award GBMF4560 and the DARPA GRAPHS program under SPAWAR Grant N66001-14-14063. Any opinions, findings, and conclusions or recommendations expressed in this publication
are those of the author(s) and do not necessarily reflect the views of DARPA, SSC Pacific, or
the Moore Foundation. We thank two anonymous reviewers for providing a simplification of
our previous branching algorithm and pointing out the result from [9] enabling us to branch
on vertices with degree greater than 7. We also thank Felix Reidl for helpful suggestions on an
earlier draft that significantly improved the presentation of the results.
References
[1] A. Björklund, T. Husfeldt, P. Kaski, and M. Koivisto. Fourier meets Möbius: fast subset
convolution. In Proceedings of STOC, pages 67–74, 2007.
[2] Z. Chen, G. Lin, and L. Wang. An approximation algorithm for the minimum co-path set
problem. Algorithmica, 60(4):969–986, 2011.
[3] Y. Cheng, Z. Cai, R. Goebel, G. Lin, and B. Zhu. The radiation hybrid map construction
problem: recognition, hardness, and approximation algorithms. Unpublished Manuscript,
2008.
[4] D. Cox, M. Burmeister, E. Price, S. Kim, and R. Myers. Radiation hybrid mapping:
a somatic cell genetic method for constructing high-resolution maps of mammalian
chromosomes. Science, 250:245–50, 1990.
[5] M. Cygan, J. Nederlof, M. Pilipczuk, M. Pilipczuk, J. van Rooij, and J. Wojtaszczyk.
Solving connectivity problems parameterized by treewidth in single exponential time. In
FOCS, pages 150–159. IEEE, 2011.
[6] Q. Feng, Q. Zhou, and S. Li. Randomized parameterized algorithms for co-path set
problem. In FAW, pages 82–93. Springer, 2014.
[7] Q. Feng, Q. Zhou, and J. Wang. Kernelization and randomized parameterized algorithms
for co-path set problem. J. Comb. Optim., 2015. in press, DOI 10.1007/s10878-015-9901-y.
[8] F. Fomin, S. Gaspers, S. Saurabh, and A. Stepanov. On two techniques of combining
branching and treewidth. Algorithmica, 54(2):181–207, April 2009.
[9] S. Gaspers. Exponential Time Algorithms - Structures, Measures, and Bounds. VDM,
2010.
[10] T. Kloks. Treewidth, Computations and Approximations, volume 842 of LNCS. Springer,
1994.
[11] K. Mulmuley, U. Vazirani, and V. Vazirani. Matching is as easy as matrix inversion. In
Proceedings of STOC, pages 345–354. ACM, 1987.
[12] M. Pilipczuk. Solving connectivity problems parameterized by treewidth in single exponential time. In MFCS, pages 520–531, 2011.
[13] C. Richard III, D. Withers, T. Meeker, S. Maurer, G. Evans, R. Myers, and D. Cox. A
radiation hybrid map of the proximal long arm of human chromosome 11 containing the
multiple endocrine neoplasia type 1 (men-1) and bcl-1 disease loci. Am. J. Hum. Genet.,
49(6):1189–1196, 1991.
[14] N. Robertson and P. Seymour. Graph minors. II. Algorithmic aspects of tree-width. J.
Algorithms, 7(3):309–322, 1986.
[15] D. Slonim, L. Kruglyak, L. Stein, and E. Lander. Building human genome maps with
radiation hybrids. J. Comp. Biol., 4(4):487–504, 1997.
[16] C. Zhang, H. Jiang, and B. Zhu. Radiation hybrid map construction problem parameterized. Journal of Combinatorial Optimization, 27(1):3–13, 2014.
13
| 8cs.DS
|
A Partial-Order Approach
to Array Content Analysis
arXiv:1408.1754v1 [cs.PL] 8 Aug 2014
Graeme Gange1 , Jorge A. Navas2 , Peter Schachte1 , Harald Søndergaard1, and
Peter J. Stuckey1
1
2
Department of Computing and Information Systems,
The University of Melbourne, Victoria 3010, Australia
{gkgange,schachte,harald,pstuckey}@unimelb.edu.au
NASA Ames Research Center, Moffett Field, CA 94035, USA
[email protected]
Abstract. We present a parametric abstract domain for array content
analysis. The method maintains invariants for contiguous regions of the
array, similar to the methods of Gopan, Reps and Sagiv, and of Halbwachs and Péron. However, it introduces a novel concept of an array
content graph, avoiding the need for an up-front factorial partitioning
step. The resulting analysis can be used with arbitrary numeric relational abstract domains; we evaluate the domain on a range of array
manipulating program fragments.
1
Introduction
Most imperative programming languages offer mutable arrays. However, owing
to the indirect relation between storage and retrieval, arrays are not particularly
amenable to static analysis. While analysis of array bounds (e.g, [15,16,2]) is well
studied, only recently has there been real progress in analyzing array content.
Early approaches involved array smashing [1], where an entire array was treated
as a single symbolic variable. This used the concept of weak updates: transfer
functions for array assignment that can only weaken the previous abstract state.
Weak updates generally lead to a rapid loss of precision when different segments
of an array have different properties.
A significant improvement was Gopan, Reps and Sagiv’s use of array partitioning [7] to split arrays into symbolic intervals, or segments. Partitioning
is facilitated by an initial analysis of the array index expressions to determine
the relative order of the indices they denote. A key idea is to distinguish segments that represent single array cells from those that represent multiple cells.
This permits strong updates on the singleton segments. The array partitioning method selects a small set of partition variables, maintaining disjunctive
information about properties which hold over all feasible total orderings of the
partition variables.
Halbwachs and Péron [9] extended the
i1 := 0 ; . . . ; im := 0
approach to support relational content dox := ‹
mains and a limited form of quantified inwhile(i1 ă n ^ . . . ^ im ă n)
variants. The resulting method is precise,
p := ‹
but has its own drawbacks. First, it reif (p ă 0)
quires an initial segmentation phase, where
Ari1 s := x ` 1
the set of partition variables is identified;
i1 := i1 ` 1
and as this phase is purely syntactic, it is
else if (p ă 1)
possible for variables to be omitted which
Ari2 s := x ` 2
are critical to the invariant. Second, there
i2 := i2 ` 1
are exponentially many possible total or...
derings of the partition variables; if many
else
partition variables are identified, the analyArim s := x ` m
sis may become prohibitively expensive. For
im := im ` 1
example, on the init randm family of programs shown in Figure 1, the number of partitions at the loop head follows the progres- Fig. 1: The init randm family of
sion r6, 30, 222, 2190, 27006s as m increases program fragments.
from 1 to 5 (this is discussed in more detail
in Appendix C). Finally, the analysis does
not support arbitrary manipulation of index variables; indices may only be incremented and decremented.
Cousot et al. [3] instead maintain a single partitioning of the array, selecting
a consistent totally ordered subset from a scalar variable analysis. This does not
require a separate segmentation phase, saving considerable overhead; however, as
it considers only a single consistent ordering (and supports only value domains)
the invariants it derives are quite weak. Consider init rand2 , with x fixed to
0—so each element in r0, nq will be assigned either 1 or 2. In this case, the
relationship between i1 and i2 is not known, so we must select either 0 ď i1 ă n
or 0 ď i2 ă n. In either case, the desired invariant at the loop exit is lost.
An alternative approach to expressing array properties is to lift an abstract
domain to quantified invariants [8]. This technique is quite general but there
are two major limitations. First, it requires from the user the specification of
templates to describe when quantifiers are introduced. Second, it is expensive,
owing to the computation of under-approximations. For example, to join the formulas @U pG ñ eq and @U pG1 ñ e1 q, we must compute an under-approximation
of G \ G1 , since G and G1 are in negative positions, and this is prohibitively
expensive for many domains.
Dillig et al. [6] replace strong and weak updates with fluid updates. Their
method is a points-to and value analysis, so not relational in our sense. It builds
a points-to graph where nodes represent abstract locations that include arrays
qualified by index variables. Edges represent constraints on index variables that
identify which concrete elements in the source location point to which concrete
location in the target. A fluid update removes the dichotomy between strong and
weak updates by computing first a constraint ϕ representing the elements that
2
are modified by the update. Then it adds a new points-to edge with ϕ (strong
update) while adding the negation of ϕ to existing edges from the source (weak
update). As ϕ is an over-approximation, its negation is an under-approximation
and thus it would be unsound to add it directly to other edges. Instead the
analysis produces bracketing constraints which are pairs of over- and underapproximations so that negation can be done in a sound manner. This analysis
is very expressive, avoiding the large number of explicit partitions fixed a priori
in [7,9]. However, the method can still be very expensive since whenever an array
is accessed, the points-to edges must be modified by adding possibly disjunctive
formulas.
We propose a new approach to array content analysis. We extend any existing scalar domain by introducing a pseudo-variable to refer to segments of
each array, selecting index expressions as nodes in a graph, and annotating the
graph edges with the properties that hold in the segments of the arrays between
those index expressions. These array content graphs offer greater flexibility than
other approaches, as they allow us to reason about properties that hold over contiguous array segments without committing to a single total ordering on index
expressions, while still taking advantage of available partial ordering information. The result is an array content analysis which is fully automatic, can be
used with arbitrary domains, and does not incur the up-front factorial cost of
previous methods [7,9]. In particular, it can be used for relational analyses, and
accounts for the possibility of array elements being related to array indices.
We base our presentation on a small control flow graph language.
Instructions
Array assignments
Jumps
Blocks
Programs
I
A
J
B
P
Ñ
Ñ
Ñ
Ñ
Ñ
v1 “ constant | v1 “ ˝ v2 | v1 “ v2 ˛ v3 | A
v1 “ arrrv2 s | arrrv1 s “ v2
if pv1 ’ v2 q label 1 label 2 | br label | error | end
label : I* J
B`
Each basic block is a (possibly empty) sequence of instructions, ending in a
(possibly conditional) jump. Arithmetic unary and binary operators are denoted
by ˝ and ˛ respectively, and comparison operators by ’. We assume that there is
a fixed set of arrays tA1 , . . . , Ak u, which have global scope (and do not overlap in
memory). The semantics is conventional and not discussed here. Figure 2 shows
an example program in diagrammatic form.
Our analysis assumes an abstract domain L “ xL, Ď, K, J, \, [y for analysis
over the array-free fragment of the language (obtained by leaving out A) (scalar
analysis). We use this parametric domain to construct the array content analysis.
The remainder of this paper is structured as follows. Section 2 introduces
the method and its underlying ideas. Section 3 discusses computational details
and efficiency. The method has been evaluated experimentally; Section 4 gives
a report and Section 5 concludes, suggesting further work.
3
2
A Graph-Based Array Content Domain
We let V and A be sets of scalar and array variables, respectively. A state in
the concrete domain is a pair xσ, ρy, where σ : V ÞÑ Z maps scalar variables to
integer values, and ρ : A Ñ Z Ñ Z maps array cells to values.3
Let VA denote the set of variables, and CA
the set of constants, which may act as segment
bounds. We use an extended set of variables
head :
V 1 “ VA Y CA Y tv ` | v P VA Y CA u. The
i“0
pseudo-variable v ` represents the value v ` 1.
br guard
This allows us to talk about properties that
apply to single array cells. The constant vertices are often needed; because array properguard :
ties often hold over ranges bounded by a conif pi ă N q body tail
stant on one side, and array processing code
often initializes a fixed set of indices before
body :
processing the rest of the array, it is not suffitail :
v “ Aris
cient to consider only those variables and conend
Bris “ v
stants used directly as indices. Consider the
i“i`1
copy program shown in Figure 2. In this case
br guard
neither 0 nor N is ever used directly as an index, but both define boundary conditions of
i, and are needed for the invariant of interest. Fig. 2: Copying array A to B
In practice, however, VA is often considerably
smaller than V .
We wish to relate the value of elements in an array segment to variables in
the scalar domain. A state in the abstract domain has the form xϕ, Ψ y, where
ϕ expresses scalar properties and Ψ expresses array content properties. For each
array A P A, we allocate a corresponding variable a, a segment variable, which
occurs only in Ψ , never in ϕ. Relations between scalar and segment variables
are captured in Ψ . We use U “ ta, b, . . .u to denote the set of segment variables.
Sometimes, we may wish to relate the values in an array segment to the corresponding index, for example, to prove that Aris “ x ` i across a segment. To
support this, we introduce a variable idx to represent the index of a given read.
We use UI “ tidxu Y U to denote the augmented set of segment variables.
The analyses of the scalar domain and array contents are based on the same
lattice L. We represent the program state as a pair of the scalar properties ϕ and
the |V 1 | ˆ |V 1 | matrix Ψ of array properties such that ψij denotes the properties
which hold for all indices in the interval ri, jq. In a slight abuse of notation, we
use ψpℓq to denote the formula ψ with each symbolic array variable a replaced
with the corresponding array element Arℓs. That is:
ψpℓq ” ψta ÞÑ Arℓs | A P Au
3
For simplicity we assume arrays elements are integers. The extension to arbitrary
types (that may include integers) is not difficult, as the complexity of array elements
acting as indices is present in what we consider.
4
J
J
ϕ“0ďiăN
Ψ“
0
i
i`
N
0 i
i`
Ka“b J
K K a“v
K K
K
K K
K
N
J
J
J
K
a“b
0
a“v
i
J
`
i
N
i`
N
J
a“b
0
a“v
i
Fig. 3: Array content graph after assignment v “ Aris in Figure 2. Vertices and
matrix entries corresponding to 0` and N ` are omitted.
Then the reading of an edge pi, ψij , jq is: @ i ď ℓ ă j . ψij pℓq. Given some numeric
abstract domain L, a set of arrays A and scalar variables V , the array content
domain CL pA, V 1 q is a pair xϕ, Ψ y, where ϕ is a value in L, and Ψ is a |V 1 | ˆ |V 1 |
matrix of L-values. Assume we have a function evalA which constructs a new
state which treats the ℓth element of each array as a scalar variable:
evalA pσ, ρ, ℓq “ σ Y ta ÞÑ ρpAqpℓq | A P Au
The concretization function γ is then defined on the components in terms of the
concretization function γL of the scalar domain L:
γpxϕ, Ψ yq “ txσ, ρy | σ P γL pϕqu X γA pΨ q
č
γA pΨ q “
txσ, ρy | @σpiqďℓăσpjq evalA pσ, ρ, ℓq P γL pψij qu
i,jPV 1
Notice that there are no constraints on ρ in the first equation; ρ can be any array
variable assignment of type A Ñ Z Ñ Z. As any value in the content domain
CL pA, V 1 q is the Cartesian product of a fixed set of elements of L, CL pA, V 1 q also
forms a lattice, and possesses all the corresponding fixed point properties.
For each edge ψij , we can assume the corresponding interval is non-empty;
that is, rri ă jss P ψij . Note that the edge from i to i` has no such constraint,
since i ă i`1 is always true. Since the interval ri`1, iq is clearly empty, ψi` i “ K.
Given the state representation, we can take the join of two abstract states
by a piecewise application of the join from L:
fi
» 1
2
1
2
+
*
. . . ψ1n
\ ψ1n
ψ11 \ ψ11
ffi
.
.
1
1
2
2
1
2 —
.
..
..
..
xϕ , Ψ y \xϕ , Ψ y “ ϕ \ ϕ , –
fl
1
2
1
2
ψn1 \ ψn1 . . . ψnn \ ψnn
We can compute the meet [ analogously. To see how the analysis works, consider
again the program in Figure 2. Figure 3 shows the abstract state immediately
after executing v “ Aris. The array content information is given by the matrix
5
Ψ of array properties. The array content graph shown upper right is really just
a way of visualizing the matrix. Note that infeasible edges, those labelled K,
are omitted. In fact, we shall usually show only the “transitive reduction” of
the array content graph, so that an edge ik whose value is given by ψik “
Ű
j pψij \ ψjk q is omitted. We depict edges representing J without a label. This
leads to the shorthand graph in Figure 3’s lower right.
2.1
Normalization of Abstract States
Given a set C of constraints of the form e P ri, jq ñ ψij , we wish to normalize
the state by computing the strongest consequences of C, still in that form.
The critical observation is this: @x, y . x ă y |ù @z.px ă zq _ pz ă yq. That
is, any property that holds over both rx, zq and rz, yq must also hold over the
range rx, yq. To compute the strongest consequences, then, we must compute
the greatest fixed point of a rewrite system derived from the set of inequalities:
@i, j, k . ψij Ď ψik \ ψkj .
It is tempting to try to compute the fixed point using the obhead :
vious rewrite system:
x “ Ar0s
i“1
br guard
@i, j, k . ψij “ ψij [pψik \ ψkj q
But, as we shall see, this formulation does not guarantee termination for all useful domains.
Mohri [12] describes the algebraic shortest path problem, where
the operations p[, \q form a
semiring in the domain of edge
weights. This is very close to what
we need, as every distributive lattice forms a bounded semiring. But
“numeric” domains used in static
analysis generally fail to be distributive, so we cannot use Mohri’s
framework directly.
guard :
if pi ă N q body tail
body :
v “ Aris
if pv ą xq update step
step :
i“i`1
br guard
update :
x“v
br step
tail :
end
Fig. 4: Find maximum value in array A
Example 1. Consider the array-maximum program from Figure 4. Figure 5(1)
shows the program state just after v “ Aris is executed (scalar constraints on
the left and array constraints on the right). On the branch with v ď x, we simply
add the constraint to the scalar domain, resulting in (2a). If v ą x, we add the
constraint to the scalar domain (2b), then update x with v, resulting in (2b1 ).
Observe that, in both cases, we can only discover the relationship between a and
x indirectly via v. If we do not push scalar relations into the edge properties,
the underlined invariants are lost. The final result is shown at (3).
If the statement x = v in update was instead replaced by x = Aris, we would
first have to lift the invariant x “ v from the singleton property ψii` to the
6
(1)
aďx
0ďiďN
a“v
0
0ďiďN
(2a)
vďx
(2b)
(2b1 )
0ďiďN
xăv
0ďiďN
x“v
0
i
a ď x, a ă v
0
i`
N
i`
N
a“v
i
a ă v, a ă x a “ v, a “ x
0
i
aďx
i`
N
a “ v, a ď x
0
0ďiďN
x“v
N
a “ v, a ď x
aďx
(3) 0 ď i ď N
(2b: )
i`
i
i
i`
N
a ă x, a ă v a “ v, a “ x
0
i
i`
N
Fig. 5: Array maximum example
scalar domain. We then push this property into the segment ψ0i , allowing us to
derive a ă x. The state in Figure 5(2b: ) shows the program state if x “ v were
to be replaced with x “ Aris.
This illustrates that it is not sufficient to simply compute the transitive closure
of Ψ ; we must also lift properties from Ψ out to the scalar domain. For a fully
reduced state xϕ, Ψ y, the following properties must be satisfied for all i, j, k:
1. The graph of segments must be internally consistent: ψij Ď ψik \ ψkj
2. Segment properties are consistent with the scalar domain: ψij Ď ϕ [rri ă jss
3. For each non-empty segment, the scalar domain must be consistent with the
scalar properties of that segment: ϕ |ù i ă j ñ ϕ Ď DUI . ψij
Notice that we only propagate constraints to the scalar component from segments
that are known to be non-empty. If we tried to propagate information from all
segments, we would incorrectly derive K as soon as any segment was determined
to be empty. One solution is to simply apply the three rules until a fixed point is
reached. This is guaranteed to compute the fully reduced state. However, while
this direct construction is conceptually clean, it suffers from some pragmatic
issues relating to both termination and efficiency, as we shall see.
2.2
Termination
The normalization process is not guaranteed to terminate for arbitrary lattices.
Example 2. Assume the analysis uses convex polyhedra. Consider the state in
Figure 6(a). Any fixed point will satisfy the properties A Ď B \ C and B Ď A \ C.
7
A
0
i
A
B
C
B
j
C
(a)
(b)
Fig. 6: At a fixed point, we have A Ď B \ C and B Ď A \ C. The regions A and
B will be progressively reduced, indefinitely.
Let A, B and C be the gray regions shown in Figure 6(b)—the intention is
that A shares a line segment with C, as does B. Assume we start by exploiting B Ď A \ C. We compute A \ C, yielding the polygon given by the topmost
dashed line. This allows us to trim the top portion of B. We then compute
B \ C, and trim the top-left region of A. However, now A has changed, so we
re-compute A \ C and again reduce B. This process asymptotically approaches
the greatest fixed point A “ A [ C, B “ B [ C.
If we modify the equations in Example 2 slightly, to A ĎpA [ Bq \pA [ Cq (still
a valid approximation of the concrete state), convergence is immediate.
The fixed point process will clearly terminate for the interval domain, as the
possible interval end-points are drawn from the initial set. We can also show that
it is guaranteed to terminate for both octagons [11] and convex polyhedra [4];
proofs are given in Appendix A. Unfortunately, we do not yet have a more general
characterisation of the lattices for which termination is (or is not) guaranteed.
2.3
Abstract Transfer Functions
In this section, we describe the abstract transfer functions necessary to perform
array content analysis on the language described in Section 1.
Variable assignment. The effect of a scalar assignment x “ f on an abstract
state follows the behaviour of the underlying domain L. We first project out the
previous value of x (assuming x does not occur in f ) then introduce the new
constraint into the scalar domain. However, when we project x from our scalar
domain, we must also update all incoming and outgoing edges of x (and x` ).
This becomes: xϕ, Ψ yrrx “ fss “ xϕrrx “ fss, Ψ 1 y where Ψ 1 is given by, for all p, q:
"
J
if p P tx, x` u _ q P tx, x` u
1
ψpq “
Dx . ψpq otherwise
The assumption that f is free of x is not always well founded; however, we can always transform the program so that it is the case: x “ f ñ xt “ x; x “ frx{xts.
8
0
(b)
a“b“v
a“b
(a)
i
b
i
0a “
b
i`
it a “ b “ v it
0a “
`
i`
it a “ b “ v it
N
it a “ b “ v it `
b
i`
N
it a “ b “ v it
`
a“b
(f)
0
N
b
b
a“ i
(e)
N
i
(c)
0a “
N
a“b“v
b
a“
0a “
i`
i`
i
(d)
i
i`
N
`
Fig. 7: Array content graph during analysis of the program given in Figure 2,
while executing i “ i ` 1. The dashed edges indicate equality (in reality they
represent two edges each labelled K).
Example 3. Consider the array-copy program given in Figure 2, immediately
before the assignment i “ i ` 1. The state is shown in Figure 7 (a).
We must first introduce a new variable it to hold the prior value of i, transforming i “ i ` 1 to it “ i; i “ it ` 1. The normalized graph after the it “ i
statement is shown in Figure 7 (b). To handle the i “ it ` 1 statement we
must eliminate the annotations on edges corresponding to i and i` , resulting in
state (c). Note that the edge ψ0it ` , omitted from the diagram, has annotation
a “ b. State (d) completes the handling of the i “ it ` 1 statement, introducing
the new value of i into the scalar domain. The scalar domain discovers that
it ` “ i (the dashed edge). When normalizing state (d), ψit ` i “ K, so the rule
ψ0i Ď ψ0it ` \ ψit ` i results in ψ0i becoming a “ b, the desired invariant (e), which
after projecting out it and i`
t gives (f).
Array reads. An array read x “ Aris is relatively simple. As for standard
variable assignment, we must existentially quantify the variable x. But instead
of introducing a relation into the scalar domain, we add the constraint x “
a to the singleton segment ψii` . This transfer function may be formulated as
xϕ, Ψ yrrx “ Arisss “ xDx.ϕ, Ψ 1 y where Ψ 1 is given by, for all p, q:
"
ψpq rra “ xss if p “ i, q “ i`
1
ψpq “
Dx . ψpq otherwise
Normalization handles the consequences for the scalar part.
Array writes. When we store a value into an array at index i, we update the
corresponding edge property ψii` . However, this is not sufficient, as the singleton
9
a“
S
a“0
0
`
a“
i
i
a“0 a“0
0
0
j
N
a“0
0
i
i`
j
N
0
0 K
a“
SrrAris “ 1ss
0
a“1
i
...
0
i`
...
a“
0
j
N
i K
i` K
j K
N K
a “ 0\a “ 1
0 i
i`
j N
Ka“0a“0a“0 J
K K a“0a“0 J
K K
K a“0 J
K K
K
K J
K K
K
K K
`
˛¨ j ˛N
¨ i
a“0
a“0
a “ 0 ˝ \ ‚˝ \ ‚ J
a “ 1 ¨a “ 1˛
a“0
K
a “ 1 ˝ \ ‚J
a“1
K
K
a“0 J
K
K
K
J
K
K
K
K
i
Fig. 8: In state S, we know that all elements between 0 and j have been initialized
to 0. When we evaluate Aris “ 1, we update the edge ψii` . However, i may also
be covered by the edge ψ0j . For these potentially overlapping edges, we must
perform a weak update, taking the join of the previous value with the new.
i may be covered by other edges. As with previous analyses, we distinguish
between strong updates where all elements in a (generally, singleton) segment
are updated with a given property, and weak updates [1] where some elements of
a segment may be updated. An edge ψpq must be updated if p ď i ď q ^ p ă q
is possible in the current state. This is possible if and only if the edges ψpi` and
ψiq are both feasible.
Consider the array state illustrated in Figure 8, where 0 ď i ď j ď N and the
first j elements have been initialized to 0. When we store 1 at Aris, we update
the singleton ψii` with the property a “ 1. However, there are other segments
that may contain Aris. The segments ψ0i` , ψij and ψ0j are all consistent with
index i. In this case, they must all be weakly updated (annotations on ψ0i` and
ψij have been omitted; they are identical to the annotation on ψ0j ).
We can formulate this as xϕ, Ψ yrrAris “ fss “ xϕ, Ψ 1 y, where Ψ 1 is given by,
for all p, q:
1
ψpq
$
&
ψpq rra “ fss
if p “ i, q “ i`
ψpq
if ϕ |ù pp ą i _ q ă i` q
“
%
ψpq \ ψpq rra “ fss otherwise
Notice that if we have some other variable j such that ϕ |ù i “ j, we will initially
perform only a weak update of the segment ψjj ` . However, the normalization
procedure will then enforce consistency between ψii` and ψjj ` .
10
3
Improving Efficiency through Relaxation
When computing the strongest matrix entries, we perform a substantial amount
of redundant work. We include all constraints ϕ from the scalar domain in each
matrix entry ψij , and these constraints will be processed during each step in
the shortest-path computation. This is not ideal, as many are irrelevant to the
content properties, and abstract domain operations are often proportional to
the number of constraints. Hence we want to construct some relaxation ψ:ij of
ψij that discards irrelevant scalar properties. We shall use d to denote this
relaxation operation (that is, ψ:ij “ ϕ d ψij ). At a minimum, it must satisfy:
ψij Ď ϕ [ ψ:ij “ ϕ [pϕ d ψij q
We want to make ϕ [ ψ:ij as close to ψij as possible, while keeping the representation of ψ:ij concise. This is very similar to the process of constraint abduction [10].
However, even with a relaxation Ψ: such that ϕ [ ψ:ij ” ψij , we may still lose
relevant information.
Example 4. Consider a program state (using the domain of octagons) with
scalar property ϕ “ rrx ă yss, and segment properties ψik “ ϕ [rri ă k ^ x ă ass,
ψkj “ ϕ [rrk ă j ^ y “ ass. Computing the value of ψij gives the expected
ψij “ ϕ [rri ă j ^ x ă ass. However, although the relaxations
ψ:ik “ rri ă k ^ x ă ass and ψ:kj “ rrk ă j ^ y “ ass
are exact, computing ψij as before yields:
rri ă j ^ i ă k ^ x ă ass \rri ă j ^ k ă j ^ y “ ass “ rri ă jss
We could avoid this loss of information by conjoining the scalar part during each
step of the fixed point computation
ψ:ij Ď ϕ d ppϕ [ ψ:ij [ ψ:ik q \pϕ [ ψ:ij [ ψ:kj qq
but while this avoids the loss of information, it also defeats the original goal of
reducing computation cost. Instead we define a more conservative d operation
which maintains enough additional information to retain properties of interest.
Example 5. Consider again the analysis that was performed in Example 4, but
now with a modified d operation which gives us ψ:ik “ rri ă k ^ x ă ass and
ψ:kj “ rrk ă j ^ y “ a ^ x ă ass. In this case, when we compute ψ:ij , we get:
ψ:ij “ rri ă j ^i ă k ^x ă ass \rri ă j ^k ă j ^y “ a^x ă ass “ rri ă j ^x ă ass.
This maintains the property of interest, without repeatedly conjoining with ϕ
during the fixed point process.
Another observation from Examples 4 and 5 is that while the boundary
constraints for each edge (such as rri ă jss for ψij ) are not implied by ϕ, they
are typically irrelevant to the segment properties, unless:
11
rri ă nss
rri ă nss
n
i
J
i
K
i`
rri
ă
n
K
nss
i`
Fig. 9: Solving this set of constraints, we strengthen ψii` and derive rri ă nss. If
we use d, we have ψ:in “ J, so we fail to strengthen ψ:ii` .
– ϕ [rri ă jss “ K, in which case the edge must be empty, or
– the array content Arℓs is some function of the index ℓ
– ψji` “ K (or j “ k ` for some k, and ψki “ K)
This is particularly troublesome for domains that explicitly store the transitive
closure of constraints, as we then expend substantial computation maintaining
consequences of rri ă jss, which are largely irrelevant, and lost during the join.
If we choose an operator d which discards the consequences of edge boundaries, we must take particular care not to lose information in the third case
mentioned. Consider the state shown in Figure 9, where the edge from n to i` is
infeasible. Computing the original fixed point, we obtain rri ă nss at ψii` , which
is then lifted out to the scalar domain. However, rri ă nss is obviously implied
by ϕ [rri ă nss, so will typically be discarded by d. The property rri ă nss is not
obtained at ψii` , so is never lifted out to the scalar domain. We must, therefore,
add the following case to the normalization rules given in Section 2.1:
4. ϕ Ďrri ě jss if ψi,j “ K
This ensures that any scalar properties resulting from infeasible segments are
included in the scalar domain. Further, since many elements of Ψ: may be J, the
relaxation operator allows us to take advantage of sparse matrix representations.
Unfortunately we are not aware of any existing, general operations suitable
for computing Ψ: ; as they must consider both underlying lattice and the characteristics of the implementation. We complete this section by outlining suitable
relaxation operators for difference-bound matrices (DBMs) [5] (and octagons).
Relaxations for polyhedra can be found in the appendix, and it should not be
particularly difficult to define analogous operators for alternative domains.
A value in the DBM (or octagon) domain consists of a set of constraints
rrvi ´ vj ď kss (or rr˘vi ˘ vj ď kss for octagon). We can construct a relaxation
‹
of ψij by computing the transitive closure ψij
, and discarding any constraints
implied by ϕ [rri ă jss:
‹
ϕ d ψij “ tc | c P ψij
, ϕ [rri ă jss & cu
If the abstract states are stored in closed form, we can simply collect the constraints of ψij not appearing in ϕ [rri ă jss. Or we can avoid performing many
implication tests by instead collecting the constraints involving variables in U .
12
4
Experimental Evaluation
We have implemented the analysis in sparcolyzer, a prototype array content
analyser for the language described in Section 1. sparcolyzer is implemented
in ocaml, using the Fixpoint library4 . For the underlying domain, we implemented the DBM domain in C++, customized for operating on sparse graphs.
Experiments were performed on a 3.0GHz Core 2 Duo with 4Gb ram running
Ubuntu Linux 12.04. The set VA of segment bounds were pre-computed using a
simple data-flow analysis to collect all variables which may (possibly indirectly)
be involved in the computation of an array index.
We tested sparcolyzer on a number of array manipulation program fragments. Most of these were taken from Halbwachs and Péron [9]. We added several
additional fragments that illustrate interesting properties of the analysis, including members of the init rand family discussed in Section 1 (Figure 1).
Computation time for each instance is given in Table 1. For instances taken
from [9], we include the original reported runtimes, although these are not directly comparable, as the experiments in [9] were performed on a slower machine
(Core2 Duo 1.6 GHz, with 2MB of RAM) using the domain of difference bound
matrices with disequalities (dDBM) [13].5 The implementation from [9] has not
been available and we have not tried to reconstruct it.
Table 1 compares the runtimes of two variants of our content domain to the
approach of Halbwachs and Péron [9] (the hp08 column). The naive variant
uses the direct implementation, where the matrix is represented as a |V 1 | ˆ |V 1 |
array, and a copy of the scalar domain is stored in each matrix entry. The
sparse variant stores, for each row and column, a set of non-J entries so that
the normalization, \ and [ operations do not need to process entries that will
definitely remain J, and computes the fixed point on the relaxed matrix Ψ: ,
rather than directly on Ψ . sparse uses the simple relaxation step of discarding
all constraints not involving some array variable A P U . Note that sparse still
iterates over all |V 1 | ˆ |V 1 | elements when changes to the scalar domain occur,
as a change to the scalar domain may affect any matrix element.
In cases where there are very few partitions—either because there are very
few index variables, or they are highly constrained—we expect the partitionbased methods to be faster (as they do not need to compute closure over transitive edges). The performance of naive is comparable to that of [9] on instances
with few partitions, and it improves substantially on more complex instances.
sparse is faster yet, sometimes by several orders of magnitude, and still finds
the desired invariant in all but two cases.
It is interesting to compare the behaviour of sentinel and first nonnull.
These programs superficially appear quite similar; in both cases, we set up an
‘end-of-array’ marker, then scan the array to find a particular element. However,
the invariants necessary to prove the desired properties are quite different. In
4
5
http://pop-art.inrialpes.fr/people/bjeannet/bjeannet-forge/fixpoint/
Performance of the domains should be roughly equivalent, as in the absence of explicit disequalities, dDBM behaves identically to DBM.
13
program
init
init offset
init rand2
init rand3
init rand4
init rand5
arraymax
copy
partition hoare
partition hp08
sentinel
first nonnull
naive
0.11
0.28
2.48
9.59
31.14
80.58
0.13
0.13
1.51
3.50
0.14
0.60:
sparse
0.02
0.07
0.14
0.64
1.98
4.96
ă0.01
ă0.01
0.06
0.14:
ă0.01
0.01:
hp08 [9]
0.05
0.10
0.02
22.87
0.21
2.25
Table 1: Analysis times in seconds. Instances where we were unable to prove
the desired invariant are marked with :.
x := Ar0s
i := 1
j := n ´ 1
while(i ď j)
if (Aris ă x)
Ari ´ 1s := Aris
i := i ` 1
else
while(j ě i ^ Arjs ě x)
j := j ´ 1
(‹)
if (j ą i)
Ari ´ 1s := Arjs
Arjs := Aris
i := i ` 1
j := j ´ 1
Ari ´ 1s := x
x := Ar0s
i := 0
j := n ´ 1
while(i ď j)
while(Aris ď x ^ i ď j)
i := i ` 1
while(Arjs ą x ^ i ď j)
j := j ´ 1
if (i ď j)
t := Arjs
Arjs := Aris
Aris := t
(a) partition hp08
(b) partition hoare
Fig. 10: Quicksort partitioning: (a) as done in [9]; and (b) a la Hoare
the case of first nonnull, we require:
ps “ n ^ @e P r0, nq . Ares ‰ 0q _ ps ă n ^ Arss “ 0 ^ @e P r0, sq . Ares ‰ 0q
This can be expressed using the approaches of Gopan et al. and Halbwachs and
Péron [7,9], as they store a separate invariant for each total ordering amongst
the partition variables. Our approach, however, cannot handle such disjunctive
reasoning, so the segment property quickly reaches J.
Consider partition hp08 (the variant of the Quicksort partition step given
in [9], shown in Figure 10(a)). As the imperative source language we use does
14
allow loads inside conditionals, such reads are hoisted outside the corresponding loops; for example, the loop marked p‹q is transformed as shown below.
At the point marked p‹q, it is possible to determine
that rrj ă iss ñ rrArjs ă xss. Thus it is easy to
show that rrArjs ă xss holds at the loop exit. In
ej := Arjs
the hoisted version, naive method can prove the
while(j ě i ^ ej ě x)
invariant successfully, because the property rrj ă
j := j ´ 1
iss ñ rrej ă xss is derived for the edge ψji . When
ej := Arjs
we exit the loop with rrj ă iss, this property gets
extracted to the scalar domain, and we get rrej ă
xss ^ rrej “ ass at ψjj 1 . When using the sparse method, however, the property
on ψji is discarded, as it involves only scalar variables, so the invariant is lost.
If we were to use the original version without hoisting, we would be unable
to prove the invariant using either method, as we cannot express rrj ă iss ñ
rrArjs ă xss directly.
However, our method easily proves the standard version (partition hoare;
Figure 10(b)) correct, whether or not the reads are hoisted.
5
Conclusion and Future Work
We have described a new approach to automatic discovery of array properties,
inspired by algebraic shortest-path algorithms. This approach retains much of
the expressiveness of the partitioning methods of [7] and [9], but avoids the
need for syntax dependence and an up-front factorial partitioning step. The
method can successfully derive invariants for a range of interesting array program
fragments, and is substantially faster than partitioning-based approaches with
even modest numbers of index variables.
Several improvements could be made to the performance of the analysis.
The current implementation does not take advantage of liveness information,
and maintains entries in the content graph for all variables in VA at each step.
Clearly, performance could be improved by eliminating non-live variables from
the matrix.
Algorithms which maintain shortest path information can often be improved
by storing only the transitive reduction of the graph. As our domains are not
distributive, it is non-trivial to determine whether a given edge must occur in
the transitive reduction; however, it would be worth investigating whether maintaining the transitive reduction would prove beneficial.
References
1. B. Blanchet, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, D. Monniaux,
and X. Rival. Design and implementation of a special-purpose static program analyzer for safety-critical real-time embedded software. In T. Mogensen, D. Schmidt,
and I. H. Sudborough, editors, The Essence of Computation: Complexity, Analysis,
Transformation, volume 2566 of LNCS, pages 85–108. Springer, 2002.
15
2. R. Bodı́k, R. Gupta, and V. Sarkar. ABCD: Eliminating array bounds checks on
demand. In ACM Symposium on Programming Language Design and Implementation (PLDI’00), pages 321–333. ACM Press, 2000.
3. P. Cousot, R. Cousot, and F. Logozzo. A parametric segmentation functor for fully
automatic and scalable array content analysis. In Proceedings of the 38th ACM
Symposium on Principles of Programming Languages, pages 105–118. ACM Press,
2011.
4. P. Cousot and N. Halbwachs. Automatic discovery of linear constraints among
variables of a program. In Proceedings of the Fifth ACM Symposium on Principles
of Programming Languages, pages 84–97. ACM Press, 1978.
5. D. L. Dill. Timing assumptions and verification of finite-state concurrent systems.
In Automatic Verification Methods for Finite State Systems, volume 407 of LNCS,
pages 179–212. Springer, 1989.
6. I. Dillig, T. Dillig, and A. Aiken. Fluid updates: Beyond strong vs. weak updates. In A. D. Gordon, editor, Proceedings of the 19th European Symposium on
Programming, volume 6012 of LNCS, pages 246–266. Springer, 2010.
7. D. Gopan, T. Reps, and M. Sagiv. A framework for numeric analysis of array operations. In Proceedings of the 32nd ACM Symposium on Principles of Programming
Languages, pages 338–350. ACM Press, 2005.
8. S. Gulwani, B. McCloskey, and A. Tiwari. Lifting abstract interpreters to quantified
logical domains. In Proceedings of the 35th ACM Symposium on Principles of
Programming Languages, pages 235–246. ACM Press, 2008.
9. N. Halbwachs and M. Péron. Discovering properties about arrays in simple programs. In ACM Symposium on Programming Language Design and Implementation
(PLDI’08), pages 339–348. ACM Press, 2008.
10. M. J. Maher. Herbrand constraint abduction. In Proceedings of the 20th IEEE
Symposium on Logic in Computer Science, pages 397–406. IEEE Comp. Soc., 2005.
11. A. Miné. The octagon abstract domain. Higher-Order and Symbolic Computation,
19(1):31–100, 2006.
12. M. Mohri. Semiring frameworks and algorithms for shortest-distance problems.
Journal of Automata, Languages and Combinatorics, 7(3):321–350, 2002.
13. M. Péron and N. Halbwachs. An abstract domain extending difference-bound
matrices with disequality constraints. In B. Cook and A. Podelski, editors, Verification, Model Checking and Abstract Interpretation, volume 4349 of LNCS, pages
268–282. Springer, 2007.
14. A. Schrijver. Theory of Linear and Integer Programming. Wiley, 1986.
15. N. Suzuki and K. Ishihata. Implementation of an array bound checker. In Proceedings of the Fourth ACM Symposium on Principles of Programming Languages,
pages 132–143. ACM Press, 1977.
16. H. Xi and F. Pfenning. Eliminating array bound checking through dependent
types. In ACM Symposium on Programming Language Design and Implementation
(PLDI’98), pages 249–257. ACM Press, 1998.
16
Appendix A: Termination Proofs
Theorem 1. The shortest-path computation terminates for the octagon domain.
Proof. Consider an initial set of abstract states X “ rx1 , . . . , xn s, and a system of
inequalities of the form xi Ďpxi [ xj q \pxi [ xk q. Let BPpxq denote the bounding
hyperplanes of x. We then have: BPppx [ yq \px [ zqq Ď BPpx [ yq Y BPpx [ zq.
Each time one
Ű 1 is evaluated, each bounding hyperplane is an
Ť of the equations
BPp
X q As the initial set of bounding hyperplanes is
element of:
1
X ĎX
finite, and each iteration must tighten at least one bounding plane, the tightening
process must eventually
Ť terminate; in fact, the number of descending steps is
\
[
bounded by: |X| ˆ | xPX BPpxq|
In the case of octagons, projection cannot introduce new bounding hyperplanes,
so the addition of the propagation rules (ϕ Ď D UI . ϕii` and ϕij Ď ψ) does not
affect termination.
Theorem 2. The shortest-path computation terminates for the convex polyhedron domain.
Proof. We can prove termination for convex polyhedra in a similar fashion as
for octagons. Consider an initial set of abstract states X “ rx1 , . . . , xn s, and a
system of inequalities of the form xi Ďpxi [ xj q \pxi [ xk q. By the polyhedron
decomposition theorem (see, e.g. [14]), any polyhedron may be generated by a
finite set of points P , and a set of rays R. Let RV denote the set of rays of unit
length in the direction of variable v (that is, a vector with v th component 1 or
´1, and all other components 0), for each v P V . Given polyhedra x “ xPx , Rx y
and y “ xPy , Ry y, the results of the operations of interest have the following
properties:
x [ y “ xP 1 , R1 y where R1 Ď Rx Y Ry
x \ y “ xP 1 , R1 y where P 1 Ď Px Y Py , R1 Ď Rx Y Ry
D V . x “ xP 1 , R1 y where P 1 Ď Px , R1 Ď Rx Y RV
Ť
None of these operations can introduce rays not in R “ RV Y xPX Rx . Additional extreme points can only be introduced during the application of [. During
each fixed point iteration, exactly one of three cases must occur:
1. All abstract values in X remain the same.
2. The set of rays for some x P X changes.
3. All the sets of rays remain the same, and the set of extreme points for some
x P X changes.
In case 1, we terminate. As the set of rays is restricted to R, and each iteration is
strictly descending, case 2 can only occur finitely many times. Assuming the set
of
Ť rays remains fixed,Şevery introduced extreme point must be some element of
xPX 1 xu. As this set is finite, and each step is descending,
X 1 ĎX tP | xP, Ry “
this can only occur finitely many times without case 2 occurring. As case 2 must
always occur after a bounded number of steps, and can only occur finitely many
times, the fixed point process must eventually terminate.
\
[
17
We conclude that this process will terminate on the most commonly used
relational numeric domains. In the case of polyhedra, it is worth noting that
Theorem 2 does not provide any bounds on the coefficients of hyperplanes in
the resulting polyhedra. In some cases, the coefficients may grow quite large
before converging, which can cause problems for domains implemented with
fixed-precision machine arithmetic.
Appendix B: Relaxation for polyhedra
The relaxation algorithm for polyhedra follows the same intuition as that for
octagons; we wish to collect the transitive closure of ψij , then discard anything
implied separately by ϕ [rri ă jss.
However, the polyhedra domain provides two difficulties: computing the
transitive closure of a set of linear constraints is non-trivial (as the domain is
typically stored as a minimal set of generators and hyperplanes [4]); and often is
a bad idea in general: as the constraints do not have bounded arity, there may
be exponentially more constraints in the transitive closure than in the original
problem. Instead, we collect a set of constraints that is sufficient to reconstruct
the constraints that, had we computed the transitive closure, would have been
kept.
Given constraints c1 “ rrk1⊺ x ě m1 ss and c2 “ rrk2⊺ x ě m2ss, we can construct
a new constraint (that is not implied separately by either c1 or c2 ) by resolution if
there is some matched pair of coefficients k1v , k2v , such that k1v ą 0 and k2v ă 0.
1v
k2 q⊺ x ě m1 ` kk1v
m2 ss, which
We then construct a new constraint c1 “ rrpk1 ` kk2v
2v
th
has v coefficient 0.
Example 6. Consider the constraints c1 “ rrx ` y ě 7ss, c2 “ rrz ´ 2y ě 2ss, and
c3 “ rrw ` 2y ě 3ss. c1 and c2 may be resolved, as c1 contains the term y, and
c2 contains ´2y. This yields c12 “ rrx ` z2 ě 8ss. We cannot, however, construct
any new constraints by combining c1 with c3 .
Rather than computing the transitive closure explicitly, given an initial set of
interesting constraints C: and other constraints C, we find all those constraints
that are resolvable with those in C: (taking into account the direction of previous
: We then continue this process until no
resolution steps), and add them to C.
further resolvable constraints are found:
$
trans‹ pR Y tpsignpkv1 q, v 1 qu, C: Y tcu, Cq
’
’
&
if D ps, vq P R, c “ rrk ⊺ x ě mss P C, v 1 P x
: Cq “
trans‹ pR, C,
s.t. v ‰ v 1 , kv s ă 0, ppsignpkv1 q, v 1 q R R _ c R C:
’
’
%:
C otherwise
: Cq “ trans‹ ptsignpkv , vq | rrk ⊺ x ě mss P C,
: v P xu, C,
: Cq
transpC,
Given a set of linear constraints C defining a polyhedron, we can construct the
initial set C: with the elements of C that are not implied by ϕ [rri ă jss, then
18
: C z Cq.
: As with octagons, we can avoid
compute the relaxation with transpC,
performing implication tests by instead initializing C: with those constraints
containing variables in VA .
Example 7. Consider a constraint of interest c0 “ rra ´ y ě 0ss, with additional
constraints C “ tc1 “ rry ´ z ě 0ss, c2 “ rrz ` w ě 0ss, c3 “ rrx´ y ě 0ssu. Initially,
we have C: “ trra ´ y ě 0ssu, and R “ tp`, aq, p´, yqu. We can resolve with
rry ´ z ě 0ss, since p´, yq P R, but p´, zq R R. During the second step, we include
c2 , adding p´, wq to R. At this point, we have R “ tp`, aq, p´, yq, p´, zq, p`, wqu,
: so we return
and C: “ tc0 , c1 , c2 u, and we cannot add anything to either R or C;
:
the current value of C.
Appendix C: Partitions for init randm
The large number of partitions required for the init rand family is not necessarily obvious. Assuming n is non-negative, we must distinguish between the
case where the array is empty (thus n “ 0) or non-empty (n ą 0). This gives
us two base orderings: rt0, nus, and rt0u, tnus. In these descriptions, sets denote
equivalence classes, and equivalence classes are listed in increasing order.
If n “ 0, there is only one possible value for i1 : rt0, i, nu, ti`us. Otherwise,
we must distinguish all the possible relations between 0, i and n. The resulting
orderings are as follows:
rt0u, ti1 , nuti`
rt0, i1 , nu, ti`
1 us
1 us
`
rt0, i1 u, ti1 , nus rt0u, ti1u, ti`
,
1 nus
`
rt0, i1u, ti`
u,
tnus
rt0u,
ti
u,
ti
1
1
1 u, tnus
When we construct the partitions for m “ 2, we introduce i2 into all feasible
locations in each of the partitions for m “ 1.
In cases where we must explicitly distinguish the 0th element, this progression
grows substantially faster, with m “ 1, . . . , 5 yielding r9, 45, 333, 3285, 40509s.
19
| 6cs.PL
|
STAR: Spatio-Temporal Altimeter Waveform Retracking using
Sparse Representation and Conditional Random Fields
Ribana Roscher, Bernd Uebbing, Jürgen Kusche
arXiv:1709.07681v1 [physics.ao-ph] 22 Sep 2017
Institute of Geodesy and Geoinformation, University of Bonn, Nußallee 15+17, 53115 Bonn, Germany
Abstract
Satellite radar altimetry is one of the most powerful techniques for measuring sea surface height variations,
with applications ranging from operational oceanography to climate research. Over open oceans, altimeter
return waveforms generally correspond to the Brown model, and by inversion, estimated shape parameters
provide mean surface height and wind speed. However, in coastal areas or over inland waters, the waveform
shape is often distorted by land influence, resulting in peaks or fast decaying trailing edges. As a result,
derived sea surface heights are then less accurate and waveforms need to be reprocessed by sophisticated
algorithms. To this end, this work suggests a novel Spatio-Temporal Altimetry Retracking (STAR) technique. We show that STAR enables the derivation of sea surface heights over the open ocean as well as
over coastal regions of at least the same quality as compared to existing retracking methods, but for a
larger number of cycles and thus retaining more useful data. Novel elements of our method are (a) integrating information from spatially and temporally neighboring waveforms through a conditional random
field approach, (b) sub-waveform detection, where relevant sub-waveforms are separated from corrupted or
non-relevant parts through a sparse representation approach, and (c) identifying the final best set of sea
surfaces heights from multiple likely heights using Dijkstra’s algorithm. We apply STAR to data from the
Jason-1, Jason-2 and Envisat missions for study sites in the Gulf of Trieste, Italy and in the coastal region
of the Ganges-Brahmaputra-Meghna estuary, Bangladesh. We compare to several established and recent
retracking methods, as well as to tide gauge data. Our experiments suggest that the obtained sea surface
heights are significantly less affected by outliers when compared to results obtained by other approaches.
Keywords: coastal, oceans, altimetry, retracking, sea surface heights, conditional random fields, sparse
representation
1. Introduction
For several decades, radar altimetry is routinely being used for monitoring sea surface height (SSH)
variations. Observed SSHs play a key role in several applications, ranging from operational oceanography
(Chelton et al., 2001) and tidal modeling (Wang, 2004; Savcenko & Bosch, 2008) to gravity estimation
(Hwang et al., 1998), and they serve as important indicators in climate research. Recently, radar altimetry
in coastal zones (Gommenginger et al., 2011) and for inland water bodies (Birkett & Beckley, 2010) has
become a topic of increasing interest. However, in both applications one needs to mitigate the potentially
significant land influence on the altimeter return signal.
The altimeter instrument on-board a satellite emits a spherically propagating, nadir-directed radar pulse,
which is reflected at the surface. Range information can then be inferred from the two-way travel time (Fu
& Cazenave, 2001). In addition, the returned signal energy is measured over time, forming an altimeter
waveform. It can be shown that over an ideal surface, the return waveform corresponds to the theoretical
Brown model (Brown, 1977) and the estimated shape parameters of this model provide information on mean
SSH and significant wave height (SWH), while the amplitude strength of the reflected radar pulse can be
used to derive wind speed. On board the satellite, the waveform signal is sampled at discrete epochs with a
Preprint submitted to Remote Sensing of Environment
September 25, 2017
spacing of about 3 ns of two-way travel time, which are generally referred to as range gates (Chelton et al.,
1989). Altimeter measurements do not refer to an individual point directly below the satellite, but rather
to a footprint with a diameter of several kilometers, depending on SWH and the altitude of the altimetry
mission.
As illustrated in Fig. 1, the return waveform over the open ocean consists of three main parts; first,
before any return energy from the radar pulse is measured, the waveform contains only thermal noise which
is present in all radar systems. As soon as the front of the radar pulse hits the wave crests, the altimeter
footprint is defined by a single point and the measured return energy begins to rise. Afterwards, more of
the pulse illuminates the surface around the initial point and the footprint becomes a growing circle, which
corresponds to rapidly increasing signal energy in the measured altimeter waveform. The leading edge of
the altimeter return waveform is defined between the first energy return from wave crests and the return
energy after the radar pulse has reached the wave troughs (Fig. 1). At this point. the area of the footprint
circle reaches its maximum, which is defined as the pulse–limited footprint (PLF, Chelton et al., 1989).
Afterwards, the circle transforms into an annulus with increasing inner and outer radii, but with a fixed
illuminated area. The corresponding signal energy measured outside of the PLF is referred to as trailing
edge of the measured waveform (Fig. 1). The slope of the trailing edge can be utilized to derive information
on the off-nadir attitude of the altimeter satellite.
The 50%-point on the leading edge corresponds to the mean sea level between wave crests and wave
troughs, and thus represents the reference point for the range measurement. An algorithm on board of the
satellite tries to position this point inside a pre-defined range window at a fixed tracking range gate (31
for Jason-2, Quartly et al., 2001). This range window consists of a fixed number of range gates, covering
about 50 m depending on the satellite mission, and it is positioned by the onboard tracker based on prior
information on the range. However, positioning is not always perfect and the 50%-point is not located exactly
at the tracking gate. Consequently, this requires a ground-based reprocessing of the altimeter waveforms
transmitted back to Earth, a procedure which is called retracking. Over the open ocean, the shape of the
waveform will be close to the theoretical Brown model with the 50%-point being only slightly shifted from
the tracking gate position, and this can be easily corrected using an ocean model retracker (Brown, 1977;
Hayne, 1980; Deng, 2003). For ice surfaces where the waveform signal often contains two leading edges due
to the radar signal partly penetrating through the upper snow layer, specialized retracking algorithms have
been developed (Martin et al., 1983). However, in coastal areas the waveform shape is typically disturbed
by land influences in the altimeter footprint, resulting in peaks or fast decaying trailing edges.
These deviations of coastal waveforms from the Brown model lead conventional ocean retrackers to
generate diverging or strongly biased estimates of SSH, as land-induced peaks propagate along the trailing
edge towards the leading edge while the altimeter ground track approaches the coast (Lee et al., 2010). In
order to mitigate the land influences on the waveform shape, various tailored approaches have been proposed.
As an example for methods that seek to model the entire waveform, (Halimi et al., 2013) combined a 3parameter Brown ocean model with a modeled asymmetric peak to account for land influences. A different
approach for dealing with the influence of peaks on the retracked estimates is to first partition the waveform
in a pre-processing step; i.e. to identify relevant parts of the waveform, such as the leading edge, but also
possible peaks. For example, (Hwang et al., 2006) first identify relevant sub-waveforms and then apply a
threshold retracking algorithm to each of the sub-waveforms, which leads to multiple equally likely SSH
estimates at each location from which the final estimate is chosen based on comparison to a-priori height
information. In this way, peaks that appear outside of the relevant sub-waveforms are ignored. Recently, for
retracking SSHs over inland water bodies, (Uebbing et al., 2015) combined the sub-waveform approach from
(Hwang et al., 2006) and the waveform model from (Halimi et al., 2013) to suppress land-induced peaks on
the trailing edge, but also to account for possible peaks close to the leading edge of the waveform. This
could be shown to lead to improved lake heights compared to conventional methods. In a different approach,
(Passaro et al., 2014) suggested a two-step procedure, similar to a previously published approach by Sandwell
& Smith (2005), where in the first step all 3 parameters (amplitude, range and SWH) are estimated. In
the second iteration they fixed the SWH to a mean value derived from the first step and re-estimated the
amplitude and range correction, since SWH estimations are strongly correlated to the range correction. This
2
leads to improved SSHs closer to the coast.
Here, we introduce a novel method for the analysis of sea surface heights from altimetric waveforms,
which will utilize spatial information from neighboring range gates within one waveform, as well as temporal information from neighboring waveforms along the altimeter track. This Spatio-Temporal Altimetry
Retracker (STAR) can be applied to altimetry data over the open ocean, as well as in coastal areas. Our
contributions are twofold: First, our analysis includes a novel sub-waveform detection scheme, which to our
knowledge for the first time integrates spatial as well as temporal information. This differs from the conventional sub-waveform detection algorithm (Hwang et al., 2006) in that we partition the entire waveform
into separate sub-waveforms, instead of identifying possible, disjointed leading edges. Second, in order to
be largely independent of the choice of tuning (or ’hyper’) parameters within the sub-waveform detection
scheme, we derive multiple sub-waveform partitionings by varying the weight w between unary and binary
terms of the conditional random field. This leads to a range of partitionings of the entire waveform, and
subsequently to a point cloud of equally likely SSHs at each measurement position, each of which is estimated
using a 3-parameter ocean model (Halimi et al., 2013). We then employ Dijkstra algorithm (Dijkstra, 1959)
to find reasonably smooth SSHs, without resorting to fitting.
Our sub-waveform detection scheme uses a sparse representation (SR) approach, where the return power
at all range gates within one particular sub-waveform is modeled by a weighted linear combination of a
single common set of basis waveforms, which are derived from synthetic Brown waveforms. The concept of
SR has been applied to many areas of signal analysis (Wright et al., 2010), but this study appears to be the
first which uses it on radar altimetry.
SSHs and other sea surface conditions such as wave height are neither independent along tracks, nor
between neighboring tracks (Sandwell & Smith (1997)). Spatial information has been used in the analysis, for example, by Maus et al. (1998) through simultaneously processing of a sequence of waveforms for
tracking of travel times, or Halimi et al. (2016) for a smooth estimation of altimetric parameters. This
means, the integration of spatial information can be carried out in different parts of the analysis. The
latter two approaches, for example, integrate spatial information about neighboring waveform to develop
improved estimation algorithms for retracking. Here, we integrate spatio-temporal information by means of
a conditional random field (CRF, Lafferty et al., 2001): to this end, we introduce spatial relations between
the return power at range gates of temporally neighboring waveforms, i.e. neighboring waveforms within
one pass and cycle, and relations between the range gate power within a single waveform. In this way, range
gates which are relevant for SSH estimation can be distinguished from corrupted or non-relevant waveform
parts, since they are represented through a different linear combination of basis sub-waveforms (see Fig. 1).
In contrast to Halimi et al. (2016) where the conditional random field is used as part of an algorithm to
enforce a smooth estimation of the retracking parameters, we propose to use the conditional random field at
the sub-waveform detection step. Subsequently, any retracking method can be applied to the identified individual sub-waveforms for deriving the SSH, thus effectively ignoring disturbing signals outside the selected
sub-waveform. This means, our approach could be transferred in future to the analysis of Delay-Doppler
altimetry.
This paper is organized as follows: First, we introduce the altimetry data that we used for validating
our method, as well as our two study sites in Sec. 2. We describe the sub-waveform detection with SR and
the integration of spatio-temporal knowledge by employing a CRF in Sec. 3. Moreover, the estimation of
single SSHs and the determination of the best set of SSHs from multiple heights using Dijkstra’s algorithm
is explained in Sec. 4. In Sec. 5, we compare the performance of our sub-waveform detection against an
existing method and we evaluate the proposed analysis framework by means of analysing Jason-1, Jason-2
and Envisat waveforms in the coastal regions of the Gulf of Trieste, and in the northern Bay of Bengal in
the coastal waters of Bangladesh. Sec. 6 concludes the paper.
3
Figure 1: Waveform with disturbing peak caused by land influences (colored in blue). The relevant part for sea surface height,
determined by sub-waveform detection, is illustrated in green. A theoretical waveform model is depicted in orange.
2. Data and Study Sites
2.1. Data
We apply our retracking method to Jason-2 Sensor Geophysical Data Records (SGDR) of the Jason-2
mission, as well as to SGDRs from the Jason-1 and European Environmental Satellite (Envisat) missions.
The Ocean Surface Topography Mission (OSTM)/ Jason-2 was launched in mid of 2008 succeeding the
Jason-1 mission on the same orbit. The satellite flies in a near circular ∼10 day repeat orbit with an altitude
of 1336 km and inclination of 66 ◦ and a separation of the groundtracks equal to 315 km at the equator. The
main instrument is the Poseidon-3 altimeter which emits radar pulses in the Ku-Band (13.575 GHz/2.21 cm)
and C-Band (5.3 GHz/5.08 cm) (AVISO, 2015). Additional instruments are a microwave radiometer used to
derive the wet troposphere correction, as well as GPS and DORIS systems for precise orbit determination
(Rosmorduc et al., 2011). The Jason-2 SGDRs have been obtained from the Archiving, Validation and
Interpretation of Satellite Oceanographic (AVISO) team which is part of Centre National d’Etudes Spatiales
(CNES). The SGDRs are sorted by pass and cycle, including 254 passes per cycle or ∼10 day repeat orbit
and we utilize data of passes 053 and 196 from the beginning of the mission in July 2008 (cycle 0) until the
end of 2014 (cycle 239).
The composition of the Jason-1 mission is very similar to Jason-2. It was launched in December, 2001
as a successor to the Topex/Poseidon mission. After the launch of Jason-2 in June, 2008 both satellites
flew on the same orbit in close distance to allow intercalibration of the satellite missions. After 6 months,
Jason-1 was moved to an interleaved orbit located in the middle between the nominal orbit to increase the
spatial resolution of the combined Jason-1 and Jason-2 data until January, 2012. Afterwards the satellite
was moved to a drifting geodetic orbit and passivated and decommissioned in July, 2013 after losing contact. We will use data from the interleaved period from February, 2009 (cycle 263) until January, 2012
(cycle 370). The Jason-1 SGDRs for the interleaved orbit are acquired from the Physical Oceanography
Distributed Active Archive Center (PO.DAAC, ftp://podaac.jpl.nasa.gov/allData/coastal_alt/L2/
ALES/jason-2/) operated by the Jet Propulsion Laboratory (JPL) which is part of the National Aeronautics
and Space Administration (NASA).
The Envisat satellite was launched in March, 2002 succeeding the ERS-2 mission in the same orbit by
the European Space Agency (ESA). The orbit is a 35-day repeat orbit with an altitude of 800 km, about
4
an inclination of 98.55 ◦ which allows the satellite to cover higher latitude regions compared to e.g. Jason-2
with a higher spatial resolution (80 km separation at the equator) at the cost of a longer repeat period
of 35 days. The satellite carries a total of 10 instruments of which the DORIS positioning system, the
microwave radiometer and the Radar Altimeter 2 (RA2) altimeter instrument are of most importance to us
(Rosmorduc et al., 2011). The RA2 altimeter is a dual frequency altimeter emitting radar pulses in Ku-Band
(13.575 GHz/2.21 cm) and S-Band (3.2 GHz/9.37 cm) (ESA, 2007). The Envisat SGDR data are provided
by ESA (https://earth.esa.int/). We utilize data from June, 2002 (cycle 7) to September, 2010 (cycle
93).
We extract the 20 Hz (18 Hz for Envisat) tracker range, altitude and waveforms which are needed during
the retracking algorithm. Additionally, 1 Hz atmospheric model corrections for the dry and wet troposphere,
as well as the ionosphere are extracted from the SGDR data and linearly interpolated to the high rate
positions.
For validation of the retracked coastal SSHs, tide gauge data with hourly resolution from the University
of Hawaii Sea Level Center (UHSLC) are used. The hourly data are uncorrected with respect to tidal and
inverse barometric effects and we do not apply these corrections to neither the altimetry data, nor the tide
gauge data to remove the influence of these corrections on the validation. The tide gauge data for Trieste,
Italy is available from June, 2009 to December 2015 and for the tide gauge station in Chittagong, Bangladesh
we have data from July, 2007 til December 2015. Additionally, we utilize openly available GDR datasets for
the Jason-2 (cycles 1 to 239) and Envisat (cycles 7 to 93) mission that include the ALES retracked ranges
for comparison which are distributed by PO.DAAC, JPL.
For the validation of significant wave height and wind speed, we utilize model data from the ERAInterim reanalysis (Dee et al., 2011) which is distributed by the European Centre for Medium-Range Weather
Forecast (ECMWF) and interpolated to the altimetry track.
2.2. Study Sites
For investigating the quality of our proposed STAR algorithm, we selected two study sites located in
the Gulf of Triest and in the coastal regions of Bangladesh. The sites include varying conditions, including
shallow coastal waters, open ocean areas, temporally submerged sand banks and transition zones between
river estuaries and the ocean.
(a) Triest TG at the Adriatic Sea and the groundtracks of Jason-2 pass 196, as well as Jason-1 interleaved orbit, pass 161.
(b) Chittagong TG at the Bay of Bengal and the
groundtracks of Jason-2 nominal orbit, pass 053, as
well as Envisat, pass 0352.
Figure 2: Location of the altimetry tracks and the tide gauges used for validation of the retracked SSHs.
5
2.2.1. Triest
The first study site is located in the northern Adriatic Sea in the north-eastern part of the Gulf of Venice
and includes the Gulf of Triest. The descending nominal orbit pass 196 of the Jason-2 mission crosses the
study area from north-west to the south-east (Fig. 2(a)). It crosses from the Italian mainland to ocean close
to the city of Marano Lagunare at approx 45.76 ◦ N and covers about 5.9 km of the Laguna di Marano before
there is a short ocean-land-ocean transition over the Isola di Sant’ Andrea at ∼45.71 ◦ N. Then, the track
covers about 45.5 km of open ocean in the Gulf of Venice before transitioning to the Croatian mainland at
∼45.36 ◦ N. For the last 5 km from ∼45.4 ◦ N until the track is over the Croatian mainland, it runs very close
to the Croatian coast with a distance of less than 2 km.
Furthermore, we utilize data from the ascending Jason-1 pass 161 of the interleaved orbit (Fig. 2(a)).
It crosses our study area from the south-west to the north-east. The first 37 km are located over the open
ocean and the groundtrack covers the Croatian mainland from ∼45.48 ◦ N to 45.50 ◦ N. Then the track covers
open water until it reaches the Italian mainland at ∼ 45.71 ◦ N. At about 45.54 ◦ N the track is located less
than 2 km away from the Slovenian mainland.
The tide gauge station Triest is located at ∼ 13.75 ◦ E and 45.65 ◦ N in the harbor of the city of Triest in
the Gulf of Triest. The distance of the tide gauge station to the Jason-2 groundtrack is about 31 km at the
closest point.
2.2.2. Bangladesh
The second study site is located in the northern Bay of Bengal in a region right off the coast of Bangladesh.
The region is covered by the ascending Jason-2 nominal orbit pass 053 which crosses the study area from
the south-west to the north-east (Fig. 2(b)). The track coverage starts over the open ocean parts of the Bay
of Bengal and reaches Sandwip Island after ∼ 95 km at ∼ 22.41 ◦ N . From ∼ 22.53 ◦ N to 22.63 ◦ N the track
again covers a strip of about 13 km of open water related to the estuary of the Ganges-Brahmaputra-Meghna
Delta (GBMD) before it reaches the Bangladesh mainland. Between latitudes of ∼ 22.25 ◦ N and 22.35 ◦ N
there are some sand banks located along the track which are submerged during high tide, but not during
low tide.
Additionally, the descending pass 0352 of the Envisat mission runs almost parallel to the Jason-2 track
and crosses the study site from the north-east to the south-west. It first reaches open water related to the
GBMD at ∼22.67 ◦ N and after about 14 km the track reaches Sandwip Island at ∼ 22.55 ◦ N. At 22.40 ◦ N,
the track transitions back to open water and afterwards covers the remaining 104 km of open ocean in our
study area. Again, the sandbanks mentioned above might influence the data acquisition during low tide.
The Chittagong tide gauge station is located at ∼ 91.83 ◦ E and 22.25 ◦ N in the Chittagong harbor. The
distance to the Jason-2 groundtrack is about 38 km at the closest point.
3. Sub-waveform Detection
3.1. Notation
Let us consider L consecutive waveforms collected along a cycle, each of which contains return energy
for G range gates. We arrange these waveforms in xl , l = 1, . . . , L. Each waveform can be represented
through a set of overlapping windowed waveforms ξ l,g ∈ xl with windows centered at ξl,g and comprising
Nξ neighboring range gates, as illustrated in Fig. 3 for a single echo. Within this framework, sub-waveform
detection means identifying K sub-waveforms in each xl by deciding for each range gate on indices ŷ l,g ,
which define the best-fitting sparse representation (SR)-based models. The aim of the proposed approach is
to optimally detect these sub-waveforms, where the number of models per waveform is unknown and needs
to be determined during the detection process.
6
3.2. Detection Framework
The schematic of our sub-waveform detection framework is illustrated in Fig. 3. The input of the
framework are windowed waveforms ξ l,g and synthetic Brown waveforms. These Brown waveforms are the
basis waveforms collected in a dictionary D, which is used for a sparse representation-based modelling of
signals. Given the input, a conditional random field (CRF) is formulated, which consists of an unary, datadependent term computed by sparse representation, and a binary term enforcing a smooth partitioning of
the entire waveform.
w A variation of the weighting between these two terms yield various sets of optimal
w
indices Ŷ = ŷ l,g , resulting in different partitionings of the entire waveform. The framework is flexible
regarding the chosen methods, e.g. the sparse representation can be replaced by other methods such as
correlation or other similarity measures. In the following, detailed explanations to the framework will be
provided.
Figure 3: Detection Framework: The input is given by altimetric waveforms and synthetic Brown waveforms. Each range
gate is represented by a windowed waveform, i.e., the range gate’s center point and neighboring range gates, as illustrated
by blue, violet and orange areas. A conditional random field is formulated, which consists of an unary, data-dependent term
computed by sparse representation and a binary term enforcing a smooth partitioning of the waveform. In the conditional
random field, the graphical model is constructed by connecting temporally adjacent range gates as well as adjacent range gates
within one waveform. A variation of hyperparameters in the conditional random field result in different partitionings of the
entire waveform.
3.3. Conditional Random Field
In order to perform sub-waveform detection while integrating information about neighboring range gates,
we make use of a conditional random field (CRF). The range gates are represented in a graph, where each
range gate is connected to spatially adjacent range gates within one waveform and to temporally adjacent
range gates along the satellite’s ground track (see Fig. 3). The basic idea is the assignment of each range
gate, represented as windowed waveform, to the best-fitting model, for which we use sparse representation
in this framework. Neighbored range gates which are assigned to the same model are summarized to one
sub-waveform. Therefore, all range gates in one sub-waveform follow the same underlying model. In our
approach, the optimal sub-waveform partitioning minimizes the energy functional
X
X
E(Y ) =
U(ξ l,g , y l,g ) − w
B(ξ l,g , ξ l,q , y l,g , y l,q ),
(1)
l,g
l,g,q∈Q
7
where the unary term U, depending on the windowed waveforms ξ l,g , describes the agreement between
the measured windowed waveform and a sparse representation model represented by y l,g . These models are
identified by so-called non-zero activation indices Y = y l,g , indicating which synthetic basis waveforms are
used for signal reconstruction within each specific model. The binary term B depends on both the non-zero
activation indices y l,g and y l,q , as well as the windowed waveforms with q ∈ Q indicating the set of direct
neighbors of each range gate. The weight between both terms is a hyperparameter and is denoted by w. In
our approach, we assume a sub-waveform to be a set of neighboring range gates which are modeled with the
same non-zero activation indices, as explained in more detail in the next paragraphs.
3.3.1. Unary Term - Sparse Representation
Generally, in terms of sparse coding (e.g., Olshausen & Field (1997)) a (G × 1)-dimensional waveform
xl can be represented by a linear combination of a few basis waveforms, which are collected in a (G × V )dimensional dictionary D = [dv ], v = 1, . . . , V , such that xl = Dαl + with kk being the reconstruction
error. Solving this problem means finding the activation vector αl containing the optimal coefficients,
whereas most of the elements are zero.
During sub-waveform detection, instead of representing the whole waveform signal, only windowed waveforms will be sparsely represented by windowed basis waveforms. In this case, each range gate, represented by
a windowed waveform, can be assigned to a specific sparse linear combination of windowed basis waveforms
constituting the best approximation for it. In more detail, a windowed waveform ξ l,g is sparsely represented
using activations αl,g and a sub-dictionary D g that contains certain rows of the underlying dictionary D (see
identical colored entities in Fig. 3 and Fig. 4). Since the dictionary comprises synthetic Brown waveforms,
sub-dictionaries contain specific parts of the synthetic waveforms. The indices of non-zero elements in αl,g ,
i.e., the dictionary elements participating in the reconstruction, are the non-zero activation indices y l,g used
for defining the models (cmp. Sec. 3.3). It is important to note that D is fixed for L consecutive waveforms,
such that sub-dictionaries used for temporally neighboring windowed waveforms are identical.
Figure 4: Schematic illustration of sparse representation of windowed waveforms. Colors in xl indicate different windowed
waveforms ξl,g , which are independently sparsely represented. Colors in the dictionary are indicating D g , i.e. the respective
rows used for reconstructing the windowed waveform. For sub-waveform detection, neighbored range gates which are represented
with the same dictionary elements are grouped to one sub-waveform.
The optimal α̂l,g can be formulated as
b l,g = argmin kD g αl,g − ξ l,g k2
α
subject to kαl,g k0 ≤ M,
(2)
b l,g are given by the non-zero activation indices y l,g . For a set
where the indices of M non-zero elements in α
b l,g , the reconstruction error of windowed waveform ξ l,g will be
of optimal activations α
b l,g − ξ l,g k2 ,
rl,g = kD g α
(3)
and r l,g = [rl,g ] is the vector collecting the reconstruction errors for all possible sets of dictionary elements.
8
This optimization can be solved with orthogonal matching pursuit (OMP, Tropp et al. (2006)), which falls in
the class of greedy algorithms. For this, the first dictionary element is chosen to be the one that maximizes
the absolute value of the inner product between the dictionary element itself and the sample which is
meant to be reconstructed; i.e. we maximize the collinearity. Each further dictionary element is chosen
in the same way, however, using the current residual instead of the sample until the number of the used
dictionary elements exceeds M . The dictionary elements and samples are normalized. An alternative would
be an exhaustive search through all combinations of non-zero activation indices, which however, would be
computationally challenging.
In more detail, the unary term in the energy functional penalizes the reconstruction error for a given set
of non-zero activation indices and the difference of its activations’ sum to 1, both describing the agreement
between data and a specific sparse representation model:
!
X
1
1
b v,l,g ,
r l,g +
abs 1 −
α
(4)
U(ξ l,g , y l,g ) =
Z1
Z2
v
where the first term is the normalized reconstruction error obtained by the non-zero activation indices
b l,g . Both unary
y l,g , and the second term is the normalized difference to 1 of the estimated activations α
terms are normalized P
by the standard deviation ζ of the values for each range gate, given by Z1 = ζ(r l,g )
b v,l,g )), in order to ensure an equal treatment of all range gates. The sum-toand Z2 = ζ(abs (1 − v α
one penalization serves as regularization to constrain the solution space to reasonable results, which can
alternatively be directly incorporated in (2) in a more restrictive way by introducing an additional sum-toone constraint, as often used in remote sensing (Bioucas-Dias et al., 2012).
A dictionary should have the following properties for our purpose: First, the elements should have a high
approximation ability, and second the choice of used dictionary elements for reconstruction should be unique
and stable. In order to build a suitable dictionary, a set of synthetically generated waveforms is sampled
and relevant waveforms are selected to serve as dictionary elements. We choose the most relevant synthetic
waveforms by selecting these ones which are most similar to the estimated waveforms and most dissimilar
to each other. For simulating dictionary elements, the parameters of the Brown model (Brown, 1977), such
as the waveform amplitude or the epoch, are sampled randomly, where the probability density functions
are chosen to resemble empirical distributions of Brown model parameters from a large set of open-ocean
echoes from the Envisat or Jason-1/-2 missions. In this way, the proposed algorithm will be as independent
as possible of the chosen altimetry mission.
3.3.2. Binary Term
The binary term in (1) serves to incorporate prior knowledge about the spatial relations between adjacent
range gates within a single waveform and between temporally consecutive waveforms within one cycle. As
mentioned before, we assume a sub-waveform to be a set of neighboring range gates which are sparsely
represented by a common set of dictionary elements, i.e. they share the same non-zero activation indices
y l,g , which is set to be the objective in (1). Therefore, we prefer neighboring range gates with similar
characteristics to be reconstructed with a common set of dictionary elements. The binary term in (1) is
given by
cos ξ l,g , ξ l,q , if y l,g = y l,q
B(ξ l,g , ξ l,q , y l,g , y l,q ) =
.
(5)
0,
if y l,g 6= y l,q
The similarity measure cos ξ l,g , ξ l,q relaxes the constraint of the representation of neighboring range gates,
e.g., in order to consider possible adaptions of the range window by the satellite on-board tracker.
3.3.3. Sparse Representation-based Conditional Random Field
As mentioned earlier, the evaluation of all possible sets of non-zero activation indices is computationally
difficult. However, this information is needed in the CRF to find the best set of indices for all range gates.
9
To overcome this problem, we optimize the CRF in a greedy manner. In detail, the search for the optimal
estimation of non-zero activation indices in iteration j = 1 for a windowed waveform is performed over
j=1
Dl,g
= {1, . . . , V } and fixed after the CRF application. Stopping after a single iteration is identical to the
usage of a correlation similarity as unary term. For each further iteration j > 1, the optimal sets of non-zero
j
j−1
j−1
activation indices will be derived from Dl,g
= {{y l,g
, 1}, . . . , {y j−1
l,g , V }} with y l,g being the optimal set of
non-zero activation indices from the previous iteration. The final CRF that is employed to find an optimal
estimation of non-zero activation indices in each iteration is given by the minimizer of the energy
!!
X 1
X
X
1
b v,l,g
r l,g +
abs 1 −
E(Y ) =
α
−w
B(ξ l,g , ξ l,q , y l,g , y l,q ) ,
(6)
Z1
Z2
v
l,g
l,g,q∈Q
where all entities correspond to the current iteration (index j is omitted for simplicity). The final optimal
b w.
set of non-zero activation indices given a specific weight w is denoted by Y
4. Sea Surface Height Estimation
In this study, we consider only the effect of the hyperparameter w, i.e. the choice of the relative weight
between the unary and binary terms in (1). As it turned out in our numerical experiments, no single choice
of w provides optimal results for different cycles, different study sites or altimetry missions. Therefore,
we define a range of possible sub-waveforms by varying w. A variation of the hyperparameter w in (1)
leads to a variation in the number of detected sub-waveforms. Therefore, we choose a set of reasonable
W = {w1 , . . . , wP } hyperparameters. In other words, we derive multiple sub-waveform partitionings, ranging
from a very coarse partitioning which includes only a few sub-waveforms to a very fine one which also captures
small peaks in separate sub-waveforms. For SSH estimation, we use all sub-waveforms which are sufficiently
large enough and which can be used for retracking. We end up with several equally likely SSHs at each
measurement location which could be thought to form a ’point cloud’. Finally, the Dijkstra algorithm
is employed to choose the smoothest combination of these SSHs, such that finally a single SSH at each
along-track location is provided.
4.1. Sea Surface Height
In the context of this study, we define SSH e
hSSH as given by
e
hSSH = a − (R + ∆dry + ∆wet + ∆iono + ∆retr ),
(7)
where a is the satellite’s altitude and R is the tracker range related to the fixed tracking gate and provided
in the altimeter data records. The atmospheric model corrections ∆dry , ∆wet and ∆iono , extracted from the
SGDR data, refer to the influence of the dry and wet part of the troposphere, as well as the ionospheric
influence on the signal. The retracking range correction ∆retr is derived from the retracking procedure,
described below, by converting the estimated epoch t0 from two-way travel time to range using ∆retr = 0.5ct0 ,
where c is the speed of light in vacuum.
Additional tidal corrections are applied after the final heights have been selected, to reduce the impact of
noise in the tidal corrections on the final height detection. In particular, ocean tide correction will introduce
a large noise component in coastal areas which will corrupt the selection of final SSHs through Dijkstra’s
algorithm.
Furthermore, we validate our retracking results against tide gauge data with at least hourly temporal resolution without any tidal and barometric corrections applied to remove possible effects from the comparison.
For this purpose we add relevant corrections
hSSH = e
hSSH − ∆ssb − ∆set − ∆lt − 0.468∆pt ,
10
(8)
where ∆ssb is the sea state bias correction which we compute as 5% of our retracked SWH. Additionally,
∆set is the solid earth tide correction, ∆lt is the loading tide and ∆pt is the pole tide correction. The factor
0.468 only applies the solid earth part of the pole tide correction, ignoring the part resulting from ocean
tides (Fenoglio-Marc et al., 2015). In other words, we do not apply the ocean tide correction or the inverse
barometric correction to render hSSH being directly comparable to the high rate tide gauge data.
4.2. Ocean model for Retracking
For retracking we employ a weighted 3-parameter ocean model (Halimi et al., 2013) which is given by
t − t0 − ψσ 2
A
√
1 + erf
xl (t) =
2
2σ
(9)
ψσ 2
exp −ψ t − t0 −
+ NT ,
2
for each two-way travel time t, centered on the tracking gate. Here, NT is the thermal noise which is
pre-computed from the first few range gates. The three fitted model parameters represent the amplitude
A, which is related to backscatter, the rise time of the leading edge σ which can be converted to SWH, as
well as the epoch or retracking gate t0 which refers to the position of the mid point of the leading edge.
Additionally, ψ is defined by
ψ=
1
4c
cos(2φ),
γh 1 + a/Re
sin2 (θ)
,
γ=
ln 4
(10)
where c is the speed of light in vacuum, Re is the radius of the Earth (Deng, 2003) and φ is the off-nadir
pointing angle which is assumed to be zero in our application. The antenna beamwidth parameter γ is
defined in (Brown, 1977, Eq. (4)) and can be computed from the beamwidth θ of the altimeter instrument.
4.3. Shortest-Path Algorithm for Finding the Best Set of Sea Surface Heights
The problem that we discuss here can be viewed as an optimization problem with certain constraints.
Given now a ’point-cloud’ of (equally) possible SSHs for each groundtrack-point, we need to realize a
consistent SSH at each measurement position by finding the optimal SSH candidates. Before applying a
shortest-path algorithm, we remove outliers from the ’point cloud’ through the RANdom SAmple Consensus
(RANSAC) algorithm (Fischler & Bolles, 1981) while assuming a linear model of SSH change in along-track
direction. Sea level change is not linear and thus, we apply the RANSAC algorithm with a threshold of 3 m
to a moving window covering about 20 s of measured data. By selecting the moving window to cover 20 s we
make sure that we always include a relatively large portion of water, especially at land-ocean transitions.
All points that deviate too far from the linear model estimated by the RANSAC algorithm are discarded
and the resulting set of accepted SSHs is then used to find the shortest path and estimating the optimal
SSHs at each measurement location as illustrated in Fig. 5 and Fig. 6. Here, we chose Dijkstra algorithm
(Dijkstra, 1959), but other shortest-path algorithms would also be possible. Dijkstra’s method requires
one to choose edge weights between individual connected nodes. In our application, we chose the height
differences between connected nodes as edge weights, thus, favoring smaller height changes over larger ones.
For the start and end point of the Dijkstra graph we use the first and last sea surface height at the start
and end position.
11
Figure 5: Schematic illustration of optimal sea surface height estimation employing Dijkstra algorithm. Dijkstra finds the
optimal path (illustrated in blue) through equally likely successively arranged sea surface heights. where edge weights are
derived from the difference between heights of 2 nodes.
Figure 6: Exemplary SSH pointcloud from the Triest study site as part of our proposed STAR algorithm. The black points
are selected by the RANSAC algorithm to estimate the linear model in order to discard outliers. Afterwards all points within
range of the linear model are used in the Dijkstra algorithm to find the best (smoothest) set of SSHs (orange line).
5. Results
5.1. Setting
We compare the STAR algorithm to existing retracking algorithms, such as the standard range derived
from a MLE4 retracking method provided in the SGDR data, the 30%-threshold retracker (ICE1, Martin
et al. (1983)) and the equally weighted 3-parameter ocean model (W3POM, see Eq. (9)). Furthermore, the
specialized coastal algorithms, such as the Adaptive Leading Edge Subwaveform retracker (ALES, Passaro
et al. (2014)) and the Improved Threshold Retracker (ITR, (Hwang et al., 2006)), the latter combined with
a threshold of 50%, are considered.
When implementing the STAR method as described above, we chose to set the neighborhood of the
windowed waveforms ξ to Nξ = 5. For the number of non-zero elements in (2) we select M = 2. Values of
M > 2 lead to significantly increased computation times while a larger combination of the limited number
of dictionary elements might result in less clearly defined sub-waveforms from unique combinations of the
basis elements. To avoid finding an optimal weighting parameter w for each measurement region, we run
(1) for five different choices of w ∈ W = {0.1, 0.5, 1, 2, 100}. This will result in five partitionings of the total
waveform into sub-waveforms, ranging from a very fine partitioning to a very coarse one. The parameters
for the computation of our basis elements are generated based on the average estimated parameters from
12
an application of the 50%-threshold retracker to the current block of waveforms (20 waveforms per block
in our framework). The mean epoch and amplitude serve as input into the basis element generation to
produce 1000 waveforms by randomly varying the mean parameters: the amplitude is randomly varied by
about 10% of its mean threshold-retracked value. The epoch is randomly sampled using a Gaussian weighted
distribution with its maximum at the mean threshold-retracked value. The waveheight is randomly sampled
from the range 0 m to 12 m, which is a most common range of waves. The cross-correlation coefficients
between the obtained 1000 generated waveforms are computed and only the 15 waveforms which are most
distinct from each other are kept and form the dictionary elements. In the following we utilize all the
detected sub-waveforms to derive the results for STAR.
Our STAR algorithm, ICE1, W3POM and ITR have been implemented in our in-house C++ altimetry
toolbox. As ITR provides heights for each detected leading edge, we assumed the first detected leading edge
to yield the correct retracking parameters, as we do not have any prior information on the heights. The
SGDR SSHs are extracted from the GDR data and ALES retracked ranges are extracted from external
GDR data. The SSHs from all retracking methods have been processed in the same way.
5.2. Sub-waveform Detection
First, we compare the sub-waveforms detected by STAR resulting from five different weights w ∈ W in
(1) to sub-waveforms from the method by Hwang et al. (2006) for an arbitrarily chosen waveform measured
off the coast of Croatia (Fig. 7). The waveform parts that have not been detected as part of a leading edge by
the method presented by Hwang et al. (2006) are indicated by the white background. In contrast, individual
sub-waveforms are depicted in alternating shades of grey. For this waveform, five potential leading edges
have been identified by the method of Hwang et al. (2006), each of which can be utilized to derive potential
sea surface heights. When comparing the partitioning of the total waveform for the different weights w ∈ W
by STAR, lower weights lead to a significantly increased number of identified sub-waveforms. In some cases
single sub-waveforms contain only one range gate since the binary term in (1) is weighted significantly lower
than the unary term. With increased weight w, the similarity constraint is enforced and the sub-waveform
size increases. For a weight w = 100, the identified sub-waveform corresponds to the entire waveform, which
leads to including the standard case of retracking the complete waveform in our derived point clouds.
In a next step, we compare the resulting point clouds of STAR (orange points, Fig. 8) to the point cloud
derived using Hwang et al. (2006) algorithm (black points, Fig. 8) for one exemplary cycle during low tide
conditions at the Bangladesh site. A small bias over the open ocean can be identified between both methods,
which is related to the chosen threshold of 50% used for ITR. For the sandbank near 22.325 ◦ N both point
clouds agree well. However, prior to reaching the sandbank the points derived from ITR drop rapidly by
about 10 m to a level of approximately -64 m (outside the plot boundaries), while the point cloud based on
STAR becomes less dense but still preserves enough meaningful SSHs of about -55.5 m up to the beginning
of the sandbank. Between Sandwip Island and Bangladesh mainland, STAR is able to derive SSHs from the
detected sub-waveforms, while ITR is more influenced by land returns disturbing the retrieved waveform,
which results in no meaningful SSHs that can be detected during the low tide conditions in this case.
5.3. Retracked SSHs
For validation of the retracked SSHs, we compare STAR SSHs to retracked SSHs from various conventional and coastal methods. The top of Fig. 9 shows retracked SSHs derived for one arbitrarily selected
cycle 165 at almost high tide at the Trieste study site. The bottom part of Fig. 9 shows the corresponding
return waveforms at each measurement location.
Over the open ocean between ∼45.46 ◦ N and 45.58 ◦ N, we find the waveforms corresponding well to the
theoretical Brown model and the SSHs from all retracking algorithms agree well. However, ICE1 shows a
small bias compared to the other retracking algorithms which is likely due to the 30%-threshold. At about
5-7 km off the nearest coast in the northern and southern part of the study site, the SSHs based on the
W3POM and SGDR indicate a rapid drop in sea level. This drop is related to the influence from peaks,
resulting from land influence from the Croatian and Italian mainlands. The coastal methods STAR, ALES
13
Figure 7: Sub-waveform partitioning using our method for weights from the set W = {0.1, 0.5, 1, 2, 100} inserted into (1). The
waveform is an arbitrarily selected coastal waveform from Jason-2 pass 196, cycle 69. Grey shaded background depicts the
individual sub-waveforms while a white background indicates no sub-waveform in that part of the total waveform.
and ITR are not influenced by the peaks as they consider only parts of the total waveform that do not
include the peaks at these measurement positions.
In the southern coastal area (45.37 ◦ N to 45.46 ◦ N), the SSHs from W3POM, again, generally agree with
the SSHs from ICE1 and STAR with the latter showing less noisy along-track variations and providing SSHs
right up to the coast of the Croatian mainland (Fig. 9). Between 2 to 4 km off the Croatian coast the SSHs
based on ALES and ITR show significant outliers which are related to the peak, due to the land influence
of the Croatian mainland, being located very close to the leading edge resulting in biased estimations.
Additionally, smaller peaks preceding the leading edge can be detected (Fig. 9, bottom) which are identified
as potential leading edges by the ITR algorithm and, thus, result in outliers due to our assumption to utilize
the first identified leading edge in the ITR algorithm. Similar behavior can be observed from the ICE1
algorithm, e.g. near ∼43.75 ◦ N, where the peaks preceding the leading edge lead to significant outliers.
In the northern coastal area (45.64 ◦ N to 45.71 ◦ N), we find good agreement between the SSHs based
on ALES, ITR and STAR, with a few significant outliers from ALES over the last two kilometers off the
nearest coast. At about 3.5 km off the Isola di Sant’ Andrea, a drop in SSH can be detected from all three
methods with ITR showing the strongest drop by several meters while the SSHs from STAR only drop less
than 1 m (Fig. 9) due to relatively broad peaks located directly at the leading edge.
Over the relatively shallow Laguna di Marano all methods, except for SGDR, provide a sea level similar
to the open ocean level during high tide, despite of the waveforms consisting of strong specular peak shapes
(Fig. 9, bottom).
14
Figure 8: Point cloud of SSHs resulting from STAR compared to ITR (Hwang et al., 2006).
5.4. Repeatability of the Retracked SSHs
The STAR method utilizes a randomized dictionary for sub-waveform decomposition; therefore it makes
sense to investigate whether this leads to an uncertainty in the final SSHs. To answer this question we
conduct a Monte Carlo study: we ran the STAR algorithm 1000 times, utilizing the arbitrarily selected
cycle 69 of pass 196 of the Jason-2 mission. Resulting SSHs are shown in Fig. 10 (top) with additional
zoom-in regions for two coastal areas, as well as for the open ocean. The bottom part of the figure displays
the corresponding root mean square difference (RMS) ranging from about 0 to 1 m, where the open ocean
area is also given in an additional sub-plot for a range of 0 to 5 cm.
Over the open ocean area, the variability along the groundtrack is in the range of about 15-20 cm (Fig.
10, top). Here, the repeatability of STAR is given by a RMS of less than 5 cm (Fig. 10, bottom).
In the southern coastal area, the along-track variability is in the range of 60 cm revealing a slight increase
in sea level towards the Croatian coast (Fig. 10, top). Due to the land influence on the shape of the waveform
as is visible in the parabolas (Fig. 9, bottom), the retracking results are sensitive with respect to the size of
the individual detected sub-waveforms. Nonetheless, we still find good repeatability in this region with RMS
of less than 20 cm (Fig. 10, bottom), which is less compared to the along track variability and significantly
less compared to the relatively large outliers produced by other retracking algorithms in this region (see Fig.
9, top).
Part of the variability in our current state of the algorithm is due to the Dijkstra algorithm, which
is employed to find the best SSHs in our point cloud. Single points at a measurement location can have
significant influence on the chosen path since we only allow edge connections between neighboring locations.
For example at ∼45.375 ◦ N, we find a standard deviation of almost zero (Fig. 10, bottom) and in the top
plot of Fig. 10 it is possible to identify a single point where the SSH-tracks of all runs intersect. Since all
paths obtained by Dijkstra algorithm include this point, preceding and succeeding SSHs are influenced and
tend to be close to the SSH at this point.
In the northern coastal region, the general along-track variability is about in the same range of 20 cm as
compared to the open ocean area. The repeatability in the shallow waters off the Isola di Sant’ Andrea is
similar to the Croatian coast in the south with a RMS of less than 20 cm. However, we find some significant
outliers in some runs which are related to a weak leading edge and strong peaks close to the leading edge
(e.g., Fig. 9, bottom). Over the Laguna di Marano a large variability can be detected (Fig. 10) due to
strong specular peak waveforms. Small changes in the detected sub-waveforms will have a significant impact
15
Figure 9: (top) Comparison of SSH derived by various retrackers along the Jason-2 pass 196 in the Gulf of Trieste for an
arbitrarily chosen cycle (cycle 165). The distance to the nearest coastline (DTC) is provided in light gray. (bottom) Radargram
of the corresponding waveforms of inside the study area. The waveform return power is given color coded and has been capped
at a power of 300 for visual reasons.
Table 1: Median values of the different quality measures for the study site in Trieste. See also top plots in Figs. 12, 13 and 14.
Here, ρ represents the correlation, N the percentage of retained cycles and σ the RMS.
Retracker
SGDR
W3POM
ICE1
ITR
ALES
STAR
ρ [-]
0.62
0.53
0.90
0.86
0.86
0.90
Open Ocean
N [%] σ [m]
85
0.43
85
0.53
99
0.15
99
0.17
97
0.16
99
0.15
Croatian Coast
ρ [-] N [%] σ [m]
0.54
59
0.53
0.46
42
0.74
0.34
36
0.83
0.17
3
1.62
0.73
81
0.28
Italian Coast
ρ [-] N [%] σ [m]
0.27
8
1.35
0.27
7
1.58
0.63
74
0.44
0.69
88
0.32
0.26
60
0.50
0.80
98
0.22
on the derived SSHs.
5.5. Comparison to Tide Gauge Data
In this experiment we compare the SSH estimated using various retracking methods with tide gauge
data for Trieste, Italy and Chittagong, Bangladesh. For comparison we utilize cycles 1 to 230 of Jason-2
data acquired over the described study sites (see Sec. 2.1), since for this period hourly tide gauge data as
well as data obtained by ALES is available. The tide gauge data between July 2009 and December 2014
are interpolated to the times of crossing for each cycle. As the tide gauge data is not corrected for tidal or
atmospheric pressure effects, we also do not apply these corrections to the altimetry data by employing (8).
We then remove outliers from the altimetric SSHs (deviation > 3 m from mean SSH) and evaluate correlation
and difference RMS with respect to tide gauge data. Due to the large number of available Jason-2 cycles,
we set the minimum number of cycles that are required to derive reliable correlations and RMS to 50.
The percentage of all cycles which meet the criteria above is shown in Fig. 11. The median results of
the following sections are summarized in Tables 1 and 2.
16
Figure 10: Repeatability of STAR. (top) SSHs obtained from 1000 runs on cycle 69, pass 196 at our Trieste study site. Additionally, we show zoomed in sub-plots for coastal and open ocean regions. Colors are chosen randomly. (bottom) Corresponding
root mean square difference (RMS) derived from 1000 runs where open ocean region is also provided in a zoomed in sub-plot.
Table 2: Median values of the different quality measures for the study site at the coast of Bangladesh. See also bottom plots
in Figs. 12, 13 and 14. Here, ρ represents the correlation, N the percentage of retained cycles and σ the RMS.
Retracker
SGDR
W3POM
ICE1
ITR
ALES
STAR
ρ [-]
0.94
0.92
0.95
0.97
0.95
0.97
Open Ocean
N [%] σ [m]
93
0.40
96
0.56
99
0.43
99
0.37
98
0.38
99
0.33
ρ [-]
0.64
0.31
0.50
0.63
0.14
0.88
Sandbank
N [%] σ [m]
15
1.06
24
1.82
55
1.43
72
1.15
47
1.38
78
0.64
ρ [-]
0.29
0.22
0.40
0.46
0.88
Channel
N [%]
28
29
34
27
68
σ [m]
1.33
1.34
1.33
1.36
0.72
5.5.1. Overall Correlation
We investigate the overall correlation ρ with tide gauge data at each 20 Hz along-track position, using
all available cycles (Fig. 11). The results for the Jason-2 mission are shown in Fig. 12 for the study site at
the Gulf of Trieste (top) and the coast of Bangladesh (bottom).
Over the open ocean area at the Trieste study site (Fig. 12, top), SSHs obtained by ICE1, ITR, ALES
and STAR show correlations of more than 0.8 with the tide gauge time series. For W3POM and SGDR
we find correlations between 0.3 and 0.7. Over the open ocean, southward of 22.2 ◦ N, in the Bangladesh
study area (Fig. 12, bottom), the SSHs derived from the considered retracking methods agree well to each
other. In the South, the coastal shelf transitions towards the deep ocean and the correlation begins to drop
to about ρ = 0.6 at the border of the study site. Over the deep ocean, the correlation with tide gauge data
from the Chittagong station drops rapidly, as already seen by Kusche et al. (2016). Over the central coastal
shelf up to about 15 km off the coast, we find correlations of up to 100% between retracked SSHs and tide
gauge data. In the immediate coastal areas of the Trieste and Bangladesh study sites, the MLE4 algorithm
used to derive the standard ranges in the SGDR data did not converge and consequently no SGDR SSHs
are available in these regions.
Towards the Croatian coast, we find ρICE1 and ρITR to decline to a level of 0.3 to 0.5 (Fig. 12, top) which
agrees to ρW3POM in this region, while ρALES shows a rapid decline to a level of 0 to 0.2 at about 3 km off
the coast. ρSTAR also declines towards the coast to a level of 0.5 to 0.7.
At the Italian coast, we again find a decline in overall correlation of the retracked SSHs with the tide
gauge time series (Fig. 12, top). About 4 to 5 km off the Isola di Sant’ Andrea, ρICE1 and ρALES rapidly
17
Figure 11: Percentage of the total number of available Jason-2 cycles (227) after applying outlier detection and minimum
number of cycles requirement. (top) Study site at the Gulf of Trieste. (bottom) Study site at the coast of Bangladesh. The
distance to the nearest coastline (DTC) is provided in light gray.
decline to a level of 0 to 0.2. While ρALES remains at this level, ρICE1 increases again to a level of 0.8 right
at the coast, which agrees well to ρW3POM , ρITR and ρSTAR . Over the Laguna di Marano, we generally find
a lower correlation level of 0 to 0.3 or even negative correlation in the southern part of the Laguna for all
retracked SSHs.
Starting about 15 km off the coast of Sandwip Island at the Bangladesh study site, the overall correlations
for all retrackers decline (Fig. 12, bottom). Over the sandbank area, ρICE1 drops to a level of 0.3 to 0.4,
ρITR shows a more moderate drop to a level of 0.5 to 0.7 and ρSTAR indicates only a small drop to 0.8, which
is achieved due to automatically removing low tide conditions where the sandbank height does not fit to the
conditions imposed by the RANSAC algorithm before applying Dijkstra’s algorithm (see Sec. 4.3). Over the
strip of about 13 km of open water which is part of the estuary of the Ganges-Brahmaputra-Meghna Delta,
we find ρW3POM and ρICE1 at nearly zero, while for ALES, ITR and STAR, ρALES ≈ 0.7, ρITR ≈ 0.8 to 0.9
and ρSTAR ≈ 0.9 are found in the central part of this region. Lower correlations of W3POM and ICE1 while
showing similar number of retained cycles and RMS (Table 2) are related to the fact that both methods
utilize the total waveform which mainly consists of a strong peak only. For W3POM this will often lead to
divergence of the estimation. The heights from the ICE1 algorithm will be influenced by the signal shape
of the waveform. This leads to varying parts of the waveform being used to derive the amplitudes and
consequently to inconsistencies when deriving the range correction.
5.5.2. Number of Retained cycles
We compare the number of cycles of SSH derived from each retracker which can be utilized at each
measurement location by iteratively eliminating the largest difference to the tide gauge time series from the
time series above until a correlation of at least 0.9 is achieved. This enables a more direct comparison of the
individual retracking methods. For evaluation, the number of retained cycles N is plotted for each 20 Hz
along-track position in both of our study areas (Fig. 13).
Over the open ocean regions of our study sites in the Gulf of Trieste and off the coast of Bangladesh,
18
Figure 12: Correlation ρ of SSHs derived from several retracking algorithms, including our STAR method, with hourly tide
gauge data. (top) Study site at the Gulf of Trieste. (bottom) Study site at the coast of Bangladesh. The distance to the
nearest coastline (DTC) is provided in light gray.
we find good agreement between NICE1 , NALES , NITR and NSTAR with 95 − 100% of retained cycles. For
NW3POM and NSGDR we find a slightly lower level of about 90% of retainable cycles resulting from divergence
of the parameter estimation. In addition, it is obvious that the time series of altimetry derived SSHs over
the deep ocean parts of the Bay of Bengal does not agree well with time series obtained on the coastal shelf.
In the Croatian coastal area, all retracking methods experience a small downward peak near 45.425 ◦ N.
About 4 km off the coast, we find a rapid drop in NALES and NITR to a level of 0 − 10% retained cycles,
with NITR rising back to about 65% of retained cycles 1 km off the coast. For NICE1 we also detect a drop
to about 30%, followed by a rise to 80% at about 1 km off the coast. About 60 − 80% of the SSHs can be
retained from applying the W3POM model in this coastal area. NSTAR declines from nearly 100% over the
open ocean to 80% at about 1 km off the coast.
At the Italian coast, we find nearly 100% of retainable cycles up to approximately 3-4 km off the coast
for ICE1, ITR, ALES and STAR. For NICE1 , NITR , NALES and NSTAR a steep drop to less than 20% of
retained cycles can be detected near 45.675 ◦ N resulting from land influences from the small islands that
separate the Laguna di Marano from the Gulf of Trieste (see also Sec. 5.3). The median percentage of
retainable cycles for SGDR and W3POM in this coastal area (Table 1) is quite low, indicating difficulties
to reach convergence when the total waveform is used. The sub-waveform retrackers provide a larger number
of retainable cycles as they are less affected by the land influences. Over the central part of the Laguna
di Marano, generally we find a level of 10 − 20% of retained cycles for all retrackers, except STAR, which
provides NSTAR ≈ 40%.
At the Bangladesh study site, NW3POM and NSGDR start to decline near 22.15 ◦ N towards the coast of
Sandwip Island and reach a minimum level of 0 − 15% at about 15 km off the coast. In addition, a decrease
of NICE1 , NALES , NITR and NSTAR can be detected. The STAR algorithm is able to retain at least 85% of
SSHs of the available cycles in this region. In the northern Bay of Bengal, the tidal amplitude can reach
several meters which leads to significantly different land influences on the shape of the waveform between
low and high tide. Consequently, this starts to affect the retrieved SSHs at about 15 km off the coast with
NALES , NITR and NSTAR starting to decrease towards the coast. For NALES we find a strong drop to 10%
19
Figure 13: Percentage of cycles retained to achieve a correlation of at least 0.9 with hourly tide gauge data from a total number
of 227 available cycles. (top) Study site at the Gulf of Trieste. (bottom) Study site at the coast of Bangladesh. The distance
to the nearest coastline (DTC) is provided in light gray.
in front of the Sandbank region mentioned before followed by an increase to a level of NALES ≈ 50% at
the coast. NITR and NSTAR drop to a level of 70 − 80% over the sandbank area. Here, ITR and STAR
perform similar, while the overall correlations suggest a lower correlation for ITR in this region compared
to STAR, due to already removed SSHs from low tide cycles by the RANSAC algorithm as part of STAR.
Over the strip of open water between Sandwip Island and the mainland of Bangladesh, NALES , NITR and
NSTAR indicate that all three algorithms are able to retain about 85 − 95% of the available cycles in the
central parts of this open water strip, while STAR is able to provide a larger number of cycles towards the
coastlines.
5.5.3. Root Mean Square Difference
The RMS between SSHs derived from the individual retracking methods and tide gauge data computed
at each 20 Hz along-track position provides an additional quality measure (Fig. 14). We apply the same
criteria for computation as introduced in Sec. 5.4.
Over the open ocean, we generally find results similar to the analysis of the overall correlation with
σICE1 , σITR , σALES and σSTAR agreeing well at a level of 0.15-0.2 m at the Trieste site and 0.2-1.0 m at the
Bangladesh site and σW3POM and σSGDR being slightly larger compared to the other methods.
Close to the Croatian coast, we find σICE1 , σITR and σALES to increase significantly with σALES reaching
more than 1.5 m. The large RMS derived for ALES is probably related to the small peaks in front of the
leading edge. Since ALES sub-waveforms are selected from the first range gate up to the end of the leading
edge, small peaks in front of the leading edge will lead to difficulties during the parameter estimation. For
STAR we find a smaller increase in RMS relative to the open ocean area to σSTAR ≈ 0.4 m. In the northern
part of the Trieste site, ICE1, ITR, ALES and STAR keep their level of RMS from the open ocean area
up to a distance of 4-5 km to the nearest coast.
In the northern part of the Bay of Bengal (Fig. 14, bottom), the SSHs from all retracking methods
become more noisy northward of about 22.0 ◦ N, with a minimum RMS level of about 0.25 m. Over the
20
Figure 14: Root mean square difference (RMS) derived for all along-track locations after applying the selection criteria mentioned above for all available cycles (Fig. 11) for both study sites. (top) Study site at the Gulf of Trieste. (bottom) Study site
at the coast of Bangladesh. The distance to the nearest coastline (DTC) is provided in light gray.
sandbank region, the RMS obtained from SGDR, W3POM, ICE1 and ALES show a rapid increase to at
least 1 m, while the RMS based on STAR exhibits a relatively smaller increase to a level of about 70 cm up
to the coast of Sandwip Island. The same level is found for σSTAR in the small strip of open water between
Sandwip Island and the mainland of Bangladesh with σALES and σITR at a similar level in the center of this
region.
5.6. Significant Wave Height and Sigma-Nought
Besides range corrections, the use of the 3-parameter retracking model also allows to retrieve SWH and
sigma-nought σ 0 . The SWHs and sigma-nought are selected based on the heights chosen by the shortestpath algorithm. Theoretically it would be possible to run the Dijkstra algorithm independently on the
point-clouds of SWH and sigma-nought. However, we think that the values should be consistent with the
selected heights.
No independently measured time series of wave height or wind speed is available for our two study sites.
Therefore, we compare temporal median, RMS and the number of retained cycles of good SWH and wind
speed which were obtained by individual retracking algorithms. We further compare the obtained results to
ERA-Interim model data. We will focus on the Bangladesh study site since no model data is available from
the Gulf of Trieste area.
5.6.1. Significant Wave Height
Fig. 15 shows the median SWH and RMS over all 239 cycles for the Bangladesh study site. Wave heights
from SGDR and ALES agree well with the model data over the open ocean while STAR and W3POM
are biased. At the same time, the median wave height values of all retrackers decline towards the coast,
reaching a minimum over the sandbank area. On the contrary, the model data does not show this decline.
Over open ocean, the RMS values from all retrackers agree well with ALES and STAR providing the lowest
21
Figure 15: SWH compared to ERA-Interim model data. Top: median over all cycles, Bottom: RMS over all cycles.
RMS at a level of about 60 cm while the model data suggests a level of about 48 cm. At about 10 km off the
coast the RMS values from all retrackers increase due to the sandbank’s influence during low tide cycles.
The low agreement between the retracked SWH and the model data, especially closer to the coast, can
be explained by the estimation of the retracking parameters. All utilized retrackers constrain the SWH
parameter to positive values since wave heights below zero are not physically meaningful. This sometimes
leads to forcing the wave height to be zero during the estimation for individual waveforms since the leading
edge, necessary for SWH parameter estimation, is only represented by a limited amount of range gates
during calm sea state conditions. For ALES method the problem appears less severe since it is a two stage
procedure in which SWH is estimated and averaged along-track during the first stage and kept fixed during
the second stage. For sub-waveform retrackers such as STAR the problem will occur more often in case
the selected sub-waveform which contains the leading edge is relatively small. In case the zero wave heights
are not utilized to derive the median and RMS value, SGDR, ALES and STAR agree well with the model
data (Fig. 16, Table 3). Over the sandbank area and the channel between Sandwip Island and the main
land, SGDR and W3POM quality decreases due to the low number of available cycles (Fig. 13, bottom).
Zero wave heights occur for about 10-15 % of the cycles during calm sea state conditions at the Bangladesh
study site, especially in coastal areas. Similar observations are made for the Trieste study site. In future,
STAR may be extended to a two-step procedure similar to ALES in order to counter these effects. We also
applied our method to reduced SAR (RDSAR) waveforms, which are conventional altimetry like waveforms
that are derived from the SAR signal during post-processing. The RDSAR data from Cryosat-2 data showed
that the problem does not occur when zero-padding is applied during the derivation of the reduced SAR
waveforms due to doubling the number of range gates available which allows better estimation of SWH (not
shown here).
5.6.2. Wind Speed
Wind speed has been derived from the estimated amplitudes utilizing the 2-parameter model by Gourrion
et al. (2002) as described in AVISO (2015). The model input consists of the retracked sigma-nought and
SWH. Here, we utilized only non-zero SWH in order to keep the wind speed estimate unbiased from problems
in SWH estimation. Similar to SWH, we display the median and RMS for each retracker and from the ERAInterim model data at each measurement position.
Over the open ocean, median wind speed and RMS from SGDR, W3POM, ICE1, ITR and STAR
agree well with the wind speed derived from ERA-Interim model data (Fig. 17, Table 4). However, SGDR
22
Figure 16: SWH compared to ERA-Interim model data. Top: median over all cycles, Bottom: RMS over all cycles. All
zero-SWH removed.
Table 3: Median values of SWH median and RMS over different regions along the track at the Bangladesh study site. All
zero-SWH removed. M denotes the median and σ the RMS.
Retracker
SGDR
W3POM
ALES
STAR
ERA-I
Open Ocean
M [m] σ [m]
0.94
0.69
1.06
0.80
0.78
0.57
0.89
0.59
0.84
0.45
Sandbank
M [m] σ [m]
0.59
0.80
1.13
1.20
0.66
0.86
0.62
0.68
0.63
0.38
Channel
M [m] σ [m]
0.67
0.82
0.65
0.71
-
and ICE1 are slightly biased above the model data, while W3POM, ITR and STAR are slightly biased
below the model data. ALES shows a stronger bias with respect to the model data. From about 22.15◦ N
towards the coast the waveforms start to get influenced by land returns leading to peaks moving along the
trailing edge towards the leading edge. Retrackers that utilize the full waveform, such as SGDR, W3POM
and ICE1, return amplitude estimates which are biased by the peak influence and thus do no longer provide
reliable wind speed. Sub-waveform method, such as ITR, ALES and STAR also show a slight decline in
wind speed, and over the sandbank area where one finds mostly specular peak waveforms, the derived wind
speed, especially during low tide, is no longer meaningful. In the center of the channel between Sandwip
Island and the Bangladesh main land only ALES and STAR are able to provide meaningful wind speed.
The reason for the bias in wind speed between SGDR and W3POM and STAR is due to the chosen
estimation model. For the SGDR data a MLE4 estimation method is employed utilizing a 4 parameter
model (Amarouche et al., 2004). The fourth parameter is the off-nadir angle that can be derived from the
slope of the trailing edge and which is assumed to be zero for the 3-parameter models. The off-nadir angle
influences the estimation of the amplitude (Amarouche et al., 2004, Eq. 8) which will lead to a general bias
in amplitude in case the off-nadir angle is different from zero. Here, the derived off-nadir angle is slightly
positive which leads to a smaller amplitude and consequently higher estimated wind speed. Even small
changes in sigma-nought have significant influence on the derived wind speed. However, for sub-waveform
methods, such as STAR, it is not feasible to try and estimate the off-nadir angle since the sub-waveforms
do not include enough range gates from the trailing edge for a reliable estimation.
23
Figure 17: Wind Speed compared to ERA-Interim model data. Top: median over all cycles, Bottom: RMS over all cycles.
Table 4: Median values of wind speed median and RMS over different regions along the track at the Bangladesh study site. M
denotes the median and σ the RMS.
Retracker
SGDR
W3POM
ICE1
ITR
ALES
STAR
ERA-I
Open Ocean
M [m/s] σ [m/s]
4.57
2.30
3.85
1.89
4.85
2.33
4.06
2.02
3.35
1.72
3.86
1.91
4.60
2.12
Sandbank
M [m/s] σ [m/s]
2.31
1.86
0.49
1.63
0.23
1.90
1.76
1.98
2.42
1.68
2.43
1.77
3.58
1.73
Channel
M [m/s] σ [m/s]
0.10
0.50
-0.15
0.38
0.01
1.61
1.02
1.68
2.31
1.70
2.94
1.37
5.7. Application to Jason-1 and Envisat Data
We investigate the application of STAR to altimetry data from the Jason-1 interleaved period to the
Trieste site (Fig. 2(a)) and to Envisat data for the Banlgadesh site (Fig. 2(b)). However for Jason-1
interleaved, there is no ALES-data available for comparison. Here, we focus on the number of cycles
retained to reach a correlation of > 0.9 (Fig. 18 and 19, top), as well as the RMS at each along-track
position (Fig. 18 and 19, bottom). For Jason-1, we found 95 cycles that overlapped with the available time
period for the tide gauge data at the Trieste station, while we found only 32 cycles of Envisat data that
overlapped with the tide gauge period available for the Chittagong station.
Utilizing data from the Jason-1 mission over the open ocean at the Trieste study site, ICE1, ITR
and STAR show more retained cycles, as well as a significantly smaller RMS compared to SGDR- and
W3POM-based SSHs. In the coastal area over about the last 5 km in front of the Croatian coast, STAR
derived SSHs fit well to the tide gauge data (Fig. 18). After crossing parts of the Croatian mainland, the
track transitions back to the ocean and we find ICE1, ITR and STAR retaining significantly more cycles
compared to SGDR- and W3POM. The general level of retained cycles is lower due to signal losses that
happened between the coasts of Croatia and Italy during nine of the Jason-1 interleaved cycles. At the
Italian coast, all ITR, ICE1 and STAR are able to maintain a high quality of SSHs in the coastal area with
which decreases only slightly over the last 1-2 km off the coast.
Additionally, we have utilized Envisat track 416 for our experiments, which crosses our Trieste study
site from the north-east to the south-west. However, the overlapping period with the available tide gauge
24
Figure 18: (top) Percentage of cycles retained to achieve a correlation of at least 0.9 with the hourly tide gauge data from a
total number of 95 available cycles of the Jason-1 interleaved mission. (bottom) Root mean square difference (RMS) derived
for all along-track locations. The distance to the nearest coastline (DTC) is provided in light gray.
data was less than one year and deriving any correlations or RMS from this short period containing only 10
cycles of the 35-day Envisat repeat orbit would not be meaningful. Nonetheless, SSHs derived from these
10 cycles of Envisat data compared well to the available tide gauge data (not shown here).
For the study site at the coast of Bangladesh, we found an overlap period of more than 3 years between
32 cycles (2007 - 2010) of Envisat data and the Chittagong tide gauge which allows to derive meaningful
correlations and RMS differences. Over the open ocean area, the number of retained cycles and the RMS
agree well between all the considered retracking methods with W3POM bases SSHs being slightly more
noisy (Fig. 19). About 15 km before the Envisat track reaches Sandwip Island, the differences between the
individual retracking methods become evident. Ratios of retained cycles for W3POM and SGDR show a
rapid drop to less than 25% over the sandbank area and a gradual increase towards the coast of Sandwip
Island to a level of 35% and 60%, respectively. This behavior is combined with a strong increase in RMS
differences which increase to more than 1 m. Similar behavior can be observed for ICE1 and ITR, which
both show a smaller drop in the number of retained cycles over the last 15 km towards the coast to a level
of about 75%, but their corresponding RMS increases significantly to values greater than 1 m. For ALES
there is no data available over the last 15 km in front of Sandwip Island. STAR shows the smallest drop in
number of retained cycles and recovers back to a level of more than 90% close to Sandwip Island. Over the
small strip of open water between Sandwip Island and the mainland of Bangladesh, the coastal retrackers
ITR, ALES and STAR retain about 90% of the available cycles providing SSHs that fit well to the tide
gauge data. SGDR, W3POM and ICE1 are not able to provide more than 0 − 25% of retained cycles and
exhibit significantly higher RMSs in this region.
6. Conclusion
A novel method for analyzing altimetric waveforms and deriving sea surface heights, SWH and sigmanought has been suggested. The proposed technique partitions the total waveform into individual sub25
Figure 19: (top) Percentage of cycles retained to achieve a correlation of at least 0.9 with the hourly tide gauge data from
a total number of 32 available cycles of the Envisat mission. (bottom) Root mean square differences (RMS) derived for all
along-track locations. The distance to the nearest coastline (DTC) is provided in light gray.
waveforms which can be analyzed in combination with existing retracking models. The sea surface heights
provided by STAR were found to be of at least the same quality or better compared to existing conventional
and coastal retracking methods over the open ocean, as well as in coastal regions. In addition, correlations
with tide gauge data revealed generally more usable cycles close to the coast in combination with lower root
mean square differences compared to existing methods. Of course, depending on the retracking model that
is combined with the derived sub-waveforms, it is possible to derive significant wave height and backscatter
in the same way.
Comparison of the derived sub-waveforms with the Hwang et al. (2006) method reveal good correspondence between identified parts of the waveform. We found the influence of the random component of STAR
on the SSH results to be at a level of less than 5 cm over the open ocean and at about 20 cm in coastal regions.
This is in range with modifications that can be applied to conventional retracking algorithms including biases
between different retracking methods, different weighting schemes or varying estimators; these effects are
considered to be in the order of a few centimeters. Sea surface heights derived from the STAR algorithm have
been extensively validated for Jason-2 data and compared to five independent available retracking methods,
as well as hourly in-situ tide gauge measurements. At the study sites in the Gulf of Trieste, Italy and off the
coast of Bangladesh, we found varying surface conditions including (deep) open ocean and shallow coastal
waters, temporally submerged sandbanks and transition zones between river estuaries and the ocean. Consequently, deriving five partitionings of the total waveform enabled the STAR algorithm to handle a larger
variety of waveform shapes compared to existing coastal retracking algorithms. Examination of estimated
SWH and wind speed revealed good agreement to other retracking algorithms as well as model data from
ERA-Interim. Furthermore, we applied STAR to altimetry data from the Jason-1 interleaved period, as well
as Envisat , which also resulted in significant improvements in the quality of coastal sea surface heights.
We are confident that the STAR method will enable a wide range of further studies, including a more
comprehensive validation of significant wave height and sigma-nought. In this context, we will also consider
an improved selection of final retacking results by considering SWH, sigma-nought etc. instead of only utilizing the SSHs. In addition, the algorithm can be improved further by tuning the a-priori hyperparameters
26
that control, e.g., the resolution of the sub-waveform, as well as extending the Dijkstra algorithm to reduce
the impact of potentially non-optimal sea surface heights of neighboring measurement locations. The derived
sub-waveforms will be combined with different available waveform models in order to adapt and extend the
concept to other regions, such as rivers and lakes. One might also consider a slightly different approach by
using the partitioning into sub-waveform to derive weighting schemes for retracking of the whole waveform.
Acknowledgements
We acknowledge funding through the BanD-AID (KU 1207/19-1) project. We also thank the associate
editor and three anonymous reviewers for their constructive comments which helped us to significantly
improve the manuscript.
References
Amarouche, L., Thibaut, P., ZaZanife, O. Z., Dumont, J.-P., Vincent, P., & Steunou, N. (2004). Improving the Jason-1 Ground
Retracking to Better Account for Attitude Effects. Marine Geodesy, 27 , 171–197. doi:10.1080/01490410490465210.
AVISO (2015). OSTM / Jason-2 Products Handbook . CNES / NASA http://www.aviso.altimetry.fr. SALP-MU-M-OP15815-CN, Issue 1.9.
Bioucas-Dias, J. M., Plaza, A., Dobigeon, N., Parente, M., Du, Q., Gader, P., & Chanussot, J. (2012). Hyperspectral unmixing
overview: Geometrical, statistical, and sparse regression-based approaches. IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing, 5 , 354–379.
Birkett, C. M., & Beckley, B. (2010). Investigating the performance of the Jason-2/OSTM radar altimeter over lakes and
reservoirs. Marine Geodesy, 33 , 204–238.
Brown, G. S. (1977). The average impulse response of a rough surface and its applications. IEEE Trans. Antennas Propagat.,
25 , 67–74.
Chelton, D. B., Ries, J. C., Haines, B. J., Fu, L.-L., & Callahan, P. S. (2001). Satellite altimetry. In L.-L. Fu, & A. Cazenave
(Eds.), Satellite Altimetry and Earth Sciences: A Handbook of Techniques and Applications chapter 1. (pp. 1–132). Academic
Press.
Chelton, D. B., Walsh, E. J., & MacArthur, J. L. (1989). Pulse compression and sea level tracking in satellite altimetry. J.
Atmos. Oceanic Technol., 6 , 407–438.
Dee, D. P., Uppala, S. M., Simmons, A. J., Berrisford, P., Poli, P., Kobayashi, S., Andrae, U., Balmaseda, M. A., Balsamo, G.,
Bauer, P., Bechtold, P., Beljaars, A. C. M., van de Berg, L., Bidlot, J., Bormann, N., Delsol, C., Dragani, R., Fuentes, M.,
Geer, A. J., Haimberger, L., Healy, S. B., Hersbach, H., Hólm, E. V., Isaksen, L., Kållberg, P., Köhler, M., Matricardi, M.,
McNally, A. P., Monge-Sanz, B. M., Morcrette, J.-J., Park, B.-K., Peubey, C., de Rosnay, P., Tavolato, C., Thépaut, J.-N.,
& Vitart, F. (2011). The era-interim reanalysis: configuration and performance of the data assimilation system. Quarterly
Journal of the Royal Meteorological Society, 137 , 553–597. doi:http://dx.doi.org/10.1002/qj.828.
Deng, X. (2003). Improvement of Geodetic Parameter Estimation in Coastal Regions from Satellite Radar Altimetry. Ph.D.
thesis Curtin University of Technology.
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische Mathematik , 1 , 269–271.
ESA (2007). Envisat RA2/MWR Product Handbook . European Space Agency (ESA) https://earth.esa.int/handbooks/
ra2-mwr/. URL: https://earth.esa.int/handbooks/ra2-mwr issue 2.2.
Fenoglio-Marc, L., Dinardo, S., Scharroo, R., A., R., M., D. S., B., L., Becker, M., Benveniste, J., & Weiss, R. (2015). The
German Bight: A Validation of CryoSat-2 altimeter data in SAR mode. Advances in Space Research, 55 , 2641 – 2656.
doi:10.1016/j.asr.2015.02.014.
27
Fischler, M. A., & Bolles, R. C. (1981). Random sample consensus: a paradigm for model fitting with applications to image
analysis and automated cartography. Communications of the ACM , 24 , 381–395.
Fu, L.-L., & Cazenave, A. (Eds.) (2001). Satellite Altimetry and Earth Sciences: A Handbook of Techniques and Applications.
Academic Press.
Gommenginger, C., Thibaut, P., Fengolio-Marc, L., Quartly, G. D., Deng, X., Gómez-Enri, J., Challenor, P., & Gao, Y. (2011).
Retracking altimeter waveforms near the coasts: A review of retracking methods and some applications to coastal waveforms.
In S. Vignudelli, A. G. Kostianoy, P. Cipollini, & J. Benveniste (Eds.), Coastal Altimetry. Springer-Verlag Berlin Heidelberg.
Gourrion, J., Vandermark, D., Bailey, S., Chapron, B., Gommenginger, G. P., Challenor, P. G., & Srokosz, M. A. (2002). A
two-parameter wind speed algorithm for Ku-Band altimeters. J. Atmos. Oceanic Technol., 19 , 2030–2048.
Halimi, A., Mailhes, C., Tourneret, J.-Y., & Snoussi, H. (2016). Bayesian estimation of smooth altimetric parameters: Application to conventional and delay/doppler altimetry. IEEE Transactions on Geoscience and Remote Sensing, 54 , 2207–2219.
Halimi, A., Mailhes, C., Tourneret, J.-Y., Thibaut, P., & Boy, F. (2013). Parameter estimation for peaky altimetric waveforms.
IEEE Trans. Geosci. Remote Sensing, 51 , 1568–1577.
Hayne, G. S. (1980). Radar altimeter mean return waveforms from near-normal-incidence ocean surface scattering. IEEE
Trans. Antennas Propagat., 28 , 687–692.
Hwang, C., Guo, J., Deng, X., Hsu, H.-Y., & Liu, Y. (2006). Coastal gravity anomalies from retracked geosat/gm altimetry:
improvement, limitation and the role of airborne gravity data. Journal of Geodesy, 80 , 204–216.
Hwang, C., Kao, E., & Parsons, B. (1998). Global derivation of marine gravity anomalies from SEASAT, GEOSAT, ERS-1
and Topex/Poseidon altimeter data. International Geophysical Journal, 134 , 449–459.
Kusche, J., Uebbing, B., Rietbroek, R., Shum, C. K., & Khan, Z. H. (2016). Sea level budget in the bay of bengal (20022014)
from grace and altimetry. Journal of Geophysical Research: Oceans, 122 . doi:10.1002/2015JC011471.
Lafferty, J. D., McCallum, A., & Pereira, F. C. N. (2001). Conditional random fields: Probabilistic models for segmenting and
labeling sequence data. In Proc. International Conference on Machine Learning (pp. 282–289). URL: http://dl.acm.org/
citation.cfm?id=645530.655813.
Lee, H., Shum, C., Emery, W., Calmant, S., Deng, X., Kuo, C.-Y., Roesler, C., & Yi, Y. (2010). Validation of jason-2 altimeter
data by waveform retracking over california coastal ocean. Marine Geodesy, 33 , 304–316.
Martin, T. V., Zwally, H. J., Brenner, A. C., & Bindschadler, R. A. (1983). Analysis and retracking of continental ice sheet
radar altimeter waveforms. Journal of Geophysical Research: Oceans (1978–2012), 88 , 1608–1616.
Maus, S., Green, C. M., & Fairhead, J. D. (1998). Improved ocean-geoid resolution from retracked ERS-1 satellite altimeter
waveforms. Geophysical Journal International, 134 , 243–253.
Olshausen, B. A., & Field, D. J. (1997). Sparse coding with an overcomplete basis set: A strategy employed by v1? Vision
Research, 37 , 3311 – 3325. URL: http://www.sciencedirect.com/science/article/pii/S0042698997001697. doi:http://
dx.doi.org/10.1016/S0042-6989(97)00169-7.
Passaro, M., Cipollini, P., Vignudelli, S., Quartly, G. D., & Snaith, H. M. (2014). ALES: A multi-mission adaptive subwaveform
retracker for coastal and open ocean altimetry. Remote Sensing of Environment, 145 , 173–189.
Quartly, G. D., Srokosz, M. A., & McMillan, A. C. (2001). Analyzing altimeter artifacts: Statistical properties of ocean
waveforms. Journal of Atmospheric and Oceanic Technology, 18 , 2074–2091.
Rosmorduc, V., Benveniste, J., Bronner, E., Dinardo, S., Lauret, O., Maheu, C., Milagro, M., & Picot, N. (2011). Radar
altimetry tutorial. http://www.altimetry.info/. URL: http://www.altimetry.info/ benveniste, J. (ESA) and Picot, N.
(CNES) Ed.
Sandwell, D. T., & Smith, W. H. F. (1997). Marine gravity anomaly from Geosat and ERS-1 satellite altimetry. Journal of
Geophysical Research, 102 , 10,039–10,054.
Sandwell, D. T., & Smith, W. H. F. (2005). Retracking ERS-1 altimeter waveforms for optimal gravity field recovery. Geo-
28
physical Journal International, 163 , 79–89. doi:10.1111/j.1365-246X.2005.02724.x.
Savcenko, R., & Bosch, W. (2008). EOT08a empirical ocean tide model from multi-mission satellite altimetry. DGFI Report
No. 81.
Tropp, J. A., Gilbert, A. C., & Strauss, M. J. (2006). Algorithms for simultaneous sparse approximation. part i: Greedy
pursuit. Signal Processing, 86 , 572–588.
Uebbing, B., Kusche, J., & Forootan, E. (2015). Waveform retracking for improving level estimations from topex/poseidon,
jason-1, and jason-2 altimetry observations over african lakes. IEEE Transactions on Geoscience and Remote Sensing, 53 ,
2211–2224.
Wang, Y. (2004). Ocean Tide Modeling in Southern Ocean. Technical Report 471 Department of Civil and Environmental
Engineering and Geodetic Science, The Ohio State University, USA Columbus, Ohio 43210.
Wright, J., Ma, Y., Mairal, J., Sapiro, G., Huang, T. S., & Yan, S. (2010). Sparse representation for computer vision and
pattern recognition. Proceedings of the IEEE , 98 , 1031–1044.
29
| 1cs.CV
|
arXiv:1712.08328v1 [math.OC] 22 Dec 2017
A simple introduction to Karmarkar’s Algorithm
for Linear Programming
Sanjeev Saxena∗
Dept. of Computer Science and Engineering,
Indian Institute of Technology,
Kanpur, INDIA-208 016
December 21, 2017
Abstract
An extremely simple, description of Karmarkar’s algorithm with very few
technical terms is given.
1
Introduction
A simple description of Karmarkar’s algorithm[5] together with analysis is
given in this paper. Only knowledge of simple algebra, vector dot product
and matrices is assumed. Even though the method is described in several
books [8, 1, 2, 3, 7], analysis is either left out [8] or is fairly complicated.
We show that the presentation of Roos, Terlaky and Vial[6] can be further
simplified, (mainly) by using only essential notation. Allmost all results in
this paper are from standard textbooks (see References).
Let A be an m × n matrix of rank m and e is a vector of all ones.
Karmarkar’s Problem is:
min cT x subject to constraints Ax = 0, eT x = 1 and x ≥ 0.
Further, it is assumed that optimal value of cT x is zero and all ones
vector is feasible, i.e., Ae = 0. We have to either find a point of cost 0 or
show that none exist
We first scale the variables: x0 = xn, then the problem becomes min cT (x0 /n)
or equivalently min cT (x0 ) subject to Ax0 = 0, eT x0 = n and x0 ≥ 0. We will
drop the primes, and the problem is[6]:
∗
E-mail: [email protected]
1
min cT x subject to constraints Ax = 0, eT x = n and x ≥ 0.
Remark: Problem is trivial if cT e = 0, hence we assume that cT e > 0.
We will assume that all m + 1 equality constraints are linearly independent (else we eliminate redundant rows of A).
We need few definitions. Standard simplex consists of points in n
dimensions s.t. eT x = n, x ≥ 0. The centre of the simplex is e = (1, 1, . . ., 1).
Let R be the radius of outer sphere, the smallest sphere containing the
standard simplex (circumscribes standard simplex). R is the distance from
e to one of the corner point (see Figure for an example in three dimensions)
of the standard simplex (say, (0, 0, . . ., n)), thus R2 = (0p− 1)2 + . . . + (0 −
1)2 + (n − 1)2 = (n − 1)2 + (n − 1) = n(n − 1), or R = n(n − 1).
Let r be the radius of another sphere— the largest sphere centred at e
and completely inside the standard simplex (inscribed inside the standard
simplex). This sphere will be tangent to each face. Each face will have one
coordinate as 0. By symmetry, all other coordinates at the point of contact
will be same (say w). As the point of contact is on the standard simplex S,
0 + w + w + . . . + w = n or (n − 1)w = n or w = n/(n − 1). Hence,
r2 = (0−1)2 +(w−1)2 +. . .+(w−1)2 = 1+(n−1)(w−1)2 = 1+
Or
r
r=
1
n
=
n−1
n−1
n
n−1
We take e = (1, 1, . . ., 1) as the starting point (which by assumption is
feasible). Then, we minimise the objective function over a smaller sphere,
which we will call the inner sphere, having same centre e, but radius αr,
less than r (we will see in Section 5.2 that α can be chosen as 1/(r + 1)).
Let us assume that the minimum occurs at point z. Then, in next iteration,
we take the starting point as z. The problem of minimisation on sphere is
2
discussed in next section (Section 2). Finally, the point z is mapped to e and
the process repeated; the details of mapping are in Section 4. In Section 3,
it is shown that the objective value at next point z is a fraction of that at
the initial point e.
2
Mathematical Preliminaries
Let P be a point in (possible hyper) plane and n̂ a unit vector normal to it.
If P0 is any point in plane, then the (vector) dot product will be zero, i.e.,
P P0 • n̂ = 0. Writing in full,
n1 (x1 − x01 ) + n2 (x2 − x02 ) + . . . + nk (xk − x0k ) = 0 or equivalently,
n1 x1 + n2 x2 + . . . + nk xk = n1 x01 + n2 x02 + . . . + nk x0k = C Thus, for a • x = b,
normal will be in the direction of vector a.
Equation of sphere with centre Q as β is: (x − β) • (x − β) = r2 Let P
be any point in plane a • x = b, then we know that a • xP = b. If we want
QP to be perpendicular to plane, then (xP − β) = (constant)a
P
Consider the problem [4]: min cT x subject to (x − a)2 ≤ ρ2 . There
are two points on the sphere (see figure for an example in two dimensions),
where planes parallel to given plane can be tangent— one corresponding to
maxima and other minima. These are +n and −n. These points (say “P ”)
are in direction ±n̂ (a unit vector in direction of c) with length |xP − β| = r.
Thus, to minimise cT x over our sphere, we start at the centre β and take a
step of length r in direction −cT Less informally, if c = 0, all points on the
sphere are optimal. If c 6= 0, the solution is obtained by taking a step of
length ρ (radius of sphere) from the centre a in the direction −c; this can
be seen by considering “parallel” planes cT x =constant. The point at which
minimum will be attained will be the point of contact on the sphere to a
tangent plane.
3
Next [4] consider the problem: min cT x subject to Ax = b and (x −
2
a) ≤ ρ2 . If c = 0, all points common to (i.e., on the intersection of) sphere
and the plane are optimal. If c 6= 0, let c̄ be the orthogonal projection of
c onto the plane Ax = b. If c̄ = 0, then c is a linear combination of rows
of A and the objective function (on the intersection of sphere and plane)
is constant, and all “feasible” points are also optimal. If c̄ 6= 0, then the
solution is obtained by taking a step of length equal to the radius of lowerdimensional sphere (intersection of our sphere and plane Ax = b) from the
centre a in the direction −c̄.
P
3
Solution on inner Sphere
Let us consider the problem: min cT x subject to Ax = 0, i xi = n and
P
2
2 2
i (xi − 1) < α r , i.e, we are minimising only over points of the inner
sphere. As the inner sphere is completely inside the standard simplex, we
can drop the constraint eT x = n (and also x ≥ 0), and the problem becomes
P
min cT x s.t. Ax = 0 and (xi − 1)2 < α2 r2 .
As the point e is feasible, Ae = 0, e lies on the plane Ax = 0. Moreover,
as e, the centre of inner sphere, lies on Ax = 0, the intersection of Ax = 0
and the inner sphere will be a “sphere” of the same radius αr but in lower
dimension. (e.g., intersection of a sphere and a plane containing the centre
is a circle with same radius and centre).
The minimum value of a linear function on sphere will be at a point
(lower one) where linear function touches the sphere. The radius vector at
that point will be perpendicular to the plane. Let us assume that p̂ is a unit
vector in that direction. Then the point at which minimum will be attained
will be e − p̂(radius). Thus, for outer sphere it will be (say) zR = e − p̂R and
for the inner sphere it will be zα = e − p̂αr. We will see how to determine
p̂ later (see Section 4.2).
As the outer sphere completely contains the solution space, the minimum
value (of objective function cT x) will be smaller than (or equal to) the actual
optimal value which is zero. Thus, (assuming minimum value is at zR )
P
0 ≥ cT zR = cT (e − p̂R) or cT p̂R ≥ cT e
T
Thus, cT p̂ ≥ cRe
As the inner sphere is inside the simplex, value of the objective function
can not be less than that on the simplex. But, as the optimal value on
simplex is zero, optimal value on the inner sphere is non-negative. If the
4
minimum value occurs at zα on the inner sphere, then the value of objective
function is:
0 ≤ cT zα = cT (e − p̂αr) or cT p̂αr ≤ cT e
But as
cT e
R
≤ cT p̂, we get
!
T
T
T
T
c zα = c e − c p̂αr ≤ c e −
cT e
αr
αr = cT e 1 −
R
R
Thus, if we start with initial solution e, then zα (the next solution)
is
αr
an improvement (in value of objective function) by a factor of 1 − R over
the initial solution.
4
Karmarkar Transform and Algorithm
We will like to map the new point z again to e to repeat the process. Thus,
for a = (a1 , a2 , . . ., an ) and x = (x1 , x2 , . . ., xn ), we define the transform
y = Ta (x) with (for i = 1, . . ., n)
ai xi
yi = n P
j aj xj
If all ai s and xi s are non-negative (at least ai xi should be non-zero),
each component of transform will be less than n and sum of all coordinates
will be n, thus the range
is again our standard simplex. Moreover, if b =
1 1
1
−1
a = a1 , a2 , . . ., an , then Tb will be the inverse transformation. Thus
this transformation is one-to-one on our standard simplex.
Remark: This can also be seen directly. If Ta (x) = Ta (y), then xy11 =
P
P
x1
xn
xi = yi = 1, we have r = 1.
y2 = . . . = yn = r (say). But as
Moreover, if λ is any number then if y 0 = Ta (λx) then
nλai xi
nλai xi
nai xi
yi0 = P
= P
=P
= yi
λaj xj
λ aj xj
aj xj
Thus, Ta (λx) = Ta (x).
4.1
Modified Problem
Assume that we are applying transform Ta (with ai = 1/zi ), then point z
will be mapped to e. Let D be a diagonal matrix with diagonal entries:
D = diag(a1 , . . ., an ).
5
If z is any feasible point, then for any positive x > 0 satisfying eT x = 1,
we saw that there is a unique point ξ (which depends on x) s.t. x = Tz (ξ).
P
Remark: x = Tz (ξ) implies xi = zi ξi /( zi ξi ), thus ξi = ρ xzii where ρ is
P
such that ξi = 1.
P
P
Then, equation Ax = 0 is equivalently to j Aij xj = 0 or j Aij ρxj = 0
P
or equivalently, Aij ξj zj = 0 or Aξz = 0.
P
P
The objective function cT x = ci xi = ρ1 ci ξi zi . As optimal value of
cT x is zero, it follows that the optimal value of the transformed objective
P
function ci ξi zi is also 0.
Replacing ξ by x0 and using Z = diag(z1 , . . ., zn ), the transformed problem is:
min(Zc)T x0 subject to AZx0 = 0, eT x0 = n and x0 ≥ 0.
Moreover, as z is a feasible point Az = 0 or equivalently AZe = 0, thus e is
again feasible.
We can thus repeat the previous method with Zc instead of c and AZ
instead of A. In other words, we have to minimise the modified objective
function over inscribed sphere of radius αr i.e., the “modified” problem is:
min(Zc)T x subject to AZx = 0, eT x = n and ||x − e|| ≤ αr.
4.2
Result from Algebra– Projection Matrix
Assume that A is m×n matrix, then rank of A is said to be m, with m < n iff
all m rows of A are linearly independent, i.e., β1 A1 + β2 A2 + . . . + βm An = 0
(here 0 is a row vector of size n) has only one solution βi = 0. Thus, if v is
any 1 × m matrix (a column vector of size m), then vA = 0 implies v = 0.
We use the fact that the matrix AAT has rank m and is invertible1 .
Let A be an m × n matrix, then Ax = b (hence Ax = 0) represents a set
of m equations. Each equation will be a hyperplane in n dimensions. Let v
be a vector of size n, we will like to “project” v onto the (lower dimension or
intersection of) hyperplane Ax = b. If P = pv is the projection, then [9], we
wish to write p = P v, as best as possible, as p = P v = α1 a1 + α2 a2 + . . . +
αm am where a1 , . . ., am are rows of A. Projection p will again be a vector
of size n. This can be written as p = AT α.
1
As A is m × n matrix, AT will be n × m matrix. The product AAT will be an m × m
square matrix. Let y T be an m × 1 matrix (or y is a row-vector of size m).
Consider the equation (AAT )y T = 0. Pre-multiplying by y we get yAAT y T = 0 or
(yA)(yA)T = 0 or the dot product < yA, yA >= 0 which, for real vectors (matrices)
means, that each term of yA is (individually) zero, or yA = 0, which implies y is identically
zero. Thus, the matrix AAT has rank m and is invertible.
6
The error of projection (or rejection) E = v − p = v − P v will also be
a vector of size n. Then as error E can not be in these hyperplane, i.e., it
should be “perpendicular” to these hyperplanes, thus we want AE = 0. Or
A(v − p) = 0, or A(v − AT α) = 0, thus Av = AAT α, or α = (AAT )−1 Av.
T )−1 Av, or P = AT (AAT )−1 A.
Hence, p = P v = AT α, or P v = AT (AA
Or, E = v − P v = v − AT (AAT )−1 Av = I − AT (AAT )−1 A v
Thus, to summarise, projection matrix is P = AT (AAT )−1 A and Rejection matrix (to get the part perpendicular to the hyperplanes) is: I −
AT (AAT )−1 A
4.3
Algorithm
Start the algorithm with x = e and find z, the optimal value on the inscribed
sphere of radius αr (we will see more details of this step later).
In a typical iteration, we next apply a transformation Tz −1 to map z to
e (we need a transformation Tb such that z gets mapped to e, then b = 1/z).
Modify A and c and find a new z (say z 0 ), the optimal value on the inscribed
sphere of radius αr.
Remark: As we know that transformed value of z is e, transform Tz −1 is
not actually applied, only the values of A and c are updated.
0
Then apply inverse transformation Tz−1
−1 (or Tz ) to map z to original
0
space, to get the value of x for next iteration.
The formal algorithm
is (value of α is fixed to 1/(r + 1) in Section 5.2):
q
n
1
Initialise: r = n−1 , α = r+1
, x = e = (1, . . ., 1)
Main step: If cT x < , return current x as solution of desired accuracy.
Let D = diag(x1 , . . ., xn )
AD
I
P =
!
Let
cP
=
1−P
T
PP
cP
||cP ||
z = e − αrp̂
nDy
x = T
e Dy
p̂ =
7
T
−1
P (cD)T
5
Analysis—Potential Function
For analysis, we define a potential function:
T
Φ(x) = n log c x −
n
X
log xi
i=1
As eT x = n we have n1 xi = 1. But as Arithmetic mean is greater than
or equal to geometric mean:
P
P
P xi
Q
1/n
1=
or taking logs, n1 log xi ≤ 0, or log xi ≤ 0, or
n ≥ ( xi )
Φ(x) ≤ n log cT x or
cT x ≥ exp
Φ(x)
n
Next observe that
Φ(λx) = n log cT (λx) −
n
X
log(λxi )
i=1
=
T
log c x + log λ −
n
X
log xi +
= n log c x −
n
X
!
log λ
i=1
i=1
T
n
X
log xi = Φ(x)
i=1
Let x be a positive vector in our standard simplex and let y = Tx (z),
then
nxi zi
yi = P
, or Φ(y) = Φ(Tz (x)) = Φ(xz)
x i zi
Moreover, Φ(xz) = n log cT (xz) −
P
log zi
Finally,
Pn
i=1 log xi zi
= n log cT xz −
P
X
log zi
log xi −
∆Φ = Φ(x) − Φ(y) = Φ(x) − Φ(xz)
= n log cT x −
n
X
log xi − n log cT xz −
log xi −
X
i=1
= n log
cT x
cT xz
+
X
log zi
T
x
But log ccT xz
is the ratio of original
and transformed problems, which we saw
r
reduces by at least 1 − α R . Thus:
8
r
∆Φ ≥ −n log 1 − α
R
+
X
log zi
We will show that ∆ is more than a constant. But, before continuing
with analysis, we need some more results from algebra.
5.1
Function Ψ
In this section, all logs are to base e. Assume that |x| < 1. As log(1 + x) =
2
3
4
x − x2 + x3 − x4 + . . ., for −1 < y < 0, (let y = −x) we have
3
4
2
y − log(1 + y) = −x − log(1 − x) = x2 + x3 + x4 + . . . > 0
We define[6] Ψ(x) = x − log(1 + x). Observe that Ψ(0) = 0 Hence,
2
3
4
Ψ(x) = x2 − x3 + x4 − . . . And, Ψ(x) ≥ 0 for x > −1
Claim 1 If x > 0 and a2 = b2 + c2 , then if ax, bx, cx are greater than −1
(for log(1 + hx) to be defined), then
Ψ(−|a|x) > Ψ(bx) + Ψ(cx)
Proof: From a2 = b2 + c2 , we know that |a| ≥ |b| and |a| ≥ |c|. Without
loss of generality, assume a < 0, then −|ax| = ax and we have
3 3
4 4
2 2
Ψ(ax) = a 2x − a 3x + a 4x − . . .
2 2
3 3
4 4
Ψ(bx) = b 2x − b 3x + b 4x − . . .
3 3
4 4
2 2
Ψ(cx) = c 2x − c 3x + c 4x − . . .
As a2 = b2 + c2 , coefficients of x2 on both sides are equal.
3
3 3
3 3
We thus need to compare − a 3x with − (b +c3 )x . As a2 = b2 + c2 ,
a3 = ab2 + ac2 . Thus, we are comparing −x(ab2 + ac2 ) and −x(b3 + c3 ) or
equivalently 0 and xb2 (a − b) + xc2 (a − c).
As a < 0 and −a > |b| and −a > |c|, the expressions can be re-written
as: 0 and −xb2 (b − a) − xc2 (c − a). Both terms inside brackets are positive,
and as x > 0, right hand side will be negative.
Claim 2 If a2 = b2 + c2 + . . . + s2 , then if ax, bx, cx, . . . are greater than −1
(for log(1 + hx) to be defined), then if x > 0, Ψ(−|a|x) > Ψ(bx) + Ψ(cx) +
. . . + Ψ(sx)
Proof (induction): Let v 2 = c2 + . . . + s2 . Then a2 = b2 + v 2 . Then from
the above claim, Ψ(−|a|x) > Ψ(bx) + Ψ(−|v|x). By induction hypothesis,
Ψ(−|v|x) > Ψ(cx)+. . .+Ψ(sx) Proof follows by adding the two inequalities.
9
5.2
Analysis Continued
As − log(1 − x) = x +
x2
2
+
x3
3
+ . . . > x, hence,
∆Φ ≥ −n log 1 − α
r
R
+
X
log zi
X
r
+
log zi
R X
= αr2 +
log zi
= nα
But z = e − αrp̂ or zi = 1 − αrpi
And, log zi = log(1−αrpi ) = − (−αrpi − log(1 − αrpi ))−αrpi = −αrpi −
Ψ(−αrpi ).
P
P
P
But p̂ is a unit vector, p2i = 1. As z is feasible zi = n, or pi = 0.
Hence,
−
X
log zi =
X
αrpi +
X
Ψ(−αrpi ) =
X
Ψ(−αrpi ) ≤ Ψ(−αr)
The last inequality follows from Corollary 1 above.
For Ψ(−αr) to be defined, αr < 1, we thus2 choose α =
1
r+1 ,
we have
∆Φ ≥ αr2 − Ψ(−αr) = αr2 − ((−αr) − log(1 − αr))
= αr2 + αr + log(1 − αr)
r
r
r2
+
+ log 1 −
=
1+r 1+r
r+1
r
= r + log 1 −
r+1
= r − log(1 + r) = Ψ(r) but as r =
q
n/n − 1 > 1
≥ Ψ(1) = 1 − ln 2 ≈ 0.3
Hence, potential decreases by a fixed amount after each iteration.
After k, iterations, Φ(e) − Φ(x) > kΨ(1). But as Ψ(e) = n log cT e, we
get Φ(x) < n log cT e − kΨ(1). As x is inside the standard simplex,
Φ(x)
cT x ≤ exp
n
n log cT e − kΨ(1)
< exp
n
!
−r/R
−αr
Differentiating, f (α) = αr2 − Ψ(−αr) w.r.t. α, f 0 (α) = r2 − 1−αr
+ n 1−(αr/R)
1
1
1
, equating f 0 (α) to 0, we get 1 + r = 1−αr
, or 1 − αr = 1+r
or
f 0 (α) = r2 + r + (−r) 1−αr
1
r
1
1
αr = 1 − 1+r = r+1 or α = 1+r . Thus, the maximum value is at α = 1+r .
2
10
If we stop as soon as error cT x ≤ , we get
n log cT e − kΨ(1)
exp
n
!
≤ or
n log cT e − kΨ(1)
≤ log or
n
log cT e − kΨ(1) ≤ n log or
cT e
n
log
k ≥
Ψ(1)
Thus, after at most
x for which cT x ≤
n
Ψ(1)
T
log c e iterations, algorithm finds a feasible point
Acknowledgments
Many thanks to students of CS647A (2017-2018 batch) for their valuable
feedback, questions and comments on an earlier version.
References
[1] Mokhtar S. Bazaraa, John J. Jarvis, Hanif D. Sherali, Linear Programming and Network Flows, 4th Edition, Wiley 2009
[2] Donald Goldfarb and Michael J.Todd, Linear programming, Chapter
II in Optimization Edited by G.L. Nemhauser, A.H.G. Rinooy Kan
and M.J. Todd, Handbooks in Operations Research and Management
Science Volume 1, 1989, Pages 73-170, Elsevier
[3] H.Karloff, Linear Programming, Birkhauser, 1991.
[4] Katta G. Murty, chapter 11, pp 472-473 of “Linear Complementarity, Linear and Nonlinear Programming”, web book
http://www-personal.umich.edu/~murty/books/linear_complementarity_webbook/kat11.p
[5] Narendra Karmarkar, A New Polynomial Time Algorithm for Linear
Programming, Combinatorica, Vol 4, nr. 4, pp. 373-395 (1984).
[6] C.Roos,T Terlaky and J-P Vial, Interior Point Methods for Linear Optimization, 2nd Ed, 2006, Springer.
[7] R.Saigal, Linear Programming, A Modern Integrated Analysis, Kluwer,
1995.
11
[8] Hamdy A. Taha, Operations Research: An Introduction, 7th Edition,
Pearson, 2002.
[9] Ling-Hsiao Lyu, Lecture notes on Linear Algebra Lecture 3 (Chap. 4),
Projection and Projection Matrix, Institute of Space Science, National
Central University, Chung-Li, Taiwan, R. O. C., 2012 Spring
http://www.ss.ncu.edu.tw/~lyu/lecture_files_en/lyu_LA_Notes/Lyu_LA_2012/Lyu_LA_3_
12
| 8cs.DS
|
arXiv:1710.10109v4 [math.GR] 27 Nov 2017
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR
AND AUTOMATA GROUPS
LAURENT BARTHOLDI AND IVAN MITROFANOV
Abstract. We prove that the word problem is undecidable in functionally recursive groups, and that the order problem is undecidable in automata groups,
even under the assumption that they are contracting.
1. Introduction
Let A be a finite set, and consider a group G acting faithfully and “self-similarly”
on the set A˚ of words over A. This means that every g P G acts in the form
(1.1)
pa1 . . . an qg “ a11 pa2 . . . an qg
1
for some a11 P A and some g 1 P G depending only on a1 , g; we encode them as
pg 1 , a11 q “ Φpa1 , gq for a map Φ : A ˆ G Ñ G ˆ A. If furthermore G is finitely
generated (say by a finite set S, so G is a quotient FS ։ G of the free group on S),
then its action may be described by finite data, namely a lift Φ : A ˆ S Ñ FS ˆ A of
the restriction of Φ to the generators of G. A finitely generated group given in this
manner is called functionally recursive [4, §3], or self-similar ; we call G the group
presented by Φ, and write G “ xΦy. We call Φ an (asynchronous) transducer.
Large classes of finitely generated groups can be presented as functionally recursive ones; notably, all the “iterated monodromy groups” of Nekrashevych [16], and
the automata groups mentioned in §1.1 below.
Even though the map Φ completely determines the action of G, and therefore G
itself, it is unclear how much of G is known from Φ. Our first result is as negative
as can be:
Theorem A. There is no algorithm that, given Φ : A ˆ S Ñ FS ˆ A and s P S,
determines whether s “ 1 in xΦy.
1.1. Automata groups. Assume now that G is a functionally recursive group,
and that in the action (1.1) the elements g 1 have at most the length of g, in the
generating set S. Then, up to replacing S by S Y S ´1 Y t1u, the map Φ takes the
form Φ : A ˆ S Ñ S ˆ A; we call it a finite state transducer. The group G is called
an automata group; these form a notorious class of groups, containing all finitely
generated linear groups as well as infinite torsion groups such as the “Grigorchuk
group” [7] and “Gupta-Sidki groups” [10]. The Grigorchuk group is also a group
of intermediate word-growth, and was used to settle the Milnor problem on group
growth [8].
The action of S, and of G itself, may be conveniently described by a finite labeled
graph called its Moore diagram. Consider the directed graph Γ with vertex set S
Date: October 25, 2017.
This work is supported by the “@raction” grant ANR-14-ACHN-0018-01.
1
2
LAURENT BARTHOLDI AND IVAN MITROFANOV
p1, 1q
b
p0, 0q
p1, 1q
p1, 1q
d
p0, 0q
c
a
p0, 0q
p0, 0q
e
p1, 1q
p0, 1q, p1, 0q
Figure 1. The transducer generating the Grigorchuk group. Here
A “ t0, 1u and S “ ta, b, c, d, eu.
and an edge from s to t labeled pa, bq whenever Φpa, sq “ pt, bq; then the action of
s P S on A˚ is determined as follows: given a1 . . . an P A˚ , find the unique path in
Γ starting at s and whose first label letters read a1 . . . an ; let b1 . . . bn be the second
label letters; then pa1 . . . an qs “ b1 . . . bn . See Figure 1 for the graph Γ describing
the Grigorchuk group.
Every element of G (say represented by a word w of length n in S) admits
a similar description, but now using a graph with vertex set S n . The word w
represents the identity in G if and only if at every vertex reachable from w all the
outgoing edges have labels in tpa, aq | a P Au. It follows that the word problem is
decidable in G, and even belongs to LinSpace (and therefore to ExpTime); but
that is about as much as is known. We consider the “order problem” (determine
the order of an element), which was raised at the end of last century by Sidki
and by Grigorchuk, Nekrashevych and Sushchansky [9, Problem 7.2.1(a)], to which
Gillibert announced a solution in July 2017; his proof appears in [6]:
Theorem B. There is no algorithm that, given Φ : A ˆ S Ñ S ˆ A and s P S,
determines the order of s in xΦy, namely the cardinality of xsy.
Worse than that, the action is uncomputable in the following sense: consider the
natural extension of the action of xΦy to A8 . Then we have the following variants
of Theorems A and B:
Theorem A1 . There is no algorithm that, given Φ : A ˆ S Ñ FS ˆ A and a P A
and s P S, determines whether a8 is fixed by s.
Theorem B1 . There is no algorithm that, given Φ : A ˆ S Ñ S ˆ A and a P A and
s P S, determines the cardinality of the orbit of a8 under xsy.
Finally, the results in Theorems A and B can be made uniform as follows:
Theorem A2 . There is a functionally recursive group xΦy with Φ : AˆS Ñ FS ˆA
such that ts P FS | s “ 1 in xΦyu is not recursive.
Theorem B2 . There is an automata group xΦy with Φ : A ˆ S Ñ S ˆ A, and two
states s, t P S, such that the set tn P N | stn has finite orderu is not recursive.
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS
3
1.2. Contracting groups. Assume now that G is a functionally recursive group,
and that in the action (1.1) the elements g 1 are shorter than g, in the generating set
S, in the sense that there are constants λ ă 1 and C with |g 1 | ď λ|g|`C for all g P G.
Then, up to replacing S by the set of all words of length ď C{p1 ´ λq, we also have
|g 1 | ď |g|; we have thus defined a subclass of automata groups, called contracting
automata groups (see §3.4 for a more precise definition). Their word problem is
decidable in LogSpace (and therefore in PolyTime). We will see, however, that
the order and orbit order problems remain unsolvable in that restricted class:
Theorem C (= Theorem 3.4). The transducers constructed in Theorems B and B1
may be assumed to generate contracting groups.
1.3. Sketch of proofs. We encode Minsky machines in functionally recursive
groups. Minsky machines (see [15]) are restricted Turing machines with two tapes,
which may move the tapes and sense the tape’s end but may not write on them;
equivalently, they are finite state automata equipped with two counters with values
in N that may be incremented, decremented and tested for 0.
In all cases, we encode the machine, in state s with counter values pm, nq, by
m
n
the word sx2 y 2 in a functionally recursive group containing elements x, y and
an element s for each state of the machine. The action of the group is so devised
that if the machine evolves to state ps1 , m1 , n1 q then the recursive action is given
m
n
n1
m1
by s1 x2 y 2 . The image of a prescribed ray under sx2 y 2 records the computational steps of the machine when started in ps, m, nq, and in particular whether the
machine reached a final state. We construct an auxiliary element t that only acts
m
n
m
n
on sequences containing a trace of this final state, and then psx2 y 2 qtpsx2 y 2 q´1
fixes the original ray if and only if the machine never reaches the final state. Taking
the commutator of that last element acting only in the neighbourhood of the original ray yields an expression that is trivial if and only if the machine never reaches
the final state.
Inherently, sometimes the output of the transducer is longer than the input (e.g.,
if the machine increments the first counter, the transducer must replace x by x2 ).
To obtain an automata group, we have the transducer consume a power of its input
m
n
word sx2 y 2 ; e.g., the incrementation of the counter may be performed by erasing
n
m
n
every second s and every second block of y 2 ’s. The element sx2 y 2 then may be
arranged to finite order if and only if the machine reaches the final state.
1.4. Tilings. Our results on functionally recursive groups and transducers may also
be interpreted in terms of tilings. Let C be a finite set of colours, and T Ď C N,E,S,W
be a set of Wang tiles. A valid tiling is a map t : Z2 Ñ T with tpx, yqN “ tpx, y `1qS
and tpx, yqE “ tpx ` 1, yqW for all x, y P Z2 . Berger showed in [2] that it is
undecidable to determine, given T , whether there exists a valid tiling by T . This
has been improved: for λ, µ P tN, E, S, W u, call a set of tiles λµ-deterministic if
for every c, d P C there exists at most one tile u P T with uλ “ c and uµ “ d, and
λµ-complete if there exists precisely one tile u P T with these conditions. Lukkarila
showed in [14] that the undecidability result holds even under the restriction that
T is N E, N W, SE, SW -deterministic. Clearly a SW -complete tileset tiles uniquely
the first quadrant for any choice of colours on the axes.
Our result on the order problem has the following translation into tilings. We
consider tilings of the upper half-plane tpx, yq | y ě 0u. Then the following problem
is undecidable even for SE, SW -complete tilesets: “given c P C, is there an integer
4
LAURENT BARTHOLDI AND IVAN MITROFANOV
n P N such that every tiling of the upper half-plane with c8 on the horizontal axis
is horizontally n-periodic?”.
Indeed, given Φ : A ˆ S Ñ S ˆ A, set C “ A \ S and whenever Φpa, sq “
ps1 , a1 q build a tile with N, E, S, W -labels s1 , a1 , s, a respectively; also build tiles
with N, E, S, W -labels c, d, c, d for all pc, dq P C 2 zpS ˆ Aq. Then the above tiling
problem is satisfied for c P S if and only if c has finite order in xΦy.
The word problem may also be translated to a tiling problem, but now in hyperbolic space. The tileset is now T Ď C N,E,S1 ,S2 ,W . The lattice Z2 is now Λ :“ t2y pi`
xq | x, y P Zu Ă H. A tiling is a map t : Λ Ñ T with tp2y pi`xqqE “ tp2y pi`x`1qqW
and tp2y pi ` 2xqqN “ tp2y`1 pi ` xqqS1 and tp2y pi ` 2x ` 1qqN “ tp2y`1 pi ` xqqS2
for all x, y P Z. Tiles are visualized as pentagons assembling into a tiling of the
hyperbolic plane, invariant under the transformations z ÞÑ z ` 1 and z ÞÑ 2z:
The following problem is undecidable even for N E, N W -complete tilesets: “given
c P C, does every tiling of tx ` iy P H | x P r0, 1s, y ď 1uwith c on the edge from i
to i ` 1 have identical labels on the boundary half-lines tx “ 0u and tx “ 1u?”.
Indeed by subdividing and inserting the empty state we may assume that the
map Φ describing our functionally recursive group satisfies ΦpA ˆ Sq Ď S 2 ˆ A;
then tiles are defined as above.
1.5. History. Links have been established since the beginning between undecidable problems in theoretical computer science — halting of Turing machines —
and in algebra — decidability of properties of algebraic objects. Minsky machines,
because of their simplicity, have been early recognized as useful tools in this correspondence, see e.g. Gurevich’s work [11] on identities in semigroups.
Automata semigroups are defined quite similarly to automata groups; one merely
drops the requirement that the action be by invertible maps. Decision problems
have been extensively studied within the class of automata semigroups [1, 13].
Gillibert proved in [5] that the order problem is unsolvable in that class. His
proof is based on the undecidability of Wang’s tiling problem [2], and harnesses
Kari’s solution of the nilpotency problem for cellular automata [12].
There are usually serious difficulties in converting a solution in semigroups to
one in groups. In particular, the tilings at the heart of Gillibert’s construction give
fundamentally non-invertible transformations of A˚ .
On the other hand, a direct approach to the order problem succeeded for the
restricted class of “bounded automata” groups; Bondarenko, Sidki and Zapata
prove in [3] that they have solvable order problem. Gillibert announced in [6] the
undecidability of the order problem in automata groups; his work uses a simulation
of arbitrary Turing machines by transducers via cellular automata.
2. Functionally recursive groups and Minsky machines
All our theorems are proven by embedding Minsky machine computations into
functionally recursive groups. Let us recall more precisely the definition of these
machines:
Definition 2.1. A Minsky machine is a computational device M equipped with
two integer counters m, n and a finite amount of additional memory. It has a finite
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS
5
set S of states, an initial state s˚ P S, a final state s: P S, and for each state s ‰ s:
an instruction, which can by any of the following kind:
I:
II:
III:
IV:
V:
VI:
VII:
VIII:
IX:
X:
ps, m, nq ÞÑ ps1 , m ` 1, nq;
ps, m, nq ÞÑ ps1 , m, n ` 1q;
ps, m, nq ÞÑ ps1 , m ` 1, n ` 1q;
ps, m, nq ÞÑ ps1 , m ´ 1, nq, only valid if m ą 0;
ps, m, nq ÞÑ ps1 , m, n ´ 1q, only valid if n ą 0;
ps, m, nq ÞÑ ps1 , n, mq;
ps, m, nq ÞÑ pm “ 0 ? s1 : s2 , m, nq;
ps, m, nq ÞÑ pn “ 0 ? s1 : s2 , m, nq;
ps, m, nq ÞÑ m “ 0 ? ps1 , m, nq : ps2 , m ´ 1, nq;
ps, m, nq ÞÑ n “ 0 ? ps1 , m, nq : ps2 , m, n ´ 1q.
(We use the C style “?:” operator, with ‘a ? b : c’ meaning ‘if a then b else c’.)
As M is turned on, its state and counters initialize at ps0 , m0 , n0 q “ ps˚ , 0, 0q,
and then psi`1 , mi`1 , ni`1 q is determined from psi , mi , ni q using the prescribed
rules. If at some moment si “ s: then M stops; otherwise it runs forever.
△
We recall the main result on Mealy machines, namely that they are as powerful
as Turing machines:
Proposition 2.2 ([15]). (1) There is no algorithm that, given a Minksy machine
M, determine whether it stops.
(2) There is a “universal” Minsky machine M such that
tn P N | M stops when turned on in state ps˚ , 0, nqu
is not recursive.
We also note that only one of the instructions {I,II} and III is necessary, and that
in the presence of VI only one of I,II, one of IV,V, one of VII,VIII and one of IX,X is
necessary. Minimal sets of instructions are {III,IV,V,VII,VIII} and {I,IV,VI,VII}
and {III,IX,X} and {I,VI,IX}.
2.1. Proof of Theorem A1 . Let M be a Minsky machine with stateset S0 . Without loss of generality, we assume that all instructions of M are of type I, VI, IX.
We construct a functionally recursive group xΦM y presented by ΦM : A ˆ S Ñ
FS ˆ A, for sets A, S given as follows: the generating set S consists of
‚ elements x, y, s: , t and u;
‚ for each state si P S0 of type I or IX, an element si ;
‚ for each state si P S0 of type VI, three elements si , ai , bi .
The alphabet A consists of:
‚
‚
‚
‚
four letters 0,
for each state
for each state
for each state
1, :1 and :2 ;
si P S0 of type I, a letter i1 ;
si P S0 of type IX, two letters i1 and i2 ;
si P S0 of type VI, five letters i1 , i2 , . . . , i5 .
The map ΦM : A ˆ S Ñ FS ˆ A is given below, with ǫ denoting the empty word
in FS . Whenever a value of ΦM is unspecified, we take it to mean ΦM pa, sq “ ps, aq.
6
LAURENT BARTHOLDI AND IVAN MITROFANOV
‚ for the states s: and t, u we put
ΦM p0, s: q “ pǫ, :1 q;
ΦM p:1 , xq “ pǫ, :1 q;
ΦM p:1 , yq “ pǫ, :1 q;
ΦM p:2 , xq “ pǫ, :2 q;
ΦM p:2 , yq “ pǫ, :2 q;
ΦM p:1 , tq “ pǫ, :2 q;
ΦM p0, uq “ pu, 1q;
ΦM p:2 , s: q “ pǫ, :2 q;
ΦM p:2 , tq “ pǫ, :1 q
ΦM p1, uq “ pu, 0q.
ΦM p:1 , s: q “ pǫ, 0q;
‚ for all g P Sztuu we put ΦM p1, gq “ pǫ, 1q, and for all a P Azt0, 1u we put
ΦM pa, uq “ pǫ, aq.
‚ for each instruction psi , m, nq ÞÑ psj , m ` 1, nq of type I we put
ΦM pi1 , si q “ pǫ, 0q; ΦM pi1 , xq “ px2 , i1 q;
ΦM p0, si q “ psj , i1 q;
ΦM pi1 , yq “ py, i1 q.
‚ for each instruction psi , m, nq ÞÑ psj , n, mq of type VI, ΦM pa, sq is written
at position pa, sq of the following table:
element of S
0
x
y
si pai bi x, i1 q
ai
pai , i2 q
bi
input letter
i2
i3
pǫ, i3 q
px, i2 q
py, i2 q
i1
pxbi x , i1 q
py x , i1 q
pǫ, 0q
i4
i5
px2 , i4 q py, i5 q
py, i4 q px, i5 q
pǫ, 0q
pbi , i4 q pa´1
i sj , i 5 q
pǫ, i2 q
pǫ, i3 q
‚ for each instruction psi , m, nq ÞÑ pm “ 0 ? sj : sk , maxp0, m ´ 1q, nq of type
IX we put
ΦM pi1 , xq “ ps´1
k sj x, i2 q;
ΦM p0, si q “ psk , i1 q;
ΦM pi1 , si q “ pǫ, 0q;
ΦM pi2 , xq “
ΦM pi1 , yq “ py, i1 q;
px´1 s´1
j sk x, i1 q;
ΦM pi2 , yq “ py, i2 q.
Theorem A1 follows from the undecidability of the halting problem for the Minsky
machines (Proposition 2.2) and the following
Proposition 2.3. Consider the infinite sequence W “ 08 . Then the Minsky machine M does not halt if and only if the action of xΦM y satisfies
W ps˚ xyqtps˚ xyq
´1
“ W.
m
n
m
n
Proof. We encode the states of M by elements of FS . The word psi x2 y 2 qαpsi x2 y 2 q´1
corresponds to the state psi , m, nq.
It is convenient to write ΦM pa, gq “ pg 1 , a1 q in the form a ¨ g “ g 1 ¨ a1 . In this
manner, the computation of the functionally recursive action is given by a sequence
of exchanges of letters with words in FS . We check the following equalities:
If psi , m, nq Ñ psj , m ` 1, nq is an instruction of type I, then
(2.1)
m
n
m
n
m`1
0 ¨ psi x2 y 2 qtpsi x2 y 2 q´1 “ psj x2
m
n
m
n
n
m`1
y 2 qtpsj x2
m`1
n
y 2 q´1 ¨ 0.
n
Indeed 0 ¨ si x2 y 2 “ sj ¨ i1 ¨ x2 y 2 “ sj x2
y 2 ¨ i1 ; the claim follows from
m
n
m`1
n
i1 ¨ t “ t ¨ i1 and the reverse i1 ¨ psi x2 y 2 q´1 “ psj x2
y 2 q´1 ¨ 0.
If psi , m, nq Ñ psj , n, mq is an instruction of type VI, then
(2.2)
m
n
m
n
n
m
n
m
0m`2 ¨ psi x2 y 2 qtpsi x2 y 2 q´1 “ psj x2 y 2 qtpsj x2 y 2 q´1 ¨ 0m`2 .
m
n
m
n
m
n
Indeed we first check 0 ¨ si x2 y 2 “ ai bi xpxbi x q2 py x q2 ¨ i1 “ ai x2 bi y 2 x ¨ i1 .
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS
m
7
0
We obtained a word with two “blocks” of x: the blocks x2 and x2 . Each time
a ‘0’ letter is multiplied on the left of that word, the size of the first block will halve
and the size of the second one will double: for m, n, p P N, we have
0 ¨ ai x2m bi y n xp “ ai xm bi y n x2p ¨ i4
m
n
n
m
so 0m`1 ¨ si x2 y 2 “ ai xbi y 2 x2 ¨ pi4 qm i1 . Then
n
m
n
m
2
2
0 ¨ ai xbi y 2 x2 “ ai pa´1
¨ i5 ,
i sj qx y
m
n
n
m
so 0m`2 ¨ si x2 y 2 “ sj x2 y 2 ¨ i5 pi4 qm i1 . Recalling that we have a ¨ t “ t ¨ a for all
a “ i1 , . . . , i5 , the claim is proven.
If ps, m, nq Ñ pm “ 0 ? sj : sk , maxpm ´ 1, 0q, nq is an instruction of type IX,
then if m “ 0 we have
(2.3)
m
n
m
n
m
n
m
n
n
m
n
m´1
y 2 qtpsk x2
0 ¨ psi x2 y 2 qtpsi x2 y 2 q´1 “ psj x2 y 2 qtpsj x2 y 2 q´1 ¨ 0
while if m ą 0 we have
(2.4)
m
0 ¨ psi x2 y 2 qtpsi x2 y 2 q´1 “ psk x2
n
m´1
n
y 2 q´1 ¨ 0.
Indeed in the first case we have
n
n
2
0 ¨ si xy 2 “ sk ps´1
¨ i2 ,
k sj xqy
while in the second case we have
m
n
m´1
´1 ´1
0 ¨ si x2 y 2 “ sk ps´1
sk sj xq2
j sk x ¨ x
n
m´1
y 2 ¨ i1 “ sk x2
n
y 2 ¨ i1 .
Recalling that we have a ¨ t “ t ¨ a for all a “ i1 , i2 , the claim is proven.
From (2.1)–(2.4) it follows that if M does not halt then W ps˚ xyqtps˚ xyq
Conversely, if M halts then there exist k, m, n P N such that
m
n
m
´1
“ W.
n
0k ¨ ps˚ xyqtps˚ xyq´1 “ ps: x2 y 2 qtps: x2 y 2 q´1 ¨ 0k .
Then
m
n
m
n
m
n
m
n
0 ¨ s: x2 y 2 tps: x2 y 2 q´1 “ :1 ¨ tps: x2 y 2 q´1 “ :2 ¨ ps: x2 y 2 q´1 “ :2 .
In that case, we have W ps˚ xyqtps˚ xyq
´1
“ 0 k :2 0 8 ‰ W .
The computations are best carried on ΦM ’s dual Moore diagram ∆, see Figure 2
: this is the directed labeled graph with vertex set A and with for all a P A, s P S an
edge from a to b labeled ps, tq whenever ΦM pa, sq “ pt, bq. One checks an equality
‘ΦM pa, sq “ pt, bq’ by finding in ∆ a path starting at a with input label s; the
endpoint of the path is b, and the output label is t.
2.2. Proof of Theorem A. We have not yet used the letter 1 and the state u of
ΦM . Theorem A follows now from the following
Proposition 2.4. The Minsky machine M halts if and only if
rps˚ xyqtps˚ xyq´1 , us ‰ 1 in xΦM y.
Proof. The element u acts on Aω as follows: it scans X P Aω for its longest prefix
in t0, 1u˚ , and exchanges all 0 and 1 in that prefix. Write g “ ps˚ xyqtps˚ xyq´1 ;
from Proposition 2.3 we know that g fixes 08 if and only if M does not halt.
Assume first that M does not halt; then g in fact also fixes t0, 1u8 , so the
supports of g and u are disjoint and rg, us “ 1 in xΦM y.
8
LAURENT BARTHOLDI AND IVAN MITROFANOV
pg, ǫq@g
py, yq
px, x2 q
px, yq
py, ǫqpx, ǫq
:1
1
i4
pt, tq
px, ǫq
pu, uq
pbi , bi q
pt, tq
i2
px, xq
0
pai , ai q
ps: , ǫq
ps: , ǫq
px, x2 q
i3
psi , sj q
i1
pt, tq
py, y x q
pt, tq Type I
py, yq
Type VI
px, ǫqpsi , ai bi xq
pbi , a´1
i sj q
:2
pt, ǫq
py, xq
i5
py, ǫq
pt, tq
i1
py, yq
px, xbi x q
psi , sk q Type IX
px, s´1
k sj xq
i1
i2
´1 ´1
px, x sj sk xq
pt, tq
py, yq
Figure 2. The dual Moore diagram of ΦM , used in the proof of Theorem A
Assume next that M does halt; without loss of generality, we may assume M
does not stop immediately, so there is k ě 1 such that p0k`1 qg “ 0k :2 . Since
p0k`1 qu “ 1k`1 and p0k :2 qu “ 1k :2 and p1k`1 qg “ 1k`1 and p1k :2 qg “ 1k :2 ,
the commutator rg, us acts as a 2-2-cycle p0k`1 , 0k :2 qp1k`1 , 1k :2 q and in particular
rg, us ‰ 1 in xΦM y:
g
1k`1
u
0k`1
g
0 k :2
u
1 k :2
g
2.3. Proof of Theorem A2 . Consider a universal Minsky machine Mu , namely
one that emulates arbitrary Turing machines encoded in an integer n, when started
in state ps˚ , 0, nq. The set of 2n such that Mu halts when started in state ps˚ , 0, nq
is not recursive, see Proposition 2.2. Therefore, Theorem A2 follows by considering
n
n
in the group xΦMu y the elements rps˚ xy 2 qtps˚ xy 2 q´1 , us; this set of words is
recursive, but the subset of those that equal 1 in xΦMu y is not recursive.
3. Automata groups and Minsky machines
3.1. Proof of Theorem B. Let M by a Minsky machine with stateset S0 . Without loss of generality, we assume that all instructions of M are of type III,IV,V,VII,VIII,
as defined in the beginning of Section 2,
S0 “ SIII \ SIV \ SV \ SVII \ SVIII \ ts: u
We consider the transducer with stateset S :“ S0˘1 \ tǫ, x, x´1 , y, y ´1 u and alphabet
A “ tIIIi , IVi , Vi , VIIj , VIIIj | i “ 1, 2, 1, 2; j “ 1, . . . , 4, 1, . . . , 4u.
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS
9
in state
The structure of the transducer is given by its map ΦM : A ˆ S Ñ S ˆ A, first
described as a table, with ΦM pa, sq at position pa, sq. The state ǫ is the identity,
and ΦM pa, ǫq “ pǫ, aq for all a P A.
For all instructions ps, m, nq ÞÑ ps1 , m ` 1, n ` 1q of type III and for all t P S0 zSIII
we have
input letter
III2
III1
III2
III1
´1
´1
x px, III1 q px, III2 q px , III1 q px , III2 q
py ´1 , III1 q
y py, III1 q py, III2 q py ´1 , III1 q
1
s ps , III2 q pǫ, III1 q
pǫ, III2 q
pps1 q´1 , III1 q
pǫ, III1 q
pǫ, III2 q
t pǫ, III1 q pǫ, III2 q
in state
in state
For all instructions ps, m, nq ÞÑ ps1 , m ´ 1, nq of type IV and for all t P S0 zSIV , we
have
input letter
IV2
IV1
IV2
IV1
x px, IV2 q pǫ, IV1 q
pǫ, IV2 q
px´1 , IV1 q
py ´1 , IV2 q
y py, IV1 q py, IV2 q
py ´1 , IV1 q
1
1
1 ´1
s ps , IV1 q ps , IV2 q pps q , IV1 q pps1 q´1 , IV2 q
pǫ, IV1 q
pǫ, IV2 q
t pǫ, IV1 q pǫ, IV2 q
The same applies for an instruction of type V, with the roles of x, y switched.
For an instruction ps, m, nq ÞÑ pm “ 0 ? s1 : s2 , m, nq of type VII and for all
t P S0 zSVII , we have
x
y
s
t
VII1
VII2
px, VII4 q pǫ, VII3 q
py, VII1 q pǫ, VII2 q
pǫ, VII2 q ps2 , VII1 q
pǫ, VII1 q pǫ, VII2 q
input letter
VII3
VII4
VII1
pǫ, VII2 q px, VII1 q
px´1 , VII4 q
pǫ, VII3 q py, VII4 q
py ´1 , VII1 q
1
ps , VII4 q pǫ, VII4 q pps2 q´1 , VII2 q
pǫ, VII1 q
pǫ, VII3 q pǫ, VII4 q
VII2
pǫ, VII3 q
pǫ, VII2 q
pǫ, VII1 q
pǫ, VII2 q
VII3
VII4
pǫ, VII2 q px´1 , VII1 q
pǫ, VII3 q py ´1 , VII4 q
pǫ, VII3 q pps1 q´1 , VII3 q
pǫ, VII3 q
pǫ, VII4 q
The same applies for an instruction of type VIII, with the roles of x, y switched.
Note that s: is treated as a state t in all tables above.
Theorem B follows from the undecidability of the halting problem for Minsky
machines, and the following
Proposition 3.1. The Minsky machine M constructed above halts if and only if
the element s˚ xy has finite order in xΦM y.
Proof. Set G “ xΦM y. For g P G, denote by Cpgq its symmetric conjugacy class:
Cpgq :“ tg ˘x | x P Gu.
Given a symmetric conjugacy class C, choose a representative g in it, let A “
A1 \ ¨ ¨ ¨ \ Aℓ be the decomposition of A into cycles for the action of g, and choose
representatives ai P Ai . We have ΦM pai , g #Ai q “ phi , ai q for some hi P G, and it is
easy to see that the collection of symmetric conjugacy class tCphi q | i “ 1, . . . , ℓu
is independent of the choice of g and the ai .
We construct an integer-labeled, directed graph1 whose vertices are symmetric
conjugacy classes in G; for a conjugacy class C as above, there are ℓ edges starting
at C, ending respectively at Cph1 q, . . . , Cphℓ q with labels #A1 , . . . , #Aℓ .
1This graph essentially appears in the solution of [3] to the order problem in bounded automata.
10
LAURENT BARTHOLDI AND IVAN MITROFANOV
Lemma 3.2. For g P G, its order (in N Y t8u) is the least common multiple, along
all paths starting at Cpgq, of the product of the labels along the path.
Proof. Consider a path starting at Cpgq, with labels n1 , . . . , ns , and going through
vertices Cpg1 q, . . . , Cpgs q. Then g has an orbit of length n1 on A, so the order of g
is a multiple of n1 . Furthermore, g n1 fixes that orbit, and acts as g1 on sequences
that start by that orbit. Recursively, the order of g1 is a multiple of n2 ¨ ¨ ¨ ns , so the
order of g is a multiple of n1 ¨ ¨ ¨ ns . In particular, if there are paths with arbitrarily
large product of labels then g has infinite order.
Conversely, if g has infinite order then there are arbitrarily long orbits of g on
A˚ , so there are paths with arbitrarily large product of labels; and if m be the least
common multiple of all path labels then all edges on paths starting at Cpg m q are
labeled 1 so g m fixes every sequence and therefore g m “ 1.
Let us compute the subgraph spanned by Cps˚ xyq. For the computations, it is
helpful to picture the operation of the transducer ΦM by means of its dual Moore
diagram ∆, see Figure 3. Given g P G, we compute all primitive cycles in ∆ whose
input label is a power of g, and read the corresponding output label; these are the
hi in the map on symmetric conjugacy classes Cpgq
tCphi qu.
We first note, by direct inspection, that x and y commute in G. This follows by tracing the path x´1 y ´1 xy in the graphs above, and noting that they
always induce the trivial permutation of A with output either trivial or conjugate
to px´1 y ´1 xyq˘1 .
We now claim that, if ps, m, nq Ñ ps1 , m1 , n1 q is a transition of the machine M,
n1
m
n
m1
then the conjugacy class Cpsx2 y 2 q has at least one arrow to Cps1 x2 y 2 q, and
possibly other arrows, all of them to Cp1q. We also claim that if s is not of type
n1
m1
IV or V, then arrows to Cps1 x2 y 2 q are with labels ą 1; and all arrows from
m
n
Cps: x2 y 2 q are arrows to Cp1q. We see that if the machine halts then every path
starting at Cps˚ xyq has only a finite number of labels ą 1, and this shows that the
order of s˚ xy is finite.
On the other hand, if the machine does not halt then there is an path with
infinitely many labels ą 1 (because no Minsky machine can decrease its counters
infinitely many times in a row) so s˚ xy has infinite order.
Note that our transducer has the property ΦM pLi , gq “ ΦM pLi , g ´1 q, for all g P S
and all L P tIII, IV, V, VII, VIIIu. Also note that ΦM pLi , tq “ ΦM pLi , t´1 q “ pǫ, Li q
whenever t is any instruction not of type L.
Using this, we can prove that if t is not of type L, then tgn gn´1 ¨ ¨ ¨ g1 tg1 g2 ¨ ¨ ¨ gn
fixes the orbit tLi u with output ǫ.
Indeed, tgn gn´1 ¨ ¨ ¨ g1 tg1 g2 ¨ ¨ ¨ gn “ pgn tgn t´1 ¨ tgn´1 gn´2 ¨ ¨ ¨ g1 tg1 g2 ¨ ¨ ¨ gn´1 qgn ,
and we use induction on n. It follows that ptxm y n q2 fixes Li with outputs ǫ, i.e.,
there is an arrow from Cptxm y n q to Cp1q with label 2.
Let us first restrict to the orbit tIIIi u of G on A. We consider g :“ sxm y n with s
an instruction of type III. It acts as a product of two cycles pIII1 , III2 qpIII1 , III2 q; the
output label of g 2 on the first cycle, starting at III1 , is s1 xm y n ǫxm y n “ s1 x2m y 2n ,
and the output of g 2 starting on the second cycle at III1 is ǫx´m y ´n ps1 q´1 x´m y ´n P
m
n
m`1
n`1
Cps1 x2m y 2n q. There are therefore two arrows from Cpsx2 y 2 q to Cps1 x2
y 2 q,
as required. We do not consider g :“ txm y n with t an instruction of different type
or s: , because it was considered above (there are some arrows to Cp1q with labels
2).
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS 11
ps, s1 q
III1
ps, ǫq
III2
px, xq
pt, ǫq
py, yq
´1
III1
px, x
q
ps, ǫq
py, yq
pt, ǫq
IV1
px, ǫq
px, x´1 q
III2
py, yq
px, xq
IV2
IV1
px, x´1 q
pt, ǫq
ps, ps1 q´1 q
px, xq
ps, ps1 q´1 q
ps, s1 q
ps, s1 q
py, y ´1 q
IV2
pt, ǫq
py, y ´1 q
px, ǫq
py, y ´1 q
py, yq
py, y ´1 q
ps, ps1 q´1 q
pt, ǫq
py, yq
py, yq
VII1
ps, s2 q
py, ǫq
px, xq
ps, ǫq
VII2
px, ǫq
VII4
ps, s1 q
pt, ǫq
ps, ǫq
py, y ´1 q
VII4
ps, ps1 q´1 q
VII3
py, ǫq
ps, ǫq
pt, ǫq
VII3
px, x´1 q
ps, ps2 q´1 q
ps, ǫq
px, ǫq
py, y ´1 q
VII1
VII2
py, ǫq
py, ǫq
pt, ǫq
Figure 3. The dual Moore diagram of ΦM , used in the proof of Theorem B
We restrict next to the orbit tIVi u of G and consider the case g “ sx2m y n . (We
do not need to consider cases g “ sx2m`1 y n or g “ txm y n , for the first because we
suppose that if m “ 0 then M does not perform an instruction of type IV, and for
the second because it was already considered above.)
An element g “ sx2m y n fixes IV1 , IV2 , IV1 and IV2 , its outputs are respectively
1
s pxǫqm y n “ s1 xm y n , s1 xm y n , ps1 q´1 x´m y ´n and ps1 q´1 x´m y ´n . Hence there are
m
n
m´1
n
4 arrows from Cpsx2 y 2 q to Cps1 x2
y 2 q, all with labels 1.
We restrict next to the orbit tVIIi u of G and perform the same computations,
the result is in the following table:
gPG
sx2m y n
sx2m`1 y n
cycles of g
pVII1 , VII2 q
pVII3 , VII4 , VII4 , VII3 q
pVII1 , VII2 q
pVII1 , VII3 q
pVII2 , VII4 , VII1 , VII3 q
pVII2 , VII4 q
output, starting at first element of the cycle
ǫs2 x2m y n
s1 x2m y n ǫx´2m y ´n ps1 q´1 ǫ “ 1
ps2 q´1 ǫx´2m y ´n
ǫs1 x2m`1 y n
s2 x2m`1 y n ǫx´2m´1 y ´n ps2 q´1 ǫ “ 1
ǫx´2m´1 y ´n ps1 q´1 ǫ
12
LAURENT BARTHOLDI AND IVAN MITROFANOV
m
n
m
n
If m ą 0 then there are therefore two arrows from Cpsx2 y 2 q to Cps2 x2 y 2 q
with labels 2 and an arrow to Cp1q with label 4; if m “ 0 then there are two arrows
m
n
m
n
from Cpsx2 y 2 q to Cps2 x2 y 2 q with label 4 and an arrow to Cp1q with label 4.
The orbits tVi u and tVIIIi u are investigated in the same way as tIVi u and tVIIi u
respectively.
3.2. Proof of Theorem B2 . Consider a universal Minsky machine Mu , namely
one that emulates arbitrary Turing machines encoded in an integer n, when started
in state ps˚ , 0, nq. The set of 2n such that Mu halts when started in state ps˚ , 0, nq
is not recursive, see Proposition 2.2. Therefore, Theorem B2 follows by considering
in the group xΦMu y the elements s “ s˚ x and t “ y.
3.3. Proof of Theorem B1 . Let M be a Minsky machine with stateset S0 . Without loss of generality, we assume that all instructions are of type III,IX,X.
We associate to it the transducer with stateset Q :“ S0 \ tǫ, x, yu and alphabet
A “ t0, IIIi , IXj , Xj | i “ 1, 2; j “ 1, . . . , 4u.
in state
The structure of the transducer is given by its map ΦM : A ˆ Q Ñ Q ˆ A.
The state ǫ is the identity, and ΦM pa, ǫq “ pǫ, aq for all a P A.
‚ for all instructions psi , m, nq ÞÑ ps1i , m ` 1, n ` 1q of type III we have
input letter
0
III1
III2
x
pǫ, 0q
px, III1 q px, III2 q
y
pǫ, 0q
py, III1 q py, III2 q
si ps1i , III1 q pǫ, III2 q
pǫ, 0q
in state
and every instruction t of another type acts as ΦM pt, IIIℓ q “ pIIIℓ , tq.
‚ for all instructions psj , m, nq ÞÑ m “ 0 ? ps1j , m, nq : ps2j , m ´ 1, nq of type
IX we have
input letter
0
IX1
IX2
IX3
IX4
x
pǫ, 0q
pǫ, IX2 q
pǫ, IX1 q px, IX4 q pǫ, IX3 q
y
pǫ, 0q
pǫ, IX1 q
pǫ, IX2 q py, IX3 q py, IX4 q
sj pǫ, IX1 q ps2j , IX4 q ps1j , IX3 q ps1j , IX2 q
pǫ, 0q
and every instruction t of another type acts as ΦM pIXℓ , tq “ pt, IXℓ q.
‚ The same applies for every instruction psk , m, nq ÞÑ n “ 0 ? ps1k , m, nq :
ps2k , m, n ´ 1q of type X, with the roles of x and y switched.
‚ for all a P A we have ΦM pa, s: q “ pǫ, aq.
We claim that the orbit of 08 under s˚ xy is finite if and only if the machine M
stops.
Set G “ xΦM y. We construct an integer-labeled, directed graph whose vertices
are elements of G. For g P G consider its action on A and the minimal pg such that
g pg fixes 0, i.e., 0 ¨ g pg “ g 1 ¨ 0. In our graph we put an edge g Ñ g 1 with label pg
on it.
The size of the the orbit of 08 under s˚ xy is a finite number or 8 and it is equal
to product of the labels along the path starting at s˚ xy.
We claim that for any instruction ps, m, nq Ñ ps1 , m1 , n1 q there is an edge from
m
n
m1
n1
m
n
sx2 y 2 to s1 x2 y 2 with label 3, and an edge from s: x2 y 2 to 1. This is checked
on the dual Moore diagram of ΦM , see Figure 4:
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS 13
py, ǫq
IX2
px, ǫq
py, ǫq
px, ǫq
IX1
X1
px, ǫq
py, ǫq
X2
psk , ǫq
ps: , ǫq
IX4
si1q
III1
px, xq
X3
px, xq
py, yq
px, xq
ǫq
psi
,
py, yq
py, yq
py, ǫq
X4
,
ps i
px, ǫq
psk , s1k q
0
psj , ǫq
psk , s2k q
px, xq
IX3
py, yq
psj , s2j q
psj , s1j q
psj , ǫq px, ǫqpy, ǫq psk , ǫq
psi , ǫq
III2
py, yq
px, xq
Figure 4. The dual Moore diagram of ΦM , used in the proof of Theorem B1
We first note that x and y commute in G. If g “ si xm y n and si is an instruction
of type III, then the orbit of 0 under the action of g is p0, III1 , III2 q. There is an
edge labeled 3 from g to s1i xm y n ǫxm y n “ s1i x2m y 2n .
Consider next sj an instruction of type IX. There are two cases. if g “ sj xy n
then the orbit of 0 is p0, IX2 , IX4 q and the output is ǫs1j xy n ǫ; if g “ sj x2m y n then
the orbit of 0 is p0, IX1 , IX4 q and the output is ǫs2j xm y n ǫ.
m
n
This means that if m “ 0 then there is an edge labeled 3 from sj x2 y 2 to
m´1
n
m
n
m
n
y2 .
s1j x2 y 2 , and if m ą 0 then there is an edge labeled 3 from sj x2 y 2 to s2j x2
The same naturally applies to instructions of type IX. Finally, for all m, n P N
the element s: xm y n fixes 0, and there is an edge labeled 1 from s: xm y n to 1.
3.4. Contracting automata: proof of Theorem C. We finally explain how to
make the transducers ΦM of the previous subsections contracting. We expand the
definition from the introduction:
Definition 3.3 ([16, Definition 2.11.1]). Let G “ xΦy be a self-automata group
with Φ : A ˆ S Ñ S ˆ A and Φ : A ˆ G Ñ G ˆ A. For g P G and u P A˚ , the state
g@u is the unique element of G such that puvqg “ ug v g@u ; namely, the action of g
on tails of sequences starting with u.
The group G is contracting if there exists a finite subset N Ď G such that, for
all g P G, there exists npgq P N such that g@u P N whenever |u| ě npgq.
△
The minimal subset N satisfying the definition is called the nucleus. In particular, one has n@a P N for all pa, nq P A ˆ N , so Φ induces an automaton still written
r :“ An for n
Φ : A ˆ N Ñ N ˆ A. Up to replacing S by Sr :“ S Y N and A by A
larger than maxgPSr2 npgq, thus making the transducer process n letters at a time,
14
LAURENT BARTHOLDI AND IVAN MITROFANOV
one may also assume
r ˆ Sr2 q Ď Sr ˆ A.
r
ΦpA
A transducer Φ with this extra property is called nuclear.
Note that is is probably undecidable whether a self-similar group xΦy is contracting; but it is easy to decide whether a transducer Φ : A ˆ S Ñ S ˆ A is nuclear: by
minimizing the composite transducer A ˆ S 3 Ñ S ˆ A ˆ S 2 Ñ S 2 ˆ A ˆ S Ñ S 3 ˆ A,
find the set R of all words s1 s2 s3 P S 3 that equal 1 in G. Then Φ is nuclear if and
only if for all a P A, s1 , s2 P S there exists s3 P S such that if Φpa, s1 q “ ps11 , bq and
Φpb, s2 q “ ps12 , cq then s11 s12 s´1
3 P R. The more precise form of Theorem C is:
Theorem 3.4. There is no algorithm that, given a nuclear transducer Φ : A ˆ S Ñ
S ˆ A and a P A and s P S, determines the cardinality of the orbit of a8 under xsy.
There is no algorithm that, given a nuclear transducer Φ : A ˆ S Ñ S ˆ A and
s P S, determines the order of s in xΦy.
Note that the group is not changed by these operations of replacing S by N and
A by An . If Φ is nuclear, then xΦy is contracting in the sense of the introduction,
since |g 1 | ď p|g|`1q{2 in the word metric defined by N . Conversely, if |g 1 | ď λ|g|`C
then one may take N “ tg P G | C{p1 ´ λq ě |g|u to see that G is contracting in
the sense of Definition 3.3.
Lemma 3.5. Let Φ : A ˆ S Ñ S ˆ A be a transducer. If there is a constant N P N
such that every reduced path of length ě N in the dual Moore diagram of Φ contains
an ǫ letter along its output, then xΦy is contracting.
Proof. Consider g P xΦy, and represent it by a word w P S ˚ of length ℓ “ |g|.
Factor w “ w1 . . . wt with |wi | “ N for all i “ 1, . . . , t ´ 1 and |wt | ă N .
Then every g 1 as in (1.1) is computed by following, in the dual Moore diagram,
the path starting at a1 with label w on its input. The output label along that path
is g 1 , and by hypothesis each time a segment wi is read, for i ă t, an ǫ letter is
produced for g 1 ; so |g 1 | ď ℓ ´ t ` 1. Now t “ rℓ{N s, so
|g 1 | ď ℓ ´ rℓ{N s ` 1 ď p1 ´ 1{N q|g| ` 1.
We shall modify the transducers ΦM by composing them with appropriate machines. We recall the general definition: let Φ : AˆS Ñ SˆA and Ψ : BˆS Ñ SˆB
be transducers with same stateset S. Their composition is the transducer Φ ˝ Ψ
with alphabet A ˆ B, given by
AˆΨ
ΦˆB
Φ˝Ψ : pAˆBqˆS “ AˆpBˆSq ÝÑ AˆpSˆBq “ pAˆSqˆB ÝÑ pSˆAqˆB “ SˆpAˆBq.
We are given a transducer Φ with stateset S “ ts1 , . . . , sℓ , x, yu and alphabet
A. We write G “ xΦy, and freely identify words in S ˚ with their value in G. We
require that x, y commute.
For every i P t1, . . . , ℓu, consider the transducer Φi with alphabet Ai “ t0, 1u
and transitions Φi pa, qq “ pa “ 0 ? q : ǫ, q “ si ? 1 ´ a : aq.
Note (by drawing the dual Moore diagram and deleting the transitions with ǫ
output) that the only paths with input and output of same length are of the form
b
s´a
i wsi for some a, b P t0, 1u and w a word not involving si .
Note also that for a word w of form sj xm y n
(1) if i “ j then Φi p0, w2 q “ pw, 0q and Φi p1, w2 q “ pw1 , 1q with w1 conjugate
to w;
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS 15
(2) if i ‰ j then Φi p0, wq “ pw, 0q and Φi p1, wq “ pǫ, 1q.
Consider also a transducer Φ0 with alphabet A0 “ t0, 1u3 and transitions
Φ0 ppa, b, cq, si q “ pc “ 0 ? si : ǫ, pa, b, 1 ´ cqq for all i;
Φ0 ppa, b, cq, xq “ pa “ 0 ? x : ǫ, p1 ´ a, b, cqq;
Φ0 ppa, b, cq, yq “ pb “ 0 ? y : ǫ, pa, 1 ´ b, cqq.
Note that, in the dual Moore diagram of Φ0 , all paths with input label of the form
m n b
s´a
i x y sj have shorter output label as soon as |m| ` |n| ě 3. Note also that if w
is a word of the form si xm y n then for all pa, b, cq P A0 we have Φ0 ppa, b, cq, w2 q “
pw1 , pa, b, cqq for some permutation w1 of w; so in particular w1 is conjugate to w.
Furthermore, for pa, b, cq “ p0, 0, 0q we get w1 “ w in G.
Proposition 3.6. Under the hypotheses above, the transducer Φ1 :“ Φ ˝ Φ0 ˝ Φ1 ˝
¨ ¨ ¨ ˝ Φℓ generates a contracting group, and whenever we have Φpa, psi xm y n qt q “
1
1
ps1i xm y n , aq in the original transducer we have for all j P t0, 1uℓ`3 the relation
1
1
Φ1 ppj, aq, psi xm y n q4t q “ pw, pj, aqq, with w either equal to 1 or conjugate to s1i xm y n .
1
1
Furthermore, if j “ 0ℓ`3 then w “ s1i xm y n .
Proof. After applying the transducers Φ1 , . . . , Φℓ , the only words that don’t get
b
shortened are of the form s´a
i wpx, yqsj for some i, j P t1, . . . , ℓu and some a, b P
t0, 1u. These get shortened by Φ0 as soon as |w| ě 3, using the fact that x and y
commute. It follows that xΦ1 y is contracting.
Consider the transitions of psi xm y n q4 in transducer Φ1 ˝ ¨ ¨ ¨ ˝ Φℓ . On input letter
ℓ
0 it produces psi xm y n q2 , on input letter 0 ¨ ¨ ¨ 1 ¨ ¨ ¨ 0 with the ‘1’ in position i it
produces a conjugate of psi xm y n q2 and on all other input letters it produces ǫ.
Feed then psi xm y n q2 to transducer Φ0 ; on input letter 000 it produces si xm y n and
on all other input letters it produces a conjugate of si xm y n . Feed finally si xm y n
to Φ to conclude the proof.
We are ready to finish the proof of Theorem 3.4. We constructed an integerlabeled graph for a transducer Φ, whose vertices are elements of G for Theorem B1
or symmetric conjugacy classes for Theorem B.
By Proposition 3.6, the transducer Φ1 is contracting. Let us check that the order
problems for xΦy and for xΦ1 y are equivalent.
A graph for Φ1 will have the same set of vertices as the graph for Φ, and Proposition 3.6 shows that this new graph has the same set of outgoing edges for each
element of form si xm y n , with labels multiplied by 4 and, possibly, some new edges
to 1 (or to Cp1q). Since in the old graph there were no loops at non-identity elements, s˚ xy has infinite order in xΦ1 y if and only if it has infinite order in xΦy, and
the orbit of p0, 0ℓ`3 q8 is infinite under the action of s˚ xy P xΦ1 y if and only if the
orbit of 08 is infinite under the action of s˚ xy P xΦy.
r “ An , we may
Finally, by replacing the stateset S by Sr “ S Y N and A by A
1
assume that Φ is nuclear.
4. Outlook
We proved in this article the undecidability of the order problem for automata
groups, namely groups of transformations generated by a transducer.
If the transducer belongs to a restricted class, it may well be that the order
problem becomes decidable. Of particular importance are:
16
LAURENT BARTHOLDI AND IVAN MITROFANOV
Transducers of polynomial growth. In a transducer Φ (represented by a graph
as in Figure 1), let αpnq denote the number of paths of length n that end in a
non-identity state. If αpnq is a bounded function (as is the case for the Grigorchuk
group), then the order problem is solvable in xΦy, see [3]. What happens if αpnq
is bounded by a linear function? or by a polynomial of degree d? The groups
generated by such transducers have been considered by Sidki [17].
Reset transducers. These are transducers Φ with Φpa, sq “ pφpaq, ψpa, sqq for
some functions φ, ψ; namely, the state reached by the transducer is independent of
the original state. These transducers are intimately connected to tilings, by Kari’s
construction [12]. Gillibert proved in [5] that the order problem is unsolvable for
semigroups of reset automata. Is it solvable in groups of reset automata?
Reversible transducers. These are transducers whose dual is invertible; they
should be related to reversible Turing or Minsky machines. Is the order problem
solvable for groups generated by reversible automata?
Bireversible transducers. These are transducers Φ such that all 8 transducers
obtained from Φ by inverting or permuting the stateset and alphabet remain transducers; they give another point of view on square complexes (by tiling the plane
with squares whose labels are pa, s, a1 , s1 q when Φpa, sq “ ps1 , a1 q).
We expect it to be undecidable whether a functionally recursive group is actually
an automata group (for a larger generating set), whether an automata group is
contracting, and even whether a contracting group is finite. Again, the related
questions for semigroups are known to be undecidable by constructions in or similar
to [5].
References
[1] Ali Akhavi, Ines Klimann, Sylvain Lombardy, Jean Mairesse, and Matthieu Picantin, On
the finiteness problem for automaton (semi)groups, Internat. J. Algebra Comput. 22 (2012),
no. 6, 1250052, 26, DOI 10.1142/S021819671250052X. MR2974106
[2] Robert Berger, The undecidability of the domino problem, Mem. Amer. Math. Soc. No. 66
(1966), 72. MR0216954 (36 #49)
[3] Ievgen V. Bondarenko, Natalia V. Bondarenko, Said N. Sidki, and Flavia R. Zapata, On the
conjugacy problem for finite-state automorphisms of regular rooted trees, Groups Geom. Dyn.
7 (2013), no. 2, 323–355, DOI 10.4171/GGD/184. With an appendix by Raphaël M. Jungers.
MR3054572
[4] Andrew M. Brunner and Said N. Sidki, On the automorphism group of the one-rooted binary
tree, J. Algebra 195 (1997), no. 2, 465–486. MR98i:20035
[5] Pierre Gillibert, The finiteness problem for automaton semigroups is undecidable, Internat.
J. Algebra Comput. 24 (2014), no. 1, 1–9, DOI 10.1142/S0218196714500015. MR3189662
, An automaton group with undecidable order and Engel problems (2017), available
[6]
at arXiv:1710.09733.
[7] Rostislav I. Grigorchuk, On Burnside’s problem on periodic groups, Funkcional. Anal. i
Priloжen. 14 (1980), no. 1, 53–54. English translation: Functional Anal. Appl. 14 (1980),
41–43. MR81m:20045
, On the Milnor problem of group growth, Dokl. Akad. Nauk SSSR 271 (1983), no. 1,
[8]
30–33. MR85g:20042
[9] Rostislav I. Grigorchuk, Volodymyr V. Nekrashevych, and Vitaliı̆ Ī Suščans1 kiı̆, Automata,
dynamical sysytems, and groups, PROSI 231 (2000), no. 4, 128–203.
[10] Narain D. Gupta and Said N. Sidki, On the Burnside problem for periodic groups, Math. Z.
182 (1983), 385–388.
[11] Ju. Š. Gurevič, The problem of equality of words for certain classes of semigroups, Algebra
i Logika Sem. 5 (1966), no. 5, 25–35 (Russian). MR0206079
THE WORD AND ORDER PROBLEMS FOR SELF-SIMILAR AND AUTOMATA GROUPS 17
[12] Jarkko Kari, The nilpotency problem of one-dimensional cellular automata, SIAM J. Comput.
21 (1992), no. 3, 571–586, DOI 10.1137/0221036. MR1163346
[13] Ines Klimann, Jean Mairesse, and Matthieu Picantin, Implementing computations in automaton (semi)groups, Implementation and application of automata, Lecture Notes in Comput.
Sci., vol. 7381, Springer, Heidelberg, 2012, pp. 240–252, DOI 10.1007/978-3-642-31606-7 21.
MR2993189
[14] Ville Lukkarila, The 4-way deterministic tiling problem is undecidable, Theoret. Comput. Sci.
410 (2009), no. 16, 1516–1533, DOI 10.1016/j.tcs.2008.12.006. MR2502125 (2010d:03071)
[15] Marvin L. Minsky, Recursive unsolvability of Post’s problem of “tag” and other topics in
theory of Turing machines, Ann. of Math. (2) 74 (1961), 437–455, DOI 10.2307/1970290.
MR0140405
[16] Volodymyr V. Nekrashevych, Self-similar groups, Mathematical Surveys and Monographs,
vol. 117, American Mathematical Society, Providence, RI, 2005. MR2162164 (2006e:20047)
[17] Said N. Sidki, Automorphisms of one-rooted trees: growth, circuit structure, and acyclicity,
J. Math. Sci. (New York) 100 (2000), no. 1, 1925–1943. Algebra, 12. MR1774 362
Département de Mathématiques et Applications, École Normale Supérieure, Paris
and Mathematisches Institut, Georg-August Universität zu Göttingen
E-mail address: [email protected]
E-mail address: [email protected]
| 4math.GR
|
arXiv:1502.07409v3 [math.ST] 16 Mar 2017
Threshold estimation for stochastic processes with small noise
Yasutaka Shimizu∗
Department of Applied Mathematics, Waseda University
March 16, 2017
Abstract
Consider a process satisfying a stochastic differential equation with unknown drift
parameter, and suppose that discrete observations are given. It is known that a simple
least squares estimator (LSE) can be consistent, but numerically unstable in the sense
of large standard deviations under finite samples when the noise process has jumps.
We propose a filter to cut large shocks from data, and construct the same LSE from
data selected by the filter. The proposed estimator can be asymptotically equivalent to
the usual LSE, whose asymptotic distribution strongly depends on the noise process.
However, in numerical study, it looked asymptotically normal in an example where
filter was choosen suitably, and the noise was a Lévy process. We will try to justify
this phenomenon mathematically, under certain restricted assumptions.
Key words: stochastic differential equation, semimartingale noise, small noise
asymptotics, drift estimation, threshold estimator, mighty convergence.
MSC2010: 62F12, 62M05; 60G52, 60J75
1
Introduction
Let (Ω, F, (Ft )t≥0 , P) be a stochastic basis, on which an Rd -valued stochastic process X is
defined via the stochastic integral equation
Z t
ǫ
b(Xsǫ , θ0 ) ds + ǫ · Qǫt ,
(1.1)
Xt = x +
0
where x ∈ Rd , ǫ > 0, and θ0 is an unknown parameter that belongs to a parameter space
Θ0 , which is an open bounded, convex subset of Rp ; we put Θ := Θ0 , the closure of Θ0 ; b
is a measurable function on Rd × Θ; Qǫ is a stochastic process of the form
e ǫ + rǫ B H
Qǫt := Q
t
t
e ǫ is a semimartingale, B H is a fractional Brownian motion with the Hurst paramwhere Q
eter H ∈ (0, 1), and rǫ is a real number satisfying that rǫ → 0 (ǫ → 0). We also assume for
∗
E-mail: [email protected]
1
2
Threshold estimation for small noise model
e ǫ that Q
e ǫ = 0 and supt∈[0,1] Q
eǫt − Qt → 0 as ǫ → 0 with probability
the semimartingale Q
0
one. Hence we have that
sup |Qǫt − Qt | → 0 a.s.,
t∈[0,1]
ǫ → 0,
(1.2)
As is well known, B H is not a semimartingale and neither is Qǫ , but Qǫ converges
uniformly to a semimartingale Q as ǫ → 0 almost surely.
Moreover, we suppose that the Doob-Meyer decomposition of Q is given as follows.
Qt = At + Mt
(t ≥ 0),
A0 = M0 = 0
a.s.,
(1.3)
where A is a process with finite variation, and M is an Ft -local martingale. Some typical
examples for Qǫ is given in Section 2.3. We suppose that the process X ǫ = (Xtǫ )t∈[0,1]
is observed discretely in time: {Xtnk }nk=0 with tnk = k/n (the index ǫ is omitted in the
notation), from [0, 1]-interval. We denote by ∆nk X := Xtnk − Xtnk−1 and ∆n := tnk − tnk−1 =
1/n. Our interest is to estimate the value of the parameter θ0 from the discrete samples
under that n → ∞ as well as ǫ → 0: small noise asymptotics.
There are some practical advantages in small noise asymptotics:
• Statistical point of view: the drift estimation by samples from a fixed finite time
interval is justifiable under relatively mild conditions. Since we need to observe the
process long time to achieve “good” estimation of the drift without small noise assumption, we usually assume the asymptotics that the terminal time of observations
goes to infinity, under which some technical conditions such as “ergodicity” or “uniform moment conditions” for the process need to be assumed. By a suitable scaling
technique, we can regard such a long-term model as a small noise model, approximately. Then there is no need for those conditions, which are sometimes difficult to
check in practice; see remarks in Section 2.3.3 for details.
• Computational point of view: approximation of functionals of the process as ǫ → 0
is often available in relatively easy-to-calculate form as in, e.g., Yoshida [37] and
Pavlyukevich [23] among others, which is well applied to finance and insurance; see
Takahashi [32], Kunitomo and Takahashi [12], Takahashi and Yoshida [33], Uchida
and Yoshida [36], Pavlyukevich [22] and references therein.
That is why, using the small noise model is convenient to deal with both applications and
statistical inference at the same time.
Sampling problems for stochastic differential equations with small noise have been well
studied by many authors in both theoretical and applied point of views. Some earlier works
for small-diffusion models are found in the papers by Kutoyants [13, 14], Genon-Catalot
[6] and Laredo [15], and they have been developed in some directions by several authors:
e.g., martingale estimating functions are studied by Sørensen [30], efficient estimation
is investigated by Sørensen and Uchida [31], Gloter and Sørensen [8]; see also Uchida
[34, 35], and asymptotic expansion approach is initiated by Yoshida [37, 38], see also
3
Yasutaka Shimizu
Uchida and Yoshida [36], among others. Although those works are due to diffusion noise,
more general noise model are also considered recently. For example, Long [16] and Ma
[18] investigate the drift estimation of a Lévy driven Ornstein-Uhlenbeck process; see also
Long [17], and Long et al. [20] deal with the inference for non-linear drift under the small
semimartingale noise. In our paper, we do not require that the noise is a semimartingale,
but ‘approximately’ a semimartingale in the sense of (1.2).
The goal of this paper is statistical inference for the drift of the process under the
asymptotics that noise vanishes, which is the same motivation as in Long et al. [20],
but it was seen in our numerical study that our estimator performed better than theirs
under finite samples. Long et al. [20] consider the case where Qǫ ≡ L is a Lévy process
(although it can be extended to a case of a semimartingale) independent of the dispersion
parameter ǫ, and investigate the asymptotic behavior of the least squares-type estimator
(LSE) defined by
LSE
θbn,ǫ
:= arg min Ψn,ǫ (θ),
θ∈Θ
where
Ψn,ǫ (θ) = ǫ−2 ∆−1
n
n
X
k=1
|∆nk X − b(Xtnk−1 , θ) · ∆n |2
(1.4)
(1.5)
LSE is ǫ−1 -consistent with the limit of
They show that the minimum contrast estimator θbn,ǫ
a Lévy functional as ǫ → 0 and n → ∞ with nǫ → ∞ under some mild conditions on the
function b; see Theorems 4.1 and 4.2 in [20]. The asymptotic distribution generally has a
fat-tail, that causes us unsatisfactory performance even if ǫ is small enough; see numerical
results in [20], or Section 3 below. This would be due to ‘large’ shocks by the driving noise.
It will be easy to imagine that a ‘large’ jump of Qǫ makes much impact to the direction of
drift, and make the drift estimation unstable. Therefore, cutting such ‘large’ jumps could
improve the performance. That is why, we consider the threshold-type estimator defined
as follows:
θbn,ǫ := arg min Φn,ǫ(θ),
θ∈Θ
where
Φn,ǫ (θ) = ǫ
−2
∆−1
n
n
X
k=1
|∆nk X − b(Xtnk−1 , θ) · ∆n |2 1{|∆nk X|≤δn,ǫ } ,
(1.6)
(1.7)
δn,ǫ is a positive number, which is a threshold to eliminate ‘large’ shocks causing bias
to drift estimation. That is, the indicator 1{|∆nk X|≤δn,ǫ } plays a role of a filter to split
increments with ‘large’ and ‘small’ magnitude of shocks; see Shimizu [26], or Shimizu
and Yoshida [29] for the fundamental idea of those filters. It would be intuitively clear
LSE and θ
bn,ǫ can be asymptotically equivalent. However, we
that if δn,ǫ → ∞ then θbn,ǫ
will show that the same thing holds true even if δn,ǫ → 0 by choosing the sequence δn,ǫ
4
Threshold estimation for small noise model
carefully. Moreover we can see in numerical study that θbn,ǫ has much better finite-sample
LSE . Furthermore, we will show the mighty convergence (the
performance than that of θbn,ǫ
convergence of moments) for u
bn,ǫ := ǫ−1 (θbn,ǫ − θ0 ), which is a stronger result than those
in [20]:
Z
E [f (b
un,ǫ )] →
Rp
f (z) Lu (dz),
as ǫ → 0 and n → ∞ for every continuous function f of at most polynomial growth, where
Lu is the asymptotic distribution of u
bn,ǫ .
As is described above, the asymptotic distribution Lu is generally not normal unless
the limiting process Q is a Wiener process. However, it is interesting to note that we
sometimes encounter a phenomenon that u
bn,ǫ seems asymptotically normal in numerical
study. This may indicate that our filtered LSE could also be asymptotically normal
if we choose δn,ǫ in a different way as previous, otherwise we may just observe it as
an ‘approximate’ phenomenon possibly when some appropriate conditions are satisfied.
Although a justification of this phenomena under the discrete sampling is still an open
problem, we will add some discussion on these points in Section 4.
The paper is organized as follows. In Section 2, we prepare notation and assumptions,
and present the main results under discrete samples. In particular, Section 2.4 is devoted
to investigating some technical conditions for Q. We will give some easy-to-check sufficient
conditions for those when the noise Q is a Lévy process. In Section 3, we will show an
advantage of our estimator compared with the usual LSE via numerical study, and we
further observe the asymptotic distribution seems normal. Finally, we will try to give a
theoretical explanation to those asymptotic phenomena in Section 4. All the proofs of
main theorems are given in Section 5.
2
Main results
2.1
Notation and assumptions
We use the following notation:
• For a process Y = (Yt )t∈[0,1] , kY kLp = (E|Y1 |p )1/p (p > 0) and kY k∗ := supt∈[0,1] |Yt |.
• Given a multilinear form M = {M (i1 ,...,iK ) : ik = 1, . . . , dk ; k = 1, . . . , K} ∈ Rd1 ⊗
(i)
· · · ⊗ RdK and vector uk = (uk )i≤dk ∈ Rdk , we write
M[u1 , . . . , uK ] =
d1
X
i1 =1
···
dK
X
(i )
(i )
M (i1 ,...,iK ) u1 1 . . . uKK .
iK =1
Note that the above form is well-defined when the jth dimension of M (the number
of ij ) and that of uj are the same. When some of uk is missing in “M[u1 , . . . , uK ]”,
the resulting form is regarded as a multilinear form again; e.g., M[u3 , . . . , uK ] ∈
Rd1 ⊗ Rd2 . For example, when M is a vector M = (Mi )1≤i≤d , M[x] = M ⊤ x for
5
Yasutaka Shimizu
f is a matrix M = (Mij )1≤i≤d ;1≤j≤d ,
x ∈ Rd , which is the inner product, and when M
1
2
f
M[u, v] for u = (ui )1≤i≤d1 and v = (vi )1≤i≤d2 is the quadratic form u⊤ M v, and
f = M v ∈ Rd1 or M[u]
f
M[v]
= M ⊤ u ∈ Rd2 , among others. The correspondences
of dimensions will be clear from the context. We also use the notation M[u⊗K ] :=
M[u1 , . . . , uK ] when u1 = · · · = uK .
• For a = (a1 , . . . , am )⊤ ∈ Rm , ∇a = (∂/∂a1 , . . . , ∂/∂am ). Moreover, we denote
by ∇ka := ∇a ⊗ ∇ak−1 with ∇0a ≡ 1. That is, ∇ka forms a multilinear form, e.g.,
∇2a = ∇⊤
a ∇a in a matrix form.
• For a multilinear form M, |M|2 denotes the sum of the squares of each element of
M.
• C is often used as a generic positive constant that may differ from line to line.
Moreover, we write a . b if a ≤ Cb almost surely.
• C k,l (Rd × Θ; Rq ) denotes the space of functions f (x, θ) : Rd × Θ → Rq that is k and l
times differentiable with respect to x and θ, respectively. Moreover, C↑k,l (Rd × Θ; Rq )
denotes a subclass of f ∈ C k,l (Rd × Θ; Rq ) that is of polynomial growth uniformly
in θ ∈ Θ: supθ∈Θ |∇αx ∇βθ f (x, θ)| . (1 + |x|)C for any α ≤ k and β ≤ l.
• For a function g(x, θ) : Rd × Θ → Rq , we write gk−1 (θ) := g(Xtnk−1 , θ). Moreover,
denote by χk (θ) := ∆nk X − bk−1 (θ)∆n (∈ Rd ).
• All the asymptotic symbol are described under n → ∞ and ǫ → 0 unless otherwise
noted.
Using the above notation, our estimating function given in (1.7) is rewritten as
Φn,ǫ (θ) = ǫ−2 ∆−1
n
n
X
n
I[χ⊗2
k (θ)]1{|∆k X|≤δn,ǫ } ,
k=1
where I is the d × d identity matrix.
We make the following assumptions on the model (1.1):
A1 |b(x, θ) − b(y, θ)| . |x − y| for each x, y ∈ Rd and θ ∈ Θ.
Under this assumption, the ordinary differential equation
dXt0 = b(Xt0 , θ0 ) dt,
X00 = x,
has the unique solution X 0 = (Xt0 )t≥0 .
A2 b ∈ C↑2,3 (Rd × Θ; Rd ).
A3 θ 6= θ0 ⇔ b(Xt0 , θ) 6= b(Xt0 , θ0 ) for at least one value of t ∈ [0, 1].
6
Threshold estimation for small noise model
A4 I(θ0 ) :=
R1
0
∇θ b(Xt0 , θ0 )⊤ ∇θ b(Xt0 , θ0 ) dt (∈ Rp ⊗ Rp ) is positive definite,
We further make the following conditions for the limiting process Q of Qǫ .
Q1[γ] There exists some γ > 0 such that, for any k = 1, . . . , n,
)
(
P
sup
n
t∈(tn
k−1 ,tk ]
|Qt − Qtnk−1 | > ∆γn Ftnk−1
= op (1).
Q2[q] For q > R0 and processes A and M given in (1.3), the total variation of A, say
1
T V (A) := 0 |dAt |, and the quadratic variation [M, M ] satisfy that
h
i
q/2
E [T V (A)q ] + E [M, M ]1
< ∞.
Although the condition Q1 seems ad hoc, we can give some easy-to-check conditions
in some important cases where, e.g., Q is a Lévy process satisfying Q2[q] for some q > 0;
see Section 2.4 for details.
2.2
Asymptotic behavior of threshold-type estimators
Theorem 1. Suppose A1–A3, Q1[γ], and that a sequence {δn,ǫ } satisfies that
δn,ǫ ∆−1
n → ∞,
Then
−1
ǫ ∆γn δn,ǫ
= O(1),
nǫ → ∞.
(2.1)
P
θbn,ǫ −→ θ0 .
Remark 1. Condition (2.1) ensures a kind of “negligibility”:
P |∆nk X| > δn,ǫ |Ftnk−1 → 0,
LSE ; see (5.13) in the proof.
which makes θbn,ǫ asymptotically equivalent to θbn,ǫ
Theorem 2. Suppose the same assumptions as in Theorem 1, and further A4. Then
Z 1
P
−1 b
−1
∇θ b(Xt0 , θ0 )[dQt ],
ǫ (θn,ǫ − θ0 ) −→ ζ := I (θ0 )
0
where the square bracket [dQt ] is the quadratic form of dQt for the matrix ∇θ b(Xt0 , θ0 ) in
this case; see Section 2.1.
Theorem 3. Suppose the same assumptions as in Theorem 2, and that Q2[q] holds true
for any q > 0. Then, it follows for every continuous function f : Rp → R, of polynomial
growth that
h
i
E f ǫ−1 (θbn,ǫ − θ0 ) → E[f (ζ)],
where ζ is given in Theorem 2.
7
Yasutaka Shimizu
2.3
2.3.1
Examples and remarks
α-stable noise
Suppose d = 1, and that Qǫ = σB + ρǫ S, where ρǫ → ρ ∈ [0, ∞), σ ≥ 0 is a constant, B
is a Brownian motion, and S is a standard α-stable process with stability index α ∈ (0, 2)
and the skewness parameter β ∈ [−1, 1], which is denoted by Sα (1, β, 0); see, e.g., Cont
and Tankov [4], page 94 for this notation. In this case, the limiting variable ζ in Theorem
2 becomes
ζ = I −1 (θ0 ) {σD2 · Z + ρDα · Sα (1, β, 0)} ,
R
1/α
1
where Dα = 0 {∂θ b(Xt0 , θ0 )}α dt
and Z is a standard Gaussian variable. Hence if
ρ = 0 then the estimator is asymptotically normal; see also Long et al. [20].
2.3.2
Markovian noise
Let us consider the case where
ǫ
dQǫt = ǫ1 a(Xtǫ ) dWt + ǫ2 c(Xt−
) dZt ,
(2.2)
where ǫj > 0 (j = 1, 2), a, c are some suitable functions, and Z is a pure jump Lévy
process.
• Case of ρǫ := ǫ2 /ǫ1 → 0: we can reparametrize as e
ǫ := ǫ · ǫ1 to obtain that
with
e ǫt ,
dXtǫ = b(Xtǫ , θ0 ) dt + e
ǫ · dQ
e ǫ = a(X ǫ ) dWt + ρǫ c(X ǫ ) dZt .
dQ
t
t
t−
Then e
ǫ−1 (θbn,ǫ − θ0 ) is asymptotically normal as in Sørensen and Uchida [31]. We
can verify that
Z t
ǫ
e
a(Xs0 ) dWt a.s., ǫ → 0,
Qt → Q :=
0
uniformly in t ∈ [0, 1] from the fact that the stochastic integrals are continuous
with respect to the u.c.p. topology; see Theorem II.11 by Protter [24], and the
uniform convergence of X ǫ → X 0 on compact sets; see Theorem IX.4.21 by Jacod
and Shiryaev [10].
• Case of ρǫ := ǫ2 /ǫ1 → ρ ∈ (0, ∞): we can use the same model as above, and the
asymptotic distribution of e
ǫ−1 (θbn,ǫ − θ0 ) is a convolution of stochastic integrals with
respect to W and Z.
8
Threshold estimation for small noise model
• Case of ρǫ := ǫ2 /ǫ1 → ∞: we can reparametrize as ǫ := ǫ · ǫ2 to obtain that
ǫ
dXtǫ = b(Xtǫ , θ0 ) dt + ǫ · dQt ,
with
ǫ
ǫ
ǫ
dQt = ρ−1
ǫ a(Xt ) dWt + c(Xt− ) dZt .
Then the asymptotic distribution of ǫ−1 (θbn,ǫ − θ0 ) is written by a stochastic integral
with respect to Z only.
2.3.3
Long-term observations
Consider a ‘long-term’ discretely observed model:
Z t
b(Xv ) dv + Qt ,
Xt = x +
(2.3)
0
and X is observed at 0 = tn0 < t1n <, . . . , < tnn =: T for a fixed T > 0 ‘large enough’, where
Q is an α-stable process: Q1 ∼ Sα (σ, β, 0) with index α ∈ (1, 2). To estimate the drift
function b, we sometimes assume a parametric model b(x) = b(x, θ) and estimate θ under
the assumption that T → ∞. However, in a standard theory of parametric inference, we
usually need an “ergodicity” or uniform moment conditions such as supt>0 E|Xt |m < ∞
for any m > 0, which are often restrictive conditions.
Now, transform the model by v = T u, t = T s, and divide the both sides of (2.3) by T
to obtain
Z s
1
x
b(T · Yu , θ) du + QT s , s ∈ [0, 1],
Ys = +
T
T
0
where Ys = T −1 XT s . Since QT s =d T 1/α Qs by the self-similarity of stable processes, we
can regard that Y is a (weak) solution to the following SDE:
Z s
x
es,
Ys = +
b(T · Yu , θ) du + T 1/α−1 Q
T
0
e s is also an α-stable process such that Q
e 1 ∼ Sα (σ, β, 0). Suppose that ǫ := T 1/α−1
where Q
e
is small enough for given T , and putting b(x, θ) = b(T x, θ) and x
e = ǫα/(α−1) x, we can
reformulate the model as
Z s
e
e s , s ∈ [0, 1],
Ys = x
e+
b(Yu , θ) du + ǫ · Q
(2.4)
0
with ǫ > 0 small enough, and it is interpreted as a small noise SDE.
Given a “non-ergodic” model as in (2.3) with long-term observations, we can reformulate it to a small noise model (2.4) with known constant ǫ = T 1/α−1 by plotting the data
e = Y0 in the interval [0, 1]. Then, although e
b
{Ysnk := T −1 XT snk }nk=0 with snk = tnk /T and x
depends on T , we can formally use a estimator of θ in e
b under the small noise model that
does not require any restrictive condition.
9
Yasutaka Shimizu
2.4
On conditions for Q
Let us investigate some sufficient conditions to ensure Q1 and Q2 when Q is specified.
An important
case is when
Q is a Lévy process such that the characteristic exponent:
ψ(u) := log E exp(iu⊤ Q1 ) , is given by
iu⊤ z
ν(dz), u ∈ Rd ,
(2.5)
2
1
+
|z|
d
R
R
where b ∈ Rd , ν is the Lévy measure with ν({0}) = 0 and |z|≤1 |z|2 ν(dz) < ∞.
As a simple case such that Q is a Wiener process, where b = 0, ν ≡ 0, it follows from
the property of the stationary, independent increments that
)
)
(
(
ψ(u) = ib⊤ u −
P
sup
n
t∈(tn
k−1 ,tk ]
σ2 ⊤
u u+
2
Z
eiu
⊤z
−1−
|Wt − Wtnk−1 | > ∆γn Ftnk−1
sup |Wt | > ∆γn
=P
t∈(0,∆n ]
p
= 2 1 − Φ(∆γn / ∆n ) → 0,
for any γ ∈ (0, 1/2), where Φ is a standard normal distribution function; for the last
equality, see, e.g., Doob [5], or Boukai [3].
When Q is a stable process, we have the following result.
Proposition 1. When Q is a symmetric α-stable process with α ∈ (1, 2). Then Q1[γ]
holds true for any γ ∈ (0, α−1 ).
Proof. Due to the maximal inequality (3.5) in Joulin [11], we have
(
)
(
)
P
sup
n
t∈(tn
k−1 ,tk ]
|Qt − Qtnk−1 | > ∆γn Ftnk−1
=P
sup |Qt | > ∆γn
t∈(0,∆n ]
= O ∆1−γα
→0
n
as n → ∞.
We can also present some sufficient conditions to Q1[γ] when a process Q is a more
general Lévy process: let
Z
Z
−2
ν(dz) + x
|z|2 ν(dz)
h(x) :=
|z|>x
+x
−1
|z|≤x
b+
Z
|z|≤x
z|z|2
ν(dz) −
1 + |z|2
Z
|z|>x
z
ν(dz) .
1 + |z|2
Proposition 2. Suppose that Q is a Lévy process with characteristic (2.5), and that there
exists a constant
η
β := inf η > 0 : lim sup x h(x) = 0 .
(2.6)
x→0
10
Threshold estimation for small noise model
Then the condition Q1[γ] holds true for any γ ∈ (0, γ0 ), where
γ0 =
β −1
(σ = 0)
.
(β ∨ 2)−1 (σ =
6 0)
We interpret that 1/0 = ∞.
Proof. Let Qt = σWt + Zt , where W is a Wiener process and Z is a pure jump Lévy
process with characteristic
Z
i
⊤
log E exp(iu Z1 ) = ib u +
h
⊤
Rd
iu⊤ z
e
iu⊤ z
−1−
1 + |z|2
ν(dz),
u ∈ Rd .
By the independent, stationary increments property for Lévy process Q, we see that
(
)
Rn := P
.P
sup
(
n
t∈(tn
k−1 ,tk ]
|Qt − Qtnk−1 | > ∆γn Ftnk−1
)
sup |σWt | > ∆γn /2
t∈(0,∆n ]
+P
(
)
sup |Zt | > ∆γn /2
t∈(0,∆n ]
(1)
=: Rn(1) + Rn(2) .
(1)
Note that Rn → 0 for any γ ∈ (0, 1/2) if σ 6= 0, and that Rn ≡ 0 if σ = 0. Moreover,
according to Pruitt [25], (3.2), it follows that
(2)
Rt
. ∆n h(∆γn /2) = (2xn )1/γ h(xn ),
xn := ∆γn /2.
(2)
Hence, if 1/γ > β, equivalently γ ∈ (0, β −1 ), then Rn → 0 by the definition of β ≥ 0.
Then the result follows.
Corollary 1. Suppose that
given by
R
|z|≤1 |z|ν(dz)
σ2 ⊤
u u+
ψ(u) =
2
< ∞, and that the Lévy characteristic ψ is
Z
(eiu
Rd
⊤z
− 1) ν(dz).
R
z
Note that b − |z|>x 1+|z|
2 ν(dz) = 0 in the expression (2.5). Then β in Proposition 2, (2.6)
is consistent with the Blumenthal-Getoor index:
)
(
Z
β = inf
η>0:
|z|≤1
|z|η ν(dz) < ∞
≤ 1.
Proof. Note that, under the assumption, we have
h(x) =
Z
ν(dz) + x
|z|>x
−2
Z
2
|z|≤x
|z| ν(dz) + x
−1
Z
|z|≤x
z|z|2
ν(dz) .
1 + |z|2
11
Yasutaka Shimizu
Let β0 be the Blumenthal-Getoor index:
(
β0 = inf
η ∈ [0, 2] :
Z
)
η
|z|≤1
|z| ν(dz) < ∞ .
Then we easily see by the direct computation that, for any ǫ > 0
lim xβ0 +ǫ h(x) = 0,
lim inf xβ0 −ǫ h(x) = ∞.
x→0
x→0
Indeed, for x ∈ (0, 1), we have
|x
β0 +ǫ
Z
β0 +ǫ
h(x)| ≤ |x|
ν(dz) +
+ |x|
Z
≤ |x|ǫ
|z|≤1
ν(dz)
|z|>1
x<|z|≤1
β0 +ǫ−2
Z
Z
2
|z|≤x
|z| ν(dz) + |x|
|z|β0 ν(dz) + xβ0 +ǫ
ǫ
Z
|z|>1
Z
|z|3
ν(dz)
2
|z|≤x 1 + |z|
Z
|z|β0 ν(dz)
ν(dz) + |x|ǫ
β0 +ǫ−1
|z|≤x
= O(|x| ) → 0.
The other condition: lim inf xβ0 −ǫ h(x) = ∞ is easier since
x→0
x
β0 −ǫ
h(x) > x
β0 −ǫ−2
Z
2
|z|≤x
|z| ν(dz) ≥ x
−ǫ
Z
|z|≤x
|z|β0 ν(dz) → ∞,
as x → 0. Hence we get β0 = β.
The condition Q2 can be reduced to moment conditions for the Lévy measure.
Proposition 3. Suppose that Q is a Lévy process with characteristic (2.5). Then the
condition Q2[q] holds true if
Z
|z|q ν(dz) < ∞.
|z|>1
Proof. The Lévy process Q has the following Lévy-Ito decomposition:
Z tZ
Z tZ
e (ds, dz) +
z N (ds, dz)
zN
Qt = at + bWt +
0
|z|≤1
0
|z|>1
where a and b are some constants, W is a Wiener process, N is a Poisson random measure,
e (ds, dz) = N (ds, dz) − ν(dz)ds. Then the decomposition (1.3) is given by
and N
Z tZ
Z tZ
e (ds, dz).
zN
z N (ds, dz), Mt = bWt +
At = at +
0
|z|>1
0
|z|≤1
12
Threshold estimation for small noise model
and it follows from the Burkholder-Davis-Gundy inequality that
q
Z
q
E[T V (A) ] . |a| + E
.1+E
Z
0
0
1Z
q
|z|>1
1Z
|z| N (ds, dz)
q
|z|>1
e (ds, dz) +
|z| N
Z
|z|>1
|z|q ν(dz).
According to the argument as in Bichteler and Jacod [2]; see also the proofs of lemma 4.1
and Proposition 3.1 by Shimizu and Yoshida [29], we see that
E
Z
0
1Z
q
|z|>1
Hence
e (ds, dz)
|z| N
q
E[T V (A) ] .
Z
|z|>1
.
Z
|z|>1
|z|q ν(dz).
|z|q ν(dz) < ∞.
By the similar argument, it is easy to see by Hölder’s inequality that, for any m > 1,
E|[M, M ]|q/2
1/m
≤ E|[M, M ]|mq/2
.1+
Z
|z|≤1
!1/m
|z|mq/2 ν(dz)
< ∞,
if we take mq/2 ≥ 2. This completes the proof.
3
3.1
Numerical study
2-dim model
We consider the following 2-dimensional Lévy driven SDE:
b(x, θ) =
q
θ1 +
x21
+
x22 , − p
θ 2 x2
1 + x21 + x22
!⊤
,
Qt =
Vtκ,ξ + Bt
,
Stα
(3.1)
where B is a standard Brownian motion, S α is a standard symmetric α-stable process
Sα (1, 0, 0), and V κ,ξ is a variance gamma process with Lévy density
pV (z) =
κ −ξ|z|
e
,
|z|
z ∈ R, κ, ξ > 0,
which is obtained by Brownian subordination with a gamma process Gt ∼ Γ(shape =
ct, scale = 1/λ) (c, λ > 0) as follows:
√
2κ
λ2
κ,ξ
, ξ=
Vt = σWGt (σ > 0), κ =
c
σ
13
Yasutaka Shimizu
where W is the standard Brownian motion independent of G; see, e.g., Cont and Tankov
[4] for details. Assume that W, S α and V κ,ξ are independent of each other.
In the sequel, we set values of parameters as
(1)
(2)
(X0 , X0 ) = (1, 1),
(θ1 , θ2 ) = (2, 1),
(κ, ξ, α) = (5, 3, 3/2).
A sample path is given in Figure 5. Then both X (1) and X (2) are unbounded variation
jump-processes with finite activity of jumps for X (1) and infinite for X (2) . We will compare
our threshold-type estimator to the LSE by Long et al. [20].
3.2
LSE and threshold estimator
LSE , θ
bLSE ) is a solution to
Note that the LSE θbLSE = (θb1,n,ǫ
2,n,ǫ
n
X
k=1
r
LSE
θb2,n,ǫ
=−
= 1;
LSE + (X (1) )2 + (X (2) )2
θb1,n,ǫ
tn
tn
k−1
k−1
n
and that our estimator θb = (θb1,n,ǫ , θb2,n,ǫ ) is a solution to
k=1
r
(∆nk X (2) )Xt(2)
n
k=1
∆nk X (1)
n
X
Pn
∆nk X (1)
(1)
(2)
θb1,n,ε + (Xtn )2 + (Xtn )2
k−1
k=1
θb2,n,ǫ = −
n
Pn
−1
k−1
(2)
(1)
1+(Xtn
k−1
)2 +(Xtn
)2
k−1
(2)
(Xtn
)2
.
k−1
k=1 1+(X (1) )2 +(X (2) )2
tn
tn
k−1
k−1
n
1X
1{|∆nk X|≤δn,ǫ }
n
k=1
(∆nk X (2) )Xt(2)
n
r
k−1
(2)
(1)
1+(Xtn
Pn
−1
1{|∆nk X|≤δn,ǫ } =
k−1
Pn
r
k−1
)2 +(Xtn
)2
1{|∆nk X|≤δn,ǫ }
k−1
(2)
(Xtn
)2
.
k−1
k=1 1+(X (1) )2 +(X (2) )2 1{|∆n
k X|≤δn,ǫ }
tn
tn
k−1
k−1
In the simulations, we tried n = 1000, 3000, 5000 and ǫ = (0.4, 0.3, 0.05). The results
for the LSE given in (1.4) are in Table 1, and those for the threshold-type estimator in
(1.6) are in 2 and 3, in which the threshold δn,ǫ = ǫ/5 and ǫ/10 are used, respectively. The
simulations are iterated 10000 times, and mean and standard deviation (s.d.) of estimators
are given in those tables. In Table 2 and 3, the values of nδn,ǫ are also included.
3.3
Discussion
From numerical results, we can observe that the threshold-type estimator improves the
accuracy of estimation in the sense of the standard deviation compared with the usual
LSE by Long et al. [20]. Especially, the improvement for θ2 is more drastic than that for
14
Threshold estimation for small noise model
θ1 . This would be because θ2 is the mean-reverting parameter for X (2) , a stable process
which has much more frequent and larger jumps than those of X (1) . In order to make the
estimation of θ1 more accurate, we need to make ǫ smaller rather than n larger.
We should note that the case where (n, ǫ, δn,ǫ ) = (1000, 0.05, ǫ/10) causes a large bias,
the reason of which would be that the asymptotic theory does not work well because
nδn,ǫ = 5 is ‘small’ although it should be large enough in the theory. From many simulations omitted here, we see that, at least, “nδn,ǫ ≥ 10” would be needed for estimation
with ‘small’ bias; see also, e.g., the case where (n, ǫ, δn,ǫ ) = (1000, 0.05, ǫ/5), which returns
better estimation.
Although there remains a problem to choose δn,ǫ in practice, we can use the method
proposed by Shimizu [27] if the parameters in noise process is known or estimable.
Finally, we observe normal QQ-plots for normalized estimators ǫ−1 (θb1,n,ǫ − θ1 ) and
−1
ǫ (θb2,n,ǫ − θ2 ) in the case where n = 5000, ǫ = 0.1 and δ = ǫ/5; see Figures 3 and 4.
According to the results, the estimators with δ = ǫ/5 seem asymptotically normal although
Theorem 2 does not necessarily say that. For your reference, see Figures 1 and 2 that are
the normal QQ-plots for the LSE without filter proposed by Long et al. [20]. The figures
show that the usual LSE’s are not necessarily asymptotically normal as the theory saying.
Therefore, the asymptotic normality-like phenomena would be due to the filter effects.
We could understand those results intuitively as follows: cutting large jumps from a
process with infinite activity jumps, the remaining small jumps will behave as a Brownian
motion. For example, suppose that the driving noise Q is a Lévy process of infinite activity
jumps with the Lévy measure ν, and put
Z tZ
(δ)
e (ds, dz), δ > 0,
zN
Qt :=
0
|z|≤δ
e is a compensated Poisson random measure as given in the proof of Proposition
where N
R
3. Then, according to Asmussen and Rosinski [1], it follows for σ 2 (δ) := |z|≤δ |z|2 ν(dz)
that
D
σ(δ)−1 Q(δ) −→ B,
in D[0, 1]-space equipped with the sup-norm under a certain assumption, where B is a
standard Brownian motion. Therefore, by an appropriate norming of estimator, the limit
in Theorem 2 might be an integral with respect to the process that is close to a Brownian
motion. This consideration indicates that the LSE with filter can be “approximately”
asymptotically normal by letting δn,ǫ converge to zero in a suitable rate, which is a great
advantage when we would like to make confidence intervals or do statistical testing.
In the next subsection, we shall try to understand this phenomenon theoretically.
4
Could a filtered LSE be asymptotically normal?
In this section, we assume that Qǫ ≡ Q is a Lévy process with Lévy measure ν. We
consider the following two cases for Q with characteristic (2.5):
15
Yasutaka Shimizu
• Finite activity case:
Z
• Infinite activity case:
|z|≤1
ν(dz) < ∞ and σ 2 > 0;
Z
|z|≤1
ν(dz) = ∞ (possibly, σ 2 = 0).
The former case, it would be possible to show the asymptotic normality of the filtered LSE
n
by separating the
R increments ∆k X’s with or without jumps as in Shimizu and Yoshida [29].
However, when |z|≤1 ν(dz) = ∞, it is known that the filter 1{|∆nk X|≤δn,ǫ } is not enough to
separate ∆nk X’s with or without jumps; see Shimizu [26], Lemma 3.3 and some remarks
on that. In this section, we shall consider the following ad hoc situation to understand
why the filtered LSE looks like asymptotically normal.
[Assumption] All the jumps of Q are observed.
Under this assumption, the following contrast function does make sense.
e n,ǫ,δ (θ) = ǫ−2 ∆−1
Φ
n
where δ > 0 is a constant, and
kQǫ k∗k =
n
X
k=1
|∆nk X − b(Xtnk−1 , θ) · ∆n |2 1{k∆Qǫ k∗ ≤δ} ,
sup
t∈(tk ,tk+1 ]
k
|∆Qǫt |,
∆Qǫt = ǫ · (Qt − Qt− ).
Under our assumption, we can specify if 1{k∆Qǫ k∗ ≤δ} = 1 or 0, and we can define the
k
estimator of θ as the minimum contrast estimator:
e n,ǫ,δ (θ).
θen,ǫ,δ = arg min Φ
θ∈Θ
Hereafter, we further use the following notation:
• For a κ > 0,
2
σ (κ) :=
Z
2
|z|≤κ
|z| ν(dz);
λ(κ) :=
Z
ν(dz),
|z|>κ
where ν is a Lévy measure of Q.
• All the asymptotic symbols are used under δ, ǫ → 0 and n → ∞.
4.1
Finite activity case
Suppose that
Z
|z|≤1
ν(dz) < ∞ and σ 2 > 0, which implies that Q is written as
Qt = σWt +
Nt
X
i=1
Yi ,
(4.1)
R
where N is a Poisson process with intensity λ := R ν(dz) and {Yi }i=1,2,... is an i.i.d.
sequence with distribution λ−1 ν. In this special case, we have the following result by
taking δn,ǫ ↓ 0 faster than the speed of ǫ that is a “magnitude of jumps”.
16
Threshold estimation for small noise model
Theorem 4. Suppose that Q is given by (4.1), and that A1–A4 hold true. Moreover,
suppose that
δ/ǫ → 0.
Then
ǫ
−1
P
(θen,ǫ,δ − θ0 ) −→ I −1 (θ0 )
Hence θen,ǫ,δ is asymptotically normal.
4.2
Z
(4.2)
1
0
∇θ b(Xt0 , θ0 )[dWt ],
Infinite activity case
R
Theorem 5. Let Qǫ ≡ Q be a Lévy process with |z|<1 ν(dz) = ∞, and suppose A1–A4,
Q1[γ], and that Q2[q] holds true for any q > 0. Moreover suppose that
λ(δ/ǫ)
→ c ∈ (0, 1),
n log n
nǫ∆γn → 0,
(4.3)
and that there exists a constant ρ ∈ (0, 1) such that
σ ρ (δ/ǫ) log n → ∞;
nǫ · σ(δ/ǫ) → ∞.
(4.4)
(4.5)
Furthermore, suppose for each κ > 0 that
σ (κσ(δ/ǫ) ∧ δ/ǫ) ∼ σ(δ/ǫ),
(4.6)
Then there exists a d-dimensional Brownian motion B, independent of X0 = x, such that
the following weak convergence holds true:
Z 1
D
−1 e
−1
∇θ b(Xt0 , θ0 ) [dBt ] ,
(σ(δ/ǫ)ǫ)
θn,ǫ,δ − θ0 −→ I (θ0 )
0
Hence θen,ǫ,δ is asymptotically normal.
We shall give a concrete example that satisfies the above situation.
Remark 2. The assumption (4.6) is to approximate a component of compensated small
jumps by a Wiener process; see Theorem 2 by Asmussen and Rosinski [1]. According to
Proposition 2.1 in [1], a simple condition that
(δ/ǫ)−1 σ(δ/ǫ) → ∞
(4.7)
is sufficient for (4.6). Moreover, note that (4.6) requires a high jump-activity, which
excludes cases where Q is a compound Poison process or a gamma process; see some
Examples 2.2–2.4 in [1]. Therefore it is assumed in (4.4) that
λ(δ/ǫ) → ∞
(δ → 0).
17
Yasutaka Shimizu
Remark 3. Although this is the result under an ideal situation that all the jumps are
observable, we can imagine that such a phenomenon ‘approximately’ occurs in simulations
presented in Figures 3 and 4. That is, when ∆n and δn,ǫ are sufficiently ‘small’, a filter
1{|∆nk X|≤δn,ǫ } can successfully cut jumps whose sizes are larger than δn,ǫ . However, it would
be hard to show the similar result when the observations are completely discrete as in our
original setting because the filter 1{|∆nk X|≤δn,ǫ } can not exactly exclude an increment that
includes the jumps whose sizes are larger than δn,ǫ ; see also Shimizu [26] about the filter
in infinite activity cases. The complete analysis in the discretely observed cases would be
an important work in the future.
ǫ = 0.4
LSE
θb1,n,ǫ
(s.d.)
LSE
θb2,n,ǫ
(s.d.)
n = 1000
2.54872
(2.4370)
1.72415
(3.5426)
n = 3000
2.53364
(2.4489)
1.79755
(4.2814)
n = 5000
2.78911
(2.5602)
1.78645
(2.9579)
True
2.0
ǫ = 0.3
LSE
θb1,n,ǫ
(s.d.)
LSE
θb2,n,ǫ
(s.d.)
n = 1000
2.31618
(1.8248)
1.50664
(2.8685)
n = 3000
2.29381
(1.7926)
1.5275
(2.7667)
n = 5000
2.34757
(1.7429)
1.53632
(2.8160)
True
2.0
ǫ = 0.05
LSE
θb1,n,ǫ
(s.d.)
LSE
θb2,n,ǫ
(s.d.)
n = 1000
2.00599
(0.2951)
1.05963
(1.3026)
n = 3000
2.01071
(0.2913)
1.04438
(0.6773)
n = 5000
2.01002
(0.2938)
1.06135
(0.7344)
True
2.0
1.0
1.0
1.0
Table 1: These are results for the LSE (without filter) based on Long et al. [20]. We
find that the standard deviation (s.d.) are large, especially, for θb2,n,ǫ , which implies an
unstability of estimation. We would like to improve the stability by using a ‘filter’.
5
5.1
Proofs
Preliminary lemmas
We shall first establish some preliminary lemmas to show the main theorems later.
18
Threshold estimation for small noise model
Normal Q−Q Plot
10
−10
0
Sample Quantiles
20
10
−30
−20
−10
−20
0
Sample Quantiles
30
20
40
30
Normal Q−Q Plot
−4
−2
0
2
4
−4
−2
Theoretical Quantiles
0
2
4
Theoretical Quantiles
LSE
Figure 1: Normal QQ-plots for θ[
1,n,ǫ . Left: ǫ = 0.1, Right: ǫ = 0.05. The results show
that the right tail is especially heavier than that of normal distribution.
Normal Q−Q Plot
−100
Sample Quantiles
−250
−300
−200
−50
−100
−200
−150
Sample Quantiles
0
0
50
100
Normal Q−Q Plot
−4
−2
0
Theoretical Quantiles
2
4
−4
−2
0
2
4
Theoretical Quantiles
LSE
Figure 2: Normal QQ-plots for θ[
2,n,ǫ . Left: ǫ = 0.1, Right: ǫ = 0.05. The results show
that the both of tails are heavier than those of normal distribution.
19
Yasutaka Shimizu
ǫ = 0.4
θb1,n,ǫ
n = 1000
2.46071
(2.1968)
1.14289
(0.8721)
n = 3000
2.45564
(2.1777)
1.15283
(0.88426)
n = 5000
2.43215
(2.1860)
1.16222
(0.8856)
True
2.0
ǫ = 0.3
θb1,n,ǫ
(s.d.)
θb2,n,ǫ
n = 1000
2.25007
(1.6149)
1.08105
(0.6498)
n = 3000
2.25972
(1.6121)
1.1047
(0.6489)
n = 5000
2.25829
(1.6249)
1.09448
(0.6563)
True
2.0
ǫ = 0.05
θb1,n,ǫ
(s.d.)
θb2,n,ǫ
(s.d.)
n = 1000
2.00619
(0.2623)
0.98972
(0.10098)
n = 3000
2.00827
(0.2594)
0.99936
(0.1031)
n = 5000
2.00681
(0.2652)
1.00039
(0.1037)
True
2.0
(s.d.)
θb2,n,ǫ
(s.d.)
(s.d.)
1.0
1.0
1.0
Table 2: Results with filter: δn,ǫ = ǫ/5. Compared with the LSE, the improvement for
s.d. of θb2,n,ǫ is drastic although the one for θb1,n,ǫ is less. To make the estimation of θ1
accurate, we need to make ǫ smaller. The values of (nδn,ǫ , δn,ǫ ǫ−1 n1/4 ) seems enough to
meet the asymptotic conditions such that they must tend to (∞, 0).
20
Threshold estimation for small noise model
ǫ = 0.4
θb1,n,ǫ
(s.d.)
θb2,n,ǫ
(s.d.)
n = 1000
2.10973
(2.0587)
1.06597
(0.6997)
n = 3000
2.24022
(2.1736)
1.10031
(0.7226)
n = 5000
2.46723
(2.2064)
1.10585
(0.7352)
True
2.0
ǫ = 0.3
θb1,n,ǫ
(s.d.)
θb2,n,ǫ
(s.d.)
n = 1000
1.92432
(1.4600)
1.02419
(0.5216)
n = 3000
2.25282
(1.6221)
1.05871
(0.5453)
n = 5000
2.24733
(1.6021)
1.05869
(0.5349)
True
2.0
ǫ = 0.05
θb1,n,ǫ
n = 1000
0.78959
(0.2165)
0.84069
(0.0754)
n = 3000
2.00133
(0.2628)
0.98834
(0.0853)
n = 5000
2.00864
(0.2653)
0.99450
(0.0865)
True
2.0
(s.d.)
θb2,n,ǫ
(s.d.)
1.0
1.0
1.0
Table 3: Results with filter: δn,ǫ = ǫ/10. The s.d. for θb2,n,ǫ is smaller than those with
δn,ǫ /5 as well as those for θb1,n,ǫ . However, we should be careful by observing the case of
(n, ǫ) = (1000, 0.05), where both estimators are negatively biased. This would be because
nδn,ǫ = 5 is too small to meet the corresponding asymptotic condition: nδn,ǫ → ∞.
21
Yasutaka Shimizu
Normal Q−Q Plot
0
Sample Quantiles
−20
−10
−10
0
Sample Quantiles
10
10
20
20
Normal Q−Q Plot
−4
−2
0
2
4
−4
−2
Theoretical Quantiles
Figure 3: Normal QQ-plots for
as ǫ = 0.05, the distribution of
0
2
4
Theoretical Quantiles
θb1,n,ǫ. Left: ǫ = 0.1, Right: ǫ = 0.05. When ǫ is small such
θb1,n,ǫ seems almost Gaussian.
Normal Q−Q Plot
0
Sample Quantiles
0
−5
−5
Sample Quantiles
5
5
Normal Q−Q Plot
−4
−2
0
Theoretical Quantiles
2
4
−4
−2
0
2
4
Theoretical Quantiles
Figure 4: Normal QQ-plots for θb2,n,ǫ. Left: ǫ = 0.1, Right: ǫ = 0.05. When ǫ is small such
as ǫ = 0.05, the distribution of θb1,n,ǫ seems almost Gaussian.
22
0.8
0.4
0.6
X2
1.0
1.2
1.0
1.5
X1
2.0
Threshold estimation for small noise model
0.0
0.2
0.4
0.6
0.8
1.0
Time
Figure 5: A sample path of Model (3.1) with (θ1 , θ2 , κ, ξ, α) = (2, 1, 5, 3, 3/2) and ε = 0.4.
To evaluate functions of discrete samples, we use the following notation as in Long et
al. [20]:
• Discretized process: Ytn,ǫ := X⌊nt⌋/n for t ≥ 0 and n ∈ N, where ⌊x⌋ stands for the
integer part of x ∈ R.
• Since Q is a semimartingale, we can consider the Doob-Meyer decomposition: Q =
A + M , where A is a process with finite variation, and M is an Ft -local martingale
with A0 = M0 = 0 a.s.
• A stopping time for localization: for m ∈ N,
n,ǫ
τm
= inf{t ≥ 0 : |Xt0 | ∧ |Ytn,ǫ | ≥ m} ∧ Tm ,
Rt
where Tm := inf{t ≥ 0 : [M, M ]t ∧ 0 |dAs | ≥ m}. As a convention, inf ∅ = ∞.
Hence, note that limm→∞ Tm = ∞ almost surely.
Lemma 1. Under A1, it holds that
Y n,ǫ − X 0
P
∗
−→ 0.
In addition, suppose Q2[q] for some q > 0. Then it holds that
n,ǫ
1
0 q
q
E Y −X ∗ =O
+ǫ .
nq
(5.1)
(5.2)
23
Yasutaka Shimizu
Proof. From (1.2), it follows for ǫ small enough that
kQǫ k∗ . kQk∗ + 1.
(5.3)
Hence we can take the same argument as in the proof of Lemma 3.1, (3.1) by Long et al.
[20] to obtain that
P
X − X0
∗
. ǫkQǫ k∗ . ǫ(kQk∗ + 1) −→ 0,
(5.4)
since kQk∗ is bounded in probability. Hence the fact that ⌊nt⌋/n → t as n → ∞ yields
that
Y n,ǫ − X 0
∗
P
≤ kY n,ǫ − Xk∗ + X − X 0
∗
−→ 0.
This is the proof of (5.1).
Moreover, note that
E
Y
n,ǫ
q
− X0 ∗
.E
.E
"
"
sup X⌊nt⌋/n − Xt
t∈[0,1]
sup
t∈[0,1]
+ ǫq E
"
Z
q
#
+E
⌊nt⌋/n
b(Xs , θ0 ) ds
t
sup Qǫ⌊nt⌋/n − Qǫt
t∈[0,1]
q
#
q#
X − X0
+E
q
∗
X − X0
q
.
∗
Noticing the linear growthness of the function b: |b(x, θ)| . 1 + |x|, and the inequalities
(5.3) and (5.4), we have that
"
q#
Z ⌊nt⌋/n
n,ǫ
0 q
E Y − X ∗ . E sup
(1 + kXk∗ ) ds
t∈[0,1]
t
q
+ E X − X0 ∗
nt − ⌊nt⌋ q
E [(1 + kXk∗ )q ] + ǫq (1 + E [kQkq∗ ])
. sup
n
t∈[0,1]
i
h
nt − ⌊nt⌋ q
+ ǫq 1 + EkAkq∗ + ǫq E [M, M ]q/2
. sup
n
t∈[0,1]
1
q
=O
+ǫ ,
nq
+ǫ
q
E [kQkq∗ ]
under Q2[q]. We used the Burkholder-Davis-Gundy inequality in the last inequality. This
completes the proof.
Form (5.2) and (5.4) in the above proof, the following corollary is obvious.
24
Threshold estimation for small noise model
Corollary 2. Suppose A1 and Q2[q] for some q > 0. Then it holds that
q
< ∞.
sup E ǫ−1 kX − X 0 k∗
(5.5)
ǫ>0
In addition, if (nǫ)−1 = O(1), then
sup E
n∈N,ǫ>0
ǫ−1 kY n,ǫ − X 0 k∗
q
< ∞.
(5.6)
Lemma 2. Under A1, it follows that
lim τ n,ǫ
m→∞ m
=∞
a.s.,
uniformly in n ∈ N and ǫ ∈ [0, 1].
Proof. Noticing (5.3), we have by the same argument as in the proof of Lemma 3.2 by
Long et al. [20] that
!
√
2
n,ǫ
sup |Yt | ≤ 2 |x| + sup |Qs | + t eCt < ∞ a.s.
n∈N,ǫ>0
s∈[0,t]
for any t > 0. Therefore we have the consequence.
Lemma 3. Let g ∈ C↑1,1 (Rd × Θ; R). Suppose A1, A2, Q1[γ], and that
δn,ǫ ∆−1
n → ∞,
−1
ǫ ∆γn δn,ǫ
= O(1),
nǫ → ∞.
Then, we have
n
1X
P
gk−1 (θ)1{|∆nk X|≤δn,ǫ } −→
n
k=1
Z
1
0
g(Xt0 , θ) dt,
uniformly in θ ∈ Θ. In addition, suppose that Q2[q] holds for some q > 0, then
!q #
"
Z 1
n
1X
0
−1
< ∞.
gk−1 (θ)1{|∆nk X|≤δn,ǫ } −
g(Xt , θ) dt
sup E
ǫ sup
n∈N,ǫ>0
θ∈Θ n
0
k=1
Proof. Since |g(x)| . (1 + |x|)C , we have that
Z 1
n
1X
g(Xt0 , θ) dt
gk−1 (θ)1{|∆nk X|≤δn,ǫ } −
n
θ∈Θ
0
k=1
Z
n
n
1
1X
1X
0
g(Xt , θ) dt +
gk−1 (θ) −
(1 + |Xtnk−1 |)C 1{|∆nk X|>δn,ǫ }
. sup
n
θ∈Θ n
0
sup
k=1
k=1
(5.7)
(5.8)
25
Yasutaka Shimizu
= sup
θ∈Θ
Z
0
n
1
g(Ytn,ǫ ) − g(Xt0 , θ) dt +
1X
(1 + |Xtnk−1 |)C 1{|∆nk X|>δn,ǫ } .
n
k=1
For the first term on the right-hand side of the inequality, it holds that
Z 1
sup
g(Ytn,ǫ ) − g(Xt0 , θ) dt
θ∈Θ
0
≤ sup
.
Z
θ∈Θ 0
Z 1
0
∇x g(Xs0 + u(Ysn,ǫ − Xs0 ), θ) · Ysn,ǫ − Xs0 duds
1 + |Xs0 | + |Ysn,ǫ |C
0
.
1Z 1
(5.9)
Ysn,ǫ − Xs0 ds
!C
1 + sup |Xt0 | + sup |Xt0 |
t∈[0,1]
t∈[0,1]
P
sup Ytn,ǫ − Xt0 −→ 0,
(5.10)
t∈[0,1]
by Lemma 1, (5.1). Hence the proof ends if we show the second term tends to zero in
probability. Let
1
ξk,n,ǫ(θ) := (1 + |Xtnk−1 |)C 1{|∆nk X|>δn,ǫ } .
n
We show that
n
i
h
X
P
(5.11)
E ξk,n,ǫ(θ)|Ftnk−1 −→ 0;
k=1
n
X
k=1
i
h
P
E |ξk,n,ǫ(θ)|2 |Ftnk−1 −→ 0.
(5.12)
P
P
which implies that nk=1 ξk,n,ǫ(θ) −→ 0 for each θ ∈ Θ from Lemma 9 by Genon-Catalot
and Jacod [7]. First, we show that (5.11). Note that
n
X
k=1
h
E ξk,n,ǫ(θ)|Ftnk−1
Since it follows that n
show that
Pn
−1
i
1X
=
(1 + |Xtnk−1 |)C P |∆nk X| > δn,ǫ |Ftnk−1
n
n
k=1
+ |Xtnk−1 |)C = Op (1) by Lemma 3.3 in [20], it suffices to
P |∆nk X| > δn,ǫ |Ftnk−1 = op (1),
k=1 (1
for any k = 1, . . . , n. Note that it holds that
sup
n
t∈(tn
k−1 ,tk ]
|Xt − Xtnk−1 | . ∆n (1 + kXk∗ ) + ǫ
sup
n
t∈(tn
k−1 ,tk ]
|Qǫt |.
Hence it follows from Q1[γ] and the assumption on δn,ǫ that, for n large enough,
!
δn,ǫ
n
sup |Qt − Qtnk−1 | >
P |∆k X| > δn,ǫ |Ftnk−1 ≤ P
Ftnk−1
n]
2ǫ
,t
t∈(tn
k−1 k
26
Threshold estimation for small noise model
δn,ǫ
Ftn
+ P (1 + kXk∗ ) >
2∆n k−1
≤P
sup
n
t∈(tn
k−1 ,tk ]
|Qt − Qtnk−1 | & ∆γn Ftnk−1
+ P (1 + kXk∗ ) & ∆−1
n δn,ǫ Ftn
k−1
!
= op (1),
(5.13)
since kXk∗ is bounded in probability. This is the proof of (5.11). The proof of (5.12)
is similar to above, which ends the proof of (5.7). The proof of (5.8) is easy from the
estimates (5.10) and Corollary 2 since we are assuming that nǫ → ∞, so we omit the
details. Then the proof is completed.
The next lemma is a version of the Toeplitz lemma; see also, e.g., Shimizu [28]. We
need this result in the proof of the next Lemma 5.
P
Lemma 4. Let {ank }nk=1 be a positive bounded sequence, and put bn := nk=1 ank . Suppose
that a sequence {xnk }nk=1 satisfies the following conditions:
sup |xnk | < ∞
for each fixed k;
(5.14)
|xnk − x| = 0
(5.15)
n∈N
lim
sup
m→∞ k,n: m≤k≤n
P
Then, for any sequence An with An ∼ b−1
n , An
for some x ∈ R;
ani xni → x as n → ∞.
Lemma 5. Let g ∈ C↑1,1 (Rd × Θ; Rd ). Under A1, it holds that
n
X
k=1
P
gk−1 (θ)[∆nk Qǫ ] −→
Z
1
0
g(Xt0 , θ) [dQt ].
Proof. Since we are assuming the uniform convergence (1.2), we have that
n
X
k=1
gk−1 (θ)[∆nk Qǫ ] −
n
X
=
k=1
n
X
≤
k=1
+
=:
Z
Z
1
0
g(Xt0 , θ) [dQt ]
gk−1 (θ)[∆nk Qǫ − ∆nk Q] +
gk−1 (θ)[∆nk Qǫ − ∆nk Q] +
1
Z
0
Z
1
1
0
g(Ytn,ǫ , θ) − g(Xt0 , θ) [|dAt |]
g(Ytn,ǫ , θ) − g(Xt0 , θ) [dMt ]
0
(1)
(2)
In,ǫ + In,ǫ
(3)
+ In,ǫ
.
g(Ytn,ǫ , θ) − g(Xt0 , θ) [dQt ]
27
Yasutaka Shimizu
(1)
As for In,ǫ , we suppose d = 1 for simplicity of notation. Putting ank := gk (θ) − gk−1 (θ)
and xnk := Qǫtk − Qtk (since ǫ depends on n), we have that
(1)
In,ǫ
=
n n
o
X
gk (θ)[Qǫtk − Qtk ] − gk−1 (θ)[Qǫtk−1 − Qtk−1 ]
k=1
n
X
−
=
k=1
{gk (θ) − gk−1 (θ)} [Qǫtk − Qtk ]
g(X1ǫ )[Qǫ1
− Q1 ] −
n
X
k=1
ank · xnk .
Now,
that g(X1ǫ )[Qǫ1 − Q1 ] → 0 a.s. by the assumption (1.2). Moreover, the sum
Pn itnis clear
n
k=1 ak · xk also converges to zero with probability one by using Lemma 4. Indeed, the
convergence (5.14) is clear from (1.2), and that
lim
sup
k→∞ j,n: k≤j≤n
|xnj | . lim kQǫ − Qk∗ = 0
ǫ→0
a.s.
Moreover, as ǫ → 0,
bn =
n
X
k=1
ank =
n
X
k=1
[gk (θ) − gk−1 (θ)] → g(X10 , θ) − g(X00 , θ) < ∞ a.s.
P
Hence we see by the above Toeplitz lemma that nk=1 ank · xnk → 0 with probability one.
(1)
Therefore we have that In,ǫ → 0 with probability one.
(2)
As for In,ǫ , it follows from Lemma 1, (5.1) that
P
(2)
In,ǫ
. (1 + sup |Xt0 | + sup |Xt |)C sup Ytn,ǫ − Xt0 −→ 0.
t∈[0,1]
t∈[0,1]
t∈[0,1]
(3)
As for In,ǫ , using Markov’s, and the Burkholder-Davis-Gundy inequalities, we have for
any η > 0 that
Z 1
n,ǫ
(3)
n,ǫ
0
n,ǫ
P(|In,ǫ | > η) ≤ P(τm < 1) + P
g(Yt , θ) − g(Xt , θ) 1{t≤τm } [dMt ] > η/2
0
Z 1
n,ǫ
−1
n,ǫ [dMt ]
≤ P(τm < 1) + 2η E
g(Ytn,ǫ , θ) − g(Xt0 , θ) 1{t≤τm
}
n,ǫ
. P(τm
< 1) + η −1 E
Z
0
1
0
2
n,ǫ d[M, M ]t
g(Ytn,ǫ , θ) − g(Xt0 , θ) 1{t≤τm
}
1/2
.
n,ǫ
From the definition of τm
, the integrand in the last term is bounded in n and ǫ. Hence,
taking the limit n → ∞ and ǫ → 0, we see from the dominated convergence theorem that
(3)
P(|In,ǫ | > η) → 0, which completes the proof.
28
Threshold estimation for small noise model
Lemma 6. Let g ∈ C↑1,1 (Rd × Θ; Rd ). Assume A1, A2, Q1[γ], and that
δn,ǫ ∆−1
n → ∞,
−1
ǫ ∆γn δn,ǫ
= O(1),
nǫ → ∞.
Then we have
n
X
k=1
P
gk−1 (θ)[χk (θ0 )]1{|∆nk X|≤δn,ǫ } −→ 0,
uniformly in θ ∈ Θ. In addition, assume Q2[q] for any q > p = dim(Θ). Then
"
!q #
n
X
E
ǫ−1 sup
gk−1 (θ)[χk (θ0 )]1{|∆nk X|≤δn,ǫ }
< ∞.
(5.16)
(5.17)
θ∈Θ k=1
Proof. The proof of (5.16) is similar to the one of Lemma 3.5 by Long et al. [20] with a
slight extension to semimartingale version; see also Remark 4.3 in [20]. It is clear from
their proof that the indicator 1{|∆nk X|≤δn,ǫ } is not essential to the proof. So it is omitted.
As for (5.17), we note that
ǫ
−1
sup
n
X
θ∈Θ k=1
=ǫ
−1
≤ ǫ−1
sup
n Z
X
θ∈Θ k=1
Z 1
1
0
g(Ysn,ǫ , θ) [b(Xs , θ0 ) − b(Ysn,ǫ , θ0 ) + ǫ · dQǫt ] ds · 1{|∆nk X|≤δn,ǫ }
sup |g(Ysn,ǫ , θ) [b(Xs , θ0 ) − b(Ysn,ǫ , θ0 )]| ds
0 θ∈Θ
Z 1
+ sup
=:
gk−1 (θ)[χk (θ0 )]1{|∆nk X|≤δn,ǫ }
g(Ysn,ǫ , θ)[dQǫt
θ∈Θ 0
(1)
(2)
In,ǫ + In,ǫ
− dQt ] + sup
θ∈Θ
(3)
+ In,ǫ
.
Z
1
0
g(Ysn,ǫ , θ)[dQt ]
By the assumption A1 and the condition for g, we have
Z 1
(1)
−1
In,ǫ . ǫ
(1 + |Ysn,ǫ |)λ |Xs − Ysn,ǫ | ds
0
. 1 + kY n,ǫ − X 0 kλ∗ + kX 0 kλ∗ ǫ−1 kX − X 0 k + ǫ−1 kX 0 − Y n,ǫ k
Hence, under the assumption that nǫ → ∞, Corollary 2 yields that
(1) q
E|In,ǫ
| < ∞.
(2)
P
We have already shown that In,ǫ −→ 0 in the proof of Lemma 5. Hence the proof ends
(3)
P
if we show that In,ǫ −→ 0. Noticing that a bounded convex set Θ admits the following
Sobolev inequality;
sup |u(θ)| . ku(θ)kLq (Θ) + k∇θ u(θ)kLq (Θ) ,
θ∈Θ
29
Yasutaka Shimizu
for q > p = dim(Θ), we see for any q > p that
Z 1
Z 1
(3)
(1 + |Ysn,ǫ , θ|)λ · |dAs | + sup
g(Ysn,ǫ , θ)[dM ]
In,ǫ
≤
θ∈Θ
0
0
Z
n,ǫ
0 λ
0 λ
. 1 + kY − X k∗ + kX k∗ T V (A) +
Θ
Z
0
1
q
g(Ysn,ǫ , θ)[dM ǫ ]
dθ
Then, by using the Burkholder-Davis-Gundy inequality, we have
Z 1
i Z
h
q
(3) q
E
g(Ysn,ǫ , θ)[dM ]
T
V
(A)
+
E|In,ǫ
| . E 1 + kY n,ǫ − X 0 kλq
∗
0
q Θ
i
h
n,ǫ
0 λ
0 λ
. 1 + E 1 + kY − X k∗ + kX k∗ |[M, M ]1 |q/2 < ∞,
1/q
.
q
dθ
under Q2[q] for any q > p. This completes the proof of (5.17).
Lemma 7. Let g ∈ C↑1,1 (Rd × Θ; Rd ). Assume A1, A2, Q1[γ], and that
δn,ǫ ∆−1
n → ∞,
−1
ǫ ∆γn δn,ǫ
= O(1),
nǫ → ∞.
Then we have
n
X
gk−1 (θ)[χk (θ0 )]1{|∆nk X|>δn,ǫ } = op (ǫ),
k=1
for each θ ∈ Θ.
Proof. Using ∆nk X =
ǫ−1
n
X
R tnk
tn
k−1
b(Xt , θ0 ) dt + ǫ∆nk Qǫ , we have that
gk−1 (θ)[χk (θ0 )]1{|∆nk X|>δn,ǫ }
k=1
=
n
X
k=1
n
X
+
=:
ǫ
−1
Z
tn
k
tn
k−1
gk−1 (θ)[b(Xt , θ0 ) − bk−1 (θ0 )] dt · 1{|∆nk X|>δn,ǫ }
gk−1 (θ)[∆nk Qǫ ] · 1{|∆nk X|>δn,ǫ }
k=1
(1)
Hn,ǫ
(θ)
(2)
+ Hn,ǫ
(θ).
(1)
As for Hn,ǫ , we easily find it converges to zero in probability as nǫ → ∞ by the same
(1)
argument as for Hn,ǫ in Lemma 3.6 by Long et al. [20].
(2)
As for Hn,ǫ , it follows from Lemma 3 and 5 that
(2)
Hn,ǫ
(θ) =
n
X
k=1
gk−1 (θ)[∆nk Qǫ ] −
n
X
k=1
P
gk−1 (θ)[∆nk Qǫ ]1{|∆nk X|≤δn,ǫ } −→ 0,
for each θ ∈ Θ. Hence the proof is completed.
30
Threshold estimation for small noise model
5.2
Proof of Theorem 1
LSE given in (1.4). Let
We shall show that θbn,ǫ is asymptotically equivalent to θbn,ǫ
e n,ǫ(θ) := ǫ2 (Ψn,ǫ(θ) − Ψn,ǫ (θ0 )) ,
Ψ
e n,ǫ (θ) := ǫ2 (Φn,ǫ (θ) − Φn,ǫ(θ0 )) .
Φ
LSE and θ
bn,ǫ are respectively
where Ψn,ǫ and Φn,ǫ are given in (1.5) and (1.7). Then θbn,ǫ
minimum contrast estimators for contrast functions Ψn,ǫ and Φn,ǫ .
By the same argument as in the proof of Theorem 2.1 with Remark 4.3 by Long et al.
[20], all we need to show is
P
e n,ǫ (θ) − F (θ0 )| −→ 0,
sup |Φ
θ∈Θ
where F (θ) :=
R1
0
P
e n,ǫ (θ)−F (θ0 )| −→ 0, Therefore,
|b(Xt0 , θ)−b(Xt0 , θ0 )|2 dt. Since supθ∈Θ |Ψ
e n,ǫ(θ) − F (θ0 ) ≤ sup Φ
e n,ǫ (θ) − Ψ
e n,ǫ (θ) + sup Ψ
e n,ǫ (θ) − F (θ0 )
sup Φ
θ∈Θ
θ∈Θ
θ∈Θ
= sup n
θ∈Θ
k=1
n
X
|χk (θ)|2 1{|∆nk X|>δn,ǫ } + op (1)
(bk−1 (θ) − bk−1 (θ0 ))[χk (θ0 )]1{|∆nk X|>δn,ǫ }
θ∈Θ k=1
n
1X
|bk−1 (θ) − bk−1 (θ0 )|2 1{|∆nk X|>δn,ǫ } + op (1)
sup
θ∈Θ n k=1
. sup
+
n
X
The last first and second terms converges to zero in probability by Lemmas 3 and 6. This
completes the proof.
5.3
Proof of Theorem 2
We use the following notation:
• Gn,ǫ (θ) = 2−1 ∇θ Φn,ǫ(θ) ∈ Rp ;
• Kn,ǫ (θ) = ∇θ Gn,ǫ (θ) = ∇2θ Φn,ǫ (θ) ∈ Rp ⊗ Rp ;
• K(θ) =
R1
0
∇2θ b(Xt0 , θ)[b(Xt0 , θ0 ) − b(Xt0 , θ)] dt − I(θ0 ) ∈ Rp ⊗ Rp .
Then it follows by Taylor’s formula that, for some ρ ∈ (0, 1),
Z
0
1
Kn,ǫ θ0 + u(θbn,ǫ − θ0 ) du · ǫ−1 θbn,ǫ − θ0 = ǫ−1 Gn,ǫ (θbn,ǫ ) − ǫ−1 Gn,ǫ (θ0 ) .
31
Yasutaka Shimizu
Let us show that
P
ǫ−1 Gn,ǫ (θ0 ) −→
Z
1
0
∇θ b(Xt0 , θ0 ) [dQt ];
(5.18)
P
sup |Kn,ǫ (θ) − K(θ)| −→ 0.
(5.19)
θ∈Θ
Then the result follows by the same argument as in the proof of Theorem 2.2 by Long et
al. [20]; see also Uchida [35].
As for (5.18): it follows that
ǫ−1 Gn,ǫ (θ0 ) = ǫ−1
n
X
k=1
= ǫ−1
n
X
k=1
=ǫ
−1
−ǫ
n
X
∇θ bk−1 (θ0 )[χk (θ0 )]1{|∆nk X|≤δn,ǫ }
∇θ bk−1 (θ0 )[χk (θ0 )] − ǫ−1
∇θ bk−1 (θ0 )
k=1
n
X
−1
k=1
"Z
tn
k
tn
k−1
n
X
k=1
∇θ bk−1 (θ0 )[χk (θ0 )]1{|∆nk X|>δn,ǫ }
#
{b(Xt , θ0 ) − bk−1 (θ0 )} dt +
n
X
k=1
∇θ bk−1 (θ0 )[∆nk Q]
∇θ bk−1 (θ0 )[χk (θ0 )]1{|∆nk X|>δn,ǫ }
Then we can show that the last first term converges to zero in probability by the same
(1)
evaluation as for Hn,ǫ (θ0 ) in the proof of Lemma 3.6 in [20], the second term converges to
R1
0
0 ∇θ b(Xt , θ0 )[dQt ] in probability by Lemma 5, and that the third term goes to zero in
probability by Lemma 7. Similarly, as for (5.19), it follows that
Kn,ǫ (θ) =
=
n
X
k=1
n
X
k=1
+
n
∇2θ bk−1 (θ)[χk (θ)]1{|∆nk X|≤δn,ǫ } −
1X
∇θ bk−1 (θ)⊤ ∇θ bk−1 (θ)1{|∆nk X|≤δn,ǫ }
n
k=1
∇2θ bk−1 (θ)[χk (θ0 )]1{|∆nk X|≤δn,ǫ }
n
1X 2
∇θ bk−1 (θ)[bk−1 (θ0 ) − bk−1 (θ)]1{|∆nk X|≤δn,ǫ }
n
k=1
n
1X
∇θ bk−1 (θ)⊤ ∇θ bk−1 (θ)1{|∆nk X|≤δn,ǫ } ,
−
n
k=1
ij
cf. the expression of Kn,ǫ
(θ) in the proof of Lemma 3.7 by Long et al. [20]. Hence Lemmas
3 and 6 yield (5.19). Using the facts (5.18) and (5.19), and the consistency result: Theorem
1, we can show that
ǫ−1 θbn,ǫ − θ0 ∼p −K −1 (θ0 ) · ǫ−1 Gn,ǫ (θ0 ), n → ∞, ǫ → 0,
by completely the same argument as in the proof of Theorem 2.2 by Long et al. [20].
Therefore, the proof is completed.
32
5.4
Threshold estimation for small noise model
Proof of Theorem 3
P
Denote by u
b := ǫ−1 (θbn,ǫ − θ0 ). Since u
b −→ ζ from Theorem 2, the proof ends if we show
that u
b is Lp -bounded: supn,ǫ E|b
u|p < ∞ for any p > 0. For this proof, let Un,ǫ (θ0 ) =
p
{u ∈ R : θ0 + ǫu ∈ Θ0 }, and define random fields Zn,ǫ : Un,ǫ (θ0 ) → R+ by
Zn,ǫ(u) = exp {−Φn,ǫ(θ0 + ǫu) + Φn,ǫ (θ0 )} ,
u ∈ Un,ǫ (θ0 ).
Then, since θ0 ∈ Θ0 , we see that
Zn,ǫ (b
u) ≥
sup
u∈Un,ǫ (θ0 )
Zn,ǫ (u) ≥ Zn,ǫ (0) = 1.
(5.20)
Setting Vn,ǫ (r) := Un,ǫ (θ0 ) ∩ {u ∈ Rp : |u| ≥ r}, we consider the following condition for
the random fields Zn,ǫ : for every L > 0 and r > 0,
!
P
sup
u∈Vn,ǫ (r)
. r −L ,
Zn,ǫ (u) ≥ e−r
(5.21)
which is called the polynomial type large deviation inequality (PLDI), and is investigated
by Yoshida [39] in details. If this PLDI holds true then, for any L > p,
Z ∞
p
r p−1 P (|b
u| ≥ r) dr
sup E|b
u| = sup p
n∈N,ǫ>0
n∈N,ǫ>0
≤
.
sup p
0
Z
n∈N,ǫ>0
0
Z ∞
p−1
r
0
∞
r p−1
(
1∧P
sup
u∈Vn,ǫ (r)
!)
Zn,ǫ (u) ≥ 1
dr
(1 ∧ r −L ) dr < ∞,
here we used (5.20) in the first inequality. Therefore the proof ends if we show (5.21), some
sufficient conditions for which are found in the paper by Yoshida [39]. Here we shall verify
the conditions [A1′′ ], [A4′ ], [A6], [B1] and [B2] given in Theorem 3, (c) in [39]. See also
Ogihara and Yoshida [21] or Masuda [19] for simplified descriptions for those conditions.
Applying Taylor’s formula with the notation Gn,ǫ (θ), Kn,ǫ (θ) and K(θ) given in the
proof of Theorem 2, we have
log Zn,ǫ (u) = −Φn,ǫ (θ0 + ǫu) + Φn,ǫ(θ0 )
= −ǫGn,ǫ (θ0 )[u] −
ǫ2
{−K(θ0 )} [u⊗2 ] + Rn,ǫ (u),
2
where
Z
1
(s − 1) K(θ0 )[u⊗2 ] − Kn,ǫ (θ0 + s · ǫu)[u⊗2 ] ds
0
Z 1
Z 1
ǫ2
⊗2
3
{Kn,ǫ (θ0 ) − K(θ0 )} [u ] − ǫ
∇θ Kn,ǫ (θ0 + ts · ǫu)[u⊗3 ] dtds.
(s − 1)
=
2
0
0
Rn,ǫ (u) = ǫ
2
33
Yasutaka Shimizu
This means that Zn,ǫ could be Partially Locally Asymptotically Quadratic (PLAQ), which
is a starting point of [39]. According to Theorem 3, (c) in [39], if we take some “tuning
parameters” given in [A4′ ] in [39] such as β1 ≈ 1/2, ρ1 , ρ2 , β, β2 ≈ 0, then the PLDI
(5.21) holds true if the following [A1′′ ], [A6], [B1] and [B2] are satisfied; we use the same
conditioning numbers as in [39] to make those correspondences clear.
[A1′′ ] For every q > 0,
sup E
n∈N,ǫ>0
2
ǫ sup
θ∈Θ
∇3θ Φn,ǫ (θ)
q
< ∞.
Moreover, for given L > 0 and any δ > 0 small enough,
h
L−δ i
sup E ǫ−1 |Kn,ǫ (θ0 ) − K(θ0 )|
< ∞.
(5.22)
(5.23)
n∈N,ǫ>0
[A6] For any δ > 0 small enough,
i
h
sup E |ǫGn,ǫ (θ0 )|L+δ < ∞;
(5.24)
n∈N,ǫ>0
L+δ
−1 e
Φn,ǫ (θ) − F (θ)
sup E sup ǫ
< ∞,
n∈N,ǫ>0
e n,ǫ and F (θ) =
where Φ
1.
R1
0
(5.25)
θ∈Θ
|b(Xt0 , θ)−b(Xt0 , θ0 )|2 dt are given in the proof of Theorem
[B1] The matrix −K(θ0 ) (= I(θ0 )) is deterministic and positive definite.
[B2] There exists a deterministic positive number χ such that
− {F (θ) − F (θ0 )} ≤ −χ|θ − θ0 |2 .
(Note that the notational correspondence between [39] and ours is: aT = ǫ; bT = ǫ−2 ;
HT = −Φn,ǫ ; Y = −F and Γ = −K).
Now we can easily check that the conditions (5.22) and (5.24) are true by Lemma 6,
(5.17), and that (5.23) and (5.25) are also true by Lemma 3, (5.8). Moreover the conditions
[B1] and [B2] are clear from the assumptions A1 and A4, respectively, the proof ends if
we show (5.22)–(5.25). Hence the proof is completed.
5.5
Proof of Theorem 4
First, we shall show the consistency:
P
θen,ǫ,δ −→ θ0 .
34
Threshold estimation for small noise model
Since we suppose that the jumps are specified, the following “negligibility” is obtained:
)
(
P (k∆Qǫ k∗k > δ) = 1 − P
|∆Qt | ≤ δ/ǫ
sup
n
t∈(tn
k−1 ,tk ]
= 1 − e−λ(δ/ǫ)∆n → 0
(5.26)
since λ(δ/ǫ)∆n → 0. Then we have the following lemma, which is the same type of results
as Lemmas 3 and 6.
Lemma 8. Let g ∈ C↑1,1 (Rd × Θ; R). Suppose A1, A2, and that
δ/ǫ → 0.
Then, we have
n
1X
P
gk−1 (θ)1{k∆Qǫ k∗ ≤δ} −→
k
n
k=1
n
X
k=1
Z
1
0
g(Xt0 , θ) dt,
P
gk−1 (θ)[χk (θ0 )]1{k∆Qǫ k∗ ≤δ} −→ 0,
k
uniformly in θ ∈ Θ.
The proof of this lemma is an obvious modification of the proofs of Lemmas 3 and 6
using the negligibility condition (5.26). Then by the same argument as in the proof of
Theorem 1, the consistency follows.
Next, note that, as in the proof of Theorem 2,
Z 1
Kn,ǫ,δ θ0 + u(θen,ǫ,δ − θ0 ) du · θen,ǫ,δ − θ0 = Gn,ǫ,δ (θbn,ǫ ) − Gn,ǫ,δ (θ0 ) .
0
where
e n,ǫ,δ (θ); Kn,ǫ,δ (θ) = ∇θ Gn,ǫ,δ (θ);
Gn,ǫ,δ (θ) = 2−1 ∇θ Φ
Z 1
∇2θ b(Xt0 , θ)[b(Xt0 , θ0 ) − b(Xt0 , θ)] dt − I(θ0 ).
K(θ) =
0
If we show that
ǫ
−1
P
Gn,ǫ,δ (θ0 ) −→
Z
0
1
∇θ b(Xt0 , θ0 ) [dWt ] ,
then we obtain the consequence because the convergence
P
sup |Kn,ǫ,δ (θ) − K(θ)| −→ 0
θ∈Θ
(5.27)
35
Yasutaka Shimizu
holds true due to Lemma 8 and the same argument as in the proof of Lemma 3.7 in [20].
Note that
#
"Z n
n
tk
X
{b(Xs , θ0 ) − bk−1 (θ0 )} dt 1{k∆Qǫ k∗ ≤δ}
∇θ bk−1 (θ0 )
ǫ−1 Gn,ǫ,δ (θ0 ) = ǫ−1
k=1
n
X
+
k=1
n
X
+
=:
k=1
(1)
In,ǫ,δ
∇θ bk−1 (θ0 )
k
tn
k−1
Nt n
k
X
i=Ntn
k−1
+1
Yi 1{|Yi |≤δ}
∇θ bk−1 (θ0 ) [∆nk W ] 1{k∆Qǫ k∗ ≤δ}
k
(3)
(2)
+ In,ǫ,δ + In,ǫ,δ .
(1)
(1)
Note that In,ǫ,δ → 0 by the same argument as in the proof of Lemma 3.6, Hn,ǫ (θ0 ) in [20]
by the assumption (4.5). Moreover, it is easy to see that
E
(2)
In,ǫ,δ
n
λδ X
=
E|∇θ bk−1 (θ0 )| → 0.
nǫ
k=1
Furthermore, note that
(3)
In,ǫ,δ =
n
X
k=1
∇θ bk−1 (θ0 ) [∆nk W ] −
n
X
k=1
∇θ bk−1 (θ0 ) [∆nk W ] 1{k∆Qǫ k∗ >δ}
k
Now, observe a measurability that
{ω ∈ Ω : k∆Qk∗k ≤ δ} ∈ σ N (tnk−1 , s], [−δ, δ] : s ∈ (tnk−1 , tnk ] ,
which is independent of ∆nk W and Xtnk−1 . Therefore,
E
≤
≤
n
X
∇θ bk−1 (θ0 ) [∆nk W ] 1{k∆Qǫ k∗ >δ}
k=1
n
X
k=1
n
X
k=1
k
E |∇θ bk−1 (θ0 ) [∆nk W ]| P (k∆Qǫ k∗k > δ)
E
Z
tn
k
tn
k−1
trace ∇θ b⊗2
(θ
)
dt
0
k−1
!1/2
1 − e−λ(δ/ǫ)∆n → 0
from (5.26). As a result, Lemma (5) yields that
Z 1
P
(3)
∇θ b(Xt0 , θ0 ) [dWt ] .
In,ǫ,δ −→
0
This completes the proof of (5.27).
36
5.6
Threshold estimation for small noise model
Proof of Theorem 5
Note that, under the asymptotic conditions, it follows that
P (k∆Qǫ k∗k ≤ δ) = e−λ(δ/ǫ)∆n → 0;
P (k∆Qǫ k∗k > δ) → 1,
(5.28)
which is a different situation in the previous theorem.
Under this setting, we can show the following lemma.
Lemma 9. Let g ∈ C↑1,1 (Rd × Θ; R). Suppose A1, A2, Q1[γ], and that
λ(δ/ǫ)
→ c ∈ (0, 1),
n log n
nǫ∆γn → 0.
Then it follows that, for ηn := e−λ(δ/ǫ)∆n ,
Z 1
n
1 X
P
g(Xt0 , θ) dt,
gk−1 (θ)1{k∆Qǫ k∗ ≤δ} −→
k
nηn
0
(5.29)
k=1
n
1 X
P
gk−1 (θ)[χk (θ0 )]1{k∆Qǫ k∗ ≤δ} −→ 0,
k
ηn
(5.30)
k=1
uniformly in θ ∈ Θ.
Proof of Lemma 9.
As for (5.29): note that
Z 1
n
1 X
P
sup
g(Xt0 , θ) dt
gk−1 (θ)1{k∆Qǫ k∗ ≤δ} −→
k
θ∈Θ nηn
0
k=1
n
n
1X
1 X
gk−1 (θ)1{k∆Qǫ k∗ ≤δ} −
gk−1 (θ)
k
n
θ∈Θ nηn k=1
k=1
Z 1
n
X
1
+ sup
g(Xt0 , θ) dt .
gk−1 (θ) −
θ∈Θ n
0
≤ sup
k=1
and the second term on the right-hand side of the inequality converges to zero
P in probability
from Lemma 3.3 by Long et al. [20]. The last first term is rewritten as nk=1 ξkn (θ) with
1
1
n
−1 .
1
ξk (θ) = gk−1 (θ)
ǫ ∗
n
ηn {k∆Q kk ≤δ}
Then we immediately see that, since Ftnk−1 and k∆Qǫ kk are independent each other,
n
X
k=1
E[ξkn (θ)|Ftnk−1 ]
n
1X
=
gk−1 (θ) ηn−1 P (k∆Qǫ k∗k ≤ δ) − 1 = 0.
n
k=1
37
Yasutaka Shimizu
Moreover
n
X
k=1
E[|ξkn (θ)|2 |Ftnk−1 ]
n
2
1 X 2
1
1{k∆Qǫ k∗ ≤δ} − 1
= 2
gk−1 (θ)E
k
n
ηn
k=1
1
1
P
= Op
−→ 0,
= Op
nηn
e(1−c2 ) log n
P
P
by the assumption. Hence we have that nk=1 ξkn (θ) −→ 0 for every θ ∈ Θ from Lemma 9
by Genon-Catalot and Jacod [7].
PTo prove the uniformity of convergence, we have to show the tightness of the sequence
{ nk=1 ξkn (·)}n . We shall use Theorem 20 in Appendix 1 by Ibragimov and Has’minskii
[9], that is, we shall show that, for some H > 0 and any N ∈ N,
E
n
X
2N
ξkn (θ)
< H;
(5.31)
≤ H|θ1 − θ2 |2N
(5.32)
k=1
E
n
X
[ξkn (θ1 )
k=1
−
2N
ξkn (θ2 )]
As for (5.31): using the independent property of Xtnk−1 and k∆Qǫ kk , we have that
E
n
X
2N
ξkn (θ)
k=1
n
X
"
"
2N
1
|Ftnk−1
1{k∆Qǫ k∗ ≤δ} − 1
=
E |gk−1 (θ)| E
k
ηn
k=1
#
"
n
1 X
1
2N
+1
.E
sup |g(Xt )|
n2N
ηn2N −1
θ∈Θ,t∈[0,1]
k=1
1
−(1−c)
=O
→ 0.
=
O
n
(nηn )2N −1
2N
##
Therefore it is bounded. Inequality (5.32) is similarly proved since g ∈ C↑1,1 (Rd × Θ; R).
Hence (5.29) is proved.
Finally we shall show (5.30). Note that
n
1 X
gk−1 (θ)[χk (θ0 )]1{k∆Qǫ k∗ ≤δ}
k
θ∈Θ ηn k=1
n Z n
1 X tk
sup |gk−1 (θ) [b(Xs , θ0 ) − bk−1 (θ0 )]| ds · 1{k∆Qǫ k∗ ≤δ}
≤
k
ηn
θ∈Θ
tn
k−1
sup
+
=:
k=1
n
X
ǫ
ηn
sup |gk−1 (θ) [∆nk Q]| 1{k∆Qǫ k∗ ≤δ}
k=1 θ∈Θ
Jn(1) + Jn(2) .
k
38
Threshold estimation for small noise model
(1)
As for Jn : it follows from the assumption A1 that, for Y n,ǫ given in Lemma 1,
n Z n
1 X tk
sup |g(Ysn,ǫ ) [b(Xs , θ0 ) − b(Ysn,ǫ , θ0 )]| ds · 1{k∆Qǫ k∗ ≤δ}
k
n
ηn
k=1 tk−1 θ∈Θ
!C
n
1 X
sup |Xt − Ytn,ǫ |1{k∆Qǫ k∗ ≤δ}
. 1 + sup |Xt |
·
k
n
n
nη
n
t∈[0,1]
k=1 t∈(tk−1 ,tk ]
!C
n
1 X
· kX − X 0 k∗ + kX 0 − Y n,ǫ k∗ ·
. 1 + sup |Xt |
1{k∆Qǫ k∗ ≤δ}
k
nηn
|Jn(1) | =
t∈[0,1]
k=1
Since
E
n
1 X
1{k∆Qǫ k∗ ≤δ} = 1
k
nηn
k=1
⇒
n
1 X
1{k∆Qǫ k∗ ≤δ} = Op (1),
k
nηn
and Lemma 1 we have that
(5.33)
k=1
P
|Jn(1) | −→ 0.
(2)
As for Jn : it follows for some C > 0 that
|Jn(2) | .
sup
(1 + |Xt |)C ·
θ∈Θ,t∈[0,1]
≤ Op
!
n
nǫ X 1
·
sup |Qs − Qtnk−1 |1{k∆Qǫ k∗ ≤δ}
k
ηn
n s∈(tnk−1 ,tnk ]
k=1
nǫ sup |Qs | ,
s∈(0,∆n ]
since sups∈(tn ,tn ] |Qs − Qtnk−1 | ∼d sups∈(0,∆n ] |Qs | and (5.33). Moreover, noticing under
k−1 k
Q1[γ] that
sup |Qs | = op (∆γn ),
s∈(0,∆n ]
we obtain that
Jn(2) = Op (nǫ∆γn ) → 0.
This completes the proof of (5.30).
Now, putting
Ψn,ǫ (θ) :=
ǫ2 ηn−1
e n,ǫ,δ (θ) − Ψ
e n,ǫ,δ (θ0 ) ;
Ψ
F (θ) :=
Z
1
0
and using Lemma 9, we can easily show that
P
sup Ψn,ǫ (θ) − F (θ) −→ 0.
θ∈Θ
|b(Xt0 , θ) − b(Xt0 , θ0 )|2 dt,
39
Yasutaka Shimizu
P
This and the identifiability condition A3 yield the consistency: θen,ǫ,δ −→ θ0 .
Suppose that the Lévy process Q is of the form
Z tZ
Z tZ
e
z N (dt, dz),
z N (dt, dz) +
Qt = at + cWt +
0
0
|z|≤1
|z|>1
where a ∈ Rd , c ≥ 0, W is a d-dimensional Wiener process, N is a Poisson random measure
e (dt, dz) = N (dt, dz) − ν(dz)dt, and note that, for any
associated with jumps of Q, and N
δ > 0,
Z Z
Z Z
t
Qt = αδ/ǫ t + Wt +
0
R
where aδ/ǫ = a − δ/ǫ<|z|≤1 z ν(dz).
Hereafter, we put
|z|≤δ/ǫ
e (dt, dz) +
zN
t
z N (dt, dz),
0
|z|>δ/ǫ
ζ := δ/ǫ.
Using the same notation as in the previous theorem, the proof ends if we show that
Z 1
D
∇θ b(Xt0 , θ0 ) [dBt ] .
(5.34)
(σ(ζ)ǫ)−1 Gn,ǫ,δ (θ0 ) −→
0
Note that
(σ(ζ)ǫ)−1 Gn,ǫ,δ (θ0 ) = (σ(ζ)ǫ)
n
X
−1
+ σ −1 (ζ)
+σ
=:
−1
(1)
In,ǫ,δ
(ζ)
k=1
n
X
∇θ bk−1 (θ0 )
"Z
tn
k
tn
k−1
#
{b(Xs , θ0 ) − bk−1 (θ0 )} dt 1{k∆Qǫ k∗ ≤δ}
k
∇θ bk−1 (θ0 ) aδ ∆n + ∆nk W 1{k∆Qǫ k∗ ≤δ}
k=1
n
X
k=1
(2)
+ In,ǫ,δ
k
∇θ bk−1 (θ0 )
"Z
tn
k
tn
k−1
Z
|z|≤ζ
(3)
+ In,ǫ,δ .
#
e (dt, dz)
zN
(2)
As for In,ǫ,δ : Since 1{k∆Qǫ k∗ ≤δ} is independent of ∆nk W and Xtnk−1 , we have that
k
n
(2)
E In,ǫ,δ ≤ σ −1 (ζ)
1X
E |∇θ bk−1 (θ0 )aδ | P (k∆Qk∗k ≤ ζ)
n
+ σ −1 (ζ)
k=1
n
X
k=1
≤σ
−1
(ζ)
=O σ
n
X
k=1
−1
E ∇θ bk−1 (θ0 ) ∆nk W P (k∆Qk∗k ≤ δ)
E
Z
tn
k
tn
k−1
trace
(ζ)e−λ(ζ)∆n .
∇θ b⊗2
k−1 (θ0 )
dt
!1/2
P (k∆Qk∗k ≤ ζ)
40
Threshold estimation for small noise model
The last equality is due to (5.28). Therefore we have
!−1
∞
k
X
(∆n λ(ζ))
(2)
.
E In,ǫ,δ = O σ(ζ)
k!
(5.35)
k=0
For an integer M such that 1/M ≤ ρ, we see that,
M
σ(ζ)(∆n λ(ζ))M = σ 1/M (ζ)∆n λ(ζ)
≥ (σ ρ (ζ)∆n λ(ζ))M → ∞
(5.36)
due to the the condition (4.4). Hence we have
(2)
E In,ǫ,δ → 0.
(1)
(1)
As for In,ǫ,δ : By the same argument as in the proof of Lemma 3.6, Hn,ǫ (θ0 ) in [20], we
can obtain the following inequality:
(1)
|In,ǫ,δ | .
n
1 1X
|∇θ bk−1 (θ0 )| · |bk−1 (θ0 )|
nǫσ(ζ) n
+
i=1
n
X
1
nσ(ζ)
i=1
|∇θ bk−1 (θ0 )|
sup
n
s∈(tn
k−1 ,tk ]
|Qt − Qtnk−1 |1{k∆Qǫ k∗ ≤δ}
k
(2)
Using the same estimates for Jn
in the proof of Lemma 9, we have that
1
(1)
+ op σ −1 (ζ)e−λ(ζ)∆n ,
|In,ǫ,δ | = Op
nǫσ(ζ)
(1)
P
Then, by the same estimates as for (5.35) and (5.36) above, we see that |In,ǫ,δ | −→ 0.
(3)
As for In,ǫ,δ : Let
Lδt
:= σ
−1
(ζ)
Z tZ
0
|z|≤δ
e (dt, dz).
zN
Thanks to Theorem 2 by Asmussen and Rosinski [1], it follows under the assumption (4.6)
that there exists a Wiener process B, independent of W , such that
D
Lζ −→ B
in D[0, 1],
as δ → 0. Moreover, since kY n,ǫ − X 0 k∗ a.s. by Lemma (1), we have a joint convergence
D
Lζ , Y n,ǫ −→ (B, X 0 ) in D[0, 1].
as ǫ, δ → 0 and n → ∞. Hence it follows from Theorem 5.16 by Jacod and Shiryaev [10]
that
n
i Z 1
h
X
(3)
n ζ
∇θ b (Ytn,ǫ , θ0 ) dLζt
∇θ bk−1 (θ0 ) ∆k L =
In,ǫ,δ =
k=1
0
41
Yasutaka Shimizu
D
−→
Z
1
0
∇θ b(Xt0 , θ0 ) dBt
This completes the proof of (5.34), and the statement is proved.
Acknowledgement
The author would like to thank the anonymous referees for their valuable suggestions and
proposals that significantly improve this manuscript. This research has been partially
supported by JSPS KAKENHI Grant-in-Aid for Scientific Research (C), Grant Number
JP15K05009.
References
[1] Asmussen, S. and Rosinski, J. (2001). Approximations of small jumps of Lévy processes with a view towards simulation, J. Appl. Prob., 38, 482–493.
[2] Bichteler, K. and Jacod, J. (1986) Calcul de Malliavin pour les diffusions avec sauts:
existence d’une densite dans le cas unidimensionnel. Lecture Notes in Math, 132–157.
Springer-Verlag, Berlin.
[3] Boukai, B. (1990). An explicit expression for the distribution of the supremum of
Brownian motion with a change point. Comm. Statist. Theory Methods, 19, (1), 31–
40.
[4] Cont, R. and Tankov, P. (2004). Financial modelling with jump processes. Chapman
& Hall/CRC, Boca Raton, FL.
[5] Doob, J. L. (1949). Heuristic approach to the Kolmogorov-Sminorv Theorems. Ann.
Math. Statistics, 20, 393–403.
[6] Genon-Catalot, V. (1990). Maximum contrast estimation for diffusion processes from
discrete observations, Statistics, 21. 99–116.
[7] Genon-Catalot, V. and Jacod, J. (1993). On the estimation of the diffusion coefficient
for multidimensional diffusion process, Ann. Inst. Henri Poincaré Prob. Statist. 29,
119–151.
[8] Gloter, A. and Sørensen, M. (2009). Estimation for stochastic differential equations
with a small diffusion coefficient, Stochastic Process. Appl., 119, 679–699.
[9] Ibragimov, I. A. and Has’minskii, R. Z. (1981) Statistical Estimation, Springer-Verlag,
Berlin.
42
Threshold estimation for small noise model
[10] Jacod, J. and Shiryaev, A. N. (2003). Limit theorems for stochastic processes. Second
edition. Springer-Verlag, Berlin.
[11] Joulin, A. (2007). On maximal inequalities for stable stochastic integrals. Potential
Anal., 26, (1), 57–78.
[12] Kunitomo, N. and Takahashi, A. (2001). The asymptotic expansion approach to the
valuation of interest rate contingent claims. Math. Finance, 11, 117–151.
[13] Kutoyants, Yu. A. (1984). Parameter estimation for stochastic processes. Heldermann,
Berlin.
[14] Kutoyants, Yu. A. (1994). Identification of dynamical systems with small noise.
Kluwer, Dordrecht.
[15] Laredo, C.F. (1990). A sufficient condition for asymptotic sufficiency of incomplete
observations of a diffusion process, Ann. Statist., 18, 1158–1171.
[16] Long, H. (2009) Least squares estimator for discretely observed Ornstein-Uhlenbeck
processes with small Lévy noises, Statist. Prob. Letters, 79, 2076–2085.
[17] Long, H. (2010). Parameter estimation for a class of stochastic differential equations
driven by small stable noises from discrete observations, Acta Mathematica Scientia,
30B, 645–663.
[18] Ma, C. (2010). A note on “Least squares estimator for discretely observed OrnsteinUhlenbeck processes with small Lévy noises”, Statist. Prob. Letters, 80, 1528–1531.
[19] Masuda, H. (2013). Convergence of Gaussian quasi-likelihood random fields for ergodic Levy driven SDE observed at high frequency. The Annals of Statistics, 41,
1593–1641.
[20] Long, H; Shimizu, Y. and Sun, W. (2013). Least squares estimators for discretely
observed stochastic processes driven by small Lévy noises. J. Multivariate Analysis,
116, 422–439.
[21] T. Ogihara, N. Yoshida, Quasi-likelihood analysis for the stochastic differential equation with jumps. Stat. Inference Stoch. Process., 14, 189–229.
[22] Pavlyukevich, I. (2008). Ruin probabilities of small noise jump-diffusions with heavy
tails. Appl. Stoch. Models Bus. Ind. 24, (1), 65–82.
[23] Pavlyukevich, I. (2011). First exit times of solutions of stochastic differential equations
driven by multiplicative Lévy noise with heavy tails. Stoch. Dyn., 11, (2–3), 495–519.
[24] Protter, P. E. (2004). Stochastic integration and differential equations. Second edition.
Springer-Verlag, Berlin.
Yasutaka Shimizu
43
[25] Pruitt, W. E. (1981). The growth of random walks and Lévy processes. Ann. Probability, 9, (6), 948–956.
[26] Shimizu, Y. (2006). M-estimation for discretely observed ergodic diffusion processes
with infinitely many jumps. Statist. Infer. Stochastic Proc., 9, (2), 179–225.
[27] Shimizu, Y. (2010). Threshold selection in jump-discriminant filter for discretely observed jump processes. Statist. Methods and Appl., 19, (3), 355–378.
[28] Shimizu, Y. (2012). Local asymptotic mixed normality for discretely observed nonrecurrent Ornstein-Uhlenbeck processes, Ann. Inst. Statist. Math., 64, (1), 193-211.
[29] Shimizu, Y. and Yoshida, N. (2006) Estimation of parameters for diffusion processes
with jumps from discrete observations. Statist. Infer. Stochastic Proc., 9, (3), 227–
277.
[30] Sørensen, M. (2000). Small dispersion asymptotics for diffusion martingale estimating
functions, Preprint No. 2000-2, Department of Statistics and Operation Research,
University of Copenhagen, Copenhagen.
[31] Sørensen, M. and Uchida, M. (2003). Small diffusion asymptotics for discretely sampled stochastic differential equations, Bernoulli, 9, (6), 1051–1069.
[32] Takahashi, A. (1999). An asymptotic expansion approach to pricing contingent claims.
Asia-Pacific Financial Markets, 6. 115–151.
[33] Takahashi, A. and Yoshida, N. (2004). An asymptotic expansion scheme for optimal
investment problems. Stat. Inference Stoch. Process., 7, 153–188.
[34] Uchida, M. Estimation for discretely observed small diffusions based on approximate
martingale estimating functions, Scand. J. Statist. 31 (2004) 553-566.
[35] Uchida, M. (2008). Approximate martingale estimating functions for stochastic differential equations with small noises. Stochastic Process. Appl., 118, 1706–1721.
[36] Uchida, M. and Yoshida, N. (2004). Asymptotic expansion for small diffusions applied
to option pricing. Stat. Inference Stoch. Process., 7, 189–223.
[37] Yoshida, N. (1992). Asymptotic expansion of maximum likelihood estimators for small
diffusions via the theory of Malliavin-Watanabe. Probab. Theory Relat. Fields, 92,
275–311.
[38] Yoshida, N. Asymptotic expansion for statistics related to small diffusions. J. Japan
Statist. Soc., 22, 139–159.
[39] Yoshida, N. (2011). Polynomial type large deviation inequalities and quasi-likelihood
analysis for stochastic differential equations. Ann. Inst. Statist. Math., 63, (3), 431–
479
| 10math.ST
|
Approximating Partition Functions in Constant Time
arXiv:1711.01655v2 [cs.LG] 20 Feb 2018
Vishesh Jain∗
Frederic Koehler†
Elchanan Mossel‡
Abstract
We study approximations of the partition function of dense graphical models. Partition
functions of graphical models play a fundamental role is statistical physics, in statistics and in
machine learning. Two of the main methods for approximating the partition function are Markov
Chain Monte Carlo and Variational Methods. An impressive body of work in mathematics,
physics and theoretical computer science provides conditions under which Markov Chain Monte
Carlo methods converge in polynomial time. These methods often lead to polynomial time
approximation algorithms for the partition function in cases where the underlying model exhibits
correlation decay. There are very few theoretical guarantees for the performance of variational
methods. One exception is recent results by Risteski (2016) who considered dense graphical
models and showed that using variational methods, it is possible to find an O(ǫn) additive
2
approximation to the log partition function in time nO(1/ǫ ) even in a regime where correlation
decay does not hold.
We show that under essentially the same conditions, an O(ǫn) additive approximation of the
log partition function can be found in constant time, independent of n. In particular, our results
cover dense Ising and Potts models as well as dense graphical models with k-wise interaction.
They also apply for low threshold rank models.
To the best of our knowledge, our results are the first to give a constant time approximation
to log partition functions and the first to use the algorithmic regularity lemma for estimating
partition functions. As an application of our results we derive a constant time algorithm for
approximating the magnetization of Ising and Potts model on dense graphs.
∗
Massachusetts Institute of Technology. Department of Mathematics. Email: [email protected]
Massachusetts Institute of Technology. Department of Mathematics. Email: [email protected]
‡
Massachusetts Institute of Technology. Department of Mathematics and IDSS. Supported by ONR grant N0001416-1-2227 and NSF CCF-1665252 and DMS-1737944. Email: [email protected]
†
1
Introduction
One of the major algorithmic tasks in the areas of Markov Chain Monte Carlo, in statistical inference
and in machine learning is approximating partition functions of graphical models. The partition
function or the approximate partition function are used in computing marginals and posteriors - two
of the basic inference tasks in graphical models. Moreover, there is an intimate connection between
computing the partition function and Markov Chain Monte Carlo methods. In particular, Jerrum
and Sinclair [20] showed that it is possible to approximate the partition function for “self-reducible”
models for which a rapidly mixing Markov chain exists. On the other hand, for such models, a
(1 + ε) approximation of the partition function results in a rapidly mixing chain. Some key results
in the theory of MCMC provide conditions for the existence of a rapidly mixing chains and therefore
allow for efficient approximations of the partition functions e.g. [12, 13, 14] and follow up work.
A different line of work is based on the fact that the measure associated to any graphical model
can be characterized by a variational problem. Thus, approximating the partition function can
be achieved by solving optimization problems (e.g. [4], [21], [22]). Until recently, there were very
few general guarantees for the performance of variational methods. One exception is results by
Risteski [18] who showed that using variational methods, it is possible to find an O(ε||J~||1 ) additive
2
approximation to the log partition function of dense graphical models in time n1/ε , even in an
interesting regime where correlation decay does not hold. Here, ||J~||1 denotes the total weight of all
interactions of the graph so that we are guaranteed an a priori upper bound on the log partition
~ 1 ).
function being O(n + ||J||
Following Risteski [18], we consider dense graphical models and show that it is possible to find
an O(ε||J~||1 ) additive approximation to the log partition function in (randomized) constant time.
Thus our main result stated informally is:
Theorem 1.1. For dense graphical models, it possible to compute the log partition function with
additive error O(ε||J~||1 ) in time that depends only on ε and the density parameter.
We note that the approximation guarantee in Theorem 1.1 is rather weak. For most inference
tasks, one is interested in good approximation of the partition function itself and our results (like
those of [18]) only provide an approximation of the partition function within a exp(±ǫn) multiplicative factor. Such an approximation is not useful for most inference and marginalization tasks. We
note however that
• It is easy to see that in constant time, it is impossible to get a sub-exponential approximation
of the partition function (see Theorem 2.12 for a formal statement).
• There are features of graphical models that are reflected by the log partition function. In
particular, we show in Theorem 2.11 how to utilize our results to obtain an ε-approximation
for the magnetization per-site for Ising models in constant time. Similar statements hold for
other dense graphical models as well. The exact statement of Theorem 2.11 provides the
magnetization for a model that has parameters close to the model we are interested in. We
show that this is necessary as in constant time, it is impossible to break symmetry between
phases.
We note that our results hold for very general Markov random fields. On the other hand, even
in the case of ferromagnetic Ising models, only polynomial time guarantees have been previously
provided (see [13] and recent work [3, 15]).
We also note that our results can be thought as generalizing classical results about the approximability of Max-Cut and similar problems on dense graphs – in fact, the Max-Cut problem is just
1
the special case of computing the log partition function of an antiferromagnetic Ising model (i.e. all
~ 1→∞
non-zero entries of J are negative) on a graph with equal edge weights in the limit as kJk
(so that the entropy term becomes negligible). Note that in this case, the log-partition function
is Ω(kJ~k1 ) (consider choosing the Xi according to independent Rademacher random variables) so
~ 1 -additive approximation actually gives a PTAS. The first PTAS for Max-Cut on dense
that an ǫkJk
2
graphs was given by Arora, Karger and Karpinski [2] but ran in time O(nÕ(1/ǫ ) ); this was improved
2
to O(2Õ(1/ǫ ) by Frieze and Kannan in [7] using their weak regularity lemma, and remains essentially
the fastest runtime known (our algorithm matches this runtime). More recent works [1, 17] have
shown that the sample complexity of the problem – the number of entries of the adjacency matrix
which need to be probed – is only O(1/ǫ4 ). In our case, determining the correct sample complexity
remains an interesting open problem (see Section 2.4). It is also interesting to note that in the
optimization case, algorithms based on regularity and algorithms based on convex programming
hierarchies (see [5, 23]) have the same dependence on ǫ (but not n) – our result and [18] also have
the same dependence on ǫ and ∆, showing that this phenomena extends to counting-type problems
2
as well, and suggesting that it will be difficult to beat a runtime of 2Õ(1/ǫ ∆) .
2
Overview of results
2.1
Notation and definitions
We will consider Ising models where the spins are ±1 valued. The interactions can be both ferromagnetic and anti-ferromagnetic and of varying strengths. For simplicity, we primarily restrict
ourselves to the case where there are no external fields, though it will be clear that our methods
extend to that case as well (Remark 2.8).
Definition 2.1. An Ising model is a probability distribution on the discrete cube {±1}n of the
form
X
1
1
Pr[X = x] = exp(
Ji,j xi xj ) = exp(xT Jx)
Z
Z
i,j
where the collection {Ji,j }i,j∈{1,...,n}
are the
Pnentries of an arbitrary real, symmetric matrix. The
P
exp(
normalizing constant Z =
n
i,j=1 Ji,j xi xj ) is called the partition function of the
x∈{±1}
Ising model.
We will often need to consider an m × n matrix M as an mn-dimensional vector (exactly how
we arrange the mn entries of M in a single vector will not matter). When we do so, we will denote
~ for clarity.
the resulting vector by M
Definition 2.2. For a vector v = (v1 , . . . , vn ) ∈ Rn and p ∈ [1, ∞), the Lp norm of v, denoted by
||v||p , is defined by ||v||p := [|v1 |p + · · · + |vn |p ]1/p . We also define ||v||∞ := maxi∈{1,...,n} |vi |.
~ , these vector norms will give rise to matrix norms which
By viewing a matrix M as a vector M
we will use in the sequel. The cases p = 1, 2, ∞ will be of particular importance to us.
Another matrix norm that we will need arises from viewing an m × n matrix as a linear operator
from Rn to Rm via matrix multiplication.
Definition 2.3. For an m × n matrix M , we define ||M ||∞7→1 := supx∈Rn ,||x||∞≤1 ||M x||1 .
Following [18], we will focus on ∆-dense Ising models.
Definition 2.4. An Ising model is ∆-dense if ∆||J~||∞ ≤
2
~ 1
||J||
.
n2
From a combinatorial perspective ∆-dense models are a generalization of dense graph models.
From a statistical physics perspective, they generalize complete graph models such as the CurieWeiss model and the SK model [19].
Our main results will provide algorithms which run in constant time. In order to provide such
guarantees on problems with unbounded input size, we will work under the usual assumptions on
the computational model for sub-linear algorithms (as in e.g. [1, 8, 11]). Thus, we can probe matrix
entry A(i, j) in O(1) time. Also note that by the standard Chernoff bounds, it follows that for any
set of vertices V for which we can test membership in O(1) time, we can also estimate |V |/n to
additive error ǫ w.h.p. in constant time using Õ(1/ǫ2 ) samples. This approximation will always
suffice for us; so, for the sake of clarity in exposition, we will henceforth ignore this technical detail
and just assume that we have access to |V |/n (as in e.g. [8]).
2.2
Main results
Our main theorem for ∆-dense Ising models is:
Theorem 2.5. There exists a universal constant C such that for any ǫ, δ > 0, and for any ∆-dense
2
2
Ising model J on n nodes with n ≥ 2C/ǫ ∆ there is an algorithm running in time 2Õ(1/ǫ ∆) log(1/δ)
which computes an ǫ(n + kJ~k1 ) additive approximation to log Z with probability at least 1 − δ.
Our methods extend in a straightforward manner to higher order Markov random fields on
general finite alphabets as well. We will review the definitions in Section 7.1. For simplicity, we
only state the result for the case of binary k-uniform Markov random fields.
Theorem 2.6. Fix k ≥ 3. There exists a universal constant C such that√ for any ǫ, δ > 0, and for any
C(ǫ ∆)2−2k there is an algorithm
∆-dense binary k-uniform
√ 2−2kMarkov random field J on n with n ≥ 2
running in time 2Õ(ǫ ∆)
log(1/δ) which computes an ǫ(kJ~k1 +n) additive approximation to log Z
with probability at least 1 − δ.
In the previous theorem, it is possible to improve the dependence on ǫ at the expense of introducing a factor of nk in the running time. We have:
Theorem 2.7. Fix k ≥ 3. There exists a universal constant C such that for any ǫ, δ > 0, and for
2
any ∆-dense binary k-uniform Markov random field J on n with n ≥ 2C/ǫ ∆ there is an algorithm
2
~ 1 +n) additive approximation to log Z
running in time 2Õ(1/ǫ ∆) log(1/δ)nk which computes an ǫ(kJk
with probability at least 1 − δ.
Remark 2.8. Although for simplicity, we stated Theorem 2.6 and Theorem 2.7 only for k-uniform
∆-dense Markov random fields, it is immediately seen that the results extend to general ∆-dense
Markov random fields of order K by simply applying Theorem 7.5 or Theorem 7.6 to each J (k) ,
k ∈ [K]. Again, see Section 7.1 for definitions. In particular, this directly handles dense Ising
models with external fields.
~ 1 . If kJ~k1 ≤ ηn, then
Remark 2.9. The error term in Theorem 2.5 has two terms n and kJk
T
| log exp(x Jx)| = O(ηn) and therefore, log Z = n(log 2 + O(η)). Similar statements hold for the
other theorems as well. Thus, in interesting applications of the theorem, we may assume WLOG
that the dominant term is kJ~k1 (by dominant here, we mean it is at least a small fraction of the
other term).
As in [18], we are also able to handle the case of Ising models whose interaction matrices have
low threshold ranks, although not in constant time. Again, for simplicity, we only record the result
for regular low threshold rank models. Definitions will be presented in Section 7.2. Following
Remark 2.9, we assume J ′ ≥ 1.
3
Theorem 2.10. There exists a universal constant C such that for any ǫ > 0, and for any regular
2
Ising model J on n nodes with n ≥ 2Ct/ǫ (here, t is the ǫ/2 sum-of-squares rank of the model)
2
~ 1
and J ′ ≥ 1, there is an algorithm running in time 2Õ(t/ǫ ) + poly(n, t, 1ǫ ) which computes an ǫkJk
additive approximation to log Z.
2.3
Approximating the expected total magnetization
P
P
Given an Ising model Pr[X = x] = Z1 exp{ i,j Ji,j xi xj + i hi xi }, one of the most fundamental
questions one can ask about it is how many spins are +1 and −1 in expectation. In the case of
the ferromagnetic Ising model (Ji,j ≥ 0 for all i, j), this corresponds to how strongly the system
P is
magnetized. Accordingly, we define the (expected total) magnetization of an Ising model by E[ i xi ],
where the expectation is with respect to the Ising distribution on {±1}n . Perhaps surprisingly, our
results easily show that for dense Ising models, one can obtain, in some sense, an approximation to
the expected magnetization in constant time. More precisely, we have:
Theorem 2.11. Consider a ∆-dense Ising Model
Pr[X = x] :=
X
X
1
exp{
Ji,j xi xj +
hi xi },
Z
i,j
i
Let
Pr[X = x] :=
h
∆||J~||∞ ≤
||J~||1
,
n2
∆||h||∞ ≤
||~h||1
n
X
X
1
exp{
Ji,j xi xj +
(hi + h)xi }
Z
i,j
i
and let mh denote the expected total magnetization for Prh . Then, for any ǫ > 0, we can find in
~ 1 ) additive approximation to mh , for some
constant time (depending only on ǫ and ∆), an ǫ(n + kJk
h with |h| < ε.
Proof. It is well known that one can express the moments of spin systems in termsPof derivatives
of the log partition function. In particular, for the Ising model Pr[X = x] = Z1 exp{ i,j Ji,j xi xj +
P
P
1
Pi hi xi }, consider the family of perturbed Ising models defined by Prh [X = x] = Zh exp{ i,j Ji,j xi xj +
i (hi + h)xi }. Then, for any h0 , we have
1 ∂
∂ log Zh
(h0 ) =
∂h
Zh0 ∂h
X
x∈{±1}n
=
X
x∈{±1}n
X
X
exp{
Ji,j xi xj +
(hi + h)xi }
i,j
i
!
X
X
X
1
exp{
Ji,j xi xj +
(hi + h0 )xi }
xi
Zh 0
i,j
i
i
X
= E h0 [
xi ]
i
where E h0 denotes the expectation with respect to the Ising distribution perturbed by h0 . In
Zh
particular, ∂ log
∂h (0) equals the expected total magnetization of the Ising model we started out
4
with. Moreover, since by Jensen’s inequality,
!
2
X
X
X
X
1
∂
∂ log Zh
(h0 ) =
|h=h0
exp{
Ji,j xi xj +
(hi + h0 )xi }
xi
∂h2
∂h
Z
h
0
i,j
i
i
x∈{±1}n
X
X
= E h0 [(
xi )2 ] − (E h0 [
xi ])2
i
i
≥0
we see that log Z is convex in h; in particular, for any h0 ∈ R and any δ > 0, we have
log Z(h0 ) − log Z(h0 − δ)
∂ log Z
log Z(h0 + δ) − log Z(h0 )
≤
(h0 ) ≤
δ
∂h
δ
Finally,
P
• By the mean
/RHS of the equation above are given by E h′ [ i xi ]
P value theorem, the LHS
and E h′′ [ i xi ], where h0 − δ < h′ < h0 < h′′ < h0 + δ.
• By setting ǫ = δ2 in Theorem 2.5 and Remark 2.8, we can evaluate the LHS and RHS up to
the desired error in constant time
We remark that:
• Unfortunately, it is impossible to approximate in constant time the magnetization at the
specified value of the external fields. For example, consider an Ising model on 4n vertices,
where Ji,j = C for some large C if i, j ≤ 2n and Ji,j = 0 otherwise. Let hi = 1 if i ∈ [2n+1, 3n]
and hi = −1 if i ∈ [3n + 1, 4n]. We set all the other hi to 0 except that we set hI = X, where
I is uniformly chosen in [1, 2n] and X is uniformly chosen in {0, ±1}. Note that this is a dense
Ising model as per our definition. Note also that on the nodes [1, 2n] we have the Ising model
on the complete graph with one (random) node having external field.
It is easy to see that if X = 0, the magnetization is 0. The fact that C is a large constant
implies that conditioning on one vertex taking the value ± results in a dramatic change in
magnetization on the vertices [1, 2n]. In particular, the magnetization is of order n if X = +1
and is of order −n if X = −1. It thus follows that we need Ω(n) queries in order to determine
the magnetization in this case. We note that this example
corresponds to a phase transition
P
– in particular, for every ǫ > 0, if h′ > ǫ then E h′ [ i xi ] = Ω(n) for all values of X and I.
See ([6]) for general references for the Ising model on the complete graph.
• The results for computing the magnetization readily extend to other models. For example, for
Potts models, we can compute for each color the expected number of nodes of that color (up to
error ǫkJ~k1 and for an ǫ close external field). Similarly, it is easy to check we can compute
Pother
statistics at this accuracy. For instance, for the Ising model, we can approximate E[ ai xi ]
if nηkak∞ ≤ kak1 for some η > 0.
2.4
Tightness of our results
As mentioned in the introduction, our results are qualitatively tight. We now make this precise by
proving the following information-theoretic lower bound on the accuracy of constant-time algorithms
for additively approximating log Z:
5
Theorem 2.12. Fix ǫ, ∆ ∈ (0, 1/4). For any (possibly randomized) algorithm A which probes at
1
entries of J before returning an estimate to log Z, there exists a ∆-dense input
most k := 8ǫ∆
instance J such that A makes error at least ǫkJ~k1 /4 with probability at least 1/4.
Proof. We prove the claim by reduction to a hypothesis testing problem. Specifically, we show that
′ with log-partition functions that are at
there exist two different dense Ising models JM and JM
′ k /2-far apart (where kJ
~M k > kJ~′ k) such that no algorithm which makes only k probes
least ǫkJ~M
1
M
can distinguish between the two with probability greater than 3/4. This immediately implies that
for any algorithm A to estimate log Z and for at least one of the two inputs, A must make error at
′ k /4 with probability at least 1/4 when given this input — otherwise we could use the
least ǫkJ~M
1
output of A to distinguish the two models with better than 3/4 probability by checking which log Z
the output is closer to.
Let n be an instance size to be taken sufficiently large, and consider two ∆-dense ferromagnetic
Ising models defined as follows:
• JM , for which the underlying graph is the complete graph on n vertices, ǫ∆ n2 many of the
n
are
edges are randomly selected to have weight M
∆ , and the remaining (1 − ǫ∆) 2 many edges
n M
n
M
~
~
assigned weight M . Note that since ||JM ||∞ = ∆ and ||JM ||1 = ǫ∆ 2 ∆ + (1 − ǫ∆) 2 M =
(1 + ǫ(1 − ∆))M n2 , this model is indeed ∆-dense for n sufficiently large.
′ , for which the underlying graph is the complete graph on n vertices and all edges have
• JM
weight M .
′ respectively. It is easily seen that
We denote the partition functions of these models by ZM and ZM
~M ||1
log Z ′
= (1 + ǫ(1 − ∆)) n2 ≥ (1 + 3ǫ/4) n2 , and that limM →∞ M M =
limM →∞ logMZM = limM →∞ ||JM
′ ~ ||
||JM
1
′ =
limM →∞ M
= n2 . Therefore, for M sufficiently large, log ZM ≥ (1 + ǫ/2)M n2 and log ZM
′ | ≥ (ǫ/2)M n = (ǫ/2)kJ
~′ M k1 .
M n2 , so | log ZM − log ZM
2
′
′ with probability greater
Now, we show that no algorithm A can distinguish between JM and JM
′ on our input J to algorithm
than 3/4 with only k probes. We fix a 50/50 split between JM and JM
′
′
A . Since the randomized algorithm A can be viewed as a mixture over deterministic algorithms,
there must exist a deterministic algorithm A′′ with success probability in distinguishing JM from
′ at least as large as A′ . Let (u , v ) be the first edge queried by A′′ , let (u , v ) be the next edge
JM
1 1
2 2
queried assuming Ju1 v1 = M , and define (u3 , v3 ), . . . , (uk , vk ) similarly (without loss of generality,
the algorithm uses all k of its available queries). Let E be the event that Ju1 ,v1 , . . . , Juk ,vk are all
′ we see Pr(E|J = J ′ ) ≥
equal to M . Event E always happens under JM , and when the input is JM
M
ǫ∆n(n−1)/2
≥ 1 − 2kǫ∆ for n > 4k. Thus the total variation distance between the observed
1 − k n(n−1)/2−k
′ is at most 2kǫ∆, so by the Neyman-Pearson Lemma we know A′′
distribution under JM and JM
1
fails with probability at least (1/2)(1 − 2kǫ∆). Therefore for k ≤ 4ǫ∆
we see that A′′ fails with
probability at least 1/4 which proves the result.
2
Note that on the input instances given in this theorem, our algorithm will use at most 2Õ(1/ǫ ∆)
~ 1 approximation for both models with probability ≥ 0.99. Finding the
queries to output an ǫkJk
optimal dependence of the running time on the parameters ǫ, ∆ remains an interesting open problem.
2.5
Outline of the techniques
To illustrate the main reason for the intractability of the (log) P
partition function of an Ising model,
we consider the ferromagnetic case where Pr[X = x] = Z1 exp{ i,j Ji,j xi xj }, Ji,j ≥ 0. In this case,
6
it is clear that a given magnetized state x, where almost all of the spins are either +1 or −1, is more
likely than a given unmagnetized state y, where the spins are almost evenly split between +1 and
n
, we see
−1. However, since the number of states with exactly αn spins equal to +1 is simply αn
that the total number of strongly magnetized states is exponentially smaller than the total number
of unmagnetized states. Therefore, while any given unmagnetized state is less likely than any given
magnetized state, it may very well be the case that the total probability of the system being in an
unmagnetized state is greater than that of the system being in a magnetized state.
In this paper, we present an approach to dealing with this tradeoff in dense Ising models
(Theorem 2.5) based on the algorithmic regularity lemma of Frieze and Kannan (Theorem 3.3).
Roughly speaking, this lemma allows us to efficiently partition the underlying (weighted) graph
into a small number of blocks in a manner such that “cut-like” quantities associated to the graph
approximately depend only on the numbers of edges between various blocks. In Lemma 4.1, we
show that the log partition function fits into this framework. A similar statement for MAX-CUT
(which may be viewed as the limiting case of our model when only antiferromagnetic interactions
are allowed and ||J~||∞ tends to infinity) was first obtained in [7].
The key point in the regularity lemma is that the number of blocks depends only on the desired
quality of approximation, and not of the size of the underlying graph. Together with the previous
statement, this shows (Lemma 4.2) that one can approximately rewrite the sum computing the
partition function in terms of only polynomially many nonnegative summands, as opposed to the
exponentially many nonnegative summands we started out with. This provides a way to resolve the
entropy-energy tradeoff – the log of the largest summand of this much smaller sum approximates
the log of the partition function well (Lemma 5.1).
The preceding discussion reduces the problem of estimating the log partition function to an
optimization problem, although a very different one from [18]. However, as stated, it is not a
problem we can solve in constant time. In Lemma 5.2, we show how to “granulate” the parameters
to reduce the problem to constant size, and then our Algorithm 1 solves this problem efficiently
via convex programming. The proofs of Theorem 2.6, Theorem 2.7 and Theorem 2.10 follow a
similar outline, with the application of Theorem 3.3 replaced by Theorem 7.5, Theorem 7.6 and
Theorem 7.9 respectively.
3
Preliminaries
In the case of the dense Ising model, our reduction to a problem that can be solved in constant time
will be based on the algorithmic weak regularity lemma of Frieze and Kannan [8]. Before stating it,
we introduce some terminology. Throughout this section, we will deal with m × n matrices whose
entries we will index by [m] × [n], where [k] = {1, . . . , k}.
Definition 3.1. Given S ⊆ [m], T ⊆ [n] and d ∈ R, we define the [m] × [n] Cut Matrix C =
CU T (S, T, d) by
(
d if (i, j) ∈ S × T
C(i, j) =
0 otherwise
Definition 3.2. A Cut Decomposition expresses a matrix J as
J = D (1) + · · · + D (s) + W
where D (i) = CU T (Ri , Ci , di ) for all t = 1, . . . , s. We say that such a cut decomposition has
width s, coefficient length (d21 + · · · + d2s )1/2 and error ||W ||∞7→1 .
7
We are now ready to state the algorithmic weak regularity lemma of Frieze and Kannan.
2
Theorem 3.3. [8] Let J be an arbitrary real matrix, and let ǫ, δ > 0. Then, in time 2Õ(1/ǫ ) /δ2 , we
can,
≥ 1 − δ, find a cut decomposition of width O(ǫ−2 ), coefficient length at most
√ with probability
√
~ 2 / mn and error at most 4ǫ√mn||J||
~ 2.
27||J||
4
Reducing the number of summands using weak regularity
The next lemma shows that for the purpose of additively approximating the
function,
Plog partition
(1)
(s)
′
T
(1)
we may as well work with the matrix D + · · · + D . We define Z := x exp(x (D + · · · +
D (s) )x).
Lemma 4.1. Let J be the matrix of interaction strengths of a ∆-dense Ising model. For ǫ, δ > 0,
let J = D (1) + · · · + D (s) + W be a cut decomposition of J as in Theorem 3.3.Then, for Z and Z ′
as above, we have | log Z − log Z ′ | ≤ √4ǫ∆ ||J~||1 .
Proof. Note that for any x ∈ {±1}n , we have
X
X
XX
X X
|
Ji,j xi xj −
(D (1) + · · · + D (s) )i,j xi xj | = |
(
Wi,j xj )xi | ≤ |
|
Wi,j xj |
i,j
i,j
i
j
i
j
~ 2
≤ ||W ||∞7→1 ≤ 4ǫn||J||
Therefore, for any x ∈ {±1}n , we have
X
X
exp(
Ji,j xi xj ) ∈ exp (D (1) + · · · + D (s) )i,j xi xj ) ± 4ǫn||J~||2 .
i,j
i,j
Taking first the sum of these inequalities over all x ∈ {±1}n and then the log, we get
X
log Z ∈ log
exp xT (D (1) + · · · + D (s) )x ± 4ǫn||J~||2
x∈{±1}n
Finally, noting that ||J~||22 =
P
i,j
|Ji,j |2 ≤ ||J~||∞ ||J~||1 ≤
~ 2
||J||
1
,
∆n2
and the proof follows.
Recall that for each i ∈ [s], D (i) = CU T (Ri , Ci , di ) is a cut matrix between vertex sets Ri and
Ci . We pass to a common refinement of all of the Ri ’s and Ci ’s. This gives at most r ≤ 22s disjoint
sets V1 , . . . , Vr such that that every one of the Ri ’s and Ci ’s is a union of these “atoms” Va . In terms
of these Va ’s, we can define another approximation Z ′′ to the partition function:
!
r
s
X
X
X
|Va |H(ya /|Va |)
r ′ (y)i c′ (y)i di +
Z ′′ :=
exp
y
a=1
i=1
P
where
y ranges over all elements of [|V1 |] × · · · × [|Vr |], r ′ (y)i := a : Va ⊆Ri (2ya − |Va |), c′ (y)i :=
P
a : Va ⊆Ci (2ya − |Va |) and H(p) := −p log p − (1 − p) log(1 − p) is the binary entropy function. Note
that r ′ (y)i represents the net spin in Ri , and similarly for c′ (y)i . The next lemma shows that for
approximating log Z ′ , it suffices to obtain an approximation to log Z ′′ . Combined with Lemma 4.1,
we therefore see that approximating log Z ′′ is sufficient for approximating log Z.
8
Lemma 4.2. For Z ′ and Z ′′ as above, we have | log Z ′ − log Z ′′ | = O(22s log n).
P
P
P
Proof. First,
observe that xT D (i)
x = di a∈Ri ,b∈Ci xa xb = di ( a∈Ri xa )( b∈Ci xb ). Thus, letting
P
P
ri′ (x) := a∈Ri xa and c′i (x) := b∈Ci xb , we see that that
Z′ =
X
x
exp(xT (D (1) + · · · + D (s) )x) =
X
exp
x
X
ri′ (x)c′i (x)di
i
!
.
Re-expressing the summation in terms of the possible values that ri′ (x) and c′i (x) can take, we get
that
!
X
X
X
′
′ ′
Z =
ri ci di
exp
1.
r ′ ,c′
x
r ′ (x)=r ′
′
c (x)=c′
i
where r ′ ranges over all elements of [|R1 |] × · · · × [|Rs |], r ′ := (r1′ , . . . , rs′ ), r ′ (x) := (r1′ (x), . . . , rs′ (x))
and similarly for c′ and c′ (x). Next, since
Y |Va |
X
X
1=
ya
n
a
x∈{±1}
r ′ (x)=r ′
c′ (x)=c′
it follows that Z ′ =
P
y
y∈[|V1 |]×···×[|Vr |]
r ′ (y)=r ′
c′ (y)=c′
|Va |
ya .
Q
P
exp ( i r ′ (y)i c′ (y)i di ) a
n
log
αn
Finally, we apply Stirling’s formula
= nH(α) + O(log n).
to get the desired result.
5
Approximating the reduced sum in constant time using convex
programming
So far, we have seen how the problem of estimating the log partition function can be reduced to
the problem of approximating log Z ′′ . The next simple lemma reduces the estimation of log Z ′′ to
an optimization problem.
P
P
Lemma 5.1. Let y ∗ := arg maxy exp ( i ri′ (y)c′i (y)di + a |Va |H(ya /|Va |)). Then,
!
X
X
′′
′ ∗ ′ ∗
∗
log Z −
ri (y )ci (y )di +
|Va |H(ya /|Va |) ≤ 22s log n.
i
a
2s
Proof. This follows immediately by noting that Z ′′ , which is a sum of at most n2 many nonnegative
2s
summands, is at least as large as its largest summand, and no larger than n2 times its largest
summand.
The following lemma shows that for estimating the contribution of the term corresponding to
some vector y, it suffices to know the components of y up to some constant precision. This reduces
the optimization problem to one of constant size.
9
Lemma 5.2. Let J denote the matrix of interaction strengths of a ∆-dense Ising model, and let
J = D 1 + · · · + D s + W√denote a cut decomposition as in Theorem 3.3. Then, given ri , ri′ , ci , c′i for
∆
i ∈ [s] and some γ ≤ 4ǫ√27s
such that ri , ci , ri′ , c′i ≤ n, |ri − ri′ | ≤ γn and |ci − c′i | ≤ γn for all i, we
P
~ 1 /2.
get that i di |ri′ c′i − ri ci | ≤ ǫ||J||
Proof. From Theorem 3.3, we know that for all i ∈ [s], |di | ≤
|c′i ||ri′ − ri | + |ri ||c′i − ci | ≤ 2γn2 , it follows that
X
i
di |ri′ c′i
− ri ci | ≤
X
i
√
~
27||
√J||1 .
n2 ∆
Since |ri′ c′i − ri ci | ≤
√
27||J~||1
εk|J~||1
√
di 2γn ≤
2γs ≤
2
∆
2
which completes the proof.
We now present Algorithm 1 which approximates the log of the largest summand in Z ′′ by
iterating over all possible proportions of up-spins in each block Ri , Ci . For each proportion, it
approximately solves the following convex program Cr,c (where γ is as before, λ is a parameter to
be specified, and va := |Va |/n):
X
max
va H(za /va )
a
s.t.
0 ≤ za
≤ va
X
za ≤ rt + γ
rt ≤
a:Va ⊂Rt
ct ≤
X
a:Va ⊂Ct
za ≤ ct + γ.
The infeasibility of this program means that our “guess” for the proportion of up-spins in the
various Ri , Ci is not combinatorially possible (recall that there may be significant overlap between
the various Ri , Ci . On the other hand, if this program is feasible, then we actually obtain an
approximate maximum entropy configuration with roughly the prescribed number of up-spins in
Ri , Ci , i ∈ [n]. Calculating the contribution to Z ′′ of such a configuration, and maximizing this
contribution over all iterations gives a good approximation1 to the maximum summand of Z ′′ , and
hence to Z ′′ . This is the content of the next lemma.
~ 1 + λn + O(22s log n) additive approximation to
Lemma 5.3. The √output of Algorithm 1 is an ǫ||J||
2s
4
27s2
log Z ′′ when n ≥ √∆ǫ .
Proof. By Lemma 5.1, it suffices to prove the claim with log Z ′′ replaced by
X
X
ri′ (y ∗ )c′i (y ∗ )di +
|Va |H(ya∗ /|Va |).
a
i
where y ∗ is defined as in the statement of Lemma 5.1. Let r, c correspond to the “cell” which y ∗ /n
falls into, i.e. the values of r, c such that for the corresponding r ′ , c′ , we have r ′ ≤ r ′ (y ∗ ),c′ ≤ c′ (y ∗ ),
A slightly better approximation to log Z ′′ can be found by estimating the discrete sum over each region by an
integral (using Lemma 5.4 to bound the error), and approximating this integral via the hit-and-run random walk of
[16]. However, this algorithm is slower than Algorithm 1, and it turns out the gain in accuracy is negligible for our
application.
1
10
Algorithm 1 Convex programming method to estimate log partition unction
Let S = {0, γ, 2γ, . . . , ⌊1/γ⌋γ}.
M ← 0.
for r ∈ S s , c ∈ S s do
Either find a λ-approximate solution to, or verify infeasibility of, Cr,c . (e.g. by ellipsoid
method [10])
if Cr,c is feasible then
Let Hr,c denote the λ-approximate maximum value of Cr,c .
c′i = 2nci − |Ci |.
Let ri′ = 2nri − |R
Pi | and
Let Mr,c = exp ( i ri′ c′i di + nHr,c ).
M ← max(M, Mr,c ).
end if
end for
return log M
|r ′ − r ′ (y ∗ )| ≤ γn and |c′ − c′ (y ∗ )| ≤ γn. Since Hr,c is the result of solving Cr,c up to λ additive
error and since y ∗ /n is in the feasible region of the convex program by definition, it follows that
X
X
va H(za /va ) − λn ≥
|Va |H(ya∗ /|Va |) − λn
nHr,c ≥ n sup
z
a
a
where z ranges over the feasible region for Cr,c . Combining this with Lemma 5.2 gives
X
X
ri′ (y ∗ )c′i (y ∗ )di +
|Va |H(ya∗ /|Va |) − log Mr,c
a
i
≤
≤
X
i
X
i
di (ri′ (y ∗ )i c′i (y ∗ ) − ri′ c′i ) +
′
∗
′
∗
di |r (y )c (y ) −
ri′ c′i |
X
a
|Va |H(ya∗ /|Va |) − nHr,c
+ λn ≤ ǫ||J~||1 /2 + λn.
On the other hand, we know that M = Mr,c for some r, c. Let z denote the point at which the λapproximate optimum to the convex program Cr,c was attained. Define ya to be ⌈za n⌉ if za n ≤ |Va |/2
and to be ⌊za n⌋ otherwise. Similar to the inequality above, we get
X
X
log Mr,c −
ri′ (y ∗ )c′i (y ∗ )di +
|Va |H(ya∗ /|Va |)
a
i
≤ log Mr,c −
=
X
a
X
i
ri′ (y)c′i (y)di −
X
a
|Va |H(ya /|Va |)
{|Va |H(za n/|Va |) − |Va |H(ya /|Va |} +
≤ 22s log 5 +
X
i
ri′ c′i di −
X
X
i
ri′ (zn)c′i (zn)di +
ri′ c′i di −
X
i
i
X
i
ri′ (zn)c′i (zn)di −
X
X
ǫ||J~||1
≤ 22s log 5 +
ri′ (zn)c′i (zn)di −
ri′ (y)c′i (y)di
+
2
i
ri′ (y)c′i (y)di
X
ri′ (y)c′i (y)di
i
i
where r ′ and c′ are as defined in Algorithm 1 in terms of r and c, in the second inequality, we
have used Lemma 5.4 and the triangle inequality, and in the last line, we have used Lemma 5.2.
11
√
~
J||
√ 1 by
To bound the last term, since |ya − za n| ≤ 1 for all a by definition, and since |di | ≤ n27||
2 ∆
Theorem 3.3, we get
X
X
X
X
|
ri′ (zn)c′i (zn)di −
r ′ (y)i c′ (y)i di | ≤
|di ||ri′ (y)||c′i (y) − c′i (zn)| +
|di ||c′i (zn)||ri′ (y) − ri′ (zn)|
i
i
i
≤
i
X
≤2
i
|di ||Ri ||
X
i
2| +
a : Va ⊆Ci
2s
|di |n2
X
≤
√
X
i
|di ||Ci ||
~ 1 22s+1 s
27||J||
√
n ∆
X
2|
a : Va ⊆Ci
~ 1
ǫ||J||
≤
2
provided that n ≥
√
4s √2722s
,
∆ǫ
which finishes the proof.
Lemma 5.4. Let n be a natural number, and let 0 ≤ z ≤ n. Let y = ⌈z⌉ if z ≤ n/2 and y = ⌊z⌋ if
z > n/2. Then, nH(y/n) ≥ nH(z/n) − log(5).
Proof. Observe that if both y, z ≤ n/2 or both y, z ≥ n/2 then this is immediate, since entropy is
concave and maximized at 1/2, thereby implying that H(y/n) > H(z/n). In the remaining case,
y/n and z/n are on opposite sides of 1/2. By the definition of y, we know that |y/n − z/n| ≤ 1/n,
and hence, by the Mean Value Theorem, n|H(y/n)−H(z/n)| = n|H ′ (x/n)(y/n−z/n)| ≤ |H ′ (x/n)|
for some x with |x/n − 1/2| ≤ 1/n. Thus, if n ≥ 3 we have x/n ∈ [1/6, 5/6] so that |H ′ (x/n)| =
| log((x/n)/(1 − x/n))| ≤ log 5. If n ≤ 2, then we already have nH(z/n) − nH(y/n) ≤ nH(z/n) ≤
2 log 2 < log 5.
6
Putting it all Together
We are now ready to prove Theorem 2.5.
Proof. (Theorem 2.5) First, we show how to get such an algorithm with constant success probability
(i.e. 7/8). We will then boost it to the desired probability of success using the standard median
technique. The constant success algorithm is just Algorithm 1 applied√to the weakly regular partition
generated by Theorem 3.3 with regularity parameter some ǫ′ = O(ǫ ∆).
For the correctness of this algorithm, note that if M is the output of Algorithm 1, then
| log M − log Z| ≤ | log Z − log Z ′ | + | log Z ′ − log Z ′′ | + | log M − log Z ′′ |
~ 1 /4 + O(22s log n) + ǫkJk
~ 1 /4 + λn + O(22s ).
≤ ǫkJk
where the first bound is by Lemma 4.1, the second bound is by Lemma 4.2, and the last bound is
by Lemma 5.3.
~ 1 /4 + λn ≤ ǫ(kJ~k1 + n)
ǫkJ~k1 /4 + O(22s log n) + ǫkJk
when we take λ = ǫ/2 and n sufficiently large. The runtime of computing a weakly regular partition
2
2
with success probability 7/8 in Theorem 3.3 is 2Õ(1/ǫ ∆) . The runtime of Algorithm 1 is 2Õ(1/ǫ ∆)
2
as there are 2Õ(1/ǫ ∆) many points in |S|, and solving each convex program Cr,c up to λ-error (as
2
2
well as deciding feasibility) can be done in time poly(2Õ(1/ǫ ∆) , log(1/λ), log(1/γ)) = 2Õ(1/ǫ ∆) by
standard guarantees for the ellipsoid method [10].
12
Finally, to boost our success probability to 1 − δ, we use the standard median trick, which is to
repeat the algorithm O(log(1/δ)) times independently and take the median of these outputs – that
this works is easily proven by the standard Chernoff bounds.
7
7.1
Extensions
General Markov random fields
For simplicity, we will restrict ourselves to Markov random fields over a binary alphabet. It is
readily seen that the same techniques extend to Markov random fields over general finite alphabets
as well.
Definition 7.1. A binary Markov random field of order K is a probability distribution on the
discrete cube {±1}n of the form
K
X
1
Pr[X = x] = exp(
Z
n
X
(k)
Ji1 ...ik xi1 . . . xik )
k=1 i1 ,...,ik =1
(k)
where for each k, {Ji1 ,...,ik }i1 ,...,ik ∈[n] is an arbitrary collection of real numbers. Note that the sum
P
(k)
(k) (x, . . . , x), where J (k) is a k-multilinear map or
i1 ,...,ik Ji1 ,...,ik xi1 . . . xik may also be viewed as J
(k)
k-tensor with J (k) (ei1 , . . . , eik ) = Ji1 ,...,ik . Here, ej denotes the j th standard basis vector of Rn and we
P
P
(k) (x, . . . , x))
view {±1}n as a subset of Rn . The normalizing constant Z = x∈{±1}n exp( K
k=1 J
is called the partition function of the Markov random field. Moreover, if all entries of all J (i) ,
i ∈ [K]\k vanish, then we say that the Markov random field is k-uniform.
Definition 7.2. A binary Markov random field of order K is ∆-dense if ∆||J~(k) ||∞ ≤
all k.
||J ~(k) ||1
nk
for
As indicated earlier, our approach for general Markov random fields mirrors our approach for the
Ising model. This is essentially because one has a similar algorithmic regularity lemma for general
k-dimensional matrices.
Definition 7.3. A k-dimensional matrix M on X1 × · · · × Xk is a map M : X1 × · · · × Xk → R.
Definition 7.4. For Si ⊆ Xi , i = 1, . . . , k and a real number d, we define the k-dimensional matrix
(
d e = (x1 , . . . , xk ) ∈ S1 × · · · × Sk
CU T (S1 , . . . , Sk ; d)(e) =
0 otherwise
Theorem 7.5. [8] Suppose J is an arbitrary k-dimensional matrix on X1 × · · · × Xk , where we
assume that k ≥ 3 is fixed. Let N := |X1 | × · · · × |Xk | and let ǫ, δ ∈ (0, 1]. Then, in time
2
O(kO(1) ǫ−O(log2 k) 2Õ(1/ǫ ) δ−2 ), we can, with probability at least 1 − δ, find a cut decomposition of
√
√
√ k
width O(ǫ2−2k ), coefficient length at most 27 ||J~||2 / N and error at most ǫ2k N ||J~||2 .
As in the k = 2 case, the error of the cut decomposition refers to ||J −(D (1) +· · ·+D (s) )||∞7→1 :=
sup||x||∞≤1 ||J − (D (1) + · · · + D (s) )||1 , where we view a k-dimensional matrix M as a linear operator
P
from (Rn )k−1 → Rn via the formula [M (x1 , . . . , xk−1 )]i := j1 ,...,jk−1 Mj1 ...jk−1 i x1,j1 . . . xk−1,jk−1 .
13
Given this theorem, the proof proceeds exactly as before – the same argument as Lemma 4.1
k ǫ||J||
~ 1
shows that we can replace J by D (1) + · · · + D (s) (possibly incurring an additive error of 2 √
),
∆
and once we have made this replacement, a similar optimization scheme as before lets us deduce
Theorem 2.6. The proof of Theorem 2.7 is identical, the only difference being that we can obtain a
cut decomposition of smaller width using the following theorem.
Theorem 7.6. [1] Let J be a an arbitrary k-dimensional matrix on X1 × · · · × Xk , where we assume
2
that k ≥ 3 is fixed. Let N := |X1 | × · · · × |Xk | and let ǫ > 0. Then, in time 2O(1/ǫ ) O(N ) and with
probability at √
least 0.99, we can find a cut
decomposition width at most 4/ǫ2 , coefficient length at
√
most 2||J~||2 /ǫ N and error at most ǫ2k N ||J~||2 .
7.2
Ising models with low threshold rank
As in [18], we can also consider Ising models of low threshold rank. For simplicity, we consider only
the regular case, noting that our results generalise with appropriate modifications to the non-regular
case as well.
P
′
Definition 7.7. A regular weighted Ising model is one for which
j |Ji,j | = J for all i. The
normalized adjacency matrix of a regular Ising model is the matrix with entries Ji,j /J ′ . We will
denote this matrix by JD .
Definition
P7.8. The δ-sum-of squares threshold rank of a regular Ising model is defined to be
tδ (JD ) := i:|λi |>δ λ2i , where λ1 , . . . , λn denote the eigenvalues of JD .
Note that since all eigenvalues of JD have absolute value at most 1, our criterion for low threshold rank is more general than defining tδ (JD ) to be the number of eigenvalues of JD with absolute
value strictly greater than δ.
Our definition
of tδ (JD ) differs slightly from the definition in [18]. There, the author defined
P
tδ (JD ) := i : |µi |>δ 1, where µ1 , . . . , µn denote the eigenvalues of the matrix abs(JD ) defined by
abs(JD )i,j = |(JD )i,j |. His definition and our definition are related via the standard linear algebra
P
~ D )k2 , it follows that Pn λ2 = Pn µ2 .
fact that ni=1 λ2i = ||J~D ||22 ; since ||J~D ||22 = ||abs(J
2
i=1 i
i=1 i
The same linear algebra fact shows that the low δ-sum-of-squares rank setting strictly generalises the ∆-dense setting: for any regular, ∆-dense Ising model and for any δ > 0, we have
||J1~||21
1
′
~
tδ (JD ) ≤ ||J~D ||22 = J1′2 ||J~||22 ≤ J1′2 ∆n
2 = ∆ , where we have used that ||J ||1 = nJ in the regular
case.
Our methods extend straightforwardly to the low sum-of-squares threshold rank setting due to
the following algorithmic regularity lemma of Gharan and Trevisan.
Theorem 7.9. [9] Let J be the matrix of interaction strengths of a regular Ising model, let ǫ > 0
and let t := tǫ/2 (JD ). Then, in time poly(n, t, 1ǫ ), we can find a cut decomposition of width at most
√
16t/ǫ2 , |di | ≤ t and error at most 4ǫ||J~||1 .
~ 1
||J||
Recall that in the case of the ∆-dense Ising model, the bounds
on the cut decomposition we
√
27
−2
actually used were that the width is O(ǫ ), each |di | is at most √∆n2 ||J~||1 , and the error is at most
~ 1 , which are quite similar since ||J~1 || = J ′ n. Hence, by exactly the same analysis as for the
√4ǫ ||J||
∆
∆-dense case, we can conclude Theorem 2.10. We omit further details.
14
8
Acknowledgements
We thank David Gamarnik for insightful comments, Andrej Risteski for helpful discussions related
to his work [18], and Yufei Zhao for introducing us to reference [1].
References
[1] Noga Alon, Fernandez de la Vega, Ravi Kannan, and Marek Karpinski. Random sampling and
approximation of MAX-CSPs. J. Comput. System Sci., 67:212–243, 2003.
[2] Sanjeev Arora, David Karger, and Marek Karpinski. Polynomial time approximation schemes
for dense instances of np-hard problems. In Proceedings of the Twenty-seventh Annual ACM
Symposium on Theory of Computing, STOC ’95, pages 284–293, New York, NY, USA, 1995.
ACM.
[3] Alexander Barvinok. Combinatorics and complexity of partition functions. Algorithms and
Combinatorics, 30, 2017.
[4] David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe. Variational inference: A review for
statisticians. arXiv preprint arXiv:1601.00670, 2016.
[5] Wenceslas Fernandez de la Vega and Claire Kenyon-Mathieu. Linear programming relaxations
of maxcut. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 53–61. Society for Industrial and Applied Mathematics, 2007.
[6] Richard S. Ellis. Entropy, large deviations, and statistical mechanics. Springer, 2007.
[7] Alan Frieze and Ravi Kannan. The regularity lemma and approximation schemes for dense
problems. Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer
Science, pages 12–20, 1996.
[8] Alan Frieze and Ravi Kannan. Quick approximation to matrices and applications. Combinatorica, 19(2):175–220, 1999.
[9] Shayan Oveis Gharan and Luca Trevisan. A new regularity lemma and faster approximation algorithms for low threshold rank graphs. In Approximation, Randomization, and Combinatorial
Optimization. Algorithms and Techniques, pages 303–316. Springer, 2013.
[10] Martin Grötschel, László Lovász, and Alexander Schrijver. Geometric algorithms and combinatorial optimization, volume 2. Springer Science & Business Media, 2012.
[11] Piotr Indyk. Sublinear time algorithms for metric space problems. In Proceedings of the thirtyfirst annual ACM symposium on Theory of computing, pages 428–434. ACM, 1999.
[12] M. Jerrum and A. Sinclair. Approximating the permanent. SIAM J. Comput., 18(6):1149–1178,
1989.
[13] M. Jerrum and A. Sinclair. Polynomial-time approximation algorithms for ising model (extended abstract). In Automata, Languages and Programming, pages 462–475, 1990.
[14] M. Jerrum, A. Sinclair, and E. Vigoda. A polynomial-time approximation algorithm for the
permanent of a matrix with non-negative entries. Journal of the ACM, 51(4):671-697, 2004.,
51(4):671–697, 2004.
15
[15] Jingcheng Liu, Alistair Sinclair, and Piyush Srivastava. The ising partition function: Zeros and
deterministic approximation. In FOCS, 2017.
[16] László Lovász and Santosh Vempala. Fast algorithms for logconcave functions: Sampling,
rounding, integration and optimization. In Foundations of Computer Science, 2006. FOCS’06.
47th Annual IEEE Symposium on, pages 57–68. IEEE, 2006.
[17] Claire Mathieu and Warren Schudy. Yet another algorithm for dense max cut: Go greedy. In
Proceedings of the Nineteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
’08, pages 176–182, Philadelphia, PA, USA, 2008. Society for Industrial and Applied Mathematics.
[18] Andrej Risteski. How to calculate partition functions using convex programming hierarchies:
provable bounds for variational methods. In COLT, 2016.
[19] David Sherrington and Scott Kirkpatrick. Solvable model of a spin-glass. Physical review
letters, 35(26):1792, 1975.
[20] Alistair Sinclair and Mark Jerrum. Approximate counting, uniform generation and rapidly
mixing markov chains. Information and Computation, 82(1):93–133, 1989.
[21] Martin J. Wainwright and Michael I. Jordan. Graphical models, exponential families, and
variational inference. Foundations and Trends in Machine Learning, 1(1-2):1–305, 2008.
[22] Jonathan S. Yedidia, William T. Freeman, and Yair Weiss. Understanding belief propagation
and its generalizations. 2003.
[23] Yuichi Yoshida and Yuan Zhou. Approximation schemes via sherali-adams hierarchy for dense
constraint satisfaction problems and assignment problems. In Proceedings of the 5th conference
on Innovations in theoretical computer science, pages 423–438. ACM, 2014.
16
This figure "frog.jpg" is available in "jpg" format from:
http://arxiv.org/ps/1711.01655v2
| 8cs.DS
|
Submitted to the Annals of Probability
arXiv: arXiv:1605.05819
EXPONENTIALLY CONCAVE FUNCTIONS AND A NEW
INFORMATION GEOMETRY
arXiv:1605.05819v5 [math.PR] 31 May 2017
By Soumik Pal∗,† and Ting-Kam Leonard Wong∗,‡
University of Washington† and University of Southern California‡
A function is exponentially concave if its exponential is concave.
We consider exponentially concave functions on the unit simplex.
In a previous paper we showed that gradient maps of exponentially
concave functions provide solutions to a Monge-Kantorovich optimal
transport problem and give a better gradient approximation than
those of ordinary concave functions. The approximation error, called
L-divergence, is different from the usual Bregman divergence. Using
tools of information geometry and optimal transport, we show that
L-divergence induces a new information geometry on the simplex
consisting of a Riemannian metric and a pair of dually coupled affine
connections which defines two kinds of geodesics. We show that the
induced geometry is dually projectively flat but not flat. Nevertheless,
we prove an analogue of the celebrated generalized Pythagorean theorem from classical information geometry. On the other hand, we consider displacement interpolation under a Lagrangian integral action
that is consistent with the optimal transport problem and show that
the action minimizing curves are dual geodesics. The Pythagorean
theorem is also shown to have an interesting application of determining the optimal trading frequency in stochastic portfolio theory.
1. Introduction.
Definition 1.1 (Exponential concavity). Let Ω ⊂ Rd be convex. We
say that a function ϕ : Ω → R ∪ {−∞} is exponentially concave if Φ = eϕ is
concave. (By convention we set e−∞ = 0.)
Throughout this paper we let Ω be the open unit simplex
(
)
n
X
n
(1.1)
∆n = p = (p1 , . . . , pn ) ∈ R : pi > 0,
pi = 1 ,
i=1
regarded as the collection of strictly positive probability distributions on
a set with n elements. This is due to the applications we have in mind,
∗
This research is partially supported by NSF grants DMS-1308340 and DMS-1612483.
MSC 2010 subject classifications: Primary 60E05; secondary 52A41
Keywords and phrases: information geometry, optimal transport, exponential concavity, L-divergence, generalized Pythagorean theorem, functionally generated portfolio,
stochastic portfolio theory
1
2
S. PAL AND T.-K. L. WONG
although many generalizations are possible. An interesting property of exponentially concave functions is that their gradient maps give a better firstorder approximation those of than ordinary concave functions. In [45], we
introduced the concept of L-divergence. Let ϕ : ∆n → R be a differentiable
exponentially concave function. For p, q ∈ ∆n , concavity of eϕ implies that
(1.2)
ϕ(p) + log (1 + ∇ϕ(p) · (q − p)) ≥ ϕ(q),
where ∇ϕ is the Euclidean gradient. Clearly this approximation is sharper
than the linear approximation of ϕ itself. The L-divergence of ϕ is the error
in this approximation:
(1.3)
T (q | p) := log (1 + ∇ϕ(p) · (q − p)) − (ϕ(q) − ϕ(p)) ≥ 0.
The extra concavity of exponentially concave functions have found several recent applications in analysis, probability and optimization. For example, in [22], the equivalence of entropic curvature-dimension conditions and
Bochner’s inequality on metric measure spaces is established using the notion of (K, N ) convexity. When K ≥ 0 and N < ∞, the negative of a (K, N )
convex function is exponentially concave. Better gradient approximation has
also led to better algorithms in optimization and machine learning such as
those in [29, 30, 35], although the authors tend to replace the logarithmic
term in (1.2) by a quadratic approximation.
One of our primary applications in mind is related to stochastic portfolio
theory. In [25, 23], the author considers the gradient map of an exponentially concave function as a map from ∆n to its closure ∆n . The following
restatement can be found in [45, Proposition 6]. Let ϕ be a differentiable
exponentially concave function on ∆n . For p ∈ ∆n , define π(p) ∈ Rn by
(1.4)
π i (p) = pi 1 + ∇ϕ(p) · (e(i) − p) , i = 1, . . . , n,
where e(i) = (0, . . . , 1, . . . , 0) is the ith standard basis of Rn . Then, it can
be shown that π : ∆n → ∆n . In keeping with standard definitions in the
subject we will call this map the portfolio map. In this vein also see articles
[6, 24, 26, 31, 42, 49, 53, 54].
The L-divergence of ϕ should be distinguished from the Bregman divergence of ϕ defined by
(1.5)
D (q | p) := ∇ϕ(p) · (q − p) − (ϕ(q) − ϕ(p)) .
Bregman divergence was introduced in [12] and is widely applied in statistics
and optimization. To see the difference consider two fundamental examples.
EXPONENTIALLY CONCAVE FUNCTIONS
3
For q, p ∈ ∆n , the Kullback-Leibler divergence (also known as relative entropy) is given by
n
X
qi
H (q | p) =
qi log .
pi
i=1
It can be shown that the relative
P entropy is the Bregman divergence of the
Shannon entropy H(p) = − ni=1 pi log pP
i . On the other hand, fix π ∈ ∆n
and consider the cross entropy ϕ(p) = ni=1 πi log pi . This is an exponentially concave function of p whose associated portfolio map (1.4) is constant:
π(p) ≡ π. The corresponding L-divergence is given by
!
n
n
X
X
qi
qi
(1.6)
T (q | p) = log
πi
−
.
πi log
pi
pi
i=1
i=1
This quantity is sometimes referred to as the free energy in statistical physics.
In finance it is called the diversification return in [9, 14, 21, 52], the excess
growth rate in [25, 27, 44], the rebalancing premium in [10], and the volatility
return in [28].
In [45] we introduced a Monge-Kantorovich optimal transport problem
on Rn−1 which can be solved using exponentially concave functions on the
unit simplex. The cost function is defined for (θ, φ) ∈ Rn−1 × Rn−1 by
!
n−1
X
exi
(1.7)
c(θ, φ) := ψ(θ − φ), where ψ(x) := log 1 +
i=1
is strictly convex on Rn−1 . We will recall the details of this transport problem
in Section 2.1 and its relationship to exponentially concave functions. It
suffices to say for now that, given a pair of Borel probability measures P
and Q on Rn−1 the optimal coupling of the two with respect to the above
cost can be expressed in terms of the portfolio map of an exponentially
concave function on the simplex. A related cost function appears in [41] in
the completely different context of finding polytopes with given geometric
data. It also appears to be related to the study of moment measures as
introduced in [15] (see page 3836 in particular).
1.1. Our contributions. In this paper we show that information geometry
provides an elegant geometric structure underlying exponential concavity, Ldivergence and the optimal transport problem. Here is a motivating question
which is the starting point of this work. Suppose ϕ is an exponentially
concave function on ∆n with its associated L-divergence T (· | ·). Can we
geometrically characterize triplets (p, q, r) ∈ (∆n )3 such that T (q | p) + T (r |
4
S. PAL AND T.-K. L. WONG
r
primal geodesic
q
dual geodesic
p
Fig 1. Generalized Pythagorean theorem for L-divergence.
q) ≤ T (r | p)? The answer to this question determines the optimal frequency
of rebalancing the portfolio generated by ϕ (see Section 5.4). Also see Section
3.3 for a transport interpretation of this inequality.
Using tools of information geometry, we show that exponentially concave
functions on ∆n and their L-divergences induce a new geometric structure on
the simplex ∆n regarded as a smooth manifold of probability distributions.
Let ϕ be an exponentially concave function on ∆n . We only require that ϕ
is smooth and the Euclidean Hessian of eϕ is strictly positive definite everywhere (see Assumption 2.5). The induced geometric structure consists of a
Riemannian metric g and a dual pair of affine connections ∇ and ∇∗ . These
connections define via parallel transports two kinds of geodesic curves on
∆n called primal and dual geodesics. Interestingly, the duality in this geometry goes hand in hand with the duality in the related Monge-Kantorovich
optimal transport problem, and this work is the first which exploits this
connection. We summarize our main results as follows. First we give the
answer of the motivating question.
Theorem 1.2 (Generalized Pythagorean theorem). Given (p, q, r) ∈
(∆n )3 , consider the dual geodesic joining q and p and the primal geodesic
joining q and r. Consider the Riemannian angle between the geodesics at q.
(See Proposition 4.4 which expresses the Riemannian metric g as a normalized Euclidean Hessian of Φ := eϕ .) Then the difference
(1.8)
T (q | p) + T (r | q) − T (r | p)
is positive, zero or negative depending on whether the angle is less than,
equal to, or greater than 90 degrees (see Figure 1).
We also prove other remarkable properties of the geodesics: (i) There exist
EXPONENTIALLY CONCAVE FUNCTIONS
5
explicit coordinate systems under which the primal and dual geodesics are
time changes of Euclidean straight lines (Theorem 5.1). In other words, the
new geometry is dually projectively flat. In particular, the primal geodesics
are Euclidean straight lines up to time reparameterization. Moreover, the
primal and dual connections have constant sectional curvature −1 with respect to the Riemannian metric, and thus satisfy an Einstein condition
(Corollary 4.10). The primal and dual geodesics can also be constructed
as time changes of Riemannian gradient flows for the functions T (r | ·) and
T (· | p) (Theorem 5.5). This is remarkable because while the geodesic equations depend only on the local properties of T (ξ | ξ 0 ) near ξ = ξ 0 , the gradient
flows are global as they involve the derivatives of T (r | ·) and T (· | p). Indeed, this relation is known only for limited families of divergence including
Bregman divergence and α-divergence [5].
As shown in [2, Chapter 1], the generalized Pythagorean theorem holds
for any Bregman divergence which induces a dually flat geometry. We will
prove that the resulting geometries from L-divergences are not flat for n ≥
3 (Theorem 4.9). While some extensions of the generalized Pythagorean
theorem hold in certain non-flat spaces (see for example [2, Theorem 4.5]),
they involve some extra terms. To the best of our knowledge Theorem 1.2
is the first exact Pythagorean theorem that holds in a geometry which is
not dually flat. The difference (1.8) can also be given an optimal transport
interpretation (Section 3.3).
(ii) We extend the static transport problem (1.7) to a time-dependent
transport problem with a corresponding convex Lagrangian action. In Theorem 6.2 we show that the action minimizing curves are the (reparametrized)
dual geodesics which, in addition, satisfy the intermediate time optimality
condition. This allows for a consistent displacement interpolation formulation between probability measures on the unit simplex. Previously, such
studies focused almost exclusively on the Wasserstein spaces corresponding
to the cost functions C(x, y) = d(x, y)α (here d is a metric on a Polish space
with suitable properties and α ≥ 1). Displacement interpolation and the
related concept of displacement convexity were introduced in [38] and in the
thesis [37]. These ideas have grown to be immensely important in classical
Wasserstein transport with fundamental implications in geometry, physics,
probability and PDE. See [51, Chapter 7] for a thorough discussion. Our
Lagrangian, although convex, is not superlinear, and, therefore, is not covered by the standard theory. However, we expect it to lead to many equally
remarkable properties.
These results suggest plenty of problems for further research. Generalizing
Theorem 1.2 to more than three points is of interest in stochastic portfo-
6
S. PAL AND T.-K. L. WONG
lio theory. Displacement interpolation has become an extremely important
topic in optimal transport theory. Extensions to Riemanninan manifolds,
done in [16], have led to new functional inequalities. In another vein, [34]
defines Ricci curvatures on metric measure spaces in terms of displacement
interpolation and displacement convexity. We expect that the displacement
interpolation in this paper will lead to a new Otto calculus ([51, Chapter 15]) and related PDEs (such as Hamilton-Jacobi equations). It appears
that Bregman divergence and L-divergence are only two of an entire family
of divergences with special properties and corresponding optimal transport
problems. For example, see [43] which extends the optimal transport problem
(1.7) via the cumulant generating function of a general probability distribution. We also believe that this new information geometry will be useful in
dynamic optimization problems where the objective function is multiplicative in time. Finally, it is naturally of interest to study exponential concavity
on general convex domains.
1.2. Related literature. We have mentioned the L-divergence and the
Bregman divergence. In general, a divergence on a set (usually a manifold
of probability distributions) is a non-negative function D (q | p) such that
D (q | p) = 0 if and only if p = q. Divergences are not metrics in general
since they may be asymmetric and may not satisfy the triangle inequality. Apart from Bregman divergence, many families of divergences (such as
α-divergence and f -divergence) have been applied in information theory,
statistics and other areas; see the survey [7] for a catalog of these divergences. Among these divergences, Bregman divergence plays a special role
because it induces a dually flat geometry on the underlying space. First
studied in the context of exponential families in statistical inference by [40],
it gave rise to information geometry – the geometric study of manifolds of
probability distributions. Furthermore, Bregman divergence enjoys properties such as the generalized Pythagorean theorem and projection theorem
which led to numerous applications. See [2, 3, 13, 32, 39] for introductions
to this beautiful theory. The related concept of dual affine connection is also
useful in affine differential geometry (see [17, 33, 47]). In [36] dually projectively flat manifolds are characterized in terms of the Bartlett tensors and
conformal flatness. Here we identify a new and important class of examples
and show that they have concrete applications.
This work is motivated by our study in mathematical finance. Recently
optimal transport has been applied to financial problems such as robust asset
pricing; see, for example, [1, 8, 18]. This line of work has a somewhat different
flavor than ours although they share the same goal: development of model-
EXPONENTIALLY CONCAVE FUNCTIONS
7
free mathematical finance. Portfolios generated by exponentially concave
functions generate profit due to fluctuations of a sequence in ∆n representing
the stock market. This idea is sometimes called volatility harvesting and
leads naturally to the transport problem (1.7), as shown in [45]. In this
philosophy, our work can be interpreted as developing a notion of modelfree volatility.
1.3. Outline of the paper. In the next section we recall the optimal transport problem formulated in [45] using the exponential coordinate system.
Its relation with functionally generated portfolio is also reviewed. In Section
3 we relate exponential concavity with c-concavity and give a transportmotivated definition of L-divergence. Here duality plays a crucial role. After
reviewing the basic concepts of information geometry, we derive in Section 4
the geometric structure induced by an exponentially concave function. The
properties of this new geometry are then studied in Section 5. In particular, we characterize the primal and dual geodesics and prove the generalized
Pythagorean theorem which has an interesting application in mathematical
finance. Finally, in Section 6 we apply the geometric structure to construct
a displacement interpolation for the associated optimal transport problem.
Some technical and computational details are gathered in the Appendix.
2. Optimal transport and portfolio maps. In this section we recall
the optimal transport problem in [45] using the exponential coordinate system. We also review the definition of functionally generated portfolio and
explain how it relates to the transport problem.
2.1. Exponential coordinate system. The exponential coordinate system
defines a global coordinate system on ∆n regarded as an (n − 1)-dimensional
smooth manifold [2, Section 2.2].
Definition 2.1 (Exponential coordinate system). The exponential coordinate θ = (θ1 , . . . , θn−1 ) ∈ Rn−1 of p ∈ ∆n is given by
(2.1)
θi = log
pi
,
pn
i = 1, . . . , n − 1.
We denote this map by θ : ∆n → Rn−1 . By convention we set θn ≡ 0. The
inverse transformation p := θ −1 is given by
pi = pi (θ) = eθi −ψ(θ) , 1 ≤ i ≤ n,
P
Pn θi
θi = log
where ψ(θ) = log 1 + n−1
as defined in (1.7).
i=1 e
i=1 e
(2.2)
8
S. PAL AND T.-K. L. WONG
The exponential coordinate system is the first of several coordinate systems we will introduce on the simplex. By changing coordinate systems, any
function on ∆n can be expressed as a function on Rn−1 and vice versa. Explicitly, a function ϕ on ∆n can be expressed in exponential coordinates by
θ 7→ ϕ(p(θ)). To simplify the notations, we simply write ϕ(p)
ϕ(θ) dePor
n
pending on the coordinate system used. For example,
P if ϕ(p) = i=1 πi log pi
is the cross entropy where π ∈ ∆n , then ϕ(θ) = n−1
i=1 πi θi − ψ(θ).
2.2. The transport problem. We refer the reader to the books [4, 51] for
introductions to optimal transport and its interplay with analysis, probability and geometry. Let X = Y = Rn−1 be equipped with the standard
Euclidean metric and topology. Let P and Q be Borel probability measures
on X and Y respectively. By a coupling of P and Q we mean a Borel probability measure R on X × Y whose marginals are P and Q respectively. Let
Π(P, Q) be the set of all couplings of P and Q. This set is always non-empty
as it contains the product measure P ⊗ Q.
Given P and Q we consider the Monge-Kantorovich optimal transport
problem with cost c defined by (1.7):
(2.3)
inf
R∈Π(P,Q)
E [c(θ, φ)] .
Here the expectation is taken under the probability measure R under which
the random element (θ, φ) has distribution R. If an optimal coupling takes
the form (θ, F (θ)) for some measurable map F : X → Y, we say that F is a
Monge transport map.
In general, we may consider the optimal transport problem (2.3) with c
replaced by a general cost function denoted by C(·, ·) and X , Y are general
Polish spaces. The classical example is where X = Y and C is a power of
the underlying metric: C(x, y) = d(x, y)α (especially α = 1 and 2). For these
costs rich and delicate theories have been developed on Euclidean spaces,
Riemannian manifolds and geodesic metric measure spaces. However, we
consider the cost function c defined by (1.7).
Remark 2.2.
The cost function
n−1
e
c(θ, φ) := log
1
1 X θi −φi
+
e
n n
i=1
!
−
n−1
X
i=1
1
(θi − φi )
n
differs from c(·, ·) in a linear term which plays no role in optimal transport.
Thus we may consider e
c instead. The advantage of e
c is that it is non-negative
and, by Jensen’s inequality, equals zero if and only if θ = φ. To be consistent
with the notations in [45] we will use the cost function c in this paper.
EXPONENTIALLY CONCAVE FUNCTIONS
9
Definition 2.3 (c-cyclical monotonicity). A non-empty subset A ⊂ X ×
Y is c-cyclical monotone if and only if it satisfies the following property. For
any finite collection {(θj , φj )}m
j=1 in A and any permutation σ of the set
{1, 2, . . . , m}, we have
(2.4)
n
X
c(θj , φj ) ≤
j=1
m
X
c(θj , φσ(j) ).
j=1
It is well known that c-cyclical monotonicity is, under mild technical conditions, a necessary and sufficient solution criteria of the general optimal
transport problem (see [4, Chapter 1]). In particular, a coupling R of (P, Q)
is optimal if and only if the support of R is c-cyclical monotone.
2.3. Functionally generated portfolio. At this point it is convenient to
introduce the concept of functionally generated portfolio. Although it is
possible to present the theory without reference to finance-motivated concepts, we stress that the portfolio map gives an additional structure to the
transport problem not found in other cases. Also, the main examples of the
theory as well as the key quantities (such as the induced Riemannian metric) are best expressed in terms of portfolios. Mathematically, the portfolio
can be regarded as a normalized gradient of ϕ. In Section 5.4 we apply our
information geometry to functionally generated portfolios.
Functionally generated portfolio was introduced in [23] and the following
refined definition is taken from [45].
Definition 2.4 (Functionally generated portfolio). By a portfolio map
we mean a function π : ∆n → ∆n . Let ϕ : ∆n → R be exponentially
concave. We say that a portfolio map π : ∆n → ∆n is generated by ϕ if for
any p, q ∈ ∆n we have
(2.5)
n
X
i=1
π i (p)
qi
≥ eϕ(q)−ϕ(p) .
pi
We call ϕ the log generating function of π and Φ := eϕ (which is positive and
concave) the generating function. It is known that ϕ is unique (for a given
π) up to an additive constant. If ϕ is differentiable, then π is necessarily
given by (1.4).
Throughout this paper we impose the following regularity conditions on
the exponentially concave function ϕ.
Assumption 2.5 (Regularity conditions).
10
S. PAL AND T.-K. L. WONG
(i) The function ϕ is smooth (i.e., infinitely differentiable) on ∆n .
(ii) The (Euclidean) Hessian of Φ = eϕ is strictly negative definite everywhere on ∆n . In particular, Φ is strictly concave. Moreover, it can be
shown that the function π defined by (1.4) maps ∆n into ∆n .
Let us discuss these conditions briefly. Differentiability is needed to define
differential geometric structures on ∆n in terms of the derivatives of the Ldivergence. Our theory requires the L-divergence T (· | ·) to be three times
continuously differentiable, and for convenience we simply assume that ϕ is
smooth. Strict concavity guarantees that the L-divergence is non-degenerate,
i.e., T (q | p) = 0 only if q = p, and strict positive definiteness of the Hessian
implies that the induced Riemannian metric is non-degenerate.
Henceforth we let ϕ : ∆n → R be an exponentially concave function
satisfying Assumption 2.5 and let π : ∆n → ∆n given by (1.4) be the
portfolio map generated by ϕ. The cost function c always refers to the one
defined in (1.7), and a general cost function is denoted by C. Using (1.4), it
can be shown that the L-divergence (1.3) of ϕ can be expressed in the form
!
n
X
qi
(2.6)
T (q | p) = log
π i (p)
− (ϕ(q) − ϕ(p)) .
pi
i=1
Now we give several examples of functionally generated portfolios and
their log generating functions. Many more examples can be found in [25,
Chapter 3]. In particular, the constant-weighted portfolios play a special
role and will be taken as the basic example of the theory.
Example 2.6 (Examples of functionally generated portfolios).
(i) (Market portfolio) The identity map π(p) ≡ p is generated by the
constant function ϕ(p) ≡ 0. (Here Assumption 2.5 does not hold.)
(ii) (Constant-weighted portfolio) The constant
Pn map π(p) ≡ π ∈ ∆n is
generated by the
cross-entropy
ϕ(p)
=
i=1 πi log pi . The special case
π = n1 , . . . , n1 is called the equal-weighted portfolio.
(iii) (Diversity-weighted portfolio) Let λ ∈ (0, 1) be a fixed parameter.
Consider the portfolio map π : ∆n → ∆n defined by
pλ
π i (p) = Pn i
λ
j=1 pj
,
i = 1, . . . , n.
It can be shown that the generating function is ϕ(p) =
1
λ
log
P
n
λ
j=1 pj
.
EXPONENTIALLY CONCAVE FUNCTIONS
11
(iv) (Convex combinations) It is known that the set of functionally generated portfolios is convex. Indeed, let π (1) be generated by ϕ(1) and
π (2) be generated by ϕ(2) . Then for λ ∈ (0, 1), the portfolio map
π = (1−λ)π (1) +λπ (2) is generated by ϕ = (1−λ)ϕ(1) +λϕ(2) . Its gen1−λ (2) λ
.
erating function Φ = eϕ is then the geometric mean Φ(1)
Φ
This fact was used in [53, 54] to formulate and study nonparametric
estimation of functionally generated portfolio.
The following result is taken from [45].
Proposition 2.7. For any portfolio map π : ∆n → ∆n the following
statements are equivalent.
(i) There exists an exponentially concave function ϕ on ∆n which generates π in the sense of (2.5).
(ii) The portfolio map is multiplicatively cyclical monotone (MCM) in the
following sense: for any sequence {µ(t)}m+1
t=0 in ∆n satisfying µ(m +
1) = µ(0), we have
!
m
n
Y
X
µi (t + 1)
π i (µ(t))
≥ 1.
µi (t)
t=0
i=1
(iii) Define a map θ 7→ φ by
(2.7)
φi = θi − log
π i (θ)
,
π n (θ)
i = 1, . . . , n − 1.
Here π is regarded as a function of the exponential coordinate. In
words, we define φ in such a way that the exponential coordinate of
π(θ) is θ − φ. Then the graph of this map is c-cyclical monotone.
Using this result, we showed in [45] how the optimal transport problem
(2.3) can be solved in terms of functionally generated portfolios. Here is a
simple but interesting explicit example which is a direct generalization of
the one-dimensional case treated in [45, Section 4].
Example 2.8 (Product of Gaussian distributions). In the transport
problem (2.3), let P be a product of one-dimensional Gaussian distributions:
n−1
O
P =
N (ai , σi2 ),
i=1
12
S. PAL AND T.-K. L. WONG
where ai ∈ R and σi > 0. Also let
Q=
n−1
O
N (bi , (1 − λ)σi2 ),
i=1
where bi ∈ R and 0 < λ < 1. Then the optimal transport map for the
measures P and Q is given by the map (2.7), where the portfolio map π is
the following variant of the diversity-weighted portfolio discussed in Example
2.6(iii):
n
λ
X
1
wi p
(2.8)
π i (p) = Pn i λ , ϕ(p) = log
wj pλj ,
λ
w
p
j
j=1
j
j=1
where the coefficients wi are chosen such that (1 − λ)ai − log wwni = bi for all
i.
3. Optimal transport and duality.
3.1. c-concavity and duality. Now we make use of the notion of c-concavity
in optimal transport theory. The definitions we use are standard and can be
found in [4, Chapter 1]. Again, c refers to our cost function (1.7). Also recall
that X = Rn−1 and Y = Rn−1 are the underlying spaces of the variables θ
and φ respectively. For f : X → R ∪ {±∞} we define its c-transform by
f ∗ (φ) := inf (c(θ, φ) − f (θ)) ,
θ∈X
φ ∈ Y.
Similarly, the c-transform of a function g : Y → R ∪ {±∞} is defined by
g ∗ (θ) := inf (c(θ, φ) − g(φ)) ,
φ∈Y
θ ∈ X.
We say that f : X → R∪{−∞} is c-concave if there exists g : Y → R∪{−∞}
such that f = g ∗ (similar for c-concave functions on Y). A function h (on X
or Y) is c-concave if and only if h∗∗ = h.
If f : X → R ∪ {−∞} is c-concave, its c-superdifferential is defined by
(3.1)
∂ c f := {(θ, φ) ∈ X × Y : f (θ) + f ∗ (φ) = c(θ, φ)} .
For θ ∈ X we define ∂ c f (θ) := {φ ∈ Y : (θ, φ) ∈ ∂ c f }. If this set is a singleton
{φ}, we call φ the c-supergradient of f at θ and write φ = ∇c f (θ). Similar
definitions hold for a c-concave function g on Y.
13
EXPONENTIALLY CONCAVE FUNCTIONS
Let f : X → R ∪ {−∞} be c-concave. By definition of f ∗ , we have
(3.2)
f (θ) + f ∗ (φ) ≤ c(θ, φ)
for every pair (θ, φ) ∈ X × Y, and equality holds in (3.2) if and only if
(θ, φ) ∈ ∂ c f . This is a generalized version of Fenchel’s identity (see [46,
Section 12]) and will be used frequently in this paper.
Our first lemma relates exponential concavity on ∆n with c-concavity on
X and on Y. Note that the cost function is asymmetric, and c-concavity on
X is equivalent to c-concavity on Y after a change of variable.
Lemma 3.1 (Exponential concavity and c-concavity).
R ∪ {−∞} the following statements are equivalent.
For ϕ : ∆n →
(i) ϕ is exponentially concave on ∆n .
(ii) The function f : X → R ∪ {−∞} defined by
f (θ) = ϕ(p(θ)) + ψ(θ)
is c-concave on X .
(iii) The function g : Y → R ∪ {−∞} defined by
g(φ) = ϕ(p(−φ)) + ψ(−φ),
where −φ is the exponential coordinate, is c-concave on Y.
Proof. We prove the implication (i) ⇒ (ii) and the others can be proved
similarly. Suppose (i) holds and consider the non-negative concave function
Φ = eϕ on ∆n . By [46, Theorem 10.3], we can extend Φ continuously up to
∆n , the closure of ∆n in Rn . We further extend Φ to the affine hull H of ∆n
in Rn by setting Φ(p) = −∞ for p ∈
/ ∆n . The extended function Φ is then a
closed concave function on H. By convex duality (see [46, Theorem 12.1]),
there exists a family C of affine functions on H such that
(3.3)
Φ(p) = inf `(p),
`∈C
p ∈ ∆n .
Since Φ is non-negative on ∆n , each ` ∈ C is non-negative on ∆n . Replacing
` by the sequence `k = `+ k1 , k ≥ 1, we may assume without loss of generality
that each ` ∈P
C is strictly positive on ∆n . We parameterize each ` ∈ C in the
form `(p) = ni=1 ai pi where a1 , . . . , an are positive constants. (Note that
an extra constant term is not required since p1 + · · · + pn = 1.) Writing
14
S. PAL AND T.-K. L. WONG
φi := − log aani , i = 1, . . . , n − 1 and switching to exponential coordinates, we
have
!
n
X
log `(p) = log
ai p i
i=1
= log 1 +
= log 1 +
n−1
X
i=1
n−1
X
ai p i
an pn
!
+ log pn + log an
!
eθi −φi
− ψ(θ) + log an
i=1
= c(θ − φ) − ψ(θ) + log an ,
It follows from (3.3) that
(3.4)
f (θ) = ϕ(θ) + ψ(θ) = inf (c(θ − φ) + log an ) .
`∈C
Define h : Y → R ∪ {−∞} by setting
(
)
n
X
ai
h(φ) = inf − log an : ∃ `(p) =
ai pi ∈ C s.t. φi = − log
∀i ,
an
i=1
where the infimum of the empty set is −∞. From (3.4), we have
f (θ) = inf (c(θ − φ) − h(φ)) = h∗ (θ)
φ∈Y
which shows that f is c-concave on X .
The following is the c-concave analogue of the classical Legendre transformation [46]. Its proof is standard but lengthy and will be given in the
Appendix.
Theorem 3.2 (c-Legendre transformation). Let ϕ be an exponentially
concave function ϕ satisfying Assumption 2.5, and let π, defined by (1.4), be
the portfolio map generated by ϕ. Given ϕ, consider the c-concave function
(3.5)
f (θ) := ϕ(θ) + ψ(θ)
defined on X = Rn−1 via the exponential coordinate system.
(i) The c-supergradient of f is given by (2.7), i.e.,
π i (θ)
c
(3.6)
∇ f (θ) = θi − log
,
π n (θ) 1≤i≤n−1
Moreover, the map ∇c f : X → Y is injective.
θ ∈ X.
EXPONENTIALLY CONCAVE FUNCTIONS
primal Euclidean
dual Euclidean
p∗
∆n
θ
15
∆n
p
p ◦ −Id
φ
−θ
∇c f
X =R
Y 0 ⊂ Rn−1
n−1
primal exponential
∇c f ∗
dual exponential
Fig 2. Coordinate systems on ∆n .
(ii) Let Y 0 ⊂ Y be the range of ∇c f . Then the c-supergradient of f ∗ is
given on Y 0 by
∇c f ∗ (φ) = (∇c f )−1 (φ),
φ ∈ Y 0.
In fact, the map ∇c f is a diffeomorphism from X to Y 0 whose inverse is
∇c f ∗ . Also, the function f ∗ is smooth on the open set Y 0 .
Although Y 0 is in general a strict subset of Y, by Theorem 3.2 the dual
variable φ = ∇c f (θ) defines a global coordinate system of the manifold
∆n . In Theorem 5.1 we will use another coordinate system on ∆n called the
dual Euclidean coordinate system. Thus we have four coordinate systems on
∆n : Euclidean, primal, dual and dual Euclidean (see Definition 3.3). In the
following we will frequently switch between coordinate systems to facilitate
computations. To avoid confusions let us state once for all the conventions
used. We let ϕ and f = ϕ + ψ be given.
Definition 3.3 (Coordinate systems).
by (1.1)) we call the identity map
p = (p1 , . . . , pn ),
For the unit simplex ∆n (defined
pi > 0,
n
X
pi = 1
i=1
the (primal) Euclidean coordinate system with range ∆n . We let
p1
pn−1
θ = θ(p) = log , . . . , log
pn
pn
be the primal (exponential) coordinate system with range X and
φ = φ(p) := ∇c f (θ)
16
S. PAL AND T.-K. L. WONG
be the dual (exponential) coordinate system with range Y 0 . The dual Euclidean coordinate system is defined by the composition
p∗ = p∗ (p) := p(−φ(p)).
See Figure 2 for an illustration. From now on p, p∗ , θ and φ always represent
the same point of ∆n . In particular, unless otherwise specified θ and φ are
dual to each other in the sense that φ = ∇c f (θ). By convention we let
θn = φn = 0 for any p ∈ ∆n .
Notation 3.4 (Switching coordinate systems). We identify the spaces
∆n , X and Y 0 using the coordinate systems in Definition 3.3. If h is a function
on any one of these spaces, we write h(p) = h(θ) = h(φ) = h(p∗ ) depending
on the coordinate system used.
We also record a useful fact. A formula analogous to the first statement
is derived in [49].
Lemma 3.5.
For 1 ≤ i ≤ n − 1, we have
∂
f (θ) = π i (θ), θ ∈ X ,
∂θi
∂ ∗
f (φ) = −π i (φ), φ ∈ Y 0 .
∂φi
Proof. The first statement is derived in the proof of Theorem 3.2. The
second statement can be proved by differentiating f ∗ (φ) = c(θ, φ) − f (θ)
(Fenchel’s identity).
3.2. c-divergence. By duality, we show that a pair of natural divergences
on ∆n can be defined for the c-concave functions f and f ∗ . Moreover, they
coincide with L-divergence. Clearly we can consider other cost functions
C other than c. When C is the squared Euclidean distance, the analogue
of Definition 3.6 below gives the classical Bregman divergence. This covers
both L-divergence and Bregman divergence under the same framework. To
the best of our knowledge these definitions, which depend crucially on the
interplay between transport and divergence, are new. We will use the triple
representation (p, θ, φ) for each point in ∆n .
Definition 3.6 (c-divergence). Consider the c-concave function f defined by (3.5) and its c-transform f ∗ .
17
EXPONENTIALLY CONCAVE FUNCTIONS
(i) The c-divergence of f is defined by
(3.7) D p | p0 = c(θ, φ0 ) − c(θ0 , φ0 ) − (f (θ) − f (θ0 )),
(ii) The c-divergence of f ∗ is defined by
(3.8) D∗ p | p0 = c(θ0 , φ) − c(θ0 , φ0 ) − (f ∗ (φ) − f ∗ (φ0 )),
p, p0 ∈ ∆n .
p, p0 ∈ ∆n .
From Fenchel’s identity (3.2) we see that D and D∗ are non-negative
and non-degnereate, i.e., they vanish only on the diagonal of ∆n × ∆n .
The following is a generalization of the self-dual expression of Bregman
divergence (see [2, Theorem 1.1]).
Proposition 3.7 (Self-dual expressions). We have
(3.9)
D p | p0 = c(θ, φ0 ) − f (θ) − f ∗ (φ0 ),
(3.10)
D∗ p | p0 = c(θ0 , φ) − f ∗ (φ) − f (θ0 ).
In particular, for p, p0 ∈ ∆n we have D (p | p0 ) = D∗ (p0 | p).
Proof. To prove (3.9), we use the Fenchel identity f (θ0 ) + f ∗ (φ0 ) =
c(θ0 , φ0 ). Starting from (3.7), we have
D p | p0 = c(θ, φ0 ) − c(θ0 , φ0 ) − (f (θ) − f (θ0 ))
= c(θ, φ0 ) − f (θ) − f ∗ (φ0 ).
The proof of (3.10) is similar.
Now we show that L-divergence is a c-divergence where c(θ, φ) = ψ(θ−φ).
Theorem 3.8 (L-divergence as c-divergence). The c-divergence of f is
the L-divergence of ϕ. Namely, for p, p0 ∈ ∆n we have
D p | p0 = T p | p0 .
Proof. Using the primal-dual relation (3.6), we have
!
n
π (θ 0 )
X
p
pn
θi −θi0 +log π i (θ0 )
0
n
ψ(θ − φ ) = log
e
= log π(p0 ) · 0 − log π n (p0 ) 0 .
p
pn
i=1
Next, by Fenchel’s identity (see (3.2)), we have
f ∗ (φ0 ) = ψ(θ0 − φ0 ) − f (θ0 ) = ψ(θ0 − φ0 ) − ϕ(θ0 ) − ψ(θ0 ).
18
S. PAL AND T.-K. L. WONG
Using these identities and (2.6), we compute
D p | p0 = ψ(θ − φ0 ) − f (θ) − f ∗ (φ0 )
p
0 pn
0
= log π(p ) · 0 − log π n (p ) 0
p
pn
0
0
− (ϕ(θ) + ψ(θ)) − ψ(θ − φ ) − ϕ(θ0 ) − ψ(θ0 )
p
0
= log π(p ) · 0 − ϕ(θ) − ϕ(θ0 )
p
0
=T p|p .
For computations it is convenient to express T (p | p0 ) solely in terms of
either the primal or dual coordinates. We omit the details of the computations.
Lemma 3.9 (Coordinate representations). For p, p0 ∈ ∆n we have
!
n
X
0 θ` −θ`0
0
π ` (θ )e
− f (θ) − f (θ0 ) ,
T p | p = log
`=1
T p | p0 = log
n
X
!
φ` −φ0`
π ` (φ)e
− f ∗ (φ0 ) − f ∗ (φ) .
`=1
3.3. Transport interpretation of the generalized Pythagorean theorem. Using Proposition 3.8 we give an interesting transport interpretation of the expression (1.8) in the generalized Pythagorean theorem (Theorem 1.2). Let
p, q, r ∈ ∆n be given. Let (θ(j) , φ(j) )1≤j≤3 be the primal and dual coordinates
of p, q and r respectively. By Proposition 2.7, the coupling (θ, φ = ∇f c (θ))
is c-cyclical monotone. Hence coupling θ(j) with φ(j) is optimal.
Consider two (suboptimal) perturbations of the optimal coupling:
(i) (Cyclical perturbation) Couple θ(1) with φ(3) , θ(2) with φ(1) , and θ(3)
with φ(2) . The associated cost is
c(θ(1) , φ(3) ) + c(θ(2) , φ(1) ) + c(θ(3) , φ(2) ).
(ii) (Transposition) Couple θ(1) with φ(3) , θ(3) with φ(1) , and keep the
coupling (θ(2) , φ(2) ). The associated cost is
c(θ(1) , φ(3) ) + c(θ(3) , φ(1) ) + c(θ(2) , φ(2) ).
EXPONENTIALLY CONCAVE FUNCTIONS
19
Now we ask which perturbation has lower cost. The difference (i) − (ii) is
c(θ(2) , φ(1) ) + c(θ(3) , φ(2) ) − c(θ(3) , φ(1) ) − c(θ(2) , φ(2) ).
By Proposition 3.8, this is nothing but the difference T (q | p) + T (r | q) −
T (r | p). Thus the generalized Pythagorean theorem gives an information
geometric characterization of the relative costs of the two perturbations.
3.4. Examples. We consider the portfolios in Example 2.6.
Example 3.10 (Constant-weighted portfolio). Let π ∈ ∆n be a constantweighted portfolio. Then ϕ is the cross entropy and we have
f (θ) = ϕ(θ) + ψ(θ) =
n−1
X
π i θi ,
i=1
which is an affine function on X . Its c-transform is also affine. Indeed, we
have
n−1
X
f ∗ (φ) =
πi (−φi ) + H(π),
i=1
Pn
where H(π) := − i=1 πi log πi is the Shannon entropy of π. For this reason
we say that the constant-weighted portfolios are self-dual.
The transport
π1
.
map in this case is given by a translation: φ = θ − log πn , . . . , log ππn−1
n
Its L-divergence (1.6) is given in primal coordinates (see Lemma 3.9) by
!
n
n
X
X
0
θ` −θ`
0
π` (θ` − θ`0 ),
π` e
−
T p | p = log
`=1
`=1
which is translation invariant. This property is equivalent to the following
numéraire invariance property [44, Lemma 3.2]: for any w1 , . . . , wn > 0, we
have T (q | p) = T (e
q | pe) under the mapping
wi p i
p 7→ pe =
.
w1 p1 + · · · + wn pn 1≤i≤n
In fact, it is not difficult to show that this property characterizes the constantweighted portfolios among L-divergences of exponentially concave functions.
Also see [44, Proposition 4.6] for a chain rule analogous to that of relative
entropy.
20
S. PAL AND T.-K. L. WONG
We have f (θ) = λ1 ψ(λθ).
Example 3.11 (Diversity-weighted portfolio).
Since
π i (θ)
= λθi ,
log
π n (θ)
the map θ 7→ φ is a scaling: φ = (1 − λ)θ. For the generalized diversityweighted portfolio in Example 2.8 the transport map θ 7→ φ is the composition of a scaling and a translation.
4. Geometric structure induced by L-divergence. In this section
we derive the geometric structure induced by a given L-divergence T (· | ·).
As always we impose the regularity conditions in Assumption 2.5. Using the
primal and dual coordinate systems (Definition 3.3), we compute explicitly
the Riemannian metric g, the primal connection ∇ (not to be confused with
the Euclidean gradient) and the dual connection ∇∗ . We call (g, ∇, ∇∗ ) the
induced geometric structure. An important fact in information geometry is
that the Levi-Civita connection ∇(0) is not necessarily the right one to use.
Nevertheless, by duality we always have ∇(0) = 12 (∇ + ∇∗ ).
4.1. Preliminaries. For differential geometric concepts such as Riemannian metric and affine connection we refer the reader to [2, Chapters 5] whose
notations are consistent with ours. For computational convenience we define
the geometric structure in terms of coordinate representations. The geometric structure is determined by the L-divergence T (· | ·) and is independent
of the choice of coordinates; for intrinsic formulations we refer the reader to
[13, Chapter 11]. The following definition (which makes sense for a general
divergence on a manifold) is taken from [2, Section 6.2].
Definition 4.1 (Induced geometric structure). Given a coordinate system ξ = (ξ1 , . . . , ξn−1 ) of ∆n , the coefficients of the geometric structure
(g, ∇, ∇∗ ) are given as follows.
(i) The Riemannian metric is given by
(4.1)
gij (ξ) = −
∂ ∂
T ξ | ξ0
0
∂ξi ∂ξj
i, j = 1, . . . , n − 1.
,
ξ=ξ 0
By Assumption 2.5 the matrix (gij (ξ)) is strictly positive definite. The
Riemannian inner product and length are denoted by h·, ·i and k · k
respectively.
(ii) The primal connection is given by
(4.2) Γijk (ξ) = −
∂ ∂ ∂
T ξ | ξ0
0
∂ξi ∂ξj ∂ξk
,
ξ=ξ 0
i, j, k = 1, . . . , n − 1.
21
EXPONENTIALLY CONCAVE FUNCTIONS
(iii) The dual connection is given by
(4.3) Γ∗ijk (ξ) = −
∂ ∂ ∂
T ξ | ξ0
0
0
∂ξk ∂ξi ∂ξj
,
i, j, k = 1, . . . , n − 1.
ξ=ξ 0
For a general divergence the above definitions were first introduced in
[19, 20]. If we define the dual divergence by T ∗ (p | p0 ) := T (p0 | p), the dual
connection of T is the primal connection of T ∗ . The primal and dual connections are dual to each other with respect to the Riemannian metric g (see [2,
Theorem 6.2]). While any divergence induces a geometric structure, it may
not enjoy nice properties. For the geometric structure induced by a Bregman divergence, it can be shown that the Riemann-Christoffel curvatures of
the primal and dual connections both vanish. Thus we say that the induced
geometry is dually flat [2, Chapter 1]. We will show that L-divergence gives
rise to a different geometry with many interesting properties.
4.2. Notations. We begin by clarifying the notations. Following our convention (see Notation 3.4), we write T (p | p0 ) = T (θ | θ0 ) = T (φ | φ0 ) depending on the coordinate system used. The primal and dual coordinate
representations have been computed in Lemma 3.9.
The Riemannian metric will be computed using both the primal and dual
coordinate systems. To be explicit about the coordinate system we use gij (θ)
∗ (φ) for its coefficients
to denote its coefficients in primal coordinates, and gij
in dual coordinates:
gij (θ) := −
∂2
T θ | θ0
0
∂θi ∂θj
,
∗
gij
(φ) := −
θ=θ0
∂2
T φ | φ0
0
∂φi ∂φj
.
φ=φ0
∗ (φ) are denoted by g ij (θ)
The inverses of the matrices (gij (θ)) and gij
and g ∗ij (φ) respectively.
The primal connection ∇ will be computed using the primal coordinate
system:
∂3
0
T
θ
|
θ
Γijk (θ) := −
∂θi ∂θj ∂θk0
,
Γkij (θ)
:=
θ=θ0
n−1
X
Γijm (θ)g mk (θ).
m=1
The dual connection ∇∗ will be computed using the dual coordinate system:
Γ∗ijk (φ) := −
∂3
T φ | φ0
0
0
∂φk ∂φi ∂φj
,
φ=φ0
Γ∗k
ij (φ) :=
n−1
X
m=1
Γ∗ijm (φ)g ∗mk (φ).
22
S. PAL AND T.-K. L. WONG
The following notations are useful. For 1 ≤ i ≤ n we define
0
(4.4)
π i (θ0 )eθi −θi
Πi (θ, θ ) := Pn
,
0 θ` −θ`0
`=1 π ` (θ )e
0
0
Π∗i (φ, φ0 )
π i (φ)eφi −φi
:= Pn
.
φ` −φ0`
`=1 π ` (φ)e
As always we adopt the convention θn = θn0 = φn = φ0n = 0. Note that
Πi (θ, θ0 ) involves the portfolio at θ0 (the second variable) while Π∗i (φ, φ0 ) involves the portfolio at φ (the first variable). The partial derivatives of Πi and
Π∗i are given in the next lemma and can be verified by direct differentiation.
We let δij be the Kronecker delta and δijk = δij δjk .
Lemma 4.2 (Derivatives of Πi and Π∗i ).
(i) For 1 ≤ i ≤ n and 1 ≤ j ≤ n − 1, we have
∂Πi (θ, θ0 )
= Πi (θ, θ0 ) δij − Πj (θ, θ0 ) ,
∂θj
∂Πi (θ, θ0 )
= −Πi (θ, θ0 ) δij − Πj (θ, θ0 )
0
∂θj
0
+ Πi (θ, θ )
!
n
X
1 ∂π i 0
1 ∂π ` 0
0
Π` (θ, θ )
(θ ) −
(θ ) .
π i (θ0 ) ∂θj0
π ` (θ0 ) ∂θj0
`=1
(ii) For 1 ≤ i ≤ n and 1 ≤ j ≤ n − 1, we have
∂Π∗i (φ, φ0 )
= Π∗i (φ, φ0 )(δij − Π∗j (φ, φ0 ))
∂φj
+
Π∗i (φ, φ0 )
!
n
X
1 ∂π `
1 ∂π i
0
∗
(φ) −
Π` (φ, φ )
(φ) ,
π i (φ) ∂φj
π ` (φ) ∂φj
`=1
∂Πi (φ, φ0 )
= −Π∗i (φ, φ0 )(δij − Π∗j (φ, φ0 )).
∂φ0j
We P
also note the following
easy fact which is used several times below.
P
P
SinceP ni=1 Πi (θ, θ0 ) ≡ ni=1 Π∗i (φ, φ0 ) ≡ 1, the partial derivatives of i Πi
and i Π∗i are all zero.
Lemma 4.3 (Derviatives of π(·)).
For 1 ≤ i ≤ n and 1 ≤ j ≤ n − 1, we
EXPONENTIALLY CONCAVE FUNCTIONS
23
have
(4.5)
∂π i
(θ) = π i (θ)(δij − π j (θ)) − π i (θ)
∂θj
!
n−1
X
∂φi
∂φ`
(θ) −
(θ) ,
π ` (θ)
∂θj
∂θj
∂π i
(φ) = −π i (φ)(δij − π j (φ)) + π i (φ)
∂φj
`=1
n−1
X
∂θi
(φ) −
∂φj
`=1
!
∂θ`
(φ) .
π ` (φ)
∂φj
Proof. We prove the second formula and the proof of the first is similar.
Using (2.7), we write
eθi −φi
π i (φ) = Pn
θ` −φ`
`=1 e
and regard θ as a function of φ (recall that θn = φn = 0). Then
θi −φi ∂θi (φ) − δ
n
e
X
ij
∂φ
∂π i
eθi −φi
∂θ`
j
θ` −φ`
Pn
−
e
(φ) =
(φ)
−
δ
P
`j
2
θ` −φ`
∂φj
∂φj
( n`=1 eθ` −φ` ) `=1
`=1 e
!
n−1
X
∂θ`
∂θi
π ` (φ)
(φ) −
(φ) .
= −π i (φ)(δij − π j (φ)) + π i (φ)
∂φj
∂φj
`=1
Note that the nth term of the sum is omitted because θn = 0.
Thanks to these formulas, computations in the primal and dual coordinates are very similar except for a change of sign. In the following we will
often give details for one coordinate system and leave the other one to the
reader.
∂φi
Last but not least, let ∂φ
(θ)
=
(θ)
be the Jacobian of the change of
∂θ
∂θj
∂θi
∂θ
coordinate map θ 7→ φ. Similarly, we let ∂φ
(φ) = ∂φ
(φ)
be the Jacobian
j
of the inverse map φ 7→ θ. The two Jacobians are inverses of each other, i.e.,
(4.6)
∂φ
∂θ
(θ) (φ) = I,
∂θ
∂φ
where I is the identity matrix. We denote the transpose of a matrix A by
A> .
4.3. Riemannian metric. For intuition, we first compute the Riemannian
inner product using Euclidean coordinates. We let Tp ∆n be the tangent space
at p.
24
S. PAL AND T.-K. L. WONG
Proposition 4.4. Let u, v ∈ Tp ∆n be represented in Euclidean coordinates, i.e., u = (u1 , . . . , un ) ∈ Rn and u1 + · · · + un = 0, and similarly for
v. Then
hu, vi = u> −Hess ϕ(p) − ∇ϕ(p)∇ϕ(p)> v
(4.7)
−1
>
=u
Hess Φ(p) v,
Φ(p)
where Φ = eϕ .
Proof. By [13, Proposition 11.3.1] we have kvk2 =
where
d2
T
dt2
(p + tv | p)
t=0
,
T (p + tv | p) = log (1 + t∇ϕ(p) · v) − (ϕ(p + tv) − ϕ(p)) .
Differentiating two times and setting t = 0 gives the first equality in (4.7)
when u = v, and polarizing gives the general case. The second equality
follows from the chain rule.
Theorem 4.5 (Riemannian metric).
(i) Under the primal coordinate system, the Riemannian metric is given
by
(4.8)
gij (θ) = π i (θ)(δij − π j (θ)) −
∂π i
(θ).
∂θj
Its inverse is given by
n−1
(4.9)
g ij (θ) =
1 ∂θi
1 X ∂θi
(φ).
(φ) +
π j (θ) ∂φj
π n (θ)
∂φ`
`=1
(ii) Under the dual coordinate system, the Riemannian metric is given by
(4.10)
∗
gij
(φ) = π i (φ)(δij − π j (φ)) +
∂π i
(φ).
∂φj
Its inverse is given by
n−1
(4.11)
g ∗ij (φ) =
1 ∂φi
1 X ∂φi
(θ) +
(θ).
π j (φ) ∂θj
π n (φ)
∂θ`
`=1
EXPONENTIALLY CONCAVE FUNCTIONS
25
Proof. (i) By Lemma 3.9 and Lemma 4.2, we compute
(4.12)
(4.13)
∂
T θ | θ0 = Πi (θ, θ0 ) − π i (θ),
∂θi
n
X
∂
1 ∂π ` 0
0
0
0
T
θ
|
θ
=
−Π
(θ,
θ
)
+
π
(θ
)
+
(θ ).
Π` (θ, θ0 )
i
i
0
∂θi
π ` (θ0 ) ∂θi0
`=1
Differentiating (4.12) with respect to θj0 , we have
(4.14)
∂2
T θ | θ0 = −Πi (θ, θ0 ) δij − Πj (θ, θ0 )
0
∂θi ∂θj
0
+ Πi (θ, θ )
!
n
X
1 ∂π i 0
1 ∂π ` 0
0
(θ ) −
Π` (θ, θ )
(θ ) .
π i (θ0 ) ∂θj0
π ` (θ0 ) ∂θj0
`=1
i
Setting θ = θ0 , we get gij (θ) = π i (θ)(δij − π j (θ)) − ∂π
∂θj (θ).
By Lemma 4.3, we have the alternative expression
!
n−1
X
∂φ`
∂φi
(θ) −
π ` (θ)
(θ) .
(4.15)
gij (θ) = π i (θ)
∂θj
∂θj
`=1
Expressing (4.15) in matrix form, we have
(4.16)
(gij (θ)) = diag(π(θ))(I − 1π > (θ))
∂φ
(θ),
∂θ
where here π(θ) = (π 1 (θ), . . . , π n−1 (θ))> , 1 = 1n−1 = (1, . . . , 1)> and I =
In−1 is the identity matrix.
To invert (4.16) we use the fact that
(I − 1π > (θ))−1 = I +
1π > (θ)
.
π n (θ)
This can be verified directly or seen as a special case of the ShermanMorrison formula. Using (4.6), we have
∂θ
1π > (θ)
1
ij
g (θ) =
(φ) I +
diag
.
∂φ
π n (θ)
π(θ)
Now (4.9) follows by expanding the matrix product.
26
S. PAL AND T.-K. L. WONG
(ii) The proofs of (4.10) and (4.11) follow the same lines. For later use we
record the following formulas:
n
(4.17)
X 1 ∂π `
∂
T φ | φ0 = Π∗i (φ, φ0 ) − π i (φ) +
(φ)Π∗` (φ, φ0 ),
∂φi
π ` (φ) ∂φi
(4.18)
∂
T φ | φ = −Π∗i (φ, φ0 ) + π i (φ0 ),
∂φ0i
`=1
0
(4.19)
∂2
T φ | φ0 = −Π∗j (φ, φ0 )(δij − Π∗i (φ, φ0 ))
0
∂φi ∂φj
− Π∗j (φ, φ0 )
!
n
X
1 ∂π j
1 ∂π `
(φ) −
(φ)Π∗` (φ, φ0 ) .
π j (φ) ∂φi
π ` (φ) ∂φi
`=1
Remark 4.6.
(4.20)
By Lemma 3.5 we have
∂π j
∂π i
∂2
(θ) =
f (θ) =
(θ).
∂θj
∂θi ∂θj
∂θi
Thus the right hand side of (4.8) is symmetric in i and j despite its ap∂π j
i
pearence. Similarly, we have ∂π
∂φj = ∂φi .
4.4. Primal and dual connections.
Theorem 4.7 (Primal and dual connections).
(i) Under the primal coordinate system, the coefficients of the primal connection ∇ is given by
(4.21)
(4.22)
Γijk (θ) = δij gik (θ) − π i (θ)gjk (θ) − π j (θ)gik (θ),
Γkij (θ) = δijk − δik π j (θ) − δjk π i (θ).
(ii) Under the dual coordinate system, the coefficients of the dual connection ∇∗ is given by
(4.23)
(4.24)
∗
∗
∗
Γ∗ijk (φ) = −δij gik
(φ) + π i (φ)gjk
(φ) + π j (φ)gik
(φ),
Γ∗k
ij (φ) = −δijk + δik π j (φ) + δjk π i (φ).
EXPONENTIALLY CONCAVE FUNCTIONS
27
Proof. We prove (ii) and leave (i) to the reader. By (4.19), we have
∂2
T φ | φ0 = −Π∗i (φ, φ0 )(δik − Π∗k (φ, φ0 ))
0
∂φk ∂φi
− Π∗i (φ, φ0 )
!
n
X
1 ∂π i
1 ∂π `
(φ) −
(φ)Π` (φ, φ0 ) .
π i (φ) ∂φk
π ` (φ) ∂φk
`=1
For notational convenience we momentarily suppress φ and φ0 in the computation (later we will do so without comment). Differentiating one more
time, we have
∂3
T φ | φ0
0
0
∂φk ∂φi ∂φj
= −δik
∗
∗
∂Π∗i
∗ ∂Πk
∗ ∂Πi
+
Π
+
Π
i
k
∂φ0j
∂φ0j
∂φ0j
n
∂Π∗i
−
∂φ0j
=
1 ∂π i X 1 ∂π ` ∗
−
Π
π i ∂φk
π ` ∂φk `
δik Π∗i (δij
−
Π∗j )
−
`=1
Π∗i Π∗k (δjk
−
!
Π∗j )
n
X
1 ∂π ` ∂Π∗`
+ Πi
π ` ∂φk ∂φ0j
−
`=1
∗ ∗
Πk Πi (δij
n
+
Π∗i (δij
−
Π∗j )
1 ∂π i X 1 ∂π ` ∗
−
Π
π i ∂φk
π ` ∂φk `
!
`=1
Evaluating at φ =
(4.25)
φ0
− Π∗j )
− Π∗i
n
X
1 ∂π ` ∗
Π (δ`j − Π∗j ).
π ` ∂φk `
`=1
and simplifying, we get
Γ∗ijk (φ) = −δijk π i − 2π i π j π k + δij π i π k + δjk π j π i + δki π k π j
∂π j
∂π i
∂π i
− δij
+ πj
+ πi
.
∂φk
∂φk
∂φk
∗
i
By (4.10), we have ∂π
∂φj = gij −π i (δij −π j ). Plugging this into (4.25) and sim∗ (φ) + π (φ)g ∗ (φ) + π (φ)g ∗ (φ). Finally,
∗
plifying, we have Γijk (φ) = −δij gik
j
i
ik
jk
we compute
Γ∗k
ij (φ) =
n−1
X
∗
∗
∗
−δij gim
(φ) + π i (φ)gjm
(φ) + π j (φ)gim
(φ) g ∗mk (φ)
m=0
= −δijk + δik π j (φ) + δjk π i (φ).
Remark 4.8. It is interesting to note that although the connections
are defined in terms of the third order derivatives of T (· | ·), the coefficients
Γkij (θ) and Γ∗k
ij (φ) are given in terms of the portfolio π which is a normalized
gradient of ϕ.
28
S. PAL AND T.-K. L. WONG
4.5. Curvatures. It is well known (see [2, Chapter 1]) that the induced
geometric structure of any Bregman divergence is dually flat. This is not
the case for the geometry of L-divergence whenever n ≥ 3 (when n = 2 the
simplex ∆2 is one-dimensional). To verify this we compute the RiemannChristoffel curvature tensors of the primal and dual connections. In this
(and only this) subsection we adopt the Einstein summation notation (see
[2, p.20]) to avoid writing a lot of summation signs.
The Riemann-Christoffel (RC) curvature tensor of a connection ∇ is defined for smooth vector fields X, Y and Z by
R(X, Y )Z = ∇X (∇Y Z) − ∇Y (∇X Z) − ∇[X,Y ] Z,
where [X, Y ] is the Lie bracket. Given
a coordinate
system ξ, its coefficients
P ` ∂
∂
∂
∂
`
are given by Rijk which satisfy R ∂ξi , ∂ξj ∂ξk = ` Rijk ∂ξ` . By [2, (5.66)],
we have
∂ `
∂ `
`
`
m
Rijk
=
Γjk −
Γ + Γ`im Γm
jk − Γjm Γik .
∂θi
∂θj ik
Theorem 4.9 (Primal and dual Riemann-Christoffel curvatures). Let
R and R∗ be the RC curvature tensors of the primal and dual connections
respectively.
(i) In primal coordinates, the coefficients of R are given by
(4.26)
`
Rijk
(θ) = δ`j gik (θ) − δ`i gjk (θ).
(ii) In dual coordinates, the coefficients of R∗ are given by
(4.27)
∗`
∗
∗
Rijk
(φ) = δ`j gik
(φ) − δ`i gjk
(φ).
In particular, for n ≥ 3 both R and R∗ are nonzero everywhere on ∆n .
Proof. We prove the statements for R. Using (4.21) and suppressing the
argument, we have
∂π j
∂ `
∂π k
Γ = −δ`j
− δ`k
,
∂θi jk
∂θi
∂θi
∂ `
∂π k
∂π i
Γ = −δ`i
− δ`k
.
∂θj ik
∂θj
∂θj
From (4.20) it follows that
(4.28)
∂ `
∂ `
∂π k
∂π k
Γjk −
Γik = −δ`j
+ δ`i
.
∂θi
∂θj
∂θi
∂θj
Next we compute (with some work)
(4.29)
`
m
Γ`im Γm
jk − Γjm Γik = −δ`i δjk π j + δ`j δik π i + δ`i π j π k − δ`j π i π k .
EXPONENTIALLY CONCAVE FUNCTIONS
29
Combining (4.28) and (4.29), we have
∂π k
∂π k
`
+ δ`i
− δ`i δjk π j + δ`j δik π i + δ`i π j π k − δ`j π i π k
Rijk
(θ) = −δ`j
∂θi
∂θj
∂π k
∂π k
= δ`j δik π i − π i π k −
− δ`i δjk π j − π j π k −
∂θi
∂θj
= δ`j gik − δ`i gjk .
To see that R does not vanish for n ≥ 3, suppose on the contrary that
` (θ) = δ g (θ) − δ g (θ) = 0 for all values of i, j, k, `.
R(θ) = 0. Then Rijk
`j ik
`i jk
Fix i and k. Letting ` = j, we have gik (θ) = δij gjk (θ). Next let j 6= i (here
we need dim ∆n = n − 1 ≥ 2). Then we get gik (θ) = 0. Since i and k are
arbitrary, we have g(θ) = 0 which is a contradiction.
We end this section by showing that the primal and dual connections have
constant sectional curvature −1. See [13, Definition 7.10.5] for the definition
of Ricci curvature.
Corollary 4.10 (Primal and dual sectional curvatures). The primal
and dual connections have constant sectional curvature −1 with respect to
g. In particular, the primal and dual Ricci curvatures satisfy the Einstein
condition
Ric = Ric∗ = −(n − 2)g.
Proof. The primal connection ∇ has constant sectional curvature k with
respect to g if and only if
R(X, Y )Z ≡ k(hY, ZiX − hX, ZiY )
for all smooth vector fields X, Y and Z (see [13, (9.7.41)]). For the primal
Riemann-Christoffel curvature tensor we have
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
`
R
,
= Rijk
=−
,
−
,
,
∂θi ∂θj ∂θk
∂θ`
∂θj ∂θk ∂θi
∂θi ∂θk ∂θj
which implies that the sectional curvature is k = −1. The claim for Ricci
curvature follows immediately by taking trace (see for example [48, (4.31)]).
The proof for the dual curvatures is the same.
5. Geodesics and generalized Pythagorean theorem. Armed with
the primal and dual connections we can formulate the primal and dual
geodesic equations. Their solutions are the primal and dual geodesics which
will be studied in this section. The highlight of this section is the generalized Pythagorean theorem (Theorem 1.2). Along the way we will prove some
remarkable properties of the geometric structure (g, ∇, ∇∗ ).
30
S. PAL AND T.-K. L. WONG
5.1. Primal and dual geodesics. Note that in Figure 1 the primal geodesic
is drawn as a straight line in ∆n . We now prove that this is indeed the case.
The same is true for the dual geodesic in dual Euclidean coordinates.
Let γ : [0, 1] → ∆n be a smooth curve. We denote time derivatives by
γ̇(t). Let θ(t) and φ(t) be the primal and dual coordinate representations of
γ. We say that γ is a primal geodesic if its satisfies
θ̈k (t) +
n−1
X
Γkij (θ(t))θ̇i (t)θ̇j (t) = 0,
k = 1, . . . , n − 1.
i,j=1
It is a dual geodesic if its satisfies
φ̈k (t) +
n−1
X
Γ∗k
ij (φ(t))φ̇i (t)φ̇j (t) = 0,
k = 1, . . . , n − 1.
i,j=1
By Theorem 4.7, the primal geodesic equation in primal coordinates is
(5.1)
θ̈k (t) + 2θ̇k (t)
n−1
X
k = 1, . . . , n − 1.
π ` (θ(t))θ̇` (t) = 0,
`=1
The dual geodesic equation in dual coordinates is
(5.2)
φ̈k (t) − 2φ̇k (t)
n−1
X
π ` (φ(t))φ̇` (t) = 0,
k = 1, . . . , n − 1.
`=1
Theorem 5.1 (Primal and dual geodesics).
(i) Let γ : [0, 1] → ∆n be a primal geodesic. Then the trace of γ in ∆n is
the Euclidean straight line in ∆n joining γ(0) and γ(1).
(ii) Let γ ∗ : [0, 1] → ∆n be a dual geodesic. For each t, let p∗ (t) be the
dual Euclidean coordinate of γ(t). Then the trace of p∗ in ∆n is the
Euclidean straight line in ∆n joining p∗ (0) and p∗ (1).
We will prove (i) and leave (ii) to the reader. The proof makes use of the
following lemmas.
Lemma 5.2. Let q, r ∈ ∆n and let θq and θr their primal coordinates
respectively. Consider the differential equation
(5.3)
00
0
2
h (t) − 2(h (t))
n−1
X
`=1
π ` (θ(t))
θ`r − θ`q
q
r
(1 − h(t))eθ` + h(t)eθ`
= 0,
31
EXPONENTIALLY CONCAVE FUNCTIONS
where θ(t) is defined by
θkq
θkr
(5.4)
θk (t) = log (1 − h(t))e + h(t)e
,
k = 1, . . . , n − 1,
Then there exists a unique solution h : [0, 1] → R satisfying h(0) = 0,
h0 (t) > 0 for t ∈ [0, 1], and h(1) = 1.
Proof. First we note that if h(t) satisfies (5.3), then for any c > 0 the
map t 7→ h(ct) also does (with domain scaled by 1/c). Also, if h0 (t0 ) = 0
then h(t) = h(t0 ) for all t ≥ t0 . Let h0 (t) be a maximal solution to (5.3)
defined on an interval [0, tmax ) with h0 (0) = 0 and h00 (0) > 0. Since the
equation can be solved in a neighborhood of t = 0, we have tmax > 0. By the
previous remark, h0 is strictly increasing on [0, tmax ). If h0 (t) hits 1 at some
t = t0 < tmax , the function h(t) = h0 (t/t0 ) is a solution with the desired
properties. In fact, we claim that
lim h0 (t) = sup h0 (t) = M,
(5.5)
t↑tmax
where M is defined by
M := inf s > 0 :
t<tmax
min (1 − s)e
1≤`≤n−1
θ`q
+ se
θ`r
q
=
min
1≤`≤n−1
eθ`
q
r
eθ` − eθ`
=0
!
1{θq >θr } + ∞ · 1{θq ≤θr }
`
`
`
`
> 1.
Thus as h(t) approaches M from below, at least one of the fractions in (5.3)
blows up to +∞. It follows that M 0 := supt<tmax h(t) ≤ M . Suppose on the
contrary that M 0 < M . Let h1 (t), t ∈ (−, ) be a solution to (5.3) satisfying
h1 (0) = M 0 and h01 (0) > 0. Note that h1 exists because the fractions in (5.3)
are finite and continuous near M 0 < M . Then the range of h1 contains an
open interval containing M 0 . Thus there exists t0 < tmax , c > 0 and t1 < 0
d
such that ct1 > −, h0 (t0 ) = h1 (ct1 ) and h00 (t0 ) = dt
h1 (ct) t=t1 . This allows
0
us to extend the range of h0 beyond M which contradicts the maximality
of M 0 . By the uniqueness theorem for ODE (note that (5.3) has smooth
coefficients), the solution h is unique.
Lemma 5.3. Let h be the solution in Lemma 5.2, and consider the curve
γ : [0, 1] → ∆n given in exponential coordinates by (5.4). Then γ is a primal
geodesic from q to r. Moreover, the trace of γ in ∆n is the Euclidean straight
line in ∆n joining q = γ(0) and r = γ(1).
32
S. PAL AND T.-K. L. WONG
Proof. That γ is a primal geodesic from q to r can be verified directly
by differentiating (5.4) and plugging into the primal geodesic equation (5.1).
We omit the computational details.
To see that the trace of γ is a Euclidean straight line in ∆n , consider its
Euclidean representation p(t) = (p1 (t), . . . , pn (t)). By (5.4) we have
eψ(θ(t)) =
1
q
r
= (1 − h(t))eψ(θ ) + h(t)eψ(θ ) .
pn (t)
Solving for h(t) gives
q
eψ(θ(t)) − eψ(θ )
h(t) = ψ(θr )
.
e
− eψ(θq )
(5.6)
Expressing (5.4) in Euclidean coordinates and using (5.6), we get after some
algebra that
r
θkq
pk (t) = e pn (t) + 1 − e
ψ(θq )
pn (t)
q
eθk − eθk
,
eψ(θr ) − eψ(θq )
k = 1, . . . , n − 1.
Hence there exists ak , bk such that
(5.7)
pk (t) = ak + bk pn (t),
k = 1, . . . , n − 1.
Together with the identity p1 (t) + · · · + pn (t) ≡ 1, (5.7) shows that γ is a
time change of the Euclidean straight line from q to r.
Proof of Theorem 5.1. (i) We have shown in Lemma 5.3 that for any
pair of points (q, r) in ∆n , there exists a primal geodesic from q to r which
is a time-changed Euclidean straight line. It remains to observe that the
geodesic is unique. Indeed, let γ be any primal geodesic. Then it solves
the primal geodesic equation (5.1). Consider the initial velocity γ 0 (0). Let
q = γ(0). By varying r ∈ ∆n as well as the initial speed, there exists a
primal geodesic γ
e inSthe form (5.4) from q to r such that γ̇(0) = γ
e˙ (0). (This
is because Tq ∆n = r∈∆n ,κ>0 κ(r − q).) By the uniqueness theorem of ODE
we have γ(·) = γ
e(·).
Using the dual coordinate system φ and dual Euclidean coordinate system
p∗ , (ii) can be proved in a similar way by considering the curve defined by
!
1
(5.8)
φk (t) = log
, k = 1, . . . , n − 1.
q
p
(1 − h(t))e−φk + h(t)e−φk
EXPONENTIALLY CONCAVE FUNCTIONS
33
A connection is projectively flat if there is a coordinate system under
which the geodesics are straight lines up to time reparameterizations. We
say that the geometric structure (g, ∇, ∇∗ ) is dually projectively flat if both
∇ and ∇∗ are projectively flat. In view of Theorem 5.1 we have the following
corollary.
Corollary 5.4. The manifold ∆n equipped with the geometric structure
(g, ∇, ∇∗ ) is dually projectively flat, but is not flat for n ≥ 3.
5.2. Gradient flows and inverse exponential maps. Motivated by the recent paper [5] we relate the primal and dual geodesics with gradient flows
under the L-divergence. Fix p, q, r ∈ ∆n . Consider the following gradient
flows starting at q:
(
γ̇(t) = −grad T (r | ·) (γ(t))
(5.9)
(primal flow)
γ(0) = q
and
(5.10)
(
γ̇ ∗ (t) = −grad T (· | p) (γ ∗ (t))
γ ∗ (0) = q
(dual flow)
Here grad denotes the Riemannian gradient with respect to the metric g.
We call (5.9) the primal flow and (5.10) the dual flow.
It can be verified easily that
d
T (r | γ(t)) = −kγ̇(t)k2
dt
and
d
T (γ ∗ (t) | p) = −kγ̇ ∗ (t)k2 .
dt
Since T (q | p) = 0 if and only if p = q, by standard ODE theory it can be
shown that the solutions γ(t) and γ ∗ (t) are defined for t ∈ [0, ∞) and
lim γ(t) = r,
t→∞
lim γ ∗ (t) = p.
t→∞
In other words, both gradient flows converge to the unique minimizers.
Theorem 5.5 (Gradient flows).
(i) The primal flow γ(t) is a time change of the primal geodesic from q to
r.
(ii) The the dual flow γ ∗ (t) is a time change of the dual geodesic from q
to p.
34
S. PAL AND T.-K. L. WONG
Recall the concept of exponential map. For q ∈ ∆n and v ∈ Tq ∆n , consider
the primal geodesic γ starting at q with initial velocity v. If γ is defined up to
time 1, we define expq (v) = γ(1). The dual exponential map exp∗ is defined
analogously. As a corollary of Theorems 5.1 and 5.5 we have the following
characterization of the primal and dual inverse exponential maps.
Corollary 5.6 (Inverse exponential maps). Let exp and exp∗ be the exponential maps with respect to the primal and dual connections respectively.
For p, q ∈ ∆n we have
(i) exp−1
q (p) ∝ −grad T (p | ·) (q).
−1
(ii) exp∗q
(p) ∝ −grad T (· | p) (q).
To prove Theorem 5.5 we begin by computing the Riemannian gradients
of T (r | ·) and T (· | p). The computation is somewhat tricky and will be
given in the Appendix. Recall the notations in (4.4).
Lemma 5.7 (Riemannian gradients).
Let p, q, r ∈ ∆n .
(i) Under the primal coordinate system, we have
(5.11)
grad T (r | ·) (q) =
n−1
X
π i (θr , θq ) π n (θr , θq )
−
+
π i (θq )
π n (θq )
i=1
n−1
X
1
= Pn
`=1 π ` (θ
q )e
θ`r −θ`q
∂
∂θiq
q
r
−eθi −θi + 1
i=1
∂
.
∂θiq
(ii) Under the dual coordinate system, we have
grad T (· | p) (q) =
n−1
X
i=1
(5.12)
π ∗i (φq , φp ) π ∗n (φq , φp )
−
π i (φq )
π n (φq )
= Pn
n−1
X
1
`=1 π ` (φ
q
p
q )eφ` −φ`
q
p
∂
∂φqi
eφi −φi − 1
i=1
∂
.
∂φqi
Proof of Theorem 5.5. We prove (i) and leave (ii) to the reader. Let
θ(t) be the primal representation of the primal flow starting at q. By Lemma
5.7, at any time t we have
r
θ̇k (t) ∝ eθk −θk (t) − 1 =
1
eθk (t)
r
eθk − eθk (t) ,
EXPONENTIALLY CONCAVE FUNCTIONS
35
where the constant of proportionality depends on θ(t) but is independent of
k. It follows that
d θk (t)
r
(5.13)
e
∝ eθk − eθk (t) , k = 1, . . . , n − 1.
dt
Comparing (5.13) and (5.4) we see that the primal flow is a time change of
the primal geodesic.
5.3. Generalized Pythagorean theorem. Having characterized the primal
and dual geodesics, we are ready to prove the generalized Pythagorean theorem. Our proof makes use of the Riemannian gradients given in Lemma
5.7. The reason is that these gradients appear to have the correct scaling
which is easier to handle, as can be seen in the proof (see (5.16)).
Proof of Theorem 1.2. Given p, q, r ∈ ∆n , consider the primal geodesic
from q to r and the dual geodesic from q to p. Let
u = −grad T (· | p) (q)
(5.14)
and v = −grad T (r | ·) (q).
By Theorem 5.5 u and v are proportional to the initial velocities of the two
geodesics. Thus, it suffices to prove that the sign of (1.8) is the same as that
of hu, vi. This claim will be established by the following two lemmas.
Lemma 5.8.
that of
The sign of T (q | p) + T (r | q) − T (r | p) is the same as
1−
(5.15)
n
X
Πk (q, p)Πk (r, q)
k=1
π k (q)
.
Proof. By Lemma 3.9, the sign of T (q | p) + T (r | q) − T (r | p) is the
same as that of
! n
n
n
X
X
X
q
q
p
p
r
r
π i (θp )eθi −θi
π j (θq )eθj −θj −
π i (θp )eθi −θi .
i=1
j=1
i=1
Rearranging, we have
n
X
−
r
q
q
p
π i (θp )(δij − π j (θq ))eθj −θj eθi −θi .
i,j=1
Since scaling does not change sign, we may consider instead the quantity
−
n
X
π i (θp )(δij − π j (θq ))
i,j=1
We get (5.15) by expanding.
Πj (θr , θq ) Πi (θq , θp )
.
π j (θq )
π i (θp )
36
S. PAL AND T.-K. L. WONG
Lemma 5.9.
Then
Consider the tangent vectors u and v defined by (5.14).
hu, vi = 1 −
(5.16)
n
X
Πk (θq , θp )Πk (θr , θq )
π k (θq )
k=1
.
Proof. For this computation we use the primal coordinate system. We
have
u = −grad T (· | p) (q) = −
n−1
X
g ik (q)
∂
q ∂
q T (· | p) (θ )
∂θk
∂θiq
g j` (q)
∂
q ∂
.
q T (r | ·) (θ )
∂θ`
∂θjq
i,k=1
v = −grad T (r | ·) (q) = −
n−1
X
j,`=1
Using the definition of the Riemannian inner product, we compute
n−1
X
hu, vi =
gij (q)
i,j=1
n−1
X
=
n−1
X
g ik (q)g j` (q)
k,`=1
g k` (θq )
k,`=1
∂
q ∂
T (r | ·) (θq )
q T (· | p) (θ )
∂θk
∂θ`q
∂
q ∂
T (r | ·) (θq ).
q T (· | p) (θ )
∂θk
∂θ`q
By (4.9), (4.18) and (A.6), we have
n−1
k`
g (q) =
X ∂θ`
1 ∂θ` q
1
(φ ) +
(φq ),
π k (q) ∂φk
π n (θq )
∂φα
α=1
∂
T (· | p) (θq ) = Πk (θq , θp ) − π k (θq ),
∂θkq
n−1
X ∂φβ
∂
q
(θq ) (π β (θq ) − Πβ (θr , θq )) .
q T (r | ·) (θ ) =
∂θ`
∂θ`q
β=1
Claim. We have
(5.17)
n−1
X
δk`
1
q p
q
q
r q
hu, vi =
(Πk (θ , θ ) − π k (θ ))(π ` (θ ) − Π` (θ , θ ))
+
.
π k (θq ) π n (θq )
k,`=1
EXPONENTIALLY CONCAVE FUNCTIONS
37
To see this, write
hu, vi =
n−1
X
(Πk (θq , θp ) − π k (θq ))(π β (θq ) − Πβ (θr , θq ))
k,β=1
·
n−1
X
`=1
!
n−1
1 X ∂φβ q ∂θ` q
1 ∂φβ q ∂θ` q
(φ ) .
(θ )
(φ ) +
(θ )
π k (q) ∂θ`q
∂θk
π n (q)
∂φα
∂θ`q
α=1
The last expression can be simplified using the identities
∂φβ q ∂θ` q
(θ )
(φ ) = δβk ,
∂θk
∂θ`q
∂φβ q ∂θ` q
(φ ) = δαβ ,
(θ )
∂φα
∂θ`q
and this gives the claim.
Finally, expanding and simplifying (5.17), we obtain the desired identity
(5.16).
With these two lemmas the proof is complete.
5.4. Application to mathematical finance. In this subsection we explain
how the new information geometry can be applied to finance. For financial
background and further details we refer the reader to [45]. Consider sequential investment in a stock market with n stocks. At time t = 0, 1, 2, . . .,
let µ(t) ∈ ∆n be the distribution of capital in the market. Explicitly, let
Xi (t) > 0 be the market capitalization of stock i. Then
(5.18)
µi (t) =
Xi (t)
X1 (t) + · · · + Xn (t)
is the market weight of stock i.
Let π : ∆n → ∆n be a portfolio map. It defines a self-financing investment
strategy such that if the current state of the market is µ(t) = p, we distribute
our capital among the stocks according to the vector π(p). For example, if
π(p) = (0.4, 0.5, 0.1), we invest 40% in stock 1, 50% in stock 2, and the
rest in stock 3. After each period, we perform the necessary trading such
that the proportions of capitals invested are rebalanced to π(µ(t + 1)). If
one chooses the identity map π(p) ≡ p, the resulting portfolio is called the
market portfolio which we denote by µ. The market portfolio plays the role
of benchmark.
Now let π be the portfolio map generated by an exponentially concave
function ϕ (Definition 2.4), and consider the value of the resulting portfolio
38
S. PAL AND T.-K. L. WONG
r
●
p
●
r●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●●
●
●●
●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●
●●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●●
●
●●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●
●●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●●●●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●
●●
●
●●
●●
●
●●
●
●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●● ●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
● ●●●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●●●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●●●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●●●
●●
●
●●
●●●
●
●
●●●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●●●
●●
●●●
●
●
●
●
●
●●
●
●
●
●
●● ●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●●
●●
●
●
●
●●
●●
●●●
●
●●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
● ●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●●●
●
●●
●
●
●
●●
●●●
●●
●
●
●
●
●●
●
●
●
●
●●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●●
●●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●●
●●●
●
●
●
●●
●
●
●
●
●
●
●
● ●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●●●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●●
●
●● ●
●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●●●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●●
●
●●
●
●● ●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●● ●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●●
●
●
●
●
●●●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●●
● ●●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●●
●●●
●●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●●●●●
●
●●
●
●●●
●●
●●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●●
●●●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●● ●
●
●●
●
●
● ●
●
●
●
●
●●●
●
●
●
●
●
●●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●●
●●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●
●
●
●●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●● ●●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●●●
●●
●●●●
●
●
●
●
●
●
●
●
●
●●
●
●●●●
●● ●
●
●●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●●
●
●
●●
●
●
●
●●
●●●●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●●
●●
●
●
●
●
●
●●
●●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●●
●
●
●
●●●
●
●
●
●
●●
●
●●
●
●
●
●
●● ●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●●●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●●
●
●
●●
●●●
●
●
●●
●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●●
●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●●●
●
●●
●
●
●
●●
●●●
●
●
●●
●
●●
●●
●
●
●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●●●●●●●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
● ●●●● ●●
●
●
●
●
● ●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
● ●●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●●●
●
●
●
●● ●●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●●●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●●●
●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●●●
●●
●
●
●
●
●
●
●
●●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●●
●
●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●●
●
● ●●
●●
●
●
●
●
●
●
●
●
●
●
● ●●●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●
●●●
●
●
●
●
●●
●
●●●
●●●
●
●
●
●●●●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●●
●
●
●
●
●●
●
●●
●●●
●
●
●
●●
●●
●
●
●
●
●●
●●
●●●
●
●
●
●
●
●●
●●
●●●
●
●
●●
●●
●
●●
●
●
●●●●
●
●
●●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●●
●
●
●
●●
●
●
●
●
●
●●
●●
●●
● ●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●● ●
●
●●●●●
●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●
●
●
●●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●●●
●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●●
●
●●
●●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●●●
●
●
●●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●●
●
●
●●●●●●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●●
●
●●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●●● ●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
● ●●
●●
●●
●
●●●●
●
●
●
●
●
●●
●
●
●
●●● ●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●●●
●●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●●●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
●
●
●●
●
●
●
●
●●●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●●●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●●
●
●●●
●●
●
●●●
●●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●●●●
●
●●
●
●
●
●
●
●
●●
●●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●●●
●
●
●
●
●●●
●
●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●
●
●●
●
●
●
●●●●
●●
●
●
●●
●
●
●
●
●
●●
●
●●●●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●●●
●●
●
●
●
●●
●●●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●●●●●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
● ●
●
●
●●
●●
●
●
●
●●
●
●
●● ●
●
● ●●
●●●●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●●● ●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●●●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●
●
●●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●●●
●
●
●
●
●●
●
●
●
●●
●●●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
● ●●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
● ●●
●
●● ●
●●●●
●
●
●
●●●●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●●●
●●●●●
●
●
●●
● ●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●●●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
● ●
●
●●
●
●●●
●
●
●
●●
●
●
●
●●
●
●●●●● ●●●
●
●●●
●
●●
●
●
●●
●
●●
●
●●●●
●
●●●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●●
●● ●
●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●●●
●
●
●
●●●
●
●● ●
●
●
●
●
●
●
●
●●●
●●
● ●●●
●
●
●
●
●●
●●
●
●●
●
●●●●
●
●
●
●
●●
●
●
●
●
●●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●●●
●
●●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●●
●●●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●●●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●●●●
●●
●
●
●●●
●●●
●
●●
●●
●
●
●● ●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●●
●
●
●
●●●
●
●
●●
●
●
●
●●●
●
●
●●●●●●
●●●
●
●
● ●●
●●
●●
●
●
●
●
●●●
●
●
●●
●●
●
●●
●●
●
●●
●
●
●
●
●●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
● ●●
●●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●●
●●
●●
●●●●
●
●
●●
●●
●
●●
●●● ●
●
●●●
●
●
●
●
● ●●
●●●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●● ●●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●
●●●●
●
●●●
●●
●
●
●●
●●
●
●●
●●
●
●
●●●●
●
●
●
●●
●●●
●
●
●●
●●
●●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●●●●
●
●
●● ●● ●●
●●
●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●●●
●
●
●
●●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●●
●
●●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●●●
●
●
●
●
●
●
●
●●●●●
●●
●●
●●●
●
●
●●●●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●●
●
●
●
●
●●
●
●
●●● ●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●
●
● ●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●●●●
●
●●●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
● ●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
● ●
●
●
●
●
●●
●
● ●●●
●●
●●●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●●
●
●
●
●
●●●
●●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●● ●
●●●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●●
●●
●
●●
●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●●
●●
●●●
●
●
●●
●●
●
●●
●●●
●
●
●●●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●●
●●●
●
●
●
●
●
●
●
●●●
●●
●
●●
●●●●
●
●
●
●
●
●
●●
●●●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●
●●
●●
●●●
●
●
●
●● ●
●●
●
●●
●●
●●
●
●
●
●
●●
●●●
●●
●
●
● ●
●
●●
●
●
●
● ●●
●●●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●●●
●
●
●●●
●●
●
●
●●●●
●●●●
●●
●●
●
●●
●
●
●●
●●●●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●● ●
●●
●
●●●
●
●
●
●
●●
●●
●
●
●
●
●●●●
●
●●
●
●
●
●
●
●
●●
●●●
●●
●
●
●
●●
●●
●
●
●● ●
●
●
●●
●
●●
●●
●
●
●●
●●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●
●●●
●
●●●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●●
●
●
●●
●●●
●●●
●●
●
●●
●
●
●●
● ●●●
●
●●
●●●
●
●●
●
●
●
●
●
●
●
●●● ●●
●●
●
●●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●●●
●
●
●
●
●
●
●●
●● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●●●
●●●●●
●●
●
●
●
●●
●
●
●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●●● ●
●
●
●●
●
●
●
●●●●
●
●
●●
●
●
●
●●●
● ●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●● ●
●● ●
●●●
●
●
●● ●
●●
●●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
● ●●●
●
●
●●
●●
●●●
●●
●
●
●●
●
●●
●
●
●
●●●
●●
●
●●
●
●
●●
●
●
●●
●●●
●
●
●
●
●
● ●●●
●
●
●
●●
●
● ●●
●
●● ●●●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●●
●
●
●●
● ●●●●
●
●
●
●
●
●
●●
●
●
●●●●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●●●
●●●●●
● ●●
●
●
●
●
●●●●
●●
●●●
●
●
●
●
●●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●
● ●●
●
●
●
●●●●
●
●●●
●●
●
●
● ●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●● ●●
●
●●
●
●
●●
●
●●●
●
●●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●● ●● ●
●●
●●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●
●●
●●●
●●●
●
●
●
●● ●
●
●●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
● ●●
●
●
●
●
●
●
●●
●
●●●●●
●
●
●
●
●
●●●
●●
●●●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●●
●
●●●
●
●
●
●●●● ●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●●●●● ●●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●●
●
●●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●●
●
●
●●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●●
●●
●●
●●
●
●●
● ●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●●
●
● ●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
●●
●
●●
●
● ●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●●
●
●
●
●●
●
●●●●
●
●●
●
●
●
●
●●●
●
●
●
●
●●●
●
●●
●
●●
●
●●
●
●●
●
●●●●
●
●●
●●
●●●●
●
●
●
●
●
●
●●
●●
●●
●●●
●●
● ●●
● ● ● ●
●
●●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●●
●
●
●●
●●
●
●
●
●●
●●●●●
●
●
●
●●
●●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●●
●●●
●●●●
●
●
● ●●
●
●
●
●
●
●●
●
●●
●●
●
●●●●●
●●
●
●
●
●
●
●●
●
●● ●●
●
●
●
●
●●●●
●●
●
●●
●●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●
●●
●
●
●
●● ●●●
●
●●●
●●
●
● ●●●●
●
●
●
●
●
● ●
●●●
●●●●
●
●●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●●
●● ●●
●
●
●●●●●
●
●
●
●●
● ●
●●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
● ●●
●●
●
●
●●
●●●
●●
●
●
●
●●
●
●
●
●
●
●
●●●●●
●
●
●
●
●●
●●
●
●●●
●
●●
●●
●
●●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●●
●●
●●
●●
●
●●
●
●● ● ●
●●
●●
●●●●●
●
●
●
●●
●●●
●●●
●
●
● ●
●●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●●
●
●
●●●
●●
●
●
●●●●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●●
●
● ●●●
●●
●●
●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●●●●●
●
●
●●
●●
●
●● ● ● ●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●●
●
●
●●●
●
●
●●
●
●
●
●
●●●●
●
●
●●●
●●
●●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●● ●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●● ●
●
●
●
●● ●
●
●
●
●
●
●
●●●●
●
●
●●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
● ●●
●●
●
●●
●
●
●
●
●●
●●
●●● ●●●●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●● ●
●
●
●
●● ●
●
●
●●
●
●●
●●
●
●
●
●
●●
●●●
●
●
●●●
●
●
●●
●
●
●
●●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●●
●●●
●
●
●●
●●●
●
●
●
●●●
●
● ●●
●●
●
●●●
●●
●
●
●
●
●●●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●●●●●
●
●
●
●
●
●
●
●
●●●●●
●
●
●●
●
●
●
●●●●
●
●
●
●
●
●●
●● ●
●
●
●●
●
●●
●
●●
●
●
●
●●●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
● ●●●●
●
●
●●●●●
●●
●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●●
●
●
●●●
●
●
●
●
●●
●●●
●●
●
●
●●
●●●●●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●●●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●●
● ●●
●
●
●●
●
●
●●●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●●
●
●●
●
●
●
● ●●
●
●
●●
●●
●
●●
●●
●●
●
●●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●●●●
●● ●●● ●
●
●
●●●●●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●●●● ●
●
●●
●●
●
●
●
●
●●
●●
●●
●●
●●
●
●●
●●
●●●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●●●●●
●
● ●●●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●●
●●●●●●
●
●
●●●
●
●
●
●
●●● ●
●
●
●
●
●●
●
●●●●●●
●● ●●
●
●
●
●●●
● ●●
●●
●
●●
●
●
●
●
●
●
●●●●
●●
●
●
●
●●
●●
●
●
●●
●
● ●●● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●● ●
●
●
●●●
●
●●●●
●
●
●
●
● ●●●
●
●
●●
●●
●
●●
●●●
●
●
●
●
●
●
●
●● ●●
● ●●
●
●
●
●●●
●
●
● ●●●●
●
●
●
●●●
●
●
●
●
●●●●
●●
●
●
●
●
●●●●
●
●
●
●● ●
●●●●●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●●●●
●
●
●● ●
●
●
●
●
●●
●
●●●
●
●
●
●● ●
●
●
●●
●●
●●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●
●
●●
●●
●●●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
● ●●
● ●●
●●
●●●●
●
●●
●
●●
● ●
●●
●
●
●●
●●
● ●●
●
●●
●●
●
●
●
●
●●
●
●●●
●●
●
●
● ●
●
●●
●●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
● ●●●●●
●
●
● ●
●
●●
●●
●
●
●
●
●
●
●
●●
● ●●
●
●●
●
●
●
●
●●●●
●●
● ●●●
●●●
●
●●
●
●
●
●
●
●●●
●
●●●●●
●●
●●●
●
●●
●
● ●●●
●●
●
●●
●●●
●●●
●
●
●●●●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●●
●●
●
●
●●●
●
●
●●● ●●
●
●
●
●●
●●●●
●
●
●●
●●●
●
●●
●●
●●
●●● ●
●
●
●
●
●
●●● ●
●●
●
●
●
●
●
●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
● ● ●●
●
●
● ●●●
●
●
●
●
●
●●
●
●●
●
●
●●●●●
●●
●●
●
● ●
● ●●●●
●●
●●●● ●
●
●
●●
●
●●
●
●
●
● ●
●●
● ●● ●
●●
●
●
● ● ●● ● ●●
●
●●
●
●
● ● ●
●●●●
●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●● ●
●
●
●
●●●
●
●
●
●
●
●●
●●
●
●
●●
●●
●
●
●
● ●
●●
●
●
●● ●
●●
●●●
●
●
●
●
●●
●
●
●
●
●●
●● ●●●
●
●●●
●
●●
●
●
●
●●●
●
● ●
●
●
●●●
●●
●
●
●
●●
●
●
●
●
●●
●●
● ●●●●●●
●
●●
●●
●●●
●
●
●
●●●
●
●● ●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●●●●
●●●
●
●
●
●●
●
●
●
●●●
●
●
●●
●●
●●●
●
●●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●●●● ●●●●
●
●
●●
●●●●●●
●●
●
●●●
●
●
●
●
●
●●
●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●● ●
●
●●
●
●
●
●
●
●●
●
●●●●
●
●
●●●
●●●
●
●
●
●●
●
● ●
●
●
●●
●
●
●
●
●
●
●
● ●●
●
●
●
●●●
●●●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●●●●
●
●●●
●
●
●●
●●
●●
●
●● ●
●
●●●
●
●
●
●
●
●
●
●● ●
●● ●●
●
●●
●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●●
●●
●●
● ●●
●●
●●
●●
●●
●
●●
●●●
●
●
●
●
●●
●
●●
● ●●
●
●
●●● ●●
●●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●●
●●●●
●
●
●
●
●● ●
●●●●●
●
●●
●
●
● ●
●
●
●●
●
●●
●●
●●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●●
●●
●●
●●
●●●●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
● ●●●●
●
●
●●
●
●●●
●
●
●
●
●
●●●
●
●●●●
●
●
●●
●
●●● ●
●
●
●
●
● ●● ●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●●●
●
●●
●
●●●
●
●
●
●
●●●
●●
●
● ●
●
●
●●
●●
●
●●●
●●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
● ●●
●
●
●
●●
●
●
●● ●
●
● ●●
●
●
●
●
●
●●
●●●●●
●●●
●●
●
●●
●●●●
●●
●
●
● ●●
●
●
●
●
●●●
●●
●
●
●
●●
●●
●
●
●
●
●●
●●●
●●●●
●
●●
●●
●
●●
●
●
●●
●
●●
●● ●
●●
●
●
●●●
●
●
● ●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●●
●
●
● ●●
●
●●
●
●
●
●●●●
●
●
●●
●
●●
●●
●●
●
●●
●
●●●
●● ●●
●
●●
●●●
●
●●
●●
●●
●
●
●
●
●
●
●●●
●● ●
●
●●
●●● ●
●●
●●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
● ●●
●●●
●
●
●
●
●●
●
●●●
●
● ● ●●●
●●
●
●
●
●●
●●
●●●
●
●
●
●
●
●
●●
●
●
●
● ●●
●
●
●
●
●● ●
●●
●●
●
●●●●●
●
●
●
●
● ●●
●●
●
●
●
●
●●
●●
●●●●
●
● ●●
●●
●
●
●●
●●●
●●●●
●
●
●●
●
●
●●
●
●●●●
●
●
●
●
●●
●●●
●●
●
●
●
●● ●●●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●● ●
●●●
●●
●
●●
●
●●● ●
●
●
●●
●●
●●●
●
●
●
●●●
●
●
●●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●
●●
●●
●
●
● ●●●
●●
●
●●
●●
●●●
●
● ●●●
●
●
●
●●
●●
●
●●●
●
● ●
●
●●●●
●
●
●
●●●
●
●●●●
●
●●
●●
●
●
●●●
●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●●
●
●
●●●●●
●●
●●
●●
●●●
●●
●
●●●●●
●
●
●●
●
●●
●●
●
●
●
●●●
●
●● ●●
●●●
●● ●
●●
●
●
●
●●
●●●
●
●
●
●●
●
●●
●●
●
●
● ● ●
●
●
●
●
●●
●
●●●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●● ●
●
●●
●
●
●
●
● ●
●●●
●●●●●
●
●
●●
●
●●
●
●
●
●●
●●
●●
●
● ●
●●●
●
●●●●
●●●
●●
●
●
●●
●
●
●
●●
●
●● ●
●●
●
●
●
●●
●
●●●
●
●
●
●
●●●●
●● ●
●●● ● ●●
● ●●●
●●
●
●
●
●
●
●
●●
●●
●
●●
●
● ●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
● ●●
●●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●●●
●●●●
●
●●●
●●●●
●
●●
●●
●●
●●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●
●●●
●
●
●●
●●
●
●
●
●
●
●
●●●●
●
● ●
●●
●●
●
●●
●
●●
●
●
●
●● ●
●●
● ●● ●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●●
●● ● ●
● ● ●●
●
●●
●
●●● ●
●
●
●
●
●
●●
●
●
●
●●
●●
●●● ●●
●●
●
●
●
●●
●
●● ●●
●
●
●
●● ●●
●●●
●●●
●
●
●●●● ●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●●●
●
●●
●
●●
●
●
●●
●●
●
●
●
●●● ●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●● ●●●
●
●
● ●●●
●
●●
● ●● ●●
●●
● ●
●
●
●●● ●●
●
●● ●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●●●●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●●
●
●● ●
●●●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●●
●●
●
●●
● ●●
●
●●
●●●
●●●●
●
●
●
●
●
●
●●
●
●●●●●●
●●
●●
●●●●●●●●●●
●
●●●
●
●
●●
●
●
●
●
●●
●
●●
●●●
●
●
●
●●●
●
●
●●
●●
●
●
●
● ●● ●
●● ●
●●
● ●●
●
●●●● ●
●●
●●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●● ●
●
●●
●
●●
●
●● ●
●● ●
● ●●●●●● ● ●
●
●
●●●●
●●
●● ●
●
●●
●●
●●●
●
●
● ●
●
●● ●
●
●
●
●●●
●
●
●● ●●
●
● ●●
●
●●
●
●
●
●
●●
●●
●
●●●●
●●●●
●
●
●
●
●●● ●
●●
●
●
●
●
●
●● ●
●●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●●●
● ●● ●● ●
●●
●
●
●
●●
●
●●
●●●
●●
●●●●●
●●●●●
●● ●●
●
● ●● ●●
● ●
●
●
●●
●●●●●●
●●●●●
● ●●●●
●●
●
●
●●
●
● ●
●●
●
●●
●
●
●●
●●●
●
● ●●
●●
●
●
●
●
●●
●
●●
●●●●
● ●●●
●
●
●●
●
●
●
●
● ●●
●
● ●
●
●
●
●●●●
●
●●
●●●●●●●
●●
●●
●●
●●●●●●
● ●●●● ●
●
●
●
●
●
●●
●●
●●
●●● ●●
●●●
●
● ●●
●●
●●●
●
●●
●
●
●
●
●●●
●
●●
●
●
●
●
● ●
●
●● ●●
●●
●●●
●●
●
●
●
●●
●
●
●● ●●●
●
●
●
●● ●
●●
●
●● ●
●
●●
●●●
●
●
●
●●
●
●
●
●
●
●
●●
●●●
●
●
● ●●● ●● ●
●
●
●
●●
●●
●●
●
●●●
●
●
●
●
●●●● ●
●●
●●
●
●
●
●●
●
● ●
●
●
●●
●
●●
●●
●
●●●
●●
●
●
●●
●●
●
●●●
●
●
●●
●●
●●●●
●
●
●
●●●
●●
●
●
●●
●
●
●
●
●● ●
● ●●●
● ●
● ●●●● ●
●●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●●
●
●●●●
●
● ●● ●
●
●●
●
●
●
●
●●●●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●●
●●●
●●●
●
●●
●
●●
●
●●
●● ●●
● ●● ●● ●●
●
●
●
●●
●
●
●●
●● ●
● ●●
●
●
●●
●●
●
●
●
●
● ●●
●
●●
●
●●●
●
●
●
●
●●
● ●●
●●
●
●●
●●
●
●
●● ●●
●● ●●●
●●●●
●●
●
●●
●
●
●
●●●●●●
●●
●●
●
●
● ●●● ●●
● ●●●
●
●●
●
●●
● ●●●
●
●●●●●●
●●●
●
●● ●
●●●
●●●
●
● ●●●
●
●
●
●
●●●●●●
●
●●
●●
●● ●
●
●●
●
●
●
●● ●
●
●●
●
●
●
●
●●● ● ●●
●
●●●
●●
●● ●●●●● ●
●
●
●
●
●●●
●
●●
●
●●
●●●●●●
●●
●
●
●●●●● ●
●
●●●
●
● ●●
●
●
●
●
●●
●
●
●●●
●●●
● ●●●● ●
●
●
●●
●
●●●
●●
●●
●
●
●
●
●
●
●
● ● ●●
●
●●
● ●●
●●
●●
●
●
●● ●●
●
●
●
●
●
●
●●●
●
●
●
●●●
●●
●
●
●
●
●
●
●●
●
●
●● ●
●● ●
●●●●●●
●
●●●
●
●●●
●●
●●
●●
●
●
●●
●
● ●●
●● ●
●●
●●
●
●●
● ●
●●●
●●●●●
●
●●
●
●
●●
●
● ●
●
●●
●●●●
●●
●
●●● ● ●●●
●●●
●
●●●
●●
●
●●
●●●●●●
●
●●
●
●
● ●●
●●
●●
●●
●●
●
●●●
●
●●● ●● ●
●
●
●
●●●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
● ●● ●
●●
●●
●●
●
●
●
●
●
● ●●
●
●
●●
●
●
●●
●●
●
●●
●●●●●
●●
●
●● ●
●● ● ●
●●
● ●●
●●
●
●●
●
●●
● ●●●●
●
●
●
●●●●●
●●
●
● ●●
●●●
●
●●●● ●●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●
●●
●● ●
●●
●●●●
●
●●
●
●
●●
●●●●●
●●●
●
●● ●●
●
●
●● ●●
●
●●
●
● ●●
●
●●
●●
●● ●
●
●●
●●●
●
●● ●
●
● ●● ●
●●
●
●●
●
●●●
●●
●●
●● ●●
●
●
●
●
●●
●
●●●
●●
●●
●●
●
●●●
●●
●●●●●
●●●
● ●
● ●●
● ●●●
● ●●● ●
●●
●●●
●
●
●●
●●●●
●● ●
●
●●● ●●
●
●
●
●
●
●●
●
● ●●
● ●
●●●
●
●
●
●
●
●
●
●●●●
●●●
●●
●
●●
●
●
●●
● ●
● ●●
●●
●●●
●
●
●
●●
●● ● ● ●
●
●●
●
●●
●
● ●●
●
●
●●●
● ●
●
●●●●
●●
●
●●
●● ●
●●●
●●●
●●
●● ●
●
●
●●
●●
●
●
●●●● ● ●●
●
●
●
●
●
●
●
●
●●●
●
●
● ●●
●●
●●●
●● ●
●
● ●
●
●
●● ●●●
●
●●●●
●
●
● ●● ●●●●
●
●●●
●●●
●●●
●
●
●●
●●
●●● ●
●● ●
●
●●
●
●
●
●
● ●
●
●●●●●
●●
●
● ●●
● ●
●●●
●
●●●●●
●
●
●
●●●●●●
●
● ●
●●●
●
●
●
●●
●●
●
●●
●● ●●●● ●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
● ●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ● ●● ●
●●●●●
●●●
●●
●
●
●
●● ●
●
●●
●●●
●
●●
●
●●
●
●●●
●
●
●
●●●
●●
●●●
●
●●
●
●● ●
●
● ●●
●●● ●●
●
●●
●
●●
● ●●
●●
●●● ●
●●● ●●●●
●
●
●
●●●●●●
●
● ● ●●● ● ● ●
●
●●
●●
●
●
●
●●
● ●●
●
●
●
●●●
●●
●
●
●●
●
●
● ●
●
●
●
●●●
●●
●●
●
●●●
●●
●●●
● ●●
● ●●
● ●● ● ●●●
●
●
● ● ●●●●
●
●●
●
● ●●●
●
● ●● ●
●●●
●
●●●
●●●●
●
●
●●
●
●
●●
●●
●●●
●
● ●
●
●● ● ●
●●●●
●● ●●● ●
● ●●
● ●●
●
●●●
●
●
●● ●●
●●
●●●●
●●
●● ●
●
●●
●●
●
●
●● ●
●●●
●
●●
●
●●
●
●
● ●●
●
●
●●
●●●
●●
●
●● ● ●●
●
●
●
●●
●● ●
●●
●
●
●● ●
● ● ●
●
●
●●●●
● ●
●●
●●●
●
●●
●●● ●●●●
●
●●●
●●
●●●
●
●●●●
● ●●
●
●●
●
●●●●
●●
●
●
●
●●
● ● ● ●●●
●
●
●
● ●●
●●
●●
● ●● ●●
●
●●
●● ●
●●
●●●
●
●
●●
●
●●●
●●●
●
●●
●
●
●
●●● ●
●
●
●●●
●●
●●
●●
●● ● ●
●
●● ●●
●●
●
●●
●●
●
●
● ●
●●●●
● ●●
●●
●
●
●
●●
●●●
●●
●●
●●
●● ●
●●
●
● ●
●
●
●
●●●●
●
●
●●
●●●●
●
●● ●
●
●●●●●●● ●●
●● ●●
●
●
●●
●● ● ●
●●●
●●●●●
●
●
●
●●
●●
●●
●●
●
●● ●
●●●●
●
●●●●
●
●
●●
● ● ●● ●
●●
● ●●
●● ●● ●
●
●
● ●●
●
●●
●
●
●●●
●●
● ● ●● ●
●●
●●●
●●
●
●●
●
●
●●
●●●
●● ●●
● ●
●
●●
●
●●
●
● ●●● ● ●●●
●
●●●●
●●
●
●●
●●
●
●
●●
●●
●●
● ●●
●
● ● ●
●
●
●
●●● ●●●
●●●●
●
● ●
● ● ●
●
●●●●
●
●●
● ●●●●
●●●
●
●
●●
● ●●
●● ●
● ●●
●●
●●
● ●●●
● ●
●
●●
●
●●●●
●
●●● ●●●
●●●●
● ●●
●
●●
● ●●●●
●●●●
●
●● ●
●
p
●●
●
●●
●●●
●●
●●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●●●
●
●
●
●
●
●●●
●●
●●●
●●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●●
●
●●
●●●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●● ●
●
●●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●●
●●
●
●
●●●
●●
●●●
●
●
●
●●●
●●
●
●
●
●●●
●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●●●
●●
●
●
●●
●
●●
●
●
●
●● ●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●●
●
●
●
●●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●
●●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
● ●●●● ●
●
●●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●●
●
●●●
●
●
●●
●
●
●
●●●
● ●●
●●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●●
●
●
●●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●●●●●
●
●●
●
●
●
●
●●
●●
●●
●
●●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●
●●
●
●
●●●●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●●●
●
● ●●●
●
●●●●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●●
●
●
●
●
●●●●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
● ●●
●
●
●●
●
●
●
●●●
●
●●
●
●●
●
●
●
●●
●
●●●
●●●
●●●
●
●
●
●●
●
●
●
●●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●
●
●●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●
●
●
● ●●● ●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●●●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●● ●●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●●●
●
●
●●
●
●
●●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●●
●
●●●
●●
●
●
●
●●●
●
●●
●●●
●
●●
●●
●
●
●
●
●
●
●
●
● ●●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●●
●
● ●
●
●●
●
●●
●
●
●
●
●●●
●●
●●●
●
●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●
●
●
●
●
●●●
●
●●
●●
●●●
●●
●
●
●
●
●●
●●●●
●
●
●
●
●●●
●●●
●
●●
●●●
●
●
●●
●
●
●●
●
●●
●●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●
● ●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●
●●●
●
●●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●●
●●
●
●
●
●●
●●
●
●
●
●●●●●
●
●
●●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●●● ●●
●●●
●●
●
●●
●
●
●
●
●
●
●●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●●
●
● ●
●●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●●●
●
●
●●
●● ●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●●
●●●
●
●
●●
●
●●
●
●
●
● ●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●● ●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●●●
●
●●●●
●
●●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●● ●
●●●
●
●●
●
●
●
●
●
●●
●
●
●
● ●
●●
●●
●
●●
●
●●
●●
●●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●●●
●
●
●●
●
●
●
●●●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
● ●
●●
● ●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
● ●●
●●●
●
●●
●●●●●
●
●●
●
●●
●●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●●
●
●●●●●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●● ●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●● ●
●
●
●
●
●
●
●●
●
●
●●●●
●
●●
● ● ●●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●●● ●
●
●
●
●
●● ●
●
●
●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●●●●
●
●●
●
●●●●
●
●●●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●●●
●●●
● ●●
●●
●●
●
●
●
●●●
●
●●●
●
●
●●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●
●
●
●
●●
●
●
●
● ● ●
●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
● ●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●●
●●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●●
●●●
●
●●●
●●
●
●●
●●
●
●
●
●
●
●● ●●●
●●●●●
●●
●
●
●
●●
●●
●●
●
●
●
●●●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●●●●
●
●
●
●
●
●
●
● ●●
●
● ●●●
●●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●●●●●
●●
●
●●●
●
●
●●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●●
●
●●
●●●
●
●
●
●
●●●
●●●
●
●● ● ●
●
●
●●
●
●●●●●
●
●
● ●●
●●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●●
●
●●
●●
●●
●
●
●
●●● ●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●●
●●●
●
●
●●●
●●
●●●●●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●●
●
●●●●●
●
●
●
●
●
●
●●● ●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●●
●●
●● ●
●
●
●
●
●●
●●
●
●
●
●
●●●●
●
●●
●
●
●●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●●●
●●
●
●
●
●
●
●
●
●●●● ●
●●
●
●●
●
●●
●
●●
●
●
●●●
●
●●
●● ●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●●
●
●
●
●
●
●●●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
● ●●
●
●
●
●●●
●
●
●
●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●●
●●● ●
●
●
●
●●
●●
●
●
●●
●●
●●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●●
●●
●●
● ●●
●
●
●
●
●●● ●
●
●
●●
●●
●
●
●●
●●
●
●
●●●
●●
●●
●
● ●
●●
●
●
●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
● ●●
●
●●
●●
●●●
●
●
●●
●
●●
●●
●
●
●
●●●
●
●
●
●
●●●●
●●
●●●●
●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●● ●
●
● ●●
●
●
●●
●●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●●
●● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●●
●●
●
●
●
●
●
●●●●
● ●●●
●●●●●
●●
●
●●
●
●
●
●
●●●●●
●●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●●●
●
●
●
●
●
●
●●
●●●
●
●
●●
●
●●
●●
●
●●
●
●
●
●●● ●
●●●
●●
●●
●
●●
●
●●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●● ●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
● ●
● ●●
●
●
●●●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
●●●●
●●
●
● ●●
●
●● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●●●
●
●
●
●
●●
●●● ●
●
●
●
●●
●
●●●
●●
●
●
●●●●
●
●
●
●●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●●
●
●●●
●●
●
●
●●
●●●
●●● ●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●
●
●
●●
●●
●
●
●
●
●●●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●●
●
●
●
●
●● ●●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●●
●●
●●
●
●● ●
●●
●
●●●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●●●
●●
●
●●●
● ●●
●
●
●●●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●●●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●● ●
●●
●
●●
●
●●
●
●
●●
●●● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●●
●●
●
●
●
●
●●
●●
●
● ●●
●
●
●
●
●●
●
●●
●●
●● ●●
●●
●
●●●●
●●●●●●●
●●
●
●● ●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●●●
●●
●●●●
●●
●
●●●
●●
●
●
● ●●
●
●
●
●
●
●
●
●●
●● ●●● ●
● ●
● ●●●
●
●
●
●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●● ●
●
●●
●●
●
●●
●●●
●
●●
●
●●
●●●
●
●
●
●
●
●
●●
●●●
●
●
● ●●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●● ●●
●●●●●
● ●●
●●
●●
●
●
●
●
●
● ●
●
●
●●
●
●●
●
●
●
●
●
●●
●●
●
●●
●●
●●●●●●●
●
●●
●
●● ●●●●
●
●
●
●
●
●● ●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
● ●●●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●● ●●
●
●
●
●
●●
●
●●●
●●
●●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●●
●●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●● ●●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●● ●
●
●● ●
●
●
●
●
●●●●
●●
●
●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●●●
●
●
●●
●
●
●●
● ● ●●●
●
●
●
●
●●
●
●
●
●●
●●● ●
●
●●
●●
●
●●
●
●
●
●
●
●
●
● ●
●
●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●●
●
●
● ●●●
●
●●●
●
●
●
●
●●●●
●●
●
●
●●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●●
●●
●●
● ● ●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●●
●●●
●
●
● ●●●
●
●
●
●
●
●
●
●●●●
● ●● ●
●●
●
●
●
●●●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
● ●●●
●●
●
●
●
●
●●
● ●●
●●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
● ●●
●●
●
●●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●●●●●●●
●●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●●●
●●
●
●
●●●●
●
●
●● ●
●●
●
●●
●●
●
●● ●
●●
●
●
●●
●●● ●●●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●● ●●● ●
●
●
●
●
●●●
●
● ●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●● ●
●
●●●●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●●●
●
●●●
●
●●
●
●
●
●●
●
●
●●●●●●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●●●●●●
●
●
●
●●
●
●●●
●
●
●
●●
●●
●
●
●●●●
●
●
●
●●
●●
●
●
●
●
●● ●
●●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●
●●
●●●●
●
●
●
●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●●
●●●
●●●
●
●
●
●
●●●●●
●
●
●
●
●
●●
●
● ●●
●
●●●
●●●
●
●
●●●
● ●
●
●
●●
●
●
●
●● ●
● ●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●●●
● ●●●
● ●●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●●
●
●
●
●
●●●
●
●
●●●●
●
●
●
●
●● ●
●
●
●
●●
●
●●
●● ● ●●
● ●
●
●●
●●
●
●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
● ●●
●
●
● ●●●●
●●
●●
●
●
●●
●
● ●
●
●
●
●● ●
●
●●
●
●●
●
●
●
●●
●● ●
●●
●
●●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●
●●●
●●
●
●●●●●
●
●
●
●
●●
●●●●
●
●
●●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●●●●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
● ●●● ●
● ●
●●
●
●
●●● ●
●● ●●●
●
●
●
●●
●● ●
●
●●
●
●
●●
●●●
●●
●
●●
●● ●
●
●
●●●
●
●●
●
●
●●●
●
●
●●●●
●●
●●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
● ● ●●
●
●
●
●●
●●
●●
●●●
●●
●●
●●
●
●●
●
●
●● ●
●●
●
●●●●
●
●
●● ●
●
●
●●
●
●
●●
●●
●●●
● ● ●
●●
●
●
●●●
●
●
●
●
●●
●●
●
●
●● ●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●●
●●
●●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●●● ●●
●●
●●●
●
●●
●
●●
●
●
●●●
●
●
●●
●
●
●
●●
●
●●●
●●●
●●●
●
●
●
●
●
●
●●
●
●
●●●
●
●●
●
●
●
●
●
●●
●
●
●●
●●
●●●●
●
●●●●
●
●●
●
●
●
●
●
●
●
●●●
●●
●
●● ●
●
●
●●
●
●
●
●●
●●●●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●● ●
●●
●
●
●
●●
●
●●
●
●●
●●●
●●
●
●●
●
●
●
●
●
●●●●●
●
●●● ●
●
●●
●●
●
●
● ●
●
●●
●
●
●
●
●●
●●
●●●●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●●
● ●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●
●
●
●●
●
●
●●●
●●●
●
●
●
●
●●●●
●●
●
●
●
●●
●
●
●
● ●●
●
● ●●●
●●●●●●
●
●
●●
●
●●
●
●●●
●
●●
●
●
●
●
●●
●
●●●●
●
●
●●●●
●●●
●●●
●● ●●
●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●●
●
●●
●●
●●●●
●
●
●
●
●
●●
●●
●●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
●
●●●●
●
●
●
●●
●
●
●
●
●●●
●
●
●
● ●●
●●●
●
● ●●
●
●●●
●●
●
●●
●
●
●
●
●
●●
●
●●
●●
●●
●●●
●●
●
●
●●
●●
●
●
●●
●●
●
●●
●
●
● ●
●
●●●
●
●
●●
●
●
●
●
●●
●●
●
●
● ●
●
●
●
●
●●●●
●
●●●
●
●●
●
●● ●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●●
●●
●
● ●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●●●
●
●●●
●
●
●●
●
●
●
●
●
●●●
●
●
●
● ●●●
● ●●
●
●
●
●
●
●
●
●●
●
●●●
● ●●●
●
●
●
●
● ●
●
●
●●
●
●
●●
●
●
●●●
●
●●
●
●●
●●●
●●
●●●
●●
●
●
●
●
●
●● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●●
●●●● ●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●●
●
●
●●
●
●
●●●●
●●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
● ●●●
●
●
●
●
●
●●●●
●
●
●●●●●
●
●●
●●●
●
●●
●● ●●●
●
●
●
●
●●●
●
●●●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●●
●
●●
●●●
●
●●
●
●
●●
●
●
●
●
●●
● ●●
●●
●●
●
●●●
●
●●
●
●
●
●
●●●
● ●
●
●●
●●
●●
●●●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●●
●●
●
●
●
●
●●
● ●●
●●●●●●●●●●
●
●
●●
●
●
●
●
●● ●●
●
●
●
●●
●●●●
●
●
●
●●
●
●●●
●●
●
●
●
●
●
●●● ●
●
●
●
●●
●●
● ●●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●●●●
●
●
●●
● ●
●
●
●●
●●
●
●●
●
●
●
●
● ●●
●
●
●
●●
● ●●
●●●
●
●●●●●
●
●
●
●
●
●●
●●●
●
●
●●●
●
●
●
●
●●
●●●
●●
●
●
● ●●
●●●●
●
●●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●●
●●
●
●●● ● ●
●
●
●
●
●●
●
●
●
●
●●
●●
●
● ● ●●●
●
● ●●
●
●
●●
●●●
●● ● ●●
●
●●
●●
●●●
●
●●●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
● ●●
●●
●●
●●●
●
● ●●●
●
●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●●
●●
●●●
●
●
●●●
●
●
●●●●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●● ●
●●●
●
●●
●●
●●
●
●
●
●
●
● ● ●
●
●
●
●
●
●●●●●●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●
●●●
●
●
●
●
● ●●●
●
●
●
●
●●
●●
●● ●●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●● ●●
●
●●
●
●
●
●
●
●
●
●●
●
●●●
●
●
● ●●
●
●●●
●
●
●
●
●●
●●
●●
●
●
●
●●●
●
●
●●
●●● ●
●
●
●●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●●●
●●
●
●
●● ●
●
●
●●
●
●
●
●●
● ●●
●
●●
●
●
●●●●●● ●● ●●
●●
●
●●
●●
●
●
●
●
● ●
●
●●
●
●
●●
●
●
●●
●●
●
●
●●●● ●●
●
●●
●●
●
●
●●
●
●●●
●
●●●
●
●
●
●●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●●●
●●
●●
●
●
●
●
●●●
●
●●●●
●●
●
●●●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●●
●●●● ●● ●●
●
●
●
● ●●●●●
●●
●
●
●● ●
●
●
●
●●●●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
●●
●●
●●
●●
●
●●
●
●
●
● ●●● ●●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●●●●●
●
●
●
●●
●
●
●
● ●●
●●
●●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●● ●
●
●
●
●
●●●●
●
●
●
●
● ●
● ●●●●●
●●●●●
●● ● ● ●
●
●●
●●
●
●
●
●
●●
● ●●●●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●●
●●
●
●
●●
●●
●
●●
●●●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
● ●●●●
●
●
● ●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●●● ●
●
●
●●●● ●
●
●
●
●● ●●●
●●●
●●
●
●●
●
●
●
●
● ●●
●
●●
●
●
●●●
●
●
●
●
●
●● ●
●
●
● ●
●
●
●
●●●
●●
●●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●●●
●
●●
●●
●
●●
●
●
●
●
●●
●
●●
● ●●●
●
●
●
●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●●
●●
●
●●
●
●
●● ●●
●
●
●
●
●
●
●●
●●●
●●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●●
●●
●●
●●
●●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●●●
●●● ●●
●
● ●●●
●●
●
● ●●
●
●●●
●●
●●●●
●
●
●
●
●●
●● ● ●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
● ●
●●
●●●
●
●
●
●●●
●
●
●●● ●
●●
●●
●
●
●
●● ●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●● ●
●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●● ●
●
●●
●●●
●
●●
●●
●●
●●●
● ● ●● ●
●
●●
●
●●
●
●
●
●
●●●
●
●●
●●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●●
●
●●
●●
● ●
●
●●
●
●
●
●●
●
●●
●
●
●●
●●
●
●
●
●●● ●
●
●
●●●
●
●●
●
● ●●
●
●●●
●
●● ● ● ●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●●● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●●●
●●●●
●
●
●●●
●●
●
●●
●
●
●●
●●
●●
●●
●
● ●
●
●●●
●
● ●
●
●
●
●
●●
●
●
●● ●●
●●
●●
●● ●
●
●
●●●
●
●
●
●●●● ●●
●●
●
●●●
●
● ●
● ●
●●●
●
●●
●●
●
●
●
●●●●
●
●●
●
●
●
●●
●●●●
●
●
●
● ●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●●●
●
●
● ●●●
●
●
●
●●
●●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●●
●
●●●
●
●●●●● ●●
●●●
●
●
●
●
●
●
●
●●●
●
●
●●
●
● ●●●
●
●
●
●
●
●
●● ●●
●
●
●
●●●●
●
●
●●
●●
●
●
●
●
● ● ●●
●
● ●●●
●
●
●●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●●
●
●
●●● ●
●
●
●
●
●
●
●●●●
●
●●
●●
●
●
●
●●●●●
●
●
●
●●
●●
●
●
●●
●●●
●
●
●
●●
●●●
●
●
●●
●●
●●
●
● ●●
●
●
●
● ●
●●
●
●●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●● ● ●
●
●
●
●
●
●
●●
●
●●
●●●
●●
●
● ●
●
●
●
●●
●
●
●
●
●●●
●
●
● ●●
●
●
●
●
●●●●●
●
●
●●
●● ●
●
●
● ●●●
●●
●
●
●●
●
●
●●
●●
●● ●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
● ●●
●●●
●●
●●
●●
●
●
●●●●●
●●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●●●
● ●
●
●
●
●●●●
●●●●
●
●
●●
●
●●
●
●
●● ●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●●●●●●●
●
●
●●●
●
●
●●
●●
●● ●●●
●●
●●
●
●
●
●
●● ●
●
●
●●●
●●● ●●
●●
●●
●
●●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●●●
●
●●● ●
●●
●●
●●
●
●
●
●
●●●●
●
●
●
●
● ●●
●●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●●●●● ● ●
●
●
●
●●
●●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●●
●●
●●
●●●
●
●
●
●
●●
●
●●
●
●●
●
●
●
● ●●
●
● ●●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
● ●●
●●
●
●●
●●
●
●●
●●
●●
●
● ●
●
●
●●
●
●●
●●
●
●
●●
●●
●●●●
● ●●
●
●●
●
●
●
●●
●
●
●●● ●
●●
●●●
●
●
●
●●● ●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●●
●●●
●
●
●●
●● ●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●●
●●
● ●●
●
●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●
● ●●
●
●
●●
●
●●
●●●
●
●●
●
●●●●
●
●●
●●
●●
●
●
●●●
●
●
●●
●
●●
● ●●
●
●
●
●
●●
●
●
●
●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●●
●●●●
●● ●●● ●
●●
●
●●
● ● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●●●
●
●
●
●
●
●●
●●
●
●● ● ●
●
●●●
●
●
●● ●●
●
●●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●●●
●
●●
●●●
●
●
●
●
●●
●
●
●●
●
●●
● ●●●● ●
●
●
●●
●
●
●●
●
●●
●
●●●
●●
●
●
●
●
●
●
●
● ●●●●●●
● ●
● ●● ●●
●
●
●●
●
●
●
●
●
●●●
●
●●● ●●
●
●
●
●
●●●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●●●
●
●●●●
●●
●
●
●● ●●
●
●●●
●
●● ●
●
●●●
●
●
●
●
●
●
●● ●
●●
●
● ●
●●
●
●
●
●
●
●●
●
●
●●
●●●
●
●●
●
●●
●
●
●●
●●
●
●
●
●●
●
●● ●● ●●●
●
●
●●
●●●
●
●●
●
●
●
●
●● ●● ●
●
●●
●
●●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●●
●●●
●●
●
●
●●●
●
●
●●
●
●●●
●
●
●●
●
● ●●●
●
●
●
●
●● ●●
●
●●
●●●
●●●●
●●
●
●
●
●●●
●
●
●●
●
● ●
●
●● ●
●● ●●●
●●
●●
●●
●●
●
●●
●●
●●●
●●
●
●●
●
●●●
●
●
●
●
●
● ●●
●
●
●
●●
●●●
●●
●
●●
●●●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●●●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●●●
●
●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
● ●
●
●●●●●●●
● ●
●● ●
●
●●
●
●●●●
●●
●●
●
●●
●
●●
●
●
●● ●
●● ●
●
●●●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●●●●
●
●
●
●●●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●●
●
●●
●
●
●●●●
●●
●
●
●
●●
●
●
●●●
● ●
●●
●● ●●
● ●
●●
●
●●●●
●
●
●
●●● ●
●●
●
●●●
●
●
●
●●
●●
●●
●
●●
●●
●●
●
●
●
●●
●
●●
●●
●
●
●
●●●●
●●
●●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●●●●
●●●●
●●●
●
● ●●
● ●●
●
●
●
●
●
●● ●●
●● ●
●●
●●●
●● ●●
●●
●● ●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●●
●
●
●●
●
●● ●
●●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
● ●
●
●●
●
●● ●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●●
●
●●●
●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●●● ●
●●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●
●
●
● ●
●● ●●
●
●
●
●●● ●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●● ●
●●
●●
●
●●●●●●
●●
●
●
●
●
●
●
● ●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
● ●●●●●
●
●
●
●
●
●
●
●●
●
●
●●●●
●
●
●
●●●● ●
●
●
●
●
●●
●
●● ●
●●●
●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●●
●
●●●●
●●
●
●●
●●
●●
●●
●
●●
●●
●
●
●
●
●
●
●● ●
●
●
●●●●
●●●
●
●
●
●
●
●
●● ●● ●
●●● ●
●
●
●
●●
●
●●
●●●● ●
●
●
●●●
●
●
●●●
●
●●●●●
●
●
●
●
●●
●
●●
●
●●
●●
●●
●
●●
●●
●●●●
●
●
●
●
●●●●
●
●
●
●●●
●
●
●
●
●
●
●
● ●●●
●
● ●●
●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●●
●●
●●●● ●●●
●●
●
●
●●
●●●
●
● ●●
●
● ●●
●●
●
●
●●●●
●
●●●
●
●
●
●
●
●
●
● ●
●●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●●
●
●●
●
●●
●
● ●●
●
●●
●●
●
●
●
●
●● ●● ●●●●
●●
●●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●●●●●
●
●
●●●
●
●●
●
●
●
●●●●
●
●
●
●
●● ●●● ●
●
●
●
● ●● ●●
●
●
●
●
●
●
●
●●
●
●●
●
●●●
●
●
●●
●
●●●
●
●
●●
●●
● ●
●
●
●●
●●
●
●●
● ●●● ● ●
●●
●
●●
●●
●
●
● ●●
●
●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●●●●●●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
● ●●
● ●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
● ●
●●
●●
●●● ●
●●●
●●● ●
●
● ●●
●●●
●
●
●
●
●●
●
●
●
●
●
●
● ●●●●
●●●●
●
●
●
● ●●
●
●
●
●●
●
●
●
●
● ●
●
●
●●
●
●●
●
●●●●
●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●●●
●● ●
●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
● ●
●●●
●
●●
●●
●
●●●
●
●●
●●●
●●
●●● ●●●
●
● ●●● ●
●● ●●● ●
●●
●●
●
●●
●
●
●●
●
●
●●●●●
●
●
●●
●
●
●
●
●
●●
●
●●●
●
●
●●
●
●
●
●●
●
●
●●●
●
●●
●
●●●●
●
●
●●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●●
●
●●●●● ● ●●●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●● ●
●● ●●● ●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●●●●●●
● ● ●● ●
●
●
●
●●●●●
●
●●
●
●
●
●
●●
● ●●
● ●
●●
●●
●
●
●●●
● ●●
●
●
●
●
●●
●
●
●
●
●●●
●●
● ●●●
●●
●
●
●●●
●●
●●●
●
●●
●
●
●●
●
●
●●
● ●
●●
● ●
●●
●
●
● ●●
●
●
●
●
●
●●
●
●●●●●
●
●● ● ●●●
●
●●●●
●●
●
●
●
●●●●
●●●
●●
●●
●
●
● ●●
●
●
●
●●
●●
●●
●
●
●
●●●●
●
●
●
●
●
● ●●
●●
●
●
●●
●
●●
●
●
●
●
●
● ●●
●
●
●
●●
●●●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●
●
●
●
●
●●
● ●●
●●
●●
●
●●
●
●
●
●● ●
●●
●
●● ●
●
●●
●
●
●
●●
●
●
●
●
● ●
●●
●●●
●
●
●
● ● ●●
●●●
●
●
●●
●
● ● ●●●● ● ●●●
●
●●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
● ● ●
●
●●●
●
●
●
●●
●●●
●
● ●
●
●●
●
●●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●● ●●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●●
●
●●●
●
●
●
●
●●
●●
●
●
●●
●●
●●
●
●●●
●●
●
●
●●
●●●
●
●
●●
●
●●●
●
●●●
● ●●●
●●
●
●●
●
●●
●
●
●
●●
●
●
●●
●●
●●●
●
●
●●
●
●
●●●● ●●●● ● ●●
●●
●
●●
●
●
●
●
●●
●
●
●●● ●●●●
●
●
●
●
●
●●
●●
●
●●
●
●
● ●●
●●
●●
● ●
● ●●●●● ●●
●
●●
●
●
●
●●●
●●●●
●
●●
●●●
●
●●●
●●
●
●
●●
● ●●●
●
●
●
●●
● ●●●
●●
●
●
●
●
●●●
●
●●
●
●
●●●
●●
●●
●●●●
●
●●
●
●
●●
●●●●
●
●
●●
●
●●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●●●
●
●
●● ●
●
●
●
●
●●
●●
●●
●●●
●
●
●
● ● ●●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●●
●
●●
●
●
●
●●● ●
●●
●
●●
●
●
●
●●●
●
●●
● ●
●
●
●
●
●● ●●
●
●
●●●
●●●
●
●●
●
●●
●
●●
●●
● ●●
●
●
●
●
●
●
● ●●
● ●●
●
●●
●
●
●●
●●●●●●
●●
●●●
●
●
●
●●
●
●
●
●●
●
●●● ●
●
●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●●●
●
●
●
●
●
●
●●
●●●
●
●●
●
●●
●●
●
●
●●
●●
●
●
●
●
● ●●
●
●●● ●●●●●
●
●
●●●
●●
●
●●
●
●
●
●●
●● ●
●● ●
●
●
●
●
●
●● ● ●
●
●
●
●●●
●●●
●●
●●
●●
●●●
●
●●
●
●
●
●
●
●
●
●
●●●
● ●●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●●
●●●
●
●
●
●
●
● ●●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●●● ●
●●●●
●
●
●
● ●
●● ●●
●●
●●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●
●
●
●
● ●● ●
●●
●●
●
●
●●
●●●
●
●●
●
●●
●
●
●
●
●
●
●
●●●
●●
●●●
● ●●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●●
●●
●
●●
●●●● ●
●●
●●
●
●●
●
●
●
●
● ●●●●
●
●●
●
●●●
●●●●
●
●●
●
●●
●
●●● ●
●
●●●
●●
●
●
●●
●
●
●
●
●●
●●●●
●●
●
●
●●
●
●
●●
●●
●●
●
● ●
●
●●
●●●
●
●●
●
●
●●
● ●●
●●
●
●●
●●●
●
●
●
●● ● ●
●●
●
●
●
●
●●
●
●●
●
●●
●●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●
●● ●
●●
●●
●●
●
●
●●
●
●
●●
●●
●
●●●●
●
●
●
●
●
●
●
●●
●
●●
●
●
● ●
●
●●
●●
●
●
● ●●●
●
●
●●●●●
● ●● ● ● ● ●●
●●
●
●
●
●
●●
●●
●●●
●
●● ●
●●
●
●
●
●
●
●
●●
●●
●
●
●●
●● ●
●
●●
●
●
●● ●●●●
●
●
●
●
●
●
●●●
●
●
● ●
●
●●●
●●
●●
●
●●
●
●●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●●
● ●●●
●
●
●
●
●
●
●●●
●● ●
●
●
●
●
●
●●●
●●
●
●
●
●
●●●
●●
●●
●●
●
●
●
●●●●●
●● ●
●
●●●
●●
●●
●
●
●
●●
●● ●●
● ●●
●●
●
●
●
●
●●
●
●
●
●
●●●●
●
●● ●
●
●
●
●
●
●
●●
●
● ●●
●●
●
●
●
●
●
●●
●●●
●●●
●
●●
●●●
● ● ● ●● ●
●●●
●
● ●●
●
●● ●
● ●
●●
●
●●●
●●
●
●
●
●
●●
●
●●
●
●
●
●●
●●●●●
●
●●
●
●
●
●●●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●●
●●
●●
●
●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●●●
●●
●
●●
●●
●●●●
●
●●
●
●
●● ●
●●●●●
●
●
●
●●
●
●
●
●
●
●
●●
● ●
●
●
●●
●●
●●
●
●
● ● ●●
●
●
● ●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
● ●●
●●
●
●
●
●
●
●● ●
● ●●
●
●
●
●
●●●●●●●
●
● ●●
●●
●●
●
● ●
● ●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●●
●●
●●
●
●●
●●
●
●
●
●●
●
●
●●
●●●
● ●●● ●
●
●
●
●
●●
●●
●
●
●
●
●●●●●●
●
●●
●
●
●
●
●●
●●●
●
●
●●
●
●
●
●●●●
●
● ●●
●
●
●●
●●
●●●
● ●●
●
●●
●●
●
●●
●
●●●
●
●
●
●
●●●
●
●●
●●
●●
●
●● ●● ●●
●
●
●
●
●
● ●●
●●
● ●●
●
●
●
●●
● ●●
●
●
●●●
●●
●
●
●
●●●
●
●
●●
●●
●●
●
●
●●
●
●●●
●●●●
●
●
●
●
●●
●
●
●
●●
●
● ●
●●
●●●
●
●● ●●
●●●●
●●●
●
●
●●
●●●
●●
●
●
●
●
●●●
●
● ●●● ● ● ●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●●● ●
●●●●
●
●
●
● ● ●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
● ● ●
●
●
●
●
●●
●
●●●
●
●●●●●
●
●
● ●●
●●
● ●
●●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●●●●
● ●●●
●●
●
● ●●
●●●
●●●●
●
●
●
●●●
●
●●
●●
●
●●
●
●
●●
●●
●●
● ●●
●
●●
●
●
● ●●●● ● ●
●
●●
●
●
●●
●
● ●●
● ●
●
● ●●
●
●
●
●
●
●
●
●●
●
●●
● ●●
●
●
●
●● ●
●
●●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
● ● ●● ●●
●
●
●
●●
●
●
●
●
● ●●
●
● ● ●●
●
●
●
●● ●●
●
●
●
●
● ●●●
●● ●
●
●
●
●●●
●
●
● ●● ●
●
●●
●●●
●
●
●
●
●●●●
●
●
● ●●●●●
●
●
●
●●●
●
● ●● ●● ● ●●●
●
●●●●●
●
●
●
●
●
●
●●
●
●●● ●
●
●●
●
●
● ●
●
●
●●
●
●
●
●●●
●●● ●●●●
●●
●
● ●●
●
●
●
●
●●
●
●●●
●
●
●
●
●●
●
●●
●
●●
●
●
● ●
●●
●●●
●
●
●●●●
●
●
●
●● ●
●
●
●●
●●
●
●
●
●●
●●●●
●
●●●●
●
●
●●
●●●●●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ● ●● ●●
●●
●●●
●
●
●
●●
● ●
●
●●
●
●
●
●
●●
●●
●
●● ●
●
●
●
● ●●●●
●
●
●●
●
● ●●
●●
●● ●●●●●●●
●
●
●●
●
●
●● ●●
●●
●●
●●
●
●●
●●
●●●●
●
●
●
●
●●●
●●
●
●●
●
●
●●●
●
●
●●●
●
●●
●
●
●●●
●
●
●● ●
●
●
●●
●
●
●
●●
●●
● ●●
●
●
●●
●●●
●●●●
●
●
●●
●
●●
●●
●
●
●
●
● ●●
●
●
●●●
●
●
●
●
●●●
●
●
● ●
●
●
●●
●●
●
●●●
●
●
●●●
●
●
●●●
●●
●
●●
●●●●
●●●
● ●●●
●
●
●●●●●
●
●●●
●
●
●
●
●
●
●
●
●●●
●●●●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●●● ●
●
●
●
●● ●●●
●
●
●
●
●
●
●
●●
●●●●●
●
●●
●
●
● ●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●●●
●●●
●●
●
●●●
●
●
●●●
●
●●
●●
●
●
● ●●●
●
●
●
●
●
●
●●
●●
●
●
●●● ●●
●● ●●
●
●●●●
●
●
●
●
●●● ●● ●
●●
●●
●●
●
●●
●
●
●●●
●
●●
●
●
●
●
●●
●●
●●●●●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●
● ●●●
●●
●●
●
●●●●●
● ●●●●●
●
●
●
●
●
●
●
●
●
● ● ●
●
● ●●●
●
●
●●●
●
●●●
●
●
●
●●
●●●
●
●
●
●
● ●●●●●
●
●
●●●
●
●●
●
●
●
●●
● ●
●
●●
● ●●●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●●
●●
●●●●●●●●●●● ● ● ●
●
●●
●
●●
●●●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●
●
●●●
●●
●●●●
●
●
●
●
●
● ●●●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●
●
●●
●
●●●
●●
●●
●
●
●
●●●●
●
●●
●●
●●●●
●
●●
●
●
●●
● ●●
●●●
●
●
●●●
●
●
●● ●
●
●
●
●
●
●
●
●
●● ●●●
● ● ●
●
●●●
● ●●
●
●●
●●
●●
●
●
●
●● ●
●● ●
●
●
●
●●
●●●
●
●
●●
●
●
●
●
●
●
●● ●
●
●●
●●
●
●
●
●●
●●
●
●●●●●
●
●
●
●●
●●
●
●
● ●●● ● ● ●●
●
●
●
●
●●
●
●●
●
●●
●
●●
●●
● ●●● ●●
●
●
●●
●
●
●●
●●
●●●
●
●
●●
●●
●●●
●●●
●●
●
●●
●●
●● ●
●
●
●
●●
●●
● ●●
●
●● ● ●
●
●●
●●
●
●
●
● ●
●●
●●●
●●●
●
●●
●●●
● ●●
●
●●●●
●
●
● ● ● ●●●●●
●●
●●●
●
●●
●
●
●●●●
●●
●
●●●●
●●
●●
●
●
●●●
●●
●●
●
●● ●●
● ●
● ●●●
●●●
●
●●
●
●
●
●
●
●●●
●●
●●
●●
● ●●
● ●●
●
●●●
●●●●●
●
●
●● ●●
●●●
●
●●●
●
●
● ●●● ● ●●●●
● ●● ●
●●●
●● ●
●
●●
● ● ●
● ●
●●
●
●●
●
●● ●●
●
● ●●
●●
●
●
●●●●●● ●
●● ● ●
●●●
●●●
● ●
●●
●● ●● ●
●
● ●●●
●●
●●
●
●
● ●●● ●
●
●
●
●
●●
●
●●
●
●●●●
●
●
●●
●
● ●
●●
●
●●
●
●
●●
●●
●●
●
●●
●
●● ● ● ●
●●
●
●
●●
●
●
●
●
●●
●
●●● ●
●●
●●●
●●
●
● ●●
●●
●
●●●●
●
●●
●● ●● ●●● ●●
●
●●
●
●
●● ●●●●
● ● ●●
●
●
●
● ●●
●●
●●●● ●
●
●●
●●●
●
●●●
●
●
●● ●● ●
●
●
●
●
●
●
●●
●●●
●●
●
●
●● ●
●
●
●●●
●●● ●●
●
●
● ●● ●
●● ●●●
●
● ●●●
●●●
●
●●
●●●●● ●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●●
●
● ●
●● ●● ●●
●
● ●●
●
●●
●●
●
●
●● ● ●●
●●
●●
●
● ●● ● ● ●●●
●
●
●●
●●●
●
●●
●
●●
●●
●●
● ● ●● ● ●●
●
●●
●●
●●●
● ●●● ●
●
● ●●●●
●
●
●
●
●●
●●
●
●
●
●
● ●●●●
●
●
●
●
●
●●
●●
●●
●
●
●●●●
●●● ●●
●
●●●
●
●●
● ● ●●●●
●●●
●
●● ●●
●
●
●
●
●●
●
●
● ●●●●
●●
●
●
●
●
●
●
●●● ●●●
●●●
●
●
●
●
● ●●
●●●● ●●
●
●●
●
●
●●●● ●
● ● ●●
●●
●
●
●
●
●
●
●
●
●●
●
●●● ●
●
●
●●●●●●●
●
●●
●
●●●●
●
●
●●
●●
●● ● ●
●●
●
●●
●
●
●
●
● ●●●
●●
●
●
●
●●●●
●●● ●
●
●●●
● ●●
●
●
●
●
●
●
●
●●
●●
●●●
●
●●
●●●
●●●●●●●
●●
●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●
●
●
● ●●
●
●
●
●
●
●●
●
●● ●●●
●
●
●
●
●
●
●
●
●●●●● ●
●
●●
●●●● ● ●
●
●
●●
●
●
●●
●
●
●●
●
●
● ●
● ●●
●
●
●●
●
●●●
●
●
●
●
●
●
●●
●
●●
●●
● ●●●
●
●
●
●
●
●●
●
●
●
●
●●
●● ●
●●●
●
●●●
●
●
●
●●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●●●●●●
●
● ●●●●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●●● ●●
●●●
●
●●●●●●●
●
●
●
●● ●
●
●
●
●●●
●
●
●
●
●
●
●●
●●●
●●
● ●●●●●●●
●●●●
●
●
●
●
●
●●
●● ●
●●
●
●
●
●
●
●
●
● ●●
●
●●
●●
●
●●
●
●
● ● ●●
●●
● ●
●
●
●●●
●
●●●
●●
●
●
●
●
●
●
● ●
●●
●●
●
●
●
● ●● ● ●● ●
●
●
●
●
●
●●
●
●●
● ●
●
●●●●
●
●
●
●
●●
●
●
●
●●●●
●
●
●●●
●
●
●●
●
●
●
●
●
●● ●●
●
●●●
●●
●
●
●●●
●● ●
●
●
● ●● ● ●● ● ●
●
●●
●●
●●
●
●●●●●●
●
●●
●
●
●
●
●●
●
●
● ●●●
●
●●
●
●
●●
●
●
●
●
●●●
●●
●● ●●
●
●●
●●
●●
● ●
●●
●●
●
●●●
●
●
●●
●●
● ●
●
●● ●
●
● ●
●●
●
●
●
●●
●
●
●
●●●●●●
●●
● ●● ● ●●
●
●●
●●●
●●
●●
●●
●
●
●●●
● ●●
● ●●
●
●
●
●
●
●
●●
●
●●
●
●●
●●●●
●
●
●●
●
●
●
●
●● ● ●
●
●
●●
●
●● ●
●
●●
●●●
●
●
●
●●● ●
●
●
●●
●
●●●
●
●●●●●
● ●
●
●●
●● ●●
●
●●
●●
●
●●
●
●
●●
●●● ●
●
●
●
●
●●●
●●● ●
●●
●
●
● ●●●●
●
●
●●●
●
●●
●
●● ●
●● ●
●●
●
●
●
●
●●●
●
●
●● ● ●
●●● ● ●
●●
●●
● ●
●
●
●
●
●
●●
●
●
●
● ●●
●
●
●
●●
●●
●
●
●●●
●
●
●●
●
● ● ●
●
●●
●●
●●
●●●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●●● ●
●●
●
●●
●●●
●
●
●
●
● ●
●●
● ●●●●●●●●● ●●●
●
●●
●●
●
●●
●
●
●●
● ● ●●●
●●●
●
●
●●●
●●●●
●
●
●●●
●
●
●●
●●
●
●●
●
●● ●●
●
●●
●
●●●●
●●●
●●
●●
●
●
●
●●●
●● ●●
● ●
●●● ● ●●
●
●
●
●
●
●●
●
●
● ●●●
●
●●●
●
●
●●● ●
●
●
●●
●
●●
●● ●
●●
●
●
●● ●
●
●●
●●
●●●
●●
●●
●
●
●●●●● ● ●
●
●●
●●
●
●
●●
●●
●
●
● ● ●●
●
● ●●
●
●
●●●●●
●●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●●
●
●
●
●●
●● ●● ●
●
●
●●●
●●● ●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●●●●●
●●
●
●
●
●
●
●●
● ●●
●
●
●
●
●
●
●● ● ●
●●
●
●
●
●
●
●●
●
●
●
●●●
●
●●●
●●
●
●
●●
●●●●●●
●●●
●●● ● ●●
● ●●●
●
●●
●●
●
●
● ●●●
●
●●
●
●
● ●
●●●
●
●●
●
●●●●
●
●● ●
●
●
●●
●
●●
●
●●
●●●
●
●
●
●
●
●
●● ● ●
●●●
●●●●
●
●
●● ●
●
●●●●
●
●●●
●●●
●
●
●● ●● ● ●●
●
●
●
● ●●● ●
●
●●●
●
●
●●●
●●●
●
●
●
●
●●●
●
●
●
●
●●
●
●●
●●●
●
●
●
●
●●
●●
●
●●
●●
●
●●
●●●●
●●
● ●● ●●●
●●
●●●
●●●●●
●
●
●
●
●●●
●●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●● ●● ●
●
●●
●
●●
●
●●
●
●
●●
●●
●●●●● ● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●● ●●● ●
●
●
●
●
●
●
●● ●● ●●
●
● ●
● ● ● ●●
●● ●
●
●●● ●● ●
●
●
●●
●
●●● ●
● ● ●
●
●● ●●●
●●● ● ●●●●●
●
●●
●● ●
●●●
●
●●● ● ●●●
●●
● ●
●
● ● ●●
● ●
●
● ●●●●
● ●●● ●
●●●●
●
●
●●
●
●
●●●
●●●
●
●
●●
●
●
●
●
●●
●● ●
●
●
● ● ●
●
●●●● ●●●
●
●
●
●
● ●●●
●●
●
●
● ●●
●
●● ●●
●●●
●
●●
●
●
●
●
●
●
●
●●
●●●●●●●
●
●●
●
●
●
●● ●●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●●
●● ● ●●●●
● ●●
● ●●●●●● ●
●
●●
●
●
●
●
●
●●
●
●● ●●
●●
●
●●
●●●●●
● ●●
●●●
●●
●●●
●
●●
●●
●●
●●
●● ●
●
● ●
●●●
●●●
●●
●●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●●●
●
●●●
●
●●
●
●●
●●
●
●
●●
●
●
●
●
●●
●●●●
●
●
●
●
● ●●
●●
●● ●●
● ● ●●●● ●●●●
●●
●
●
● ●
●● ● ●
●
●
●
●
●●●●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●● ●
●●
●
●●●
●●
●
●●●
● ●●● ●●●●●
●
●
●
●
●●● ●
●
●
●●
●
●●●●●
●
●
●
●
●●
●
●●
●●●
●
● ● ●●
●
●
●●
●
●
●
●
●
●
●● ●●●●
●●
●
●●
● ●●
●●
●● ●●
●
●
●● ●
●
●
●●
● ●●●●●●● ●
●
●●●
●
●
● ● ●● ● ●●
● ●●●
●
●
●●
●●
●●
●●●
●
● ●●
●●
●
●
●●●●●
●●
●
●
●
●
●
●
●●
●●
●
●
●●●
●●●●●
●●
●
●
●●
●
●
●● ●
●
●
●●
●●
●●
●
●●
●
●●
●
●
●
●
●
●
●●●●
●●●
● ●● ●●● ● ●
●
● ●
●●●●●
●●
●●
●●
● ●●
●●
●
●
● ●
●●●
●
●
●●
●
● ●
●●● ●
●●●
●
● ●
●●
●
● ●
●●
●●●
●●
●● ●●
●●
●
●
●
●●●
●
●
●●
●●●●
●●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
● ●●
●
●
●
●●●
●
●●
●●●●
●
●
●● ●●
●●●
●●
●●
●
●●●
●●●●
●
● ●●● ●
●
●
●
●●
●
● ● ●●●
●
●
●
●
●
●
● ●●●
●●
●
●
●
●
●
●●
●● ●● ●●●
●
●
●●● ● ●
●●
●●
●
●●
●
● ●●
●
●
●●● ●
●
●●●●
● ●● ● ● ● ● ●●
●●●
●●●
●●●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●●●
●●●
●
●
●● ●
●
●
●
●●●
●
●
●●●
●
●●
●
●
●●
●●● ●●
● ●● ●
●
●
●
●
●
●●
● ●● ●●
●
●
●●●
●●
●●
●
● ●●●
● ●● ●●●
●
●
●●
●
●●● ● ● ●●
●
●
● ●
● ●●●
●●
●● ●●
●
●
●
●
●●●
●●
●
●●
●●
●●
●
●
●
●
●●● ● ●●●●
●●
●●
● ●●
●●
●
●●●
●●●
●●
●● ●
●●
●
●
●
●
●●●●
●
●
●●
● ●
●
●●
●● ●
●●●
●● ●
●●●
●
●
●●●●
●
●
●
● ●●
●●
●
●
●●
●
●●● ●
● ●●●●●●
●
●
●
●●●
●
●●
●
●
●
●● ●
●●
●
●●●●●
●●
●
●● ●●● ● ●
● ●● ● ●● ●●●
●
●
●
●
● ●●●●●●● ●● ●●
●
●●
●
●
●
●●
●●
●●●●●●●●●
●
●
●
●
● ●●
●
●
●●
●●●
●
●
●
● ●
●
●●
●
●
●●●
●●●●
●●
●
●
●
●
●
●●●
●
●●
●●
●●
●●
●
●●●●
●●
●●
●●
●● ●● ● ●
●●●
●●●
●
●
● ● ●●●●
●
●
●
●●
●
●●
●
●
●●
●●●
● ●●
●● ● ●●●●●
●
●●●
●●
● ●●●● ●● ●
●●
● ●●
●
●
●
●●
●
●
●● ●
● ●●
●●
●
●
●
●●
●
●
●
●●
●●●
●
●
● ● ●
● ● ●●
●
●●●
●● ●●●
●
●
●
●
●
●
●●
●●●●● ●●
●
● ●●
●●
●
● ● ●●
●
●●
● ●●●● ●●●●●
●
●
● ●●
●●
●
●●●
●
●
●
●
●
● ●● ●
●
● ●●
●
●●
●
●●
●●
●●
●
●
●
●
● ●●●●
●●
●●●● ●●
●
●
●
●●
●
●●●●●●
●
●
●●● ●●
●● ●
●
●
●●●●
●
●●
● ● ● ●●
●
●●
●●
●
●●
● ●●●
●●
● ● ●●
●
●● ●
●●
●
●●
●
●●
●●●● ● ● ●●
●●
●
●●
●● ●
●
●
●
●
● ●● ●
●
●
●●
●
●●●●
●
● ●
●
●●●
●●●
●
●
●
●●
●●●● ●●●
●●
●
●●
● ●
●
●
●
●●
●
●●
●
●
●●●●
●● ●
●
●
●●●
●
●
●
●●●●
●●
●●
● ● ●● ●●
●●
●●●●
●
●
●●●● ●
●
●
●
●● ● ●
●●
●
●
●
●●
●
●
●
● ● ●●●●
●
●●
●
●●
●●
●●
●●● ● ●●
●
●
● ●●
●●●
●
●
●● ●
●
●
●●●●
●
●
● ●●●●●
●●●●●●
●
●
●
●
●●
●●
●
●●
●●
●
●●
●●
●●
●● ● ●● ●
●●●
●●
●
●
●
●●
●
● ●● ● ●
● ●
● ●
●●
●
●●
●
●
●● ●
●●
●●
●
●●
●●
●
●●
●
●●● ●
●●●
●
●
●●●●●●●
●● ●
●
●●●
● ●
●
●
●●
●●●
●
●●●
●
●
●●●
● ●
●
●
●●
●
●●●
●●●
● ●●
●
●
●●
●● ●
●
●
●●●
●
● ●●
●
●
●●
●
● ●
●
●● ●
●●● ●
●
●
●●
●●●●
●
●
● ●
●●
●
●
● ●
●●
●●●
●
●●
●
●
●●
●●
● ●
●
●●
●
●●
●
● ●
●
●
●
●●
●●●
●●
●● ●
●
●●●●●●●
●● ●
●
●●
●●
● ●●● ●
●●
●
●
●●
●●●
●
●●
●
● ●●
●●
●
● ●●
●●
●● ●●●
●●●
●●
●
●
●●
●
●●
●
●●●
●
●
●
●●●
●●
●●
●
●
●●
●
●●●
● ●
●
●
●●
●● ●
● ●●
● ●
● ●● ●
●●●
●●●
●●●
●
●
●
●
●
●
●●
●
●
●●●
●●●●
●●
●●
●
● ●
●
●●●●
● ●● ●
●●
●
●●
●
●
●
●●
●●
●
●●●●
●●
●
●●
●●
●
●
●●
●
●●
●●●●
●●
●●
●●●
●●
●
●
● ●●
●
●●
●
●
●
●
●●
●●
●● ●
●
●
●
●●●
●● ● ●
●
●
●●
●●
●
●
●
●●
●
●●
●● ●
●
●
●●
●
●●
●●
●●
●●
●
● ●●●●
●●
●
●
●
● ●●
●
●●●●
●●●●
●
●
●●●
●
●●
●●
● ●●●●● ●
●
●● ●●
●●
●●●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●
● ● ● ● ●● ●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●●● ●
●●●●● ●●●
●
●●●●●● ● ● ●● ●
●●
●
● ● ●●● ●● ● ●
●
●●
● ●● ●● ●
●
● ●●●● ●
●● ●
● ●
●
●
●
●
●● ● ● ● ●
● ●●●●●
●●● ●●●●●
●●● ●
●●
●●
●●
● ●
●●
●●
●
● ●● ●
● ●●
●
●●
● ●
●
●●
●●●
●
●
● ●
●
●●
●
●
●
●
●● ●● ● ● ●
●
● ● ●● ● ●
●●●
●
●●
●●●●
●● ● ● ● ●●● ●
●● ●
●
●
● ● ●●
●
●●
●
●● ●●
●
●●●●
●●
●
●●
●●
● ●● ● ●
●●
●
●●
●
●●
●●
●
●
●
●●
●
●
● ●●●
●●
●●
●●●
●●
●●
●
●●●
● ●●● ●● ● ●
● ●
●●●
●●●●
●
●
●●
●● ●●
●●●
●
●
●
● ●
●● ●
●●
●
●● ●
●●●
●●
●
●
● ●
● ●
●
● ●●
●
●●
●●
● ●● ●● ●
●
●●●●
● ●●
●
●●●●●
●
●● ●
●●
●
●●
●
●●
●
●●
●●
●
●●●
●
●
●●●●
●
●●
●●
●
●●
●●
●●●●
●
●●
● ●●
●●
●●
●●●
● ●●
●●
●●
●
●
●●
●●
●●
●
●
●● ●●
●●
●●
●
●
●
●
●●
●
●● ●●●●
● ● ● ●●●●
●
●
●
●●●●●●
● ●
●
●●
● ●
●●●●
●
●●
●
●
●●
● ●● ●
●
●
●
●●
●● ●●
●
● ●●
●●
● ●●●
● ● ● ●●●● ●
●
● ●●●
● ●●
●●●
●
●●●
●●
●
●
●●
●●
●●●●●●
●●●
●
●
●●●●
●
●●
●
●
●●
●●●
●
● ●●●●●
●
●●● ●
● ●●
●●
● ●●●
●●
●
●●●●
●
●
●●
●
●
●
●
●●
●●●
●
●
●● ●
●● ● ●●
●●
● ● ●● ●●
● ●
●
●
●●
●
●●
●● ●
●●●
●
●
●
● ●●●●
●● ●●●
●
● ● ●●
●●
●●●
●●
● ●
●
●●
●
●
●●
●● ●
● ●●●
●
● ●●
●
●
●
● ● ● ●
●
●
●●●●●● ● ●
●●
●●
● ● ●●
●
●● ●
●●●●
●●● ●
● ●●
●●●
●
● ●●
● ●●
●
● ●
●
●
●
●
●●
●●●
●● ●
●●● ●
●●
●
●● ●●
●● ●
●●
● ●●●●
● ●●●● ●●●●● ●●
●●
●
●●●●●
●
●
●
● ●
● ●●
●●
●●
● ●●●
●
●
●●
●
●●●● ●● ●●● ●
●
●
●● ●
●
●
● ●
●●
●● ●
●●
● ●●●●●
●
●●
●●
●● ●● ●
● ● ●● ● ●●
●●
●●
●
●
●●
●
●●●
●●
●
●
●
● ●●
● ●●●
●●
●●
●● ●
●●● ● ●
●
●
●●
●
●●
●
●●
●●●●
●
●
● ●●●
●
●●
●●●
●
●
●●
● ●
●● ●
●● ●●
●
●●
●●●
●●
●
●●
●●
●●
●● ●●
●●
●
●●●
● ●
● ●
●
●
●●● ●●●● ●
●●
● ●
● ●
●●
●
●
●●●●
●
●●
● ●●●
●●
●
● ●●
●● ●
●●
●●
●
● ●
●●●
●
●
●
●
●
● ●
●●
●
●
● ●●
●●
●
●
●
●
●●●●
● ●●
●
●●●●
●
●●●
●●
● ● ●●
● ●●●
●
●●
● ●
●● ●
●
● ● ●●● ●●
●●
● ● ●●
●
●●
●●
●●
●
●●
●● ●
● ●
●●
●
●●●
●
●
●●
●●●●●
●
●
●●
●
●
●●
● ●● ●
●
●
● ●
●
●●
●
●
●●● ●
●●●●● ● ●●●
●●●●
●●●
●●●
●
●
●
●
●
●●
●●
●
●● ●
●● ●
●
●
●
●●
● ●●
●●
●●
● ● ● ● ●●
●●
●
●
● ●●●
●
●●●
●●
●●●● ●
●
●
●
●
●● ●●
●
●●●
●
● ● ●●●● ● ●●●
● ●●●●
● ●●
●
● ●
● ●●
●●
●●●
●●●●
●
●
●●●
●
●
●●
●●●
●●●
●● ●●●
●●
● ●
●
●
●
●
●
●●●
●● ●
●● ●●●●●
● ●●
●●
● ● ●● ●●●● ●
●●
●●●● ●●
●
●●●●●●
●
●
●●
● ●●
●●●
● ●●●● ● ●● ●●
● ●●●
●●●●●
●●
●
●●
●
● ●
●
● ●●
●● ● ●
●
●●
●●
● ●●
● ●
●
●●●
● ●●
●
●●●●
●
●
●
●
●
●
●
●●
●
●
●
●●●● ●●
●
●
●
●
●●● ● ● ●●
●
●
●
●●●
●
●
●
●
●
●
●●●
●● ● ● ●●
●●●
●●
●
●●●
●
●● ●● ●●
●●
●
●●
●●
●●
●●● ●●
● ● ●●●
●●
●
● ●● ●● ●
●●
●●●●● ●
●
●
●
●●●● ●●
●●●●●
●
●
●●● ●●
●● ●
●
● ●●
●
●●
●●●
●
● ●●● ●● ●
● ●●
● ●
●
●
●●
● ●●
●
●●●●●● ●●
●
● ● ●● ● ●
● ●● ●● ●● ●
● ●●
●
r
●
p
F g 3 P ots o the reg on {q ∈ ∆n T (q p) + T (r q) ≤ T (r p)} or the equa -we ghted
or severa va ues o p and r Each po nt q on the boundary g ves
port o o π ≡ 13 13 31
a r ght-ang ed geodes c tr ang e n the sense o Theorem 1 2
beginning with $1 at time 0. Also consider the value of the market portfolio
starting with $1. Consider the ratio
V (t) :=
value of the portfolio π at time t
value of the portfolio µ at time t
called the relative value of the portfolio π. Under suitable conditions (see
[44, 45]) it can be shown that V (0) = 1 and
n
V (t + 1) X
µ (t + 1)
=
π (µ(t))
.
V (t)
µ (t)
=1
Proposition 5.10 (Fernholz’s decomposition).
we have
log V (t) = (ϕ(µ(t)) − ϕ(µ(0))) +
t−1
X
[23, 45] For a time t
T (µ(s + 1) | µ(s)) .
s=0
In this sense, the cumulative L-divergence measures the volatility harvested by the portfolio π. See [54] for optimization of this quantity.
The above discussion assumes that the portfolio rebalances every period
(say every week). In practice, due to transaction costs and other considerations, we may want to rebalance at other frequencies. Let 0 = t0 < t1 < t2
EXPONENTIALLY CONCAVE FUNCTIONS
39
be three time points and consider two ways of implementing the portfolio π:
(i) rebalance at times t0 and t1 (ii) rebalance at time t0 only. By Fernholz’s
decomposition, the relative values of the two implementations at time t2 are
log V (1) (t2 ) = (ϕ(µ(t2 )) − ϕ(µ(t0 ))) + T (µ(t1 ) | µ(t0 )) + T (µ(t2 ) | µ(t1 )) ,
log V (2) (t2 ) = (ϕ(µ(t2 )) − ϕ(µ(t0 ))) + T (µ(t2 ) | µ(t0 )) .
Letting µ(t0 ) = p, µ(t1 ) = q and µ(t2 ) = r, the difference between the two
values is
log V (1) (t2 ) − log V (2) (t2 ) = T (q | p) + T (r | q) − T (r | p) .
By the generalized Pythagorean theorem, the sign of the difference is
determined by the angle between the dual geodesic from q to p and the
primal geodesic from q to r. In Figure 3 we illustrate the result for the equalweighted portfolio of 3 stocks. In the figure, rebalancing at time t1 creates
extra profit if and only if q lies outside the region. This shows convincingly
that rebalancing more frequently is not always better, even in the absence
of transaction costs. More importantly, our framework provides a geometric
and model-independent way of saying that the sequence (p, q, r) is ‘more
volatile’ than the subsequence (p, r).
6. Displacement interpolation. In this final section we consider displacement interpolation for the optimal transport problem formulated in
Section 2.2. We refer the reader to [50, Chapter 5] and [51, Chapter 7] for
introductions to displacement interpolation.
6.1. Time dependent transport problem. Let P (0) and P (1) be Borel probability measures on Rn−1 . Consider the transport problem with cost c given
by (1.7). Suppose the transport problem is solved in terms of the exponentially concave function ϕ on ∆n . Letting f = ϕ + ψ, the (Monge) optimal
transport map is given by the c-supergradient of f . In particular, P (1) is the
pushforward of P (0) under F := ∇c f :
P (1) = F# P (0) .
The idea of displacement interpolation is to introduce an additional time
structure. We want to define an ‘action’ A(·) on curves such that the cost
function is given by
c(θ, φ) = min A(γ),
γ
where the minimum is taken over smooth curves γ : [0, 1] → Rn−1 satisfying
γ(0) = θ and γ(1) = φ. For each pair (θ, φ), a minimizing curve γ gives
40
S. PAL AND T.-K. L. WONG
a time-dependent map transporting θ to φ along a continuous path. Let
F (t) : Rn−1 → Rn−1 be defined by F (t) (θ) = γ(t), where γ is the minimizing
curve for the pair (θ, F (θ)). We want to define A in such a way that F (t) is
an optimal transport map for the probability measures P (0) and P (t) where
P0 .
P (t) = F (t)
#
classical Euclidean case with cost |x − y|2 the action is A(γ) =
R 1 For the
2
0 |γ̇(t)| dt and the optimal transport map has the form F (x) = x − ∇h(x)
where h is an ordinary concave function. The displacement interpolations
are linear interpolations:
F (t) = (1 − t)Id + tF
(See [50, Theorem 5.5, Theorem 5.6].) In particular, the individual trajectories (minimizing curves) are Euclidean straight lines which can be regarded
as the geodesics of a flat geometry. In this section we formulate and prove
an analogous statement for our transport problem.
6.2. Lagrangian action and portfolio interpolation. We begin by defining
an appropriate action. Let γ : [0, 1] → Rn−1 be a smooth curve with γ(0) =
θ. For each t, define q(t) ∈ ∆n such that its exponential coordinate is θ−γ(t),
i.e.,
qi (t)
= eθi −γi (t) ,
qn (t)
(6.1)
1 ≤ i ≤ n − 1.
Equivalently, we have qi (t) = eθi −γi (t)−ψ(θ−γ(t)) , for 1 ≤ i ≤ n−1. Intuitively,
we think of q(t) as theportfolio at time t (in the sense of interpolation). Note
that q(0) = n1 , . . . , n1 .
We define the Lagrangian action by
Z 1
1
(6.2)
A(γ) =
− log
+ q̇n (t) dt.
n
0
We take − log(·) = ∞ if the argument is not in (0, ∞). An alternative
representation of the action is
Z 1
1
d
(6.3)
A(γ) =
− log
+ e−ψ(γ(0)−γ(t)) dt.
n dt
0
Lemma 6.1.
(6.4)
For any θ, φ ∈ Rn−1 we have
c(θ, φ) = ψ(θ − φ) = min {A(γ) : γ(0) = θ, γ(1) = φ} .
EXPONENTIALLY CONCAVE FUNCTIONS
41
The action is minimized by the curve
(6.5)
γi (t) = θi − log
(1 − t) n1 + tqi (1)
,
(1 − t) n1 + tqn (1)
1 ≤ i ≤ n − 1.
In particular, for this minimizing curve we have
1
1
+ tq(1).
(6.6)
q(t) = (1 − t)
,...,
n
n
Proof. Fix a smooth curve γ : [0, 1] → Rn−1 from θ to φ. Since − log is
convex, by Jensen’s inequality we have
Z 1
Z 1
1
1
− log
q̇n (t)dt
+ q̇n (t) dt ≥ − log
+
n
n
0
0
(6.7)
1
= − log
+ qn (1) − qn (0) = − log qn (1) = ψ(θ − φ).
n
For the curve defined by (6.5), q̇(t) = q(1) − n1 is constant and so equality
holds in (6.7). Finally (6.6) follows by a direct calculation.
6.3. Displacement interpolation. We work under the following setting.
Let P (0) and P (1) be Borel probability measures on Rn−1 . Let ϕ : ∆n → R
be an exponentially concave function, satisfying Assumption 2.5, such that
F (1) := ∇c f is an optimal transport map (here f is the c-concave function
ϕ + ψ). Let π (1) : ∆n → ∆n be the portfolio map generated by ϕ(1) := ϕ.
Consider the flow (t, θ) 7→ φ(t) (θ) defined by the minimizing curves (6.5),
i.e.,
(t)
(6.8)
(t)
φi (θ) = θi − log
π i (θ)
(t)
,
1 ≤ i ≤ n − 1,
t ∈ [0, 1],
π n (θ)
where each π (t) : ∆n → ∆n is the portfolio map defined by
1
1
(t)
(6.9)
π := (1 − t)
,...,
+ tπ (1) , t ∈ [0, 1].
n
n
The following is the main result of this section. It is interesting to note that
the displacement interpolation can be interpreted naturally as the linear interpolation between the equal-weighted portfolio and the terminal portfolio.
Theorem 6.2 (Displacement interpolation).
Section 6.3.
Consider the setting of
42
S. PAL AND T.-K. L. WONG
(i) For each t ∈ [0, 1], the portfolio map π (t) is generated by the exponentially concave function ϕ(t) on ∆n defined by
ϕ(t) (p) = (1 − t)
(6.10)
n
X
1
log pi + tϕ(p),
n
p ∈ ∆n .
i=1
(ii) For each t ∈ [0, 1], let f (t) = ϕ(t) + ψ and let F (t) = ∇c f (t) . If θ is
distributed according to P (0) , then θ(t) is distributed according to P (t)
where
P (0) .
P (t) = F (t)
#
Moreover, F (t) is an optimal transport map for the transport problem
for (P (0) , P (t) ).
(iii) Endow ∆n with the geometric structure induced by the L-divergence of
ϕ. We further assume that the c-gradient F (1) = ∇c f : Rn−1 → Rn−1
is surjective. For each θ ∈ Rn−1 fixed, consider the curve t 7→ φ(t) (θ)
in dual coordinates. Then the trace of the curve is the dual geodesic
joining θ and φ(1) (θ).
Proof. (i) Follows directly from Example 2.6(iv).
(ii) It is clear that (θ, F (t) (θ)) is a coupling of (P (0) , P (t) ). By Proposition
2.7, the graph of the map F (t) is c-cyclical monotone. This proves that F (t)
is an optimal transport map.
(iii) We write (6.8) in the form
(t)
e−φi = e−θi
=
(1 − t) n1 + tπ i (θ)
(1 − t) n1 + tπ n (θ)
π (θ)
(1 − t) n1
tπ n (θ)
−θi +log π i (θ)
−θi
n
+
e
e
(1 − t) n1 + tπ n (θ)
(1 − t) n1 + tπ n (θ)
(1)
=: (1 − h(t))e−θi + h(t)e−φi
(θ)
.
By (5.8) we see that t 7→ φ(t) is a time change of a dual geodesic. The
surjectivity assumption guarantees that the curve lies within Y 0 , the range
of the dual coordinate system.
6.4. Another interpolation. From the financial perspective there is another natural interpolation, namely the linear interpolation between the
market portfolio µ (Example 2.6(i)) and the portfolio π:
(6.11)
π (t) = (1 − t)π + tµ.
EXPONENTIALLY CONCAVE FUNCTIONS
43
The corresponding log generating function is ϕ(t) = (1−t)ϕ. From the transport perspective, the market portfolio corresponds to the trivial transport
map F (θ) ≡ 0 (recall in Proposition 2.7(iii) that the portfolio has exponential coordinate given by θ − F (θ)). By the argument of Theorem 6.2 we have
the following result.
Proposition 6.3. Consider the geometric structure induced by ϕ and
assume that the range of the dual coordinate system is Rn−1 . Consider the
flow (t, θ) 7→ φ(t) (θ) in (6.8) where π (t) is given by the interpolation (6.11).
Then for each θ, in dual coordinates, the trace of the curve t 7→ φ(t) (θ) is a
time change of the dual geodesic from φ(0) (θ) to 0.
APPENDIX A: TECHNICAL PROOFS
Proof of Theorem 3.2. In this proof we treat θ and φ as independent
variables.
We prove (i) and (ii) together. We begin by observing that
(A.1)
∂
f (θ) = π i (θ),
∂θi
1 ≤ i ≤ n − 1.
To see this, write pi = eθi −ψ(θ) . Switching coordinates and using the chain
rule, we have
n
X ∂ϕ
∂
∂ϕ
f (θ) =
(p)(−pi pj ) +
(p)pi + pi
∂θi
∂pj
∂pi
j=1
n
X
∂ϕ
∂ϕ
= pi 1 +
(p) −
(p) .
pj
∂pi
∂pj
j=1
From (1.4) and a bit of computation, we see that this equals π i (θ).
Consider the c-transform of f given by
f ∗ (φ) = inf (ψ(θ − φ) − f (θ)) .
(A.2)
θ∈X
Differentiating ψ(θ − φ) − f (θ) and using (A.1), we see that θ ∈ X attains
the infimum in (A.2) if and only if
eθi −φi
Pn
= π i (θ),
θj −φj
j=1 e
i = 1, . . . , n − 1.
Rearranging, we have
(A.3)
φi = θi − log
π i (θ)
,
π n (θ)
i = 1, . . . , n − 1.
44
S. PAL AND T.-K. L. WONG
This proves that equality holds in
(A.4)
f (θ) + f ∗ (φ) ≤ ψ(θ − φ)
if and only if θ and φ satisfies the relation (A.3). In particular, for θ ∈ X
the c-supergradient ∇c f (θ) is given by (A.3).
Next we we prove that the minimizer in (A.2), if exists, is unique. Consider
instead maximization of the quantity
ef (θ)−ψ(θ−φ) = eϕ(θ)+ψ(θ)−ψ(θ−φ) = Φ(θ)
eψ(θ)
eψ(θ−φ)
.
Expanding and switching to Euclidean coordinates, this equals
Pn θi
e
1
P
(A.5)
Φ(p) n i=1θ −φ = Φ(p) Pn
,
i
i
e
i=1 ai pi
i=1
where ai = e−φi > 0. Being the quotient of a strictly concave function and
an affine function, the right hand side of (A.5) is strictly quasi-concave, i.e.,
its superlevel sets are strictly convex (see [11, Example 3.38]). This shows
that the minimizer θ in (A.2) is unique if it exists.
Let φ ∈ Y 0 . Then there exists unique θ ∈ X such that equality holds in
(A.4) and φ = ∇c f (θ). In particular, the c-supergradient ∂ c f ∗ (φ) is θ and
∇c f ∗ (∇c f (θ)) = θ. This completes the proof of (i) and (ii).
Next we prove that ∇c f ∗ = (∇c f )−1 is smooth, so that ∇c f : X → Y 0 is a
diffeomorphism. Since ∇c f : X → Y is smooth and injective, by the inverse
function theorem it suffices to show that the Jacobian of ∇c f is invertible on
X . Invertibility of the Jacobian follows from (4.16) and the fact that (gij (ξ))
is strictly positive definite.
Finally, by Fenchel’s identity we can express f ∗ in the form
f ∗ (φ) = c(∇c f ∗ (φ), φ) − f (∇c f ∗ (φ)).
Since ∇c f ∗ and the cost function c are smooth, we see that f ∗ is smooth as
well.
Proof of Lemma 5.7. (i) To prove the first formula in (5.11), we com-
EXPONENTIALLY CONCAVE FUNCTIONS
45
pute, using (4.9) and (4.12),
(grad T (· | θp ) (θq ))i
=
n−1
X
g ij (θq )
j=1
=
n−1
X
j=1
=
n−1
X
j=1
∂
T (· | θp ) (θq )
∂θjq
!
n−1
X ∂θi
1 ∂θi q
1
q
(θ ) +
(θ ) (Πj (θq , θp ) − π j (θq ))
π j (θq ) ∂φj
π n (θq )
∂φk
k=1
Πj (θq , θp ) Πn (θq , θp )
−
π j (θq )
π n (θq )
∂θi q
(θ ).
∂φj
For the second formula, we first prove a
Claim. We have
n−1
X ∂φ`
∂
r
q
(θq )(π ` (θq ) − Π` (θr , θq )).
q T (θ | θ ) =
∂θj
∂θjq
(A.6)
`=1
To see this, we use (4.13), (4.5), and compute as follows:
∂
T (θr | θq )
∂θjq
= −Πj (θr , θq ) + π j (θq ) +
n
X
`=1
1 ∂π ` q
(θ )Π` (θr , θq )
π ` (θq ) ∂θjq
= −Πj (θr , θq ) + π j (θq )+
n
X
=
`=1
n−1
X
`=1
q
δ`j − π j (θ ) −
!
!
n−1
X
∂φ` q
r q
q ∂φm q
(θ ) −
π m (θ ) q (θ ) Π` (θ , θ )
∂θjq
∂θj
m=1
∂φ` q
(θ )(π ` (θq ) − Π` (θr , θq )).
∂θjq
46
S. PAL AND T.-K. L. WONG
Now we compute, using Theorem 4.5 and the symmetry of g ij ,
(grad T (r | ·) (q))i
=
n−1
X
g ij (θq )
j=1
=
n−1
X
j=1
`=1
n−1
X
∂φ` q
(θ )(π ` (θq ) − Π` (θr , θq ))
∂θjq
(π ` (θq ) − Π` (θr , θq ))·
`=1
n−1
X
j=1
=
!
n−1
X ∂θj
1 ∂θj q
1
(φ ) +
(φq ) ·
π i (θq ) ∂φi
π n (θq )
∂φk
k=1
n−1
X
=
∂
T (θr | ·) (θq )
∂θjq
n−1
X
!
n−1
X ∂φ`
∂θ
1 ∂φ` q ∂θj q
1
j
(φ )
(φq )
(φq )
(φ ) +
π i (θq ) ∂θjq
∂φi
π n (θq )
∂φk
∂θjq
k=1
q
r
q
(π ` (θ ) − Π` (θ , θ ))
`=1
=−
1
1
δ`i +
q
π i (θ )
π n (θq )
Πi (θr , θq ) Πn (θr , θq )
+
.
π i (θq )
π n (θq )
In the second last equality we used (4.6). The proof of (ii) is similar.
ACKNOWLEDGEMENTS
The authors would like to thank Prof. Robin Graham for helpful discussions. We also thank the anonymous referee for suggestions which improved
the paper.
REFERENCES
[1] Acciaio, B., Beiglböck, M., Penkner, F. and Schachermayer, W. (2016).
A model-free version of the fundamental theorem of asset pricing and the superreplication theorem. Mathematical Finance 26 233–251.
[2] Amari, S.-i. (2016). Information Geometry and Its Applications. Springer.
[3] Amari, S.-i. and Nagaoka, H. (2007). Methods of Information Geometry 191.
American Mathematical Soc.
[4] Ambrosio, L. and Gigli, N. (2013). A users guide to optimal transport. In Modelling
and optimisation of flows on networks 1–155. Springer.
[5] Ay, N. and Amari, S.-i. (2015). A Novel Approach to Canonical Divergences within
Information Geometry. Entropy 17 8111–8129.
[6] Banner, A. D. and Fernholz, D. (2008). Short-term relative arbitrage in volatilitystabilized markets. Annals of Finance 4 445–454.
EXPONENTIALLY CONCAVE FUNCTIONS
47
[7] Basseville, M. (2013). Divergence measures for statistical data processing An
annotated bibliography. Signal Processing 93 621–633.
[8] Beiglböck, M., Henry-Labordere, P. and Penkner, F. (2013). Modelindependent bounds for option pricesa mass transport approach. Finance and
Stochastics 17 477–501.
[9] Booth, D. G. and Fama, E. F. (1992). Diversification returns and asset contributions. Financial Analysts Journal 48 26–32.
[10] Bouchey, P., Nemtchinov, V., Paulsen, A. and Stein, D. M. (2012). Volatility
Harvesting: Why Does Diversifying and Rebalancing Create Portfolio Growth? The
Journal of Wealth Management 15 26–35.
[11] Boyd, S. and Vandenberghe, L. (2004). Convex Optimization. Cambridge University Press.
[12] Bregman, L. M. (1967). The relaxation method of finding the common point of
convex sets and its application to the solution of problems in convex programming.
USSR computational mathematics and mathematical physics 7 200–217.
[13] Calin, O. and Udrişte, C. (2014). Geometric Modeling in Probability and Statistics.
Springer.
[14] Chambers, D. R. and Zdanowicz, J. S. (2014). The limitations of diversification
return. Journal of Portfolio Management 40 65–76.
[15] Cordero-Erausquin, D. and Klartag, B. (2015). Moment measures. Journal of
Functional Analysis 268 3834–3866.
[16] Cordero-Erausquin, D., McCann, R. J. and Schmuckenschläger, M. (2001).
A Riemmanian interpolation inequality à la Borell, Brascamp and Lieb. Invent. math.
146 219–257.
[17] Dillen, F., Nomizu, K. and Vranken, L. (1990). Conjugate connections and
Radon’s theorem in affine differential geometry. Monatshefte für Mathematik 109
221–235.
[18] Dolinsky, Y. and Soner, H. M. (2014). Martingale optimal transport and robust
hedging in continuous time. Probability Theory and Related Fields 160 391–427.
[19] Eguchi, S. (1983). Second order efficiency of minimum contrast estimators in a
curved exponential family. The Annals of Statistics 11 793–803.
[20] Eguchi, S. (1992). Geometry of minimum contrast. Hiroshima Math. J 22 631–647.
[21] Erb, C. B. and Harvey, C. R. (2006). The strategic and tactical value of commodity
futures. Financial Analysts Journal 62 69–97.
[22] Erbar, M., Kuwada, K. and Sturm, K.-T. (2015). On the equivalence of the
entropic curvature-dimension condition and Bochner’s inequality on metric measure
spaces. Inventiones mathematicae 201 993–1071.
[23] Fernholz, R. (1999). Portfolio generating functions. In Quantitative Analysis in
Financial Markets (M. Avellaneda, ed.) World Scientific.
[24] Fernholz, R. (2001). Equity portfolios generated by functions of ranked market
weights. Finance and Stochastics 5 469–486.
[25] Fernholz, E. R. (2002). Stochastic Portfolio Theory. Applications of Mathematics.
Springer.
[26] Fernholz, R., Karatzas, I. and Kardaras, C. (2005). Diversity and relative arbitrage in equity markets. Finance and Stochastics 9 1–27.
[27] Fernholz, R. and Shay, B. (1982). Stochastic portfolio theory and stock market
equilibrium. The Journal of Finance 37 615–624.
[28] Hallerbach, W. G. (2014). Disentangling rebalancing return. Journal of Asset
Management 15 301–316.
[29] Hazan, E., Agarwal, A. and Kale, S. (2007). Logarithmic regret algorithms for
48
S. PAL AND T.-K. L. WONG
online convex optimization. Machine Learning 69 169–192.
[30] Juditsky, A., Rigollet, P., Tsybakov, A. B. et al. (2008). Learning by mirror
averaging. The Annals of Statistics 36 2183–2206.
[31] Karatzas, I. and Ruf, J. (2016). Trading Strategies Generated by Lyapunov Functions. ArXiv e-prints 1603.08245.
[32] Kass, R. E. and Vos, P. W. (2011). Geometrical Foundations of Asymptotic Inference 908. John Wiley & Sons.
[33] Kurose, T. (1990). Dual connections and affine geometry. Mathematische Zeitschrift
203 115–121.
[34] Lott, J. and Villani, C. (2009). Ricci curvature for metric-measure spaces via
optimal transport. Annals of Mathematics 903–991.
[35] Mahdavi, M., Zhang, L. and Jin, R. (2015). Lower and upper bounds on the
generalization of stochastic exponentially concave optimization. In Proceedings of
The 28th Conference on Learning Theory 1305–1320.
[36] Matsuzoe, H. (1999). Geometry of contrast functions and conformal geometry. Hiroshima Mathematical Journal 29 175–191.
[37] McCann, R. J. (1994). A convexity theory for interacting gases and equilibrium
crystals PhD thesis, Princeton University.
[38] McCann, R. J. (1997). A convexity principle for interacting gases. Advances in
Mathematics 128 153–179.
[39] Murray, M. K. and Rice, J. W. (1993). Differential Geometry and Statistics. CRC
Press.
[40] Nagaoka, H. and Amari, S.-i. (1982). Differential geometry of smooth families of
probability distributions. Univ. Tokyo, Tokyo, Japan, METR 82–7.
[41] Oliker, V. (2007). Embedding Sn into Rn+1 with given integral Gauss curvature
and optimal mass transport on Sn . Adv. Math. 213 600-620.
[42] Pal, S. (2016). Exponentially concave functions and high dimensional stochastic
portfolio theory. ArXiv e-prints 1603.01865.
[43] Pal, S. (2017). Embedding optimal transports in statistical manifolds. To appear in
The Indian Journal of Pure and Applied Mathematics.
[44] Pal, S. and Wong, T.-K. L. (2013). Energy, entropy, and arbitrage. ArXiv e-prints
1308.5376.
[45] Pal, S. and Wong, T.-K. L. (2016). The geometry of relative arbitrage. Mathematics
and Financial Economics 10 263–293.
[46] Rockafellar, R. T. (1997). Convex Analysis. Princeton Landmarks in Mathematics. Princeton University Press.
[47] Shima, H. (2007). The Geometry of Hessian Structures. World Scientific.
[48] Sternberg, S. (2012). Curvature in Mathematics and Physics. Dover.
[49] Strong, W. (2014). Generalizations of functionally generated portfolios with applications to statistical arbitrage. SIAM Journal on Financial Mathematics 5 472–492.
[50] Villani, C. (2003). Topics in Optimal Transportation. Graduate Studies in Mathematics 58. American Mathematical Soc.
[51] Villani, C. (2008). Optimal Transport: Old and New. Springer.
[52] Willenbrock, S. (2011). Diversification return, portfolio rebalancing, and the commodity return puzzle. Financial Analysts Journal 67 42–49.
[53] Wong, T.-K. L. (2015a). Optimization of relative arbitrage. Annals of Finance 11
345–382.
[54] Wong, T.-K. L. (2015b). Universal Portfolios in Stochastic Portfolio Theory. ArXiv
e-prints 1510.02808.
EXPONENTIALLY CONCAVE FUNCTIONS
University of Washington
C-547 Padelford Hall
Seattle, Washington 98195
USA
E-mail: [email protected]
49
University of Southern California
406H Kaprielian Hall
Los Angeles, California 90089
USA
E-mail: [email protected]
| 7cs.IT
|
O N C ONVERGENCE AND S TABILITY OF GAN S
arXiv:1705.07215v5 [cs.AI] 10 Dec 2017
Naveen Kodali, Jacob Abernethy, James Hays & Zsolt Kira ∗
College of Computing
Georgia Institute of Technology
Atlanta, GA 30332, USA
{nkodali3,prof,hays,zkira}@gatech.edu
A BSTRACT
We propose studying GAN training dynamics as regret minimization, which is in
contrast to the popular view that there is consistent minimization of a divergence
between real and generated distributions. We analyze the convergence of GAN
training from this new point of view to understand why mode collapse happens. We
hypothesize the existence of undesirable local equilibria in this non-convex game
to be responsible for mode collapse. We observe that these local equilibria often
exhibit sharp gradients of the discriminator function around some real data points.
We demonstrate that these degenerate local equilibria can be avoided with a gradient
penalty scheme called DRAGAN. We show that DRAGAN enables faster training,
achieves improved stability with fewer mode collapses, and leads to generator
networks with better modeling performance across a variety of architectures and
objective functions.
1
I NTRODUCTION
Generative modeling involves taking a set of samples drawn from an unknown data generating
distribution Preal and finding an estimate Pmodel that closely resembles it. Generative adversarial
networks (GAN) (Goodfellow et al., 2014) is a powerful framework used for fitting implicit generative
models. The basic setup consists of two networks, the generator and the discriminator, playing against
each other in a repeated zero-sum game setting. The goal here is to reach an equilibrium where
Preal , Pmodel are close, and the alternating gradient updates procedure (AGD) is used to achieve this.
However, this process is highly unstable and often results in mode collapse (Goodfellow, 2017). This
calls for an deeper investigation into training dynamics of GANs.
In this paper, we propose studying GAN training dynamics as a repeated game in which both the
players are using no-regret algorithms (Cesa-Bianchi & Lugosi, 2006) and discuss how AGD 1 falls
under this paradigm. In contrast, much of the theory (Goodfellow et al., 2014; Arjovsky & Bottou,
2017) and recent developments (Nowozin et al., 2016; Arjovsky et al., 2017; Gulrajani et al., 2017)
are based on the unrealistic assumption that the discriminator is playing optimally (in the function
space) at each step and as a result, there is consistent minimization of a divergence between real and
generated distributions. This corresponds to at least one player using the best-response algorithm
(in the function space), and the resulting game dynamics can be completely different in both these
cases (Nisan et al., 2007). Thus, there is a clear disconnect between theoretical arguments used as
motivation in recent literature and what actually happens in practice.
We would like to point out that the latter view can still be useful for reasoning about the asymptotic
equilibrium situation but we argue that regret minimization is the more appropriate way to think
about GAN training dynamics. So, we analyze the convergence of GAN training from this new
point of view to understand why mode collapse happens. We start with a short analysis of the
artificial convex-concave case of the GAN game in section 2.2. This setting has a unique solution and
guaranteed convergence (of averaged iterates) using no-regret algorithms can be shown with standard
arguments from game theory literature. Here, we make explicit, the critical (previously not widely
known) connection between AGD used in GAN training and regret minimization. This immediately
∗
1
code - https://github.com/kodalinaveen3/DRAGAN
Most of our analysis applies to the simultaneous gradient updates procedure as well
1
yields a novel proof for the asymptotic convergence of GAN training, in the non-parametric limit.
Prior to our work, such a result (Goodfellow et al., 2014) required a strong assumption that the
discriminator is optimal at each step.
However, these convergence results do not hold when the game objective function is non-convex,
which is the practical case when deep neural networks are used. In non-convex games, global
regret minimization and equilibrium computation are computationally hard in general. Recent gametheoretic literature indicates that AGD can end up cycling (Mertikopoulos et al., 2017) or converging
to a (potentially bad) local equilibrium, under some conditions (Hazan et al., 2017). We hypothesize
these to be the reasons for cycling and mode collapse observed during GAN training, respectively
(section 2.3). In this work, we do not explore the cycling issue but focus our attention on the mode
collapse problem. In contrast to our hypothesis, the prevalent view of mode collapse and instability
(Arjovsky & Bottou, 2017) is that it results from attempting to minimize a strong divergence during
training. However, as we argued earlier, GAN training with AGD does not consistently minimize
a divergence and therefore, such a theory is not suitable to discuss convergence or to address the
stability issue.
Next, if mode collapse is indeed the result of an undesirable local equilibrium, a natural question
then is how we can avoid it? We make a simple observation that, in the GAN game, mode collapse
situations are often accompanied by sharp gradients of the discriminator function around some real
data points (section 2.4). Therefore, a simple strategy to mitigate mode collapse is to regularize the
discriminator so as to constrain its gradients in the ambient data space. We demonstrate that this
improves the stability using a toy experiment with one hidden layer neural networks. This gives rise
to a new explanation for why WGAN and gradient penalties might be improving the stability of GAN
training – they are mitigating the mode collapse problem by keeping the gradients of the discriminator
function small in data space. From this motivation, we propose a training algorithm involving a
novel gradient penalty scheme called DRAGAN (Deep Regret Analytic Generative Adversarial
Networks) which enables faster training, achieves improved stability and modeling performance (over
WGAN-GP (Gulrajani et al., 2017) which is the state-of-the-art stable training procedure) across a
variety of architectures and objective functions.
Below, we provide a short literature review. Several recent works focus on stabilizing the training
of GANs. While some solutions (Radford et al., 2015; Salimans et al., 2016) require the usage of
specific architectures (or) modeling objectives, some (Che et al., 2016; Zhao et al., 2016) significantly
deviate from the original GAN framework. Other promising works in this direction (Metz et al.,
2016; Arjovsky et al., 2017; Qi, 2017; Gulrajani et al., 2017) impose a significant computational
overhead. Thus, a fast and versatile method for consistent stable training of GANs is still missing in
the literature. Our work is aimed at addressing this.
To summarize, our contributions are as follows:
• We propose a new way of reasoning about the GAN training dynamics - by viewing AGD
as regret minimization.
• We provide a novel proof for the asymptotic convergence of GAN training in the nonparametric limit and it does not require the discriminator to be optimal at each step.
• We discuss how AGD can converge to a potentially bad local equilibrium in non-convex
games and hypothesize this to be responsible for mode collapse during GAN training.
• We characterize mode collapse situations with sharp gradients of the discriminator function
around some real data points.
• A novel gradient penalty scheme called DRAGAN is introduced based on this observation
and we demonstrate that it mitigates the mode collapse issue.
2
T HEORETICAL A NALYSIS OF GAN TRAINING DYNAMICS
We start with a brief description of the GAN framework (section 2.1). We discuss guaranteed
convergence in the artificial convex-concave case using no-regret algorithms, and make a critical
connection between GAN training process (AGD) and regret minimization (section 2.2). This
immediately yields a novel proof for the asymptotic convergence of GAN training in the nonparametric limit. Then, we consider the practical non-convex case and discuss how AGD can
2
converge to a potentially bad local equilibrium here (section 2.3). We characterize mode collapse
situations with sharp gradients of the discriminator function around real samples and this provides
an effective strategy to avoid them. This naturally leads to the introduction of our gradient penalty
scheme DRAGAN (section 2.4). We end with a discussion and comparison with other gradient
penalties in the literature (section 2.5).
2.1
BACKGROUND
The GAN framework can be viewed as a repeated zero-sum game, consisting of two players - the
generator, which produces synthetic data given some noise source and the discriminator, which is
trained to distinguish generator’s samples from the real data. The generator model G is parameterized
by φ, takes a noise vector z as input, and produces a synthetic sample Gφ (z). The discriminator model
D is parameterized by θ, takes a sample x as input and computes Dθ (x), which can be interpreted as
the probability that x is real.
The models G, D can be selected from any arbitrary class of functions – in practice, GANs typical
rely on deep networks for both. Their cost functions are defined as
J (D) (φ, θ) := −Ex∼preal log Dθ (x) − Ez log(1 − Dθ (Gφ (z))), and
J (G) (φ, θ) := −J (D) (φ, θ)
And the complete game can be specified as n
o
min max J(φ, θ) = Ex∼preal log Dθ (x) + Ez log(1 − Dθ (Gφ (z)))
φ
θ
The generator distribution Pmodel asymptotically converges to the real distribution Preal if updates
are made in the function space and the discriminator is optimal at each step (Goodfellow et al., 2014).
2.2
C ONVEX - CONCAVE CASE AND N O - REGRET ALGORITHMS
According to Sion’s theorem (Sion, 1958), if Φ ⊂ Rm , Θ ⊂ Rn such that they are compact and
convex sets, and the function J : Φ × Θ → R is convex in its first argument and concave in its second,
then we have min max J(φ, θ) = max min J(φ, θ)
φ∈Φ θ∈Θ
θ∈Θ φ∈Φ
That is, an equilibrium is guaranteed to exist in this setting where players’ payoffs correspond to the
unique value of the game (Neumann, 1928).
A natural question then is how we can find such an equilibrium. A simple procedure that players
can use is best-response algorithms (BRD). In each round, best-responding players play their
optimal strategy given their opponent’s current strategy. Despite its simplicity, BRD are often
computationally intractable and they don’t lead to convergence even in simple games. In contrast,
a technique that is both efficient and provably works is regret minimization. If both players
update their parameters using no-regret algorithms, then it is easy to show that their averaged iterates will converge to an equilibrium pair (Nisan et al., 2007). Let us first define no-regret algorithms.
Definition 2.1 (No-regret algorithm). Given a sequence of convex loss functions L1 , L2 , . . . :
K → R, an algorithm that selects a sequence of kt ’s, each of which may only depend on previously
)
observed L1 , . . . , Lt−1 , is said to have no regret if R(T
= o(1), where we define
T
PT
PT
R(T ) := t=1 Lt (kt ) − mink∈K t=1 Lt (k)
We can apply no-regret learning to our problem of equilibrium finding in the GAN game J(·, ·) as
follows. The generator imagines the function J(·, θt ) as its loss function on round t, and similarly
the discriminator imagines −J(φt , ·) as its loss function at t. After T rounds of play, each player
PT
PT
computes the average iterates φ̄T := T1 t=1 φt and θ̄T := T1 t=1 θt . If V ∗ is the equilibrium
value of the game, and the players suffer regret R1 (T ) and R2 (T ) respectively, then one can show
using standard arguments (Freund & Schapire, 1999) that V∗−
R2 (T )
T
≤ maxθ∈Θ J(φ̄T , θ) −
R2 (T )
T
≤ minφ∈Φ J(φ, θ̄T ) +
3
R1 (T )
T
≤V∗+
R1 (T )
T .
In other words, θ̄T and φ̄T are "almost optimal" solutions to the game, where the "almost"
2 (T )
approximation factor is given by the average regret terms R1 (T )+R
. Under the no-regret
T
condition, the former will vanish, and hence we can guarantee convergence in the limit. Next, we
define a popular family of no-regret algorithms.
Definition 2.2 (Follow The Regularized Leader). FTRL (Hazan et al., 2016) selects kt on round
Pt−1
t by solving for arg mink∈K { s=1 Ls (k) + η1 Ω(k)}, where Ω(·) is some convex regularization
function and η is a learning rate.
Remark: Roughly speaking, if you select the regularization as Ω(·) = 12 k · k2 , then FTRL becomes
the well-known online gradient descent or OGD (Zinkevich, 2003). Ignoring the case of constraint
violations, OGD can be written in a simple iterative form: kt = kt−1 − η∇Lt−1 (kt−1 ).
The typical GAN training procedure using alternating gradient updates (or simultaneous gradient
updates) is almost this - both the players applying online gradient descent. Notice that the min/max
objective function in GANs involves a stochastic component, with two randomized inputs given on
each round, x and z which are sampled from the data distribution and a standard multivariate normal,
respectively. Let us write Jx,z (φ, θ) := log Dθ (x) + log(1 − Dθ (Gφ (z))). Taking expectations with
respect to x and z, we define the full (non-stochastic) game as J(φ, θ) = Ex,z [Jx,z (φ, θ)]. But
the above online training procedure is still valid with stochastic inputs. That is, the equilibrium
computation would proceed similarly, where on each round we sample xt and zt , and follow the
updates
φt+1 ← φt − η∇φ Jxt ,zt (φt , θt ).
and
0
θt+1 ← θt + η ∇θ Jxt ,zt (φt , θt )
On a side note, a benefit of this stochastic perspective is that we can get a generalization bound on
the mean parameters φ̄T after T rounds of optimization. The celebrated "online-to-batch conversion"
(Cesa-Bianchi et al., 2004) implies that Ex,z [Jx,z (φ̄T , θ)], for
than the optimal value
h any θ, is no more
i
2 (T )
Ex,z [Jx,z (φ∗ , θ)] plus an "estimation error" bounded by E R1 (T )+R
, where the expectation is
T
taken with respect to the sequence of samples observed along the way, and any randomness in the
algorithm. Analogously, this applies to θ̄T as well. A limitation of this result, however, is that it
requires a fresh sample xt to be used on every round.
To summarize, we discussed in this subsection about how the artificial convex-concave case is easy to
solve through regret minimization. While this is a standard result in game theory and online learning
literature, it is not widely known in the GAN literature. For instance, Salimans et al. (2016) and
Goodfellow (2017) discuss a toy game which is convex-concave and show cycling behavior. But,
the simple solution in that case is to just average the iterates. Further, we made explicit, the critical
connection between regret minimization and alternating gradient updates procedure used for GAN
training. Goodfellow et al. (2014) argue that, if G and D have enough capacity (in the non-parametric
limit) and updates are made in the function space, then GAN game can be considered convex-concave.
Thus, our analysis based on regret minimization immediately yields a novel proof for the asymptotic
convergence of GANs, which does not require the discriminator to be optimal at each step.
Moreover, the connection between regret minimization and GAN training process gives a novel
way to reason about its dynamics. In contrast, the popular view of GAN training as consistently
minimizing a divergence arises if the discriminator uses BRD (in the function space) and thus, it has
little to do with the actual training process of GANs. As a result, this calls into question the motivation
behind many recent developments like WGAN and gradient penalties among others, which improve
the training stability of GANs. In the next subsection, we discuss the practical non-convex case and
why training instability arises. This provides the necessary ideas to investigate mode collapse from
our new perspective.
2.3
N ON - CONVEX CASE AND L OCAL EQUILIBRIA
In practice, we choose G, D to be deep neural networks and the function J(φ, θ) need not be convexconcave anymore. The nice properties we had in the convex-concave case like the existence of a
unique solution and guaranteed convergence through regret minimization no longer hold. In fact,
regret minimization and equilibrium computation are computationally hard in general non-convex
4
settings. However, analogous to the case of non-convex optimization (also intractable) where we
focus on finding local minima, we can look for tractable solution concepts in non-convex games.
Recent work by Hazan et al. (2017) introduces the notion of local regret and shows that if both
the players use a smoothed variant of OGD to minimize this quantity, then the non-convex game
converges to some form of local equilibrium, under mild assumptions. The usual training procedure
of GANs (AGD) corresponds to using a window size of 1 in their formulation. Thus, GAN training
will eventually converge (approximately) to a local equilibrium which is described below or the
updates will cycle. We leave it to future works to explore the equally important cycling issue and
focus here on the former case.
Definition 2.3 (Local Equilibrium). A pair (φ∗ , θ∗ ) is called an -approximate local equilibrium if
it holds that
0
0
0
∀φ , ||φ − φ∗ || ≤ η : J(φ∗ , θ∗ ) ≤ J(φ , θ∗ ) +
0
0
0
∀θ , ||θ − θ∗ || ≤ η : J(φ∗ , θ∗ ) ≥ J(φ∗ , θ ) −
That is, in a local equilibrium, both the players do not have much of an incentive to switch to any
other strategy within a small neighborhood of their current strategies. Now, we turn our attention to
the mode collapse issue which poses a significant challenge to the GAN training process. The training
is said to have resulted in mode collapse if the generator ends up mapping multiple z vectors to the
same output x, which is assigned a high probability of being real by the discriminator (Goodfellow,
2017). We hypothesize this to be the result of the game converging to bad local equilibria.
The prevalent view of mode collapse and instability in GAN training (Arjovsky & Bottou, 2017)
is that it is caused due to the supports of real and model distributions being disjoint or lying on
low-dimensional manifolds. The argument is that this would result in strong distance measures like
KL-divergence or JS-divergence getting maxed out, and the generator cannot get useful gradients to
learn. In fact, this is the motivation for the introduction of WGAN (Arjovsky et al., 2017). But, as we
argued earlier, GAN training does not consistently minimize a divergence as that would require using
intractable best-response algorithms. Hence, such a theory is not suitable to discuss convergence
or to address the instability of GAN training. Our new view of GAN training process as regret
minimization is closer to what is used in practice and provides an alternate explanation for mode
collapse - the existence of undesirable local equilibria. The natural question now is how we can avoid
them?
2.4
M ODE COLLAPSE AND G RADIENT P ENALTIES
The problem of dealing with multiple equilibria in games and how to avoid undesirable ones is
an important question in algorithmic game theory (Nisan et al., 2007). In this work, we constrain
ourselves to the GAN game and aim to characterize the undesirable local equilibria (mode collapse)
in an effort to avoid them. In this direction, after empirically studying multiple mode collapse cases,
we found that it is often accompanied by the discriminator function having sharp gradients around
some real data points (See Figure 1 2 ). This intuitively makes sense from the definition of mode
collapse discussed earlier. Such sharp gradients encourage the generator to map multiple z vectors
to a single output x and lead the game towards a degenerate equilibrium. Now, a simple strategy to
mitigate this failure case would be to regularize the discriminator using the following penalty
λ · Ex∼Preal ,δ∼Nd (0,cI) k∇x Dθ (x + δ)k2
This strategy indeed improves the stability of GAN training. We show the results of a toy experiment
with one hidden layer neural networks in Figure 2 and Figure 3 to demonstrate this. This partly
explains the success of WGAN and gradient penalties in the recent literature (Gulrajani et al., 2017;
Qi, 2017), and why they improve the training stability of GANs, despite being motivated by reasoning
based on unrealistic assumptions. However, we noticed that this scheme in its current form can be
brittle and if over-penalized, the discriminator can end up assigning both a real point x and noise
x + δ, the same probability of being real. Thus, a better choice of penalty is
λ · Ex∼Preal ,δ∼Nd (0,cI) max 0, k∇x Dθ (x + δ)k2 − k
2
At times, stochasticity seems to help in getting out of the basin of attraction of a bad equilibrium
5
Figure 1: One hidden layer networks as G and D (MNIST). On the left, we plot inception score
against time for vanilla GAN training and on the right, we plot the squared norm of discriminator’s
gradients around real data points for the same experiment. Notice how this quantity changes before,
during and after mode collapse events.
Figure 2: One hidden layer networks as G and D (MNIST). On the left, losses for both the players
are shown for vanilla GAN training and on the right, we added a regularization term to penalize the
gradients of D(x) around real data points. Notice the improved stability.
Finally, due to practical optimization considerations (this has also been observed in Gulrajani et al.
(2017)), we instead use the penalty shown below in all our experiments.
2
λ · Ex∼Preal ,δ∼Nd (0,cI) k∇x Dθ (x + δ)k − k
(1)
This still works as long as small perturbations of real data, x + δ are likely to lie off the data-manifold,
which is true in the case of image domain and some other settings. Because, in these cases, we
do want our discriminator to assign different probabilities of being real to training data and noisy
samples. We caution the practitioners to keep this important point in mind while making their choice
of penalty. All of the above schemes have the same effect of constraining the norm of discriminator’s
gradients around real points to be small and can therefore, mitigate the mode collapse situation. We
refer to GAN training using these penalty schemes or heuristics as the DRAGAN algorithm.
Additional details:
• We use the vanilla GAN objective in our experiments, but our penalty improves stability
using other objective functions as well. This is demonstrated in section 3.3.
• The penalty scheme used in our experiments is the one shown in equation 1.
• We use small pixel-level noise but it is possible to find better ways of imposing this penalty.
However, this exploration is beyond the scope of our paper.
• The optimal configuration of the hyperparameters for DRAGAN depends on the architecture,
dataset and data domain. We set them to be λ ∼ 10, k = 1 and c ∼ 10 in most of our
experiments.
2.5
C OUPLED VS L OCAL P ENALTIES
Several recent works have also proposed regularization schemes which constrain the discriminator’s
gradients in the ambient data space, so as to improve the stability of GAN training. Despite being
6
Figure 3: One hidden layer networks as G and D (MNIST). On the left, inception score plot is shown
for vanilla GAN training and on the right, we added a regularization term to penalize the gradients of
D(x) around real data points. Notice how mode collapse is mitigated.
from different motivations, WGAN-GP and LS-GAN are closely related approaches to ours. First,
we show that these two approaches are very similar, which is not widely known in the literature. Qi
(2017) introduced LS-GAN with the idea of maintaining a margin between losses assigned to real and
fake samples. Further, they also impose Lipschitz constraint on D and the two conditions together
result in a situation where the following holds for any real and fake sample pair (roughly) Dθ (x) − Dθ (Gφ (z)) ≈ ||x, Gφ (z)||
(2)
The authors argue that the resulting discriminator function would have non-vanishing gradients almost
everywhere between real and fake samples (section 6 of Qi (2017)). Next, Gulrajani et al. (2017)
proposed an extension to address various shortcomings of the original WGAN and they impose the
following condition on D ||∇x Dθ (x̂)|| ≈ 1
(3)
where x̂ = ()x+(1−)Gφ (z) is some point on the line between a real and a fake sample, both chosen
independently at random. This leads to D having norm-1 gradients almost everywhere between
real and fake samples. Notice that this behavior is very similar to that of LS-GAN’s discriminator
function. Thus, WGAN-GP is a slight variation of the original LS-GAN algorithm and we refer to
these methods as “coupled penalties”.
On a side note, we also want to point out that WGAN-GP’s penalty doesn’t actually follow from
KR-duality as claimed in their paper. By Lemma 1 of Gulrajani et al. (2017), the optimal discriminator
D∗ will have norm-1 gradients (almost everywhere) only between those x and Gφ (z) pairs which
are sampled from the optimal coupling or joint distribution π ∗ . Therefore, there is no basis for
WGAN-GP’s penalty (equation 3) where arbitrary pairs of real and fake samples are used. This
fact adds more credence to our theory regarding why gradient penalties might be mitigating mode
collapse.
The most important distinction between coupled penalties and our methods is that we only impose
gradient constraints in local regions around real samples. We refer to these penalty schemes as “local
penalties”. Coupled penalties impose gradient constraints between real and generated samples and
we point out some potential issues that arise from this:
• With adversarial training finding applications beyond fitting implicit generative models,
penalties which depend on generated samples can be prohibitive.
• The resulting class of functions when coupled penalties are used will be highly restricted
compared to our method and this affects modeling performance. We refer the reader to
Figure 4 and appendix section 5.2.2 to see this effect.
• Our algorithm works with AGD, while WGAN-GP needs multiple inner iterations to
optimize D. This is because the generated samples can be anywhere in the data space and
they change from one iteration to the next. In contrast, we consistently regularize Dθ (x)
only along the real data manifold.
To conclude, appropriate constraining of the discriminator’s gradients can mitigate mode collapse
but we should be careful so that it doesn’t have any negative effects. We pointed out some issues
7
with coupled penalties and how local penalties can help. We refer the reader to section 3 for further
experimental results.
Figure 4: Swissroll experiment (different phases of training) - Vanilla GAN (top), WGAN-GP
(middle), and DRAGAN (bottom). Real samples are marked orange and generated samples are green.
Level sets of Dθ (x) are shown in the background where yellow is high and purple is low.
3
E XPERIMENTAL R ESULTS
In section 3.1, we compare the modeling performance of our algorithm against vanilla GAN and
WGAN variants in the standard DCGAN/CIFAR-10 setup. Section 3.2 demonstrates DRAGAN’s
improved stability across a variety of architectures. In section 3.3, we show that our method also
works with other objective functions. Appendix contains samples for inspection, some of the missing
plots and additional results. Throughout, we use inception score (Salimans et al., 2016) which is a
well-studied and reliable metric in the literature, and sample quality to measure the performance.
3.1
I NCEPTION S CORES FOR CIFAR-10 USING DCGAN ARCHITECTURE
DCGAN is a family of architectures designed to perform well with the vanilla training procedure.
They are ubiquitous in the GAN literature owing to the instability of vanilla GAN in general settings.
We use this architecture to model CIFAR-10 and compare against vanilla GAN, WGAN and WGANGP. As WGANs need 5 discriminator iterations for every generator iteration, comparing the modeling
performance can be tricky. To address this, we report two scores for vanilla GAN and DRAGAN one using the same number of generator iterations as WGANs and one using the same number of
discriminator iterations. The results are shown in Figure 5 and samples are included in the appendix
(Figure 8). Notice that DRAGAN beats WGAN variants in both the configurations, while vanilla
GAN is only slightly better. A key point to note here is that our algorithm is fast compared to WGANs,
so in practice, the performance will be closer to the DRAGANd case. In the next section, we will
show that if we move away from this specific architecture family, vanilla GAN training can become
highly unstable and that DRAGAN penalty mitigates this issue.
3.2
M EASURING STABILITY AND PERFORMANCE ACROSS ARCHITECTURES
Ideally, we would want our training procedure to perform well in a stable fashion across a variety of
architectures (other than DCGANs). Similar to Arjovsky et al. (2017) and Gulrajani et al. (2017), we
remove the stabilizing components of DCGAN architecture and demonstrate improved stability &
modeling performance compared to vanilla GAN training (see appendix section 5.2.3). However, this
is a small set of architectures and it is not clear if there is an improvement in general.
To address this, we introduce a metric termed the BogoNet score to compare the stability & performance of different GAN training procedures. The basic idea is to choose random architectures for
players G and D independently, and evaluate the performance of different algorithms in the resulting
8
Algorithm
WGAN
WGAN-GP
DRAGANg
DRAGANd
Vanilla GANg
Vanilla GANd
Score
3.25
5.99
6.11
6.90
6.3
6.99
(b) Inception scores
(a) Inception score plot
Figure 5: Comparison of modeling performance on CIFAR10
Table 1: Summary of inception score statistics across 100 architectures
Algorithm
Vanilla GAN
DRAGAN
WGAN-GP
Final score
Mean Std
2.91
1.44
3.70
1.71
3.49
1.30
Area under curve
Mean
Std
277.72 126.09
312.15 135.35
300.09 100.96
Qual. score
Total
92.5
157.5
-
games. A good algorithm should achieve stable performance without failing to learn or resulting in
mode collapse, despite the potentially imbalanced architectures. In our experiment, each player is
assigned a network from a diverse pool of architectures belonging to three different families (MLP,
ResNet, DCGAN).
To demonstrate that our algorithm performs better compared to vanilla GAN training and WGAN-GP,
we created 100 such instances of hard games. Each instance is trained using these algorithms on
CIFAR-10 (under similar conditions for a fixed number of generator iterations, which gives a slight
advantage to WGAN-GP) and we plot how inception score changes over time. For each algorithm,
we calculated the average of final inception scores and area under the curve (AUC) over all 100
instances. The results are shown in Table 1. Notice that we beat the other algorithms in both metrics,
which indicates some improvement in stability and modeling performance.
Further, we perform some qualitative analysis to verify that BogoNet score indeed captures the
improvements in stability. We create another set of 50 hard architectures and compare DRAGAN
against vanilla GAN training. Each instance is allotted 5 points and we split this bounty between the
two algorithms depending on their performance. If both perform well or perform poorly, they get 2.5
points each, so that we nullify the effect of such non-differentiating architectures. However, if one
algorithm achieves stable performance compared to the other (in terms of failure to learn or mode
collapses), we assign it higher portions of the bounty. Results were judged by two of the authors in a
blind manner: The curves were shown side-by-side with the choice of algorithm for each side being
randomized and unlabeled. The vanilla GAN received an average score of 92.5 while our algorithm
achieved an average score of 157.5 and this correlates with BogoNet score from earlier. See appendix
section 5.3 for some additional details regarding this experiment.
3.3
S TABILITY USING DIFFERENT OBJECTIVE FUNCTIONS
Our algorithm improves stability across a variety of objective functions and we demonstrate this
using the following experiment. Nowozin et al. (2016) show that we can interpret GAN training
as minimizing various f -divergences when an appropriate game objective function is used. We
show experiments using the objective functions developed for Forward KL, Reverse KL, Pearson χ2 ,
Squared Hellinger, and Total Variation divergence minimization. We use a hard architecture from the
previous subsection to demonstrate the improvements in stability. Our algorithm is stable in all cases
except for the total variation case, while the vanilla algorithm failed in all the cases (see Figure 6 for
two examples and Figure 15 in appendix for all five). Thus, practitioners can now choose their game
9
objective from a larger set of functions and use DRAGAN (unlike WGANs which requires a specific
objective function).
(a) Reverse KL
(b) Pearson χ2
Figure 6: Inception score plots for two divergence measures, demonstrating superior stability for our
algorithm.
4
C ONCLUSIONS
In this paper, we propose to study GAN training process as regret minimization, which is in contrast
to the popular view that there is consistent minimization of a divergence between real and generated
distributions. We analyze the convergence of GAN training from this new point of view and
hypothesize that mode collapse occurs due to the existence of undesirable local equilibria. A
simple observation is made about how the mode collapse situation often exhibits sharp gradients of
the discriminator function around some real data points. This characterization partly explains the
workings of previously proposed WGAN and gradient penalties, and motivates our novel penalty
scheme. We show evidence of improved stability using DRAGAN and the resulting improvements in
modeling performance across a variety of settings. We leave it to future works to explore our ideas in
more depth and come up with improved training algorithms.
R EFERENCES
Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial
networks. arXiv preprint arXiv:1701.04862, 2017.
Martin Arjovsky, Soumith Chintala, and Léon Bottou.
arXiv:1701.07875, 2017.
Wasserstein gan.
arXiv preprint
Nicolo Cesa-Bianchi and Gábor Lugosi. Prediction, learning, and games. Cambridge university
press, 2006.
Nicolo Cesa-Bianchi, Alex Conconi, and Claudio Gentile. On the generalization ability of on-line
learning algorithms. IEEE Transactions on Information Theory, 50(9):2050–2057, 2004.
Tong Che, Yanran Li, Athul Paul Jacob, Yoshua Bengio, and Wenjie Li. Mode regularized generative
adversarial networks. arXiv preprint arXiv:1612.02136, 2016.
Yoav Freund and Robert E Schapire. Adaptive game playing using multiplicative weights. Games
and Economic Behavior, 29(1-2):79–103, 1999.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling,
C. Cortes, N. D. Lawrence, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp. 2672–2680. Curran Associates, Inc., 2014. URL http://papers.nips.
cc/paper/5423-generative-adversarial-nets.pdf.
10
Ian J. Goodfellow. NIPS 2016 tutorial: Generative adversarial networks. CoRR, abs/1701.00160,
2017. URL http://arxiv.org/abs/1701.00160.
Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron Courville. Improved
training of wasserstein gans. arXiv preprint arXiv:1704.00028, 2017.
Elad Hazan, Karan Singh, and Cyril Zhang. Efficient regret minimization in non-convex games.
arXiv preprint arXiv:1708.00075, 2017.
Elad Hazan et al. Introduction to online convex optimization. Foundations and Trends® in Optimization, 2(3-4):157–325, 2016.
Panayotis Mertikopoulos, Christos Papadimitriou, and Georgios Piliouras. Cycles in adversarial
regularized learning. arXiv preprint arXiv:1709.02738, 2017.
Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial
networks. CoRR, abs/1611.02163, 2016. URL http://arxiv.org/abs/1611.02163.
J. von Neumann. Zur theorie der gesellschaftsspiele. Mathematische Annalen, 100:295–320, 1928.
URL http://eudml.org/doc/159291.
Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V Vazirani. Algorithmic game theory,
volume 1. Cambridge University Press Cambridge, 2007.
Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. f-gan: Training generative neural samplers
using variational divergence minimization. In Advances in Neural Information Processing Systems,
pp. 271–279, 2016.
Guo-Jun Qi. Loss-sensitive generative adversarial networks on lipschitz densities.
abs/1701.06264, 2017. URL http://arxiv.org/abs/1701.06264.
CoRR,
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks. arXiv:1511.06434 [cs], November 2015. URL
http://arxiv.org/abs/1511.06434. arXiv: 1511.06434.
Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen.
Improved techniques for training gans. CoRR, abs/1606.03498, 2016. URL http://arxiv.
org/abs/1606.03498.
Maurice Sion. On general minimax theorems. Pacific J. Math, 8(1):171–176, 1958.
Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network. arXiv
preprint arXiv:1609.03126, 2016.
Martin Zinkevich. Online convex programming and generalized infinitesimal gradient ascent. In
Proceedings of the 20th International Conference on Machine Learning (ICML-03), pp. 928–936,
2003.
11
5
5.1
A PPENDIX
S AMPLES AND L ATENT S PACE WALKS
In this section, we provide samples from an additional experiment run on CelebA dataset (Figure 7).
The samples from the experiment in section 3.1 are shown in Figure 8. Further, Radford et al. (2015)
suggest that walking on the manifold learned by the generator can expose signs of memorization. We
use DCGAN architecture to model MNIST and CelebA datasets using DRAGAN penalty, and the
latent space walks of the learned models are shown in Figure 9 and Figure 10. The results demonstrate
that the generator is indeed learning smooth transitions between different images, when our algorithm
is used.
Figure 7: Modeling CelebA with DRAGAN using DCGAN architecture.
12
(a) Vanilla GAN
(b) DRAGAN
(c) WGAN
(d) WGAN-GP
Figure 8: Modeling CIFAR-10 using DCGAN architecture.
13
(a)
(b)
Figure 9: Latent space walk of the model learned on MNIST using DRAGAN
Figure 10: Latent space walk of the model learned on CelebA using DRAGAN
5.2
5.2.1
A DDITIONAL E XPERIMENTS
O NE HIDDEN LAYER NETWORK TO MODEL MNIST
We design a simple experiment where G and D are both fully connected networks with just one
hidden layer. Vanilla GAN performs poorly even in this simple case and we observe severe mode
collapses. In contrast, our algorithm is stable throughout and obtains decent quality samples despite
the constrained setup.
14
Inception Score
Inception Score
Generator iterations
Generator iterations
(a) Vanilla GAN
(b) DRAGAN
Figure 11: One hidden layer network to model MNIST - Inception score plots
(a) Vanilla GAN
(b) DRAGAN
Figure 12: One hidden layer network to model MNIST - Samples
5.2.2
8-G AUSSIANS E XPERIMENT
We analyze the performance of WGAN-GP and DRAGAN on the 8-Gaussians dataset. As it can be
seen in Figure 13, both of them approximately converge to the real distribution but notice that in
the case of WGAN-GP, Dθ (x) seems overly constrained in the data space. In contrast, DRAGAN’s
discriminator is more flexible.
Improved WGAN
DRAGAN
15
Figure 13: Comparing the performance of WGAN-GP and DRAGAN on the 8-Gaussians dataset.
Orange is real samples, green is generated samples. The level sets of Dθ (x) are shown in the
background, with yellow as high and purple as low.
5.2.3
S TABILITY ACROSS DCGAN A RCHITECTURE VARIATIONS
DCGAN architecture has been designed following specific guidelines to make it stable (Radford
et al., 2015). We restate the suggested rules here.
1. Use all-convolutional networks which learn their own spatial downsampling (discriminator)
or upsampling (generator)
2. Remove fully connected hidden layers for deeper architectures
3. Use batch normalization in both the generator and the discriminator
4. Use ReLU activation in the generator for all layers except the output layer, which uses tanh
5. Use LeakyReLU activation in the discriminator for all layers
We show below that such constraints can be relaxed when using our algorithm and still maintain
training stability. Below, we present a series of experiments in which we remove different stabilizing
components from the DCGAN architecture and analyze the performance of our algorithm. Specifically,
we choose the following four architectures which are difficult to train (in each case, we start with
base DCGAN architecture and apply the changes) • No BN and a constant number of filters in the generator
• 4-layer 512-dim ReLU MLP generator
• tanh nonlinearities everywhere
• tanh nonlinearity in the generator and 4-layer 512-dim LeakyReLU MLP discriminator
Notice that, in each case, our algorithm is stable while the vanilla GAN training fails. A similar
approach is used to demonstrate the stability of training procedures in Arjovsky et al. (2017) and
Gulrajani et al. (2017).
(a) tanh activation
(b) FC generator
16
(c) Constant filter
(d) FC discriminator
Figure 14: Comparing performance of DRAGAN and Vanilla GAN training in the hard variations of
DCGAN architecture.
5.2.4
S TABILITY ACROSS OBJECTIVE FUNCTIONS
Due to space limitations, we only showed plots for two cases in section 3.3. Below we show the
results for all five cases.
(a) Reverse KL
(b) Pearson χ2
(c) Forward KL
(d) Total Variation
17
(e) Squared Hellinger
Figure 15: Comparing performance of DRAGAN and Vanilla GAN training using different objective
functions.
5.3
B OGO N ET D ETAILS
We used three families of architectures with probabilities - DCGAN (0.6), ResNet (0.2), MLP (0.2).
Next, we further parameterized each family to create additional variation. For instance, the DCGAN
family can result in networks with or without batch normalization, have LeakyReLU or Tanh nonlinearities. The number and width of filters, latent space dimensionality are some other possible
variations in our experiment. Similarly, the number of layers and hidden units in each layer for MLP
are chosen randomly. For ResNets, we chose their depth randomly. This creates a set of hard games
which test the stability of a given training algorithm.
We showed qualitative analysis of the inception score plots in section 3.2 to verify that BogoNet score
indeed captures the improvements in stability. Below, we show some examples of how the bounty
splits were done. The plots in Figure 14 were scored as (averages are shown in DRAGAN, Vanilla
GAN order):
A - (5, 0), B - (3.5, 1.5), C – (2.25, 2.75), D – (2, 3)
18
| 9cs.NE
|
1
Uplink Channel Estimation and Data Transmission
in Millimeter-Wave CRAN with Lens Antenna
Arrays
arXiv:1802.02076v1 [cs.IT] 6 Feb 2018
Reuben George Stephen, Student Member, IEEE, and Rui Zhang, Fellow, IEEE
Abstract—Millimeter-wave (mmWave) communication and
network densification hold great promise for achieving highrate communication in next-generation wireless networks. Cloud
radio access network (CRAN), in which low-complexity remote
radio heads (RRHs) coordinated by a central unit (CU) are
deployed to serve users in a distributed manner, is a costeffective solution to achieve network densification. However, when
operating over a large bandwidth in the mmWave frequencies,
the digital fronthaul links in a CRAN would be easily saturated
by the large amount of sampled and quantized signals to be
transferred between RRHs and the CU. To tackle this challenge,
we propose in this paper a new architecture for mmWavebased CRAN with advanced lens antenna arrays at the RRHs.
Due to the energy focusing property, lens antenna arrays are
effective in exploiting the angular sparsity of mmWave channels,
and thus help in substantially reducing the fronthaul rate and
simplifying the signal processing at the multi-antenna RRHs
and the CU, even when the channels are frequency-selective.
We consider the uplink transmission in a mmWave CRAN with
lens antenna arrays and propose a low-complexity quantization
bit allocation scheme for multiple antennas at each RRH to
meet the given fronthaul rate constraint. Further, we propose
a channel estimation technique that exploits the energy focusing
property of the lens array and can be implemented at the CU
with low complexity. Finally, we compare the proposed mmWave
CRAN using lens antenna arrays with a conventional CRAN
using uniform planar arrays at the RRHs, and show that the
proposed design achieves significant throughput gains, yet with
much lower complexity.
Index Terms—Cloud radio access network, millimeter-wave
communication, lens antenna array, channel estimation, fronthaul constraint, antenna selection, quantization bit allocation.
I. I NTRODUCTION
N
ETWORK densification by increasing the densities of
base stations (BSs)/access points (APs) deployed, and
millimeter-wave (mmWave) communication by exploiting
large unused bandwidth in the higher frequencies of the radio
spectrum, are two key strategies for achieving the orders of
magnitude data rate improvement required for future wireless
communication networks [1]. On one hand, cloud radio access
This paper was presented in part at the IEEE Global Communications
Conference (GLOBECOM), Singapore, Dec. 4–8, 2017.
R. G. Stephen is with the National University of Singapore Graduate School
for Integrative Sciences and Engineering, National University of Singapore,
Singapore 117456, and also with the Department of Electrical and Computer
Engineering, National University of Singapore, Singapore 117583 (e-mail:
[email protected]).
R. Zhang is with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore 117583 (e-mail:
[email protected]).
network (CRAN), in which distributed low-complexity remote
radio heads (RRHs) are deployed close to the users, and
coordinated by a central unit (CU) for joint processing, provides a cost-effective way of achieving network densification.
Thanks to the centralized resource allocation and joint signal
processing for the RRHs at the CU, CRAN achieves significant improvements in both the spectral efficiency and energy
efficiency compared to the conventional cellular network [2]–
[9], while RRHs can be simple relay nodes that transmit
and receive quantized/compressed baseband signals to/from
the CU over their fronthaul links. On the other hand, with
advances in radio frequency (RF) circuits, wireless communication over mmWave bands has emerged as a promising technology to achieve high-rate communications, due to the large
bandwidth available and the beamforming gain brought about
by the possibility of deploying a large number of antennas at
the transceivers thanks to the small wavelengths [10], [11].
Thus, CRAN when integrated with mmWave communication,
achieves the double goals of network densification and ample
bandwidth at the same time for future wireless networks. However, in such a dense CRAN operating over a large mmWave
bandwidth, the digital fronthaul links would be easily saturated
by the large volume of sampled and quantized/compressed
baseband signals that need to be transmitted between the CU
and RRHs; thus, it is crucial to find cost-effective solutions to
reduce the transmission rate required for each fronthaul link.
In CRANs operating over the conventional cellular frequency bands, a considerable body of prior work (e.g. [2]–[5])
has investigated various techniques for data compression at the
RRHs to achieve fronthaul rate reduction. Further, channel estimation in networks with coordinated multi-antenna BSs has
also been considered in prior work [12], [13]. In [12], random
matrix theory was used to derive an approximate lower bound
on the uplink ergodic achievable rate in a system with multiple
single-antenna users and multi-antenna BSs, while [13] considered an estimate-compress-forward approach for the uplink
of a CRAN and proposed various schemes to optimize the
ergodic achievable sum rate subject to backhaul constraints.
Most of the above techniques for compression and/or channel estimation are applicable for relatively small bandwidth
compared to that in mmWave, and typically involve complex
signal processing and cooperative signal compression across
the RRHs, which are difficult to implement for mmWave
systems due to practical cost and complexity considerations.
A low-complexity training sequence design for CRAN was
considered in [14], where the problem of minimizing the train-
2
z-axis (wavelengths)
5
0
Antenna element
EM Lens
−5
0
2
4
x-axis (wavelengths)
6
−4
−2
0
2
4
y-axis (wavelengths)
focusing antennas that can be selected for signal sampling
and quantization. Thus, the use of lens antenna arrays can
potentially achieve significant reduction in both the fronthaul
rate requirement for transmitting the quantized signals to the
CU and the interference among the users for the joint decoding
at the CU in the uplink transmission. However, due to the
finite fronthaul rate constraint at each RRH, it is crucial to
design antenna selection and signal quantization schemes to
maximize the achievable user rates at the CU. Since the RRHs
in a CRAN are typically simple relay nodes, in this paper, we
consider that simple uniform scalar quantization (SQ) of the
sampled baseband signals is performed independently over the
antennas selected at each RRH. The major contributions of this
paper are summarized as follows.
•
Fig. 1. Illustration of lens antenna array with Dy = Dz = 10, Θ− =
π
, Θ+ = π2 , Φ− = Φ+ = π2 .
6
•
ing length, while maintaining local orthogonality among the
training sequences of the users, was considered; however, the
fronthaul constraints at the RRHs were ignored. For frequencyselective mmWave channels, channel estimation for hybrid
precoding was considered in [15], [16]. The approach is to
represent the channel taps in the angular domain corresponding
to a set of quantized angles, and then use sparse signal
processing techniques to estimate the relevant parameters.
In this paper, we propose a new architecture for mmWave
CRAN by leveraging the use of advanced lens antenna arrays [17]–[22] at the RRHs. A full-dimensional lens antenna
array [19] consists of an electromagnetic (EM) lens with
energy focusing capability integrated with an antenna array
whose elements are located on the focal surface of the
lens (see Fig. 1). The amplitude response of a lens array can
be expressed as a “sinc”-type function in terms of the angles
of arrival/departure of plane waves incident on/transmitted
from it, and the locations of the antenna elements [18], [19].
Hence, by appropriately designing the locations of the antenna
elements on the focal surface, the lens array is capable of
focusing most of the energy from a uniform plane wave
arriving in a particular direction onto a specific antenna
element or subset of elements. Moreover, due to the multipath sparsity of mmWave channels [11], a lens array can
be used to achieve the capacity of a point-to-point mmWave
channel with multiple antennas via a technique called path
division multiplexing [18], which uses simple single-carrier
modulation even for transmission over wideband frequencyselective channels, and has low signal processing complexity.
With lens antenna arrays, the angular domain sparsity of
mmWave channels is transformed to the spatial domain, which
then enables lower complexity signal processing for channel
estimation and data transmission. Moreover, in a CRAN, since
the angles of arrival of the signals from different users are
typically independent, and thus different at each RRH, the
user signals are effectively separated over a small number of
•
•
•
We introduce a new architecture for CRAN in mmWave
frequencies using full-dimensional lens antenna arrays [18], [19] at the RRHs, taking into consideration
both the elevation and azimuth angles of arrival of the
signals from the users in the uplink transmission.
We propose a simple energy-detection based antenna
selection at each RRH and a low-complexity quantization
bit allocation algorithm over the selected antennas to
minimize the total quantization noise power based only
on the estimates of the received signal power at different
antennas, subject to the fronthaul rate constraint.
When the CU has perfect channel state information (CSI),
we show that the proposed system with lens antenna
arrays can achieve better sum-rate performance compared to a conventional CRAN using uniform planar
arrays (UPAs) at the RRHs and orthogonal frequency
division multiplexing (OFDM) transmission by the users,
when the same quantization algorithm is employed in
both cases.
Under imperfect CSI, we propose a reduced-size, approximate MMSE beamforming, by exploiting the energyfocusing property of lens antenna arrays. In the proposed
scheme, for each user, only the data streams where the
estimated channel gains are larger than a certain threshold
are chosen for beamforming, while the interference on
these streams is also approximated by thresholding the
channel gain estimates.
With the proposed bit allocation at the RRHs, channel
estimation at the CU, and data transmission with singlecarrier modulation, we compare the proposed system with
the conventional CRAN with UPAs via simulations, and
show that when the fronthaul is constrained, the proposed
system can achieve significant sum-throughput gains at
much lower signal processing complexity and training
overhead compared to the benchmark.
The rest of this paper is organized as follows. In Section II we
present the system model for the proposed mmWave CRAN
with lens arrays, and carry out an analysis of the achievable
rates under perfect and imperfect CSI, with our proposed bit
allocation and channel estimation schemes. Section III gives a
brief description of the benchmark CRAN system with UPAs
and OFDM transmission. In Section IV, we compare our
proposed system with the benchmark system via simulations.
3
Finally, Section V concludes the paper.
Notation: In this paper, , denotes equality by definition,
and ∼ means “distributed as”. The cardinality of a finite set S
is denoted by |S|, while A \ B denotes the elements in A that
are not in B. Sets of M ×N real, complex, and integer matrices
are denoted by RM×N , CM×N and ZM×N respectively, while
R+ and Z++ denote the set of non-negative real numbers
and positive integers, respectively. δ[n] denotes the Kronecker
delta function. The normalized sinc function is defined as
for x 6= 0 and sinc(x) , 1 if x = 0. The
sinc(x) , sin(πx)
πx
imaginary unit is denoted by with 2 = 1. Scalars are denoted
by lower-case letters, e.g. x, while vectors and matrices are
denoted by bold-face lower-case and upper-case letters, e.g.
x and X, respectively. For x ∈ R, ⌈x⌉ denotes the smallest
integer greater than or equal to x, and ⌊x⌋ denotes the largest
integer less than or equal to x. For x ∈ C, |x| ≥ 0 denotes its
magnitude and ∠x ∈ [0, 2π) denotes its phase in radian. For
a vector x, kxk denotes its Euclidean norm. A vector with
all elements equal to 0 is denoted by 0, where the dimension is implied from the context. For vectors and matrices,
()T denotes transpose, and ()H denotes conjugate transpose
(Hermitian). For matrix X, tr(X) denotes the sum of its
diagonal elements (trace). For X with linearly independent
columns, X † , (X H X)−1 X H denotes the Moore-Penrose
pseudo-inverse. X ⊗ Y denotes the Kronecker product of two
matrices X and Y , while IN denotes the identity matrix of
dimension N . A diagonal matrix with elements x1 , . . . , xM on
the main diagonal is denoted by diag( x1 ··· xM ), and a block
diagonal matrix by blkdiag( X1 ··· XM ). CN (µ, Σ) denotes a
circularly symmetric complex Gaussian (CSCG) distribution
centered at µ with covariance Σ, and U[a, b] denotes a uniform
distribution over the interval [a, b].
II. S YSTEM M ODEL
We study the uplink transmission in a mmWave-based dense
CRAN cluster (see Fig. 2) with M multi-antenna RRHs,
denoted by M = {1, . . . , M }. The cluster is sectorized,
where each RRH has J = 3 sectors covering 120◦ each,
and each sector is served by a full-dimensional lens antenna
array with Q antenna elements denoted by the set Q(j) =
{1, . . . , Q} , j ∈ J , where J = {1, . . . , J} denotes the
set of sectors. Further, we use Q without the superscript to
denote the set of S
all antenna elements in all sectors of an
RRH, i.e., Q = j∈J Q(j) , and |Q| = JQ. Each RRH
is connected to the CU via an individual fronthaul link of
finite capacity R̄m > 0 in bits per second (bps). There are
K single-antenna users in the CRAN cluster, denoted by
K = {1, . . . , K}. All the users and RRHs share the same
bandwidth for communication. Depending on a user’s location,
its signals are incident on at most one sector of an RRH, and
we model any inter-sector/inter-cluster leakage interference as
Gaussian noise. We further assume that the total number of
sectors in the CRAN is greater than or equal to the number
of users, i.e., JM ≥ K. The users transmit over a mmWave,
frequency-selective, block-fading channel of bandwidth W Hz.
Fronthaul
link
RRH m
Central
Unit (CU)
mmWave
wireless link
User k
Lens array
CRAN
cluster
R̄m
Fig. 2. Schematic of CRAN.
The channel between user k and RRH m1 has Lm,k paths
denoted by Lm,k = {1, . . . , Lm,k }. Then the discrete-time
channel coefficient vector hm,k [n] ∈ CQ×1 at time index
n between user k and RRH m can be expressed using a
geometric channel model as
X
αm,k,ℓ a θm,k,ℓ , φm,k,ℓ δ[n − dm,k,ℓ ]
hm,k [n] =
ℓ∈Lm,k
=
X
ℓ∈Lm,k
hm,k,ℓ δ[n − dm,k,ℓ ],
(1)
where αm,k,ℓ ∈ C and dm,k,ℓ ∈ {0, 1, . . . , dmax } , D denote,
respectively, the complex gain and delay in symbol periods
corresponding to path ℓ ∈ Lm,k , and a θm,k,ℓ , φm,k,ℓ ∈
CQ×1 is the array response for the elevation and azimuth
angles of arrival of path ℓ denoted by θm,k,l and φm,k,l ,
respectively.
We consider that for each sector, each RRH is equipped
with a rectangular EM lens in the y − z plane, with
dimensions Dy × Dz normalized by the wavelength2 along
the y- and z-axes, respectively. The EM lens is followed by
a full-dimensional antenna array with Q elements placed on
the focal surface of the lens, which is a hemisphere around
the lens’ center (taken to be the origin in Fig. 1) with radius
equal to the focal length of the lens (see Fig. 1). Let each
antenna element be indexed by a pair of indexes (qe , qa ),
where qe denotes the index in the elevation direction along
the focal surface of the lens and qa denotes the index in the
azimuth direction along the focal surface. Now, for a ray
drawn from the center of the lens to an antenna element
(qe , qa ), let φqa ∈ [−Φ− , Φ+ ] denote the azimuth angle made
by the ray, where Φ− , Φ+ ∈ (0, π/2] are the maximum
azimuth angles in the negative and positive y-directions,
respectively. Similarly, let θqe ∈ [−Θ− , Θ+ ] denote the
elevation angle made by the ray, where Θ− , Θ+ ∈ (0, π/2]
are the maximum elevation angles covered by the antenna
array in the negative and positive z-directions, respectively.
Then, the antennas are placed such that the indexes qe
run from the integers −⌊Dz sin Θ− ⌋ to ⌊Dz sin Θ+ ⌋.
1 Since each user’s signals are received by only one sector of every RRH, we
refer to the channel between the user and the RRH, instead of the particular
sector of the RRH, for convenience.
2 The dimensions are assumed to be same at all RRHs for convenience.
4
Ta
Td
Tp
Power probing Guard
Guard
for antenna
interval Pilot symbols interval
selection
dmax
ADC/Quantization
1
Data symbols
ℜ{y1,m }
AGC/
Power
est.
1
RF to
baseband
Tf
Q
Fig. 3. Frame structure for uplink transmission with lens antenna arrays.
Lens
Antenna
array
AGC/
Power
est.
ℑ{y1,m }
RF
switches
dmax
SQ
ℜ{yQm ,m }
RF to
Qm baseband
To CU via
fronthaul
SQ
ℑ{yQm ,m }
y̌1,m
SQ
SQ
y̌Qm ,m
Bit allocation/
antenna
selection
Thus, the elevation angles of the antenna elements θqe
are related to the indexes qe by sin θqe = qe /Dz , qe ∈ Fig. 4. Architecture of RRH with lens antenna array.
{−⌊Dz sin Θ− ⌋, . . . , −1, 0, 1, . . . , ⌊Dz sin Θ+ ⌋}.
Next,
for each index qe , the index qa runs from the integers
A. Uniform Scalar Quantization (SQ) and Bit Allocation at
−⌊Dy cos θqe sin Φ− ⌋ to ⌊Dy cos θqe sin Φ⌋, so that the
azimuth angles of the antenna elements φqa are related RRHs
During the power-probing stage, the users transmit constant
to the indexes qa by sin φqa = qa /(Dy cos θqe ), qa ∈
amplitude signals for a duration Ta ≥ dmax + 1, and each
{−⌊Dy cos θqe sin Φ− ⌋, . . . , −1, 0, 1, . . . , ⌊Dy cos θqe sin Φ⌋} [19].
Then the amplitude response in (1) of the lens array element RRH m obtains an estimate of the average received power
(qe , qa ) ≡ q ∈ Q(j) , to a uniform plane wave incident at ρq,m on each antenna either from the AGC circuitry, or using
elevation and azimuth angles (θ, φ), can be expressed as [20] analog power estimators. We consider that each RRH performs
uniform SQ on the real and imaginary components of the
complex baseband samples received at each antenna q ∈ Q
p
aq (θ, φ) = Dz Dy sinc (qe − Dz sin θ)
and
denoted by yq,m [n], independently,3 using bq,m ∈ Z++
· sinc qa − Dy cos θ sin φ ,
(2) bits. If bq,m = 0, the symbols on antenna q are not forwarded
to the CU by RRH m (i.e., this antenna is not selected for
subsequent channel training and data transmission). Following
q ∈ Q(j) , j ∈ J . Since the RRHs have fronthaul links the design in [8], the resulting quantized samples y̌q,m [n] can
of finite capacities, they must quantize the received signals be expressed as
before forwarding to the CU. As the RRHs are typically
y̌q,m [n] = yq,m [n] + eq,m [n],
(3)
simple nodes with limited processing capability, we consider
that they perform uniform SQ independently on each antenna,
where eq,m [n] represents the quantization error, modeled as a
with the SQ bit allocation adapted in each channel coherence
random variable with mean zero and variance given by [8]
interval. As shown in Fig. 4, the RRHs perform bit allocation
based on the estimated received power levels at each antenna
ε2q,m , E[|eq,m [n]|2 ] = 3ρq,m /22bq,m , bq,m ∈ Z++ . (4)
element, which can be obtained before converting the signals
to the baseband, either using feedback from the automatic The quantization error is assumed to be uncorrelated with
gain control (AGC) circuitry, or by means of analog power yq,m [n], and as the SQ is performed independently at each
estimators [23], which can be implemented using band-pass antenna and for each sample, we have E[eq,m [n]e∗q′ ,m [n′ ]] = 0
filters and envelope detectors. We consider a frame-based for any q ′ 6= q or n′ 6= n, ∀m ∈ M. With Nyquist
transmission by the users, with frame duration Tf < Tc in rate sampling, the transmission rate required
P to forward the
quantized
signals
over
all
the
antennas
is
symbol periods, where Tc denotes the minimum coherence
q∈Q 2W bq,m bps,
time among all the user-RRH channels. Each frame is further which must not exceed the fronthaul capacity R̄m . Each RRH
m uses the estimate ρq,m of the average received power
divided into the following three stages as shown in Fig. 3:
computed from the received signals in the power-probing stage
• A power probing stage of duration Ta , where the users
to perform the SQ in the subsequent channel training and data
transmit constant amplitude signals in order to enable the transmission stages. We consider the design of quantization
RRHs to perform bit allocation and antenna selection;
bit allocation to minimize the total SQ noise power over all
• A channel training stage of duration Tp , where the users
the antennas subject to the fronthaul capacity constraint at the
transmit pilot symbols, and the CU performs channel RRH, as captured by the following optimization problem
estimation for the selected antennas using the quantized
X 3ρq,m
signals forwarded by the RRHs; and
(5)
minimize
bm
22bq,m
• A data transmission stage of duration Td , where the users
q∈Q
transmit their data, which is quantized and forwarded by
subject to
the RRHs for decoding at the CU.
X
R̄m
(5a)
bq,m ≤
2W
The above stages are separated by guard intervals of dmax
q∈Q
symbols (see Fig. 3), where dmax denotes the maximum delay
spread of all the user-RRH channels. Note that Tf = Ta +Tp +
3 This can be performed using low-cost, low-resolution analog-to-digital
Td + 2dmax . In the following, we describe each stage in detail. converters (ADCs).
5
T
where bm , [ b1,m ··· bQ,m ] ∈ ZJQ×1
++ . Since the bq,m ’s are
integers, the above problem is non-convex. However, if the
variables are relaxed so that bq,m ∈ R+ ∀q ∈ Q, we have the
following relaxed problem
X 3ρq,m
22bq,m
min
bm ∈RJQ×1
+
q∈Q
(6)
s.t. (5a)
which is convex since the objective function is convex,4 while
the constraint (5a) is linear. Thus, we have the following
proposition.
Proposition 2.1. The optimal solution to problem (6) is given
by
b′q,m = max
n1
2
log2
where λ⋆ ≥ 0 is such that
6ρ
P
q,m ln 2
λ⋆
′
q∈Q bq,m
o
, 0 , q ∈ Q,
TABLE I
A LGORITHM FOR QUANTIZATION BIT ALLOCATION
Initialize λmin = 0, λmax = 6ρmax,m ln 2, tolerance ǫ > 0
repeat
λ = (λmin + λmax )/2
Compute
bm according to (7)
P
if q∈Q bq,m ≤ R̄m /2W then
Set λmax = λ
else
Set λmin = λ
end if
until |λmax − λmin | < ǫ
Use λ⋆ = λ in (7) to compute b′m
Initialize βmin = 0, βmax = 1, and find β in (8) by bisection
similar to above until |βmax − βmin | < ǫ
13: Compute final integer solution b̃m to problem (5) according to (8)
with converged β
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
(7)
= R̄m /2W .
Proof. Please refer to the appendix.
Let Q′m denote the set of antennas with non-zero allocation
according to (7). With the optimal solution b′m to
problem (6), we proceed to construct a feasible integer solution
b̃m ∈ ZJQ×1
for the original problem (5) by rounding b′m as
++
follows
(
⌊b′q,m ⌋ if b′q,m − ⌊b′q,m ⌋ ≤ β
b̃q,m =
, q ∈ Q′m ,
(8)
⌈b′q,m ⌉ if b′q,m − ⌊b′q,m ⌋ > β
b′q,m
where β ∈ [0, 1] is an appropriate threshold. Notice that as
β is decreased, most of the b′q,m ’s would be rounded above,
making it more difficult to satisfy the constraint in (5) and
vice versa. Hence, a suitable β can be found by bisection on
the interval [0, 1], each time evaluating the constraint in (5)
with b̃m in (8) and updating β accordingly.
From (7), we observe that the optimal solution allocates
more bits to the antenna with higher estimated power ρq,m ,
which is desirable, since the antennas that receive stronger
signals from the users are more likely to have useful information to be decoded. The algorithm is summarized in Table I
and we denote the final set of selected antennas at RRH m as
Qm = {q ∈ Q | b̃q,m > 0} with |Qm | , Qm . We assume that
the fronthaul capacities are such that R̄m ≥ 2JW, ∀m ∈ M,
so that the number of selected antennas at each RRH is at
least Qm ≥ J, ∀m ∈ M. Then, since JM ≥ K, it is
always feasible to recover all K users’ signals at the CU
via linear processing with independent channels. Let Qtot ,
P
m∈M Qm denote the total number of selected antennas. For
the ease of exposition, we refer to the selected antennas as
“streams” and use the index i ∈ {1, . . . , Qtot } , I to denote
each stream so that the stream i corresponds to the antenna
(j)
qi in sector j of RRH mi .
4 For the relaxed problem, we assume that the objective function is defined
for bq,m = 0 as well.
B. Path Delay Compensation and Achievable Sum-Rate with
Perfect CSI
In this subsection, we describe the operations at the CU,
assuming perfect CSI. If the RRHs are equipped with lens
antenna arrays, we propose that the users transmit simultaneously with single-carrier modulation, so that the data
√ symbols
transmitted by user k are given by xd,k [n] = P sd,k [n],
where P ≥ 0 is the transmit power and sd,k [n] ∼ CN (0, 1)
is the complex data symbol of user k. Using (1) and (3), the
quantized symbols received at the CU corresponding to stream
i ∈ I can be expressed as
X X
y̌d,i [n] =
hi,k,ℓ xd,k [n − dmi ,k,ℓ ] + zd,i [n]
k∈K ℓ∈Lmi ,k
+ ed,i [n],
i ∈ I.
(9)
We assume that the CU decodes the users’ data symbols via
linear processing, after path delay compensation [20] on the
received signals for each stream as described below. Let ℓ⋆i,k =
arg maxℓ∈Lm ,k |hi,k,ℓ |2 denote the strongest path among all
i
those arriving on stream (selected antenna) i ∈ I from user
k. Due to the response of the lens array, the delayed versions
of a particular user’s signals with different angels of arrival
are focused on different antenna elements at each RRH in
general. Thus, to ensure that the symbols of each user that
have undergone the strongest path gain on each stream are
combined at the CU in a synchronized manner, each stream
(j)
of quantized symbols y̌d,i [n], corresponding to antenna qi of
RRH mi , is advanced by the delay dmi ,k,ℓ⋆i,k corresponding to
the path ℓ⋆i,k , to obtain the delay compensated signal y̌d,i,k n+
dmi ,k,ℓ⋆i,k , y̌¯d,i,k [n] for each user k ∈ K, as given by
y̌¯d,i,k [n]
X
=
X
k′ ∈K ℓ′ ∈Lmi ,k′
hi,k′ ,ℓ′ xd,k′ n − dmi ,k′ ,ℓ′ − dmi ,k,ℓ⋆i,k
+ zd,i n + dmi ,k,ℓ⋆i,k + ed,i n + dmi ,k,ℓ⋆i,k .
(10)
In order to write the summation over the paths in (10) in
terms of the delay differences ν = dmi ,k′ ,ℓ′ − dmi ,k,ℓ⋆i,k ∈
{0, ±1, . . . , ±dmax } , ∆ with the delay of the maximum gain
6
path dmi ,k,ℓ⋆i,k , we define for each antenna i ∈ I and user pair
k, k ′ ∈ K, the new channel coefficient
X
h̄i,kk′ [ν] ,
hi,k′ ,ℓ′ δ ν − dmi ,k′ ,ℓ′ − dmi ,k,ℓ⋆i,k ,
ℓ′ ∈Lmi ,k′
i ∈ I, k, k ′ ∈ K, ν ∈ ∆,
(11)
which is equivalent to the channel coefficient (or sum of
coefficients) corresponding to the path(s) ℓ′ ∈ Lmi ,k′ from
user k ′ to antenna i, which has (have) a delay difference of
ν with the maximum gain path ℓ⋆i,k of user k to the same
antenna. Then, (10) can be expressed as
X
h̄i,kk [ν]xd,k [n − ν]
y̌¯d,i,k [n] = h̄i,kk [0]xd,k [n] +
ν∈∆\{0}
+
X
X
k′ ∈K\{k} ν∈∆
|
|
{z
inter-symbol interference (ISI)
h̄i,kk′ [ν]xd,k′ [n − ν]
{z
inter-user interference (IUI)
}
}
+ z̄d,i,k [n] + ēd,i,k [n], i ∈ I, k ∈ K,
(12)
2
) and
where z̄d,i,k [n] , zd,i n + dmi ,k,ℓ⋆i,k ∼ CN (0, σm
i
ēd,i,k [n] , ed,i n + dmi ,k,ℓ⋆i,k , with E[ēd,i,k [n]] = 0 and
E[|ēd,i,k [n]|2 ] = ε2i defined in (4) denote the AWGN and
quantization noise samples shifted by dmi ,k,ℓ⋆i,k symbol periods. Note that y̌¯d,j,k [n] in (12) depends on user k whose
maximum gain path ℓ⋆i,k is used as reference. The second and
third terms in (12) represent user k’s own delayed symbols,
and the interfering symbols from other users, respectively.
Collecting the signals from all the streams, (12) can be written
in vector form as
X
¯d,k [n] = h̄kk [0]xd,k [n] +
h̄kk [ν]xd,k [n − ν]
y̌
ν∈∆\{0}
+
X X
k′ ∈K\k ν∈∆
h̄kk′ [ν]xd,k′ [n − ν]
+ z̄d,k [n] + ēd,k [n],
k ∈ K,
(13)
where all the vectors are of dimension Qtot × 1. We consider
¯k [n]
that the CU performs linear receive beamforming on y̌
with the beamforming vector uk ∈ CQtot ×1 to construct the
¯
estimate x̂d,k [n] = uH
k y̌d,k [n] of user k’s symbol. Treating the
inter-symbol and inter-user interference in (12) as Gaussian
noise, the signal to interference-plus-noise ratio (SINR) for
decoding xd,k [n] is given by (14) on the top of the next page.
2
2
IQM ) ∈ CQtot ×Qtot , and
IQ1 ··· σM
In (14), Σ , blkdiag(
σ1Q
2 Qtot
tot ×Qtot
, where ε2i , i ∈ I, is deΞ , diag {εi }i=1 ∈ C
fined in (4). Since the transmit powers of the users are fixed, γk
in (14) is maximized by uk = Ck−1 h̄kk [0], k ∈ K according
to the minimum mean squared error (MMSE) criterion [24],
where
X
P h̄kk [ν]h̄kk [ν]H
Ck ,
ν∈∆\{0}
+
X
X
k′ ∈K\{k} ν∈∆
P h̄kk′ [ν]h̄kk′ [ν]H + Σ + Ξ,
k ∈ K,
(15)
denotes the covariance of the noise and interference terms
in (13). In this case, the SINR in (14) becomes γk =
P h̄kk [0]H Ck−1 h̄kk [0], and by assuming the worst-case CSCG
distribution for the quantization noise, a lower bound on the
achievable sum rate over all users is thus given by
X
rlens =
log2 (1 + γk )
(16)
k∈K
in bps/Hz. In the next subsection, we consider channel estimation at the CU, and extend the achievable rate analysis to
the case with imperfect CSI.
C. Channel Estimation and Achievable Sum-Rate with Imperfect CSI
Since the CU does not know the CSI a priori, it needs
to estimate the CSI using pilot signals sent by the users and
then quantized and forwarded by the RRHs in the channel
The users transmit known pilot
estimation stage (see Fig. 3). √
symbols given by xp,k [n] = P sp,k [n], n = 0, 1, . . . , Tp −
T
1, k ∈ K. Let hi,k , [ hi,k [0] ··· hi,k [dmax ] ] ∈ C(dmax +1)×1
denote the vector of time-domain channel taps from user k to
antenna i ∈ I. Then, the vector of quantized symbols y̌p,i ,
T
[ y̌p,i [0] ··· y̌p,i [Tp −1] ] ∈ CTp ×1 received at the CU during the
channel estimation stage can be expressed as
X
y̌p,i =
Xp,k hi,k + zp,i + ep,i
k∈K
= Xp hi + zp,i + ep,i , i ∈ I,
(17)
where
Xp,k ,
xp,k [0]
xp,k [1]
0
xp,k [0]
..
.
..
.
···
···
..
.
0
0
..
.
xp,k [Tp −1] xp,k [Tp −2] ··· xp,k [Tp −1−dmax ]
∈ CTp ×(dmax +1) ,
k ∈ K,
(18)
is a Toeplitz matrix constructed from consecutive shifts of the
pilot symbols of the users, while zp,i , [ zp,i [0] ··· zp,i [Tp −1] ]T
2
I ) denotes the AWGN, and ep,i ,
with zp,i ∼ CN (0, σm
i Tp
T
2
[ ep,i [0] ··· ep,i [Tp −1] ] ∈ CTp ×1 with E[ep,i eH
p,i ] = εi ITp
X
···
X
p,K ] ∈
denotes the quantization noise. Also, Xp , [ p,1
T
CTp ×K(dmax +1) , and hi , [ hTi,1 ··· hTi,K ] ∈ CK(dmax +1)×1
in (17). It is sufficient to estimate hi , i ∈ I in (18), to find both
the channel coefficients and their corresponding tap delays.
We do not assume any prior knowledge of the probability
distribution function (pdf) of the elements of hi , i ∈ I at
the CU, i.e. they are treated as unknown constants. Then, the
least-squares (LS) estimate of hi is given by
ĥi = arg min ky̌p,i − Xp hi k2 = Xp† y̌p,i
hi
= hi + ξp,i ,
i ∈ I.
(19)
where we define ξp,i , Xp† zp,i + Xp† ep,i . Since each component of the vector X † ep,i ∈ C(dmax +1)×1 is a linear combination of the Tp independent zero-mean random variables in ep,i ,
the second term in ξp,i , X † ep,i , can be modeled as a CSCG
random vector, with mean E[X † ep,i ] = 0 and covariance
† H
2
H
−1
E[X † ep,i eH
, due to the central limit
p,i (X ) ] = εi (Xp Xp )
7
γk = P
ν∈∆\{0}
2
P |uH
k h̄kk [ν]| +
P uH
k h̄kk [0]
P
P
k′ ∈K\k
2
+ ε2i )(XpH Xp )−1 in (19).
theorem. Thus, ξp,i ∼ CN 0, (σm
i
Note that Tp ≥ K(dmax + 1) must be satisfied for the solution
in (19) to exist. The MSE of the estimate ĥi is given by
E[kĥi − hi k2 ] = E[kξp,i k2 ]
2
+ ε2i tr (XpH Xp )−1 ,
= σm
i
(20)
which is minimized if XpH Xp = cIK(dmax +1) , where c
is a constant [25, Example 4.3]. From the construction of
H
Xp , this translates to the condition Xp,k
Xp,k′ = cδ[k −
′
′
k ]I(dmax +1) , ∀k, k ∈ K, which is satisfied if each user’s
training sequence sp,k [n], n = 0, . . . , Tp − 1 is orthogonal
to that of every other user’s training sequence, and has the
“ideal” auto-correlation property [25]. This can be ensured by
using e.g., unit-amplitude Zadoff-Chu sequences [26], [27] for
sp,k [n], with equal transmit power P, ∀k ∈ K. In this case,
H
c = P Tp so that (Xp,k
Xp,k )−1 = P 1Tp I(dmax +1) , ∀k ∈ K,
and thus Xp† = P 1Tp XpH . Consequently, the LS estimate in (19)
reduces to
1
ĥi =
X H y̌p,i = hi + ξp,i ,
(21)
P Tp p
1
X H (zp,i + ep,i )
∼
where
ξp,i
=
(σ2 +ε2 )
P Tp p
CN 0, mPiTp i IK(dmax +1) . According to (21), each
tap in hi is estimated by correlating the received signal vector
y̌p,i with the training sequence of the corresponding user
shifted by the corresponding delay of the tap.
Due to the energy focusing property of the lens antenna
array in (2), and the different directions of arrival of the users’
signals at different RRHs, at any stream i, typically only one
path corresponding to a particular user corresponding to a
particular tap delay would dominate over all its the other paths.
Thus, the magnitudes of the channel coefficients in hi can
be vastly different, depending on the angle of arrival of the
users’ signals, and directly using the estimates of all the taps in
hi , i ∈ I according to (19) may be ineffective. To overcome
this issue, we propose a reduced-size, approximate linear
MMSE beamforming by exploiting the “sinc”-type response
of the lens array, where for each user, we select only the
streams which contain at least one dominant estimated tap
of its own, and then perform an approximate linear MMSE
beamforming over these selected streams by thresholding the
channel estimates, as explained below.
Notice that for the data decoding with path delay compensation detailed in Section II-B, the term h̄i,kk [0] in (12)
corresponds to the channel coefficient with maximum path
gain from user k to antenna i. Now, due to the energy focusing
property of the lens, the angle of arrival of user k’s signal is
typically such that |h̄i,kk [0]| ≫ |h̄i,kk [ν]|, ∀ν ∈ ∆ \ {0}, on
some streams i. This means that the maximum gain path of
the user would dominate over the other paths on a particular
stream i, provided the antenna corresponding to stream i has
ν∈∆
2
H
2
P |uH
k h̄kk′ [ν]| + uk (Σ + Ξ)uk
,
(14)
a response which peaks at one of the angles of arrival of the
user’s paths. Since we do not know the actual channel gains,
we aim to find the set of such streams Ik for each user k,
for which the magnitude of the maximum estimated channel
gain exceeds a certain threshold. For this, we first estimate the
tap delays of the maximum gain paths from user k to every
stream i ∈ I. These are denoted by d˘i,k , i ∈ I, and computed
from the LS estimate in (21) as follows
d˘i,k , arg max |ĥi,k [d]|, k ∈ K, i ∈ I.
(22)
d∈D
In the above, ĥi,k [d] is the ((k−1)(dmax +1)+d+1)th element
of the vector LS estimate ĥi in (21). Essentially, we choose
the tap delay corresponding to the estimate with the largest
magnitude as the estimated delay of the maximum gain path
on each stream i ∈ I. Then, for each stream i ∈ I, the estimate
of the channel coefficient corresponding to the estimated
maximum gain path, is given by ĥi,k [d˘i,k ], i ∈ I, k ∈ K,
which is the ((k − 1)(dmax + 1) + d˘i,k + 1)th element of the
vector LS estimate ĥi in (21).
Further, during the data transmission stage, the CU selects a
set of streams Ik ⊆ I, for which the above estimated channel
gains ĥi,k [d˘i,k ] corresponding to the estimated maximum
gain paths, are larger than a given threshold; i.e., we define
2
P ĥi,k [d̆i,k ]
i∈I
≥η ,
2 +ε2
σm
i
i
2
P ĥi,k d̆i,k
Ik , if ∃ at least one i ∈ I s.t.
≥ η,
2 +ε2
σm
i
i
{ik }
otherwise, where i , arg max
,
ĥ d˘
k
i∈I
i,k
i,k
(23)
k ∈ K, where η is a suitable threshold. Also, let Ik =
{j1 , . . . , jIk }, and |Ik | = Ik , so that Ik ≥ 1, i.e. at least
one stream is selected for decoding each user’s signal. Thus,
for the data transmission stage, we now have the reduced set
of streams y̌d,i , i ∈ Ik for each user k, with each y̌d,i given
by (12). On these set of streams, the CU performs path delay
compensation as in Section II-B, but now with the estimated
delays d˘i,k ’s of the maximum gain paths instead of the true
delays dmi ,k,ℓ⋆i,k ’s. The delay compensated signal can then be
expressed in terms of the delay differences of the taps with
d˘i,k , similar to (12), as
X
h̆i,kk [ν̆]xd,k [n − ν̆]
y̌˘d,i,k [n] = h̆i,kk [0̆]xd,k [n] +
ν̆∈∆\{0̆}
+
X
X
k′ ∈K\{k} ν̆∈∆
|
|
{z
ISI
h̆i,kk′ [ν̆]xd,k′ [n − ν̆]
{z
IUI
}
+ z̆d,i,k [n] + ĕd,i,k [n], i ∈ Ik , k ∈ K,
}
(24)
8
ˆ −1 ˆ
ˆ k = C̆
˘
where we have defined y̌˘d,i,k [n] , y̌d,i,k [n + d˘i,k ], z̆d,i,k [n] , beamforming ŭ
k h̆kk [0̆] on the observations y̌d,k [n]
5
zd,i [n + d̆i,k ], and ĕd,i,k [n] , ed,i [n + d̆i,k ], and also placed a˘ in (26), where
X
symbol over the delay differences ν̆ ∈ ∆ to indicate that they
ˆ
ˆ
ˆ
P h̆kk [ν̆]h̆kk [ν̆]H
C̆k ,
are w.r.t. the estimated tap delay d˘i,k . The channel coefficients
ν̆∈∆\{0̆}
h̆i,kk′ [ν̆] are defined similar to (11), and are given by
X X ˆ
ˆ
X
+
P h̆kk′ [ν̆]h̆kk′ [ν̆]H + Σ̆ + Ξ̆, (29)
˘
hi,k′ ,ℓ′ δ ν̆ − dmi ,k′ ,ℓ′ − di,k ,
h̆i,kk′ [ν̆] ,
k′ ∈K\{k} ν̆∈∆
ℓ′ ∈Lmi ,k′
i ∈ Ik , k, k ′ ∈ K, ν̆ ∈ ∆.
(25)
Note that the h̆i,kk′ [ν̆]’s in (24) and (25) denote the true
channel coefficients that correspond to a delay difference of
ν̆ with the estimated delay d˘i,k . In vector form, (24) can be
written as
X
˘d,k [n] = h̆kk [0̆]xd,k [n] +
h̆kk [ν̆]xd,k [n − ν̆]
y̌
ν̆∈∆\{0̆}
+
X
X
k′ ∈K\{k} ν̆∈∆
h̆kk′ [ν̆]xd,k′ [n − ν̆]
+ z̆d,k [n] + ĕd,k [n],
k ∈ K,
X
P h̆kk [ν̆]h̆kk [ν̆]H
ν̆∈∆\{0̆}
+
X
k′ ∈K\{k}
X
P h̆kk′ [ν̆]h̆kk′ [ν̆]H + Σ̆ + Ξ̆,
k∈K
in bps/Hz. In the following subsection we briefly describe a
benchmark scheme in the conventional CRAN when the RRHs
use UPAs and the users transmit using OFDM.
(26)
where all the vectors have dimensions Ik × 1. Then, the
reduced-size linear MMSE beamformer ŭk ∈ CIk ×1 for user
k after path delay compensation with the estimated tap delay
of the maximum gain path on each stream i ∈ Ik is given by
ŭk = C̆k−1 h̆kk [ν̆], where
C̆k ,
and the resulting SINR is given by (30) on the top of the next
page. Thus, with the proposed channel estimation, path delay
compensation, and reduced-size, approximate linear MMSE
receive beamforming for each user at the CU, a lower bound
on the effective sum-throughput over all users of the CRAN
is given by
Ta + Tp + 2dmax X
r̂lens = 1 −
log2 (1 + γ̂k ) (31)
Tf
(27)
ν̆∈∆
is the covariance
matrix ofthe interference terms in (26), with
2
2
2
2
σ
...
σm
m
Σ̆ , diag
, and Ξ̆ , εj1 ... εjIk .
j1
jI
k
However, the CU cannot compute the beamformer ŭk since
it only has the estimates of h̆i,kk′ [ν̆], computed from (21).
Now, on the selected streams i ∈ Ik , to avoid using the
noisy estimates, the CU approximates the ISI and IUI terms
in (24) by thresholding the estimates similar to (23). With the
thresholding, the estimates of h̆i,kk′ [ν̆], which we denote by
ˆ
h̆i,kk′ [ν̆], are given by
ĥ ′ [ν̆ + d˘ ], if P |ĥi,k′ [ν̆+d̆i,k ]|2 ≥ η,
i,k
i,k
2 +ε2
ˆ ′
σm
i
h̆i,kk [ν̆] ,
(28)
i
0
otherwise,
i ∈ Ik , k, k ′ ∈ K, ν̆ ∈ ∆. In (28), ĥi,k′ [ν̆ + d˘i,k ] is
the ((k ′ − 1)(dmax + 1) + (ν̆ + d˘i,k ) + 1)th element of
the vector LS estimate ĥi in (21). Note that due to the
response of the lens antenna array, for each stream i ∈ Ik ,
ˆ
the thresholded estimates h̆i,kk′ [ν̆] will be non-zero only for
a few usersh k ′ ∈ K and delay idifferences ν̆ ∈ ∆. Let
T
ˆ
h̆kk′ [ν̆] = h̆ˆ j1 ,kk′ [ν̆] ··· h̆ˆ jIk ,kk′ [ν̆]
∈ CIk ×1 denote the
vector of the thresholded channel estimates in (28). Then,
the CU applies the reduced-size, approximate linear MMSE
III. B ENCHMARK : C ONVENTIONAL CRAN
AND OFDM
WITH
UPA S
In this case, each sector of an RRH is assumed to be
equipped with a rectangular UPA of physical dimensions D̃y ×
D̃z , with adjacent antenna elements of the array separated by a
distance equal to half the wavelength. Let Q̃(j) , {1, . . . , Q̃},
denote the set of antennas
at sector j ∈ J of each RRH in
S
this case, and Q̃ = j∈J Q(j) denote the set of all antenna
elements. The number of antenna elements per sector is given
by Q̃ = Q̃y Q̃z , where Q̃y = ⌊2D̃y ⌋ denotes the number of
antennas along the y-axis of the UPA and Q̃z = ⌊2D̃z ⌋ is the
number of antennas along the z-axis. For the UPA, the array
response vector a(θ, φ) ∈ CQ̃×1 in (1) can be written as
a(θ, φ) = az (θ) ⊗ ay (θ, φ),
(32)
q
T
D̃z
where az (θ) =
[ 1 exp(π sin θ) ··· exp(π(Q̃z −1) sin θ) ] ∈
Q̃
z
CQ̃z ×1
is
ray
along
r
the
the
response
z-axis,
of
and
the
linear
ay (θ, φ)
D̃y
T
[ 1 exp(π cos θ sin φ) ··· exp(π(Q̃y −1) cos θ sin φ) ]
Q̃y
ar=
∈ CQ̃y ×1
is the response of the linear array along the y-axis.
The total channel bandwidth W is divided into N
orthogonal sub-channels (SCs) of equal width denoted
by N , {0, 1, . . . , N − 1}.
The uplink transmission protocol is similar to that in Fig. 3,
except that the users transmit their symbols in (N + µ)length symbol blocks using OFDM, where µ ≥ dmax is the
length of the cyclic prefix (CP). Hence, the guard intervals
shown in Fig. 3 are implicit in this case due to the CP,
which is discarded at the CU. We assume that all the users
transmit on all SCs for fully exploiting spatial multiplexing.
5 It is assumed that the CU knows the quantization noise variances on each
stream. These can be transmitted by the RRHs after the power probing stage
to the CU once every frame period, along with information about the number
of SQ bits on each antenna, since these quantities do not change within a
frame.
9
γ̂k = P
ν̆∈∆\{0̆}
ˆ H h̆kk [ν̆]|2 +
P |ŭ
k
ˆ H h̆kk [0̆]
P ŭ
k
P
P
k′ ∈K\k
In general, there can be multiple OFDM symbols in the power
probing, channel estimation, and data transmission stages.
For convenience, it is assumed that Ta , Tp and Td are all
a
integer multiples of N + µ, i.e. there are τa = NT+µ
∈ Z++ ,
Tp
τp = N +µ ∈ Z++ , and τd = τd ∈ Z++ OFDM symbols in
the power probing, channel estimation, and data transmission
stages of a frame, respectively. Let x̃t,k ∈ CN ×1 denote the tth
OFDM symbol transmitted by user k in the frequency-domain.
Then the corresponding time-domain signals are given by
the inverse discrete Fourier transform (IDFT) xt,k = F H x̃k ,
where F ∈ CN ×N denotes the DFT matrix with columns
T
2πn(N −1)
) ] , n ∈ N . Similar
fn , √1N [ 1 exp(− 2πn
N ) ··· exp(−
N
to Section II-A, the RRHs perform antenna selection and
SQ bit allocation on the received time-domain signals and
forward them to the CU. Let Q̃m , {q ∈ Q̃|bq,m > 0}
with |Q̃m | = Q̃m denote the set of selected antennas at each
RRH m in this case. Notice that with UPAs, the selected
set of antennas and their non-zero quantization bit allocation
at each RRH can be very different from that with the lens
antenna P
array, due to the lack of the lens focusing. Let
Q̃tot =
m∈M Q̃m and similar to Section II-A we index
the selected antennas (streams) with i ∈ {1, . . . , Q̃tot } , Ĩ,
where stream index i corresponds to antenna qi of RRH mi .
2
ν̆∈∆
ˆk
ˆ H (Σ̆ + Ξ̆)ŭ
ˆ H h̆kk′ [ν̆]|2 + ŭ
P |ŭ
k
k
k∈K
=
X√
N X̃p,t,k F0 hi,k + z̃p,t,i + ẽp,t,i
k∈K
= X̄p,t hi + z̃p,t,i + ẽp,t,i ,
(33)
where z̃p,t,i = F zp,t,i is the DFT of the AWGN so
2
I ), and ẽp,t,i = F ep,t,i is the
that z̃p,t,i ∼ CN (0, σm
i N
DFT of the quantization noise with E[ẽp,t,i ] = 0 and
2
E[ẽp,t,i ẽH
p,t,i ] = εi IN . Since each element of ẽp,t,i is a
linear combination of the N elements in ep,t,i , due to the
2
central limit
IN ). Also, in (33),
√ theorem, ẽp,t,i ∼ CN (0, εNi×K(d
max +1)
, and hi
X̄p,t , N [ X̃p,t,1 F0 ··· X̃p,t,K F0 ] ∈ C
(30)
˜p,t,i ,
is defined similar to (17). For τp > 1, stacking each of y̌
X̄p,t , z̃p,t,i , and ẽp,t,i , t ∈ {1, . . . , τp − 1} column-wise, we
have
˜p,i = X̄p hi + z̃p,i + ẽp,i .
y̌
(34)
Similar to Section II-C, the LS estimate for hi is given by6
˜p = hi + ξp,i ,
ĥi = X̄p† y̌
(35)
2
+
where ξp,i , X̄p† (z̃p,i + ẽp,i ) ∼ CN (0, (σm
i
2
H
−1
εi )(X̄p X̄p ) ). Note that the estimate in (35) exists only
if τp N ≥ K(dmax + 1), and the MSE of the estimate is
2
+ ε2i )tr (X̄pH X̄p )−1 ,
E[kĥi − hi k2 ] = E[kξp,i k2 ] = (σm
i
(36)
which is minimized if X̄pH X̄p = c̃IK(dmax +1) , where c̃ is a
constant. Due to the construction of X̄p in (34), this translates
to the requirement that
τp −1
X
H
X̄p,t,k
X̄p,t,k′ = N F0H
τX
p −1
t=0
t=0
H
X̃p,t,k
X̃p,t,k′ F0
= δ[k − k ′ ]c̃I(dmax +1)
A. Channel Estimation
In the channel estimation stage, let x̃p,t,k [n] denote the
frequency-domain pilot symbol transmitted by user k ∈ K on
SC n ∈ N in the tth OFDM symbol, where t ∈ {0, 1, . . . , τp −
1}. Also, let X̃p,t,k , diag( x̃p,t,k [0] ··· x̃p,t,k [N −1] ) ∈ CN ×N
be the diagonal matrix constructed
√ from the pilot symbols
on all the SCs, and h̃i,k =
N F0 hi,k , the vector of
frequency-domain channel coefficients from user k to antenna
i ∈ I, where hi,k ∈ C(dmax +1)×1 is the time-domain
channel coefficient vector similar to that in Section II-C, and
F0 ∈ CN ×(dmax +1) denotes the first (dmax +1) columns of F .
Then, the frequency-domain received signal vector at the CU
corresponding to the OFDM symbol t at antenna i ∈ Ĩ, after
removing the CP and applying the DFT can be expressed as
X
˜p,t,i =
y̌
X̃p,t,k h̃i,k + z̃p,t,i + ẽp,t,i
.
k, k ′ ∈ K. (37)
Now, let Ψ , V ⊗Sp,k ∈ Cτp N ×τp N , where V ∈ Cτp ×τp is a
unitary matrix and Sp,k , diag( sp,k [0] ··· sp,k [N −1] ) ∈ CN ×N
is a diagonal matrix with unit amplitude training symbols,
H
H
satisfying Sp,k
Sp,k = IN . Then, ΨH Ψ = V H V ⊗Sp,k
Sp,k =
Iτp N . The√condition
(37)
can
then
be
ensured
by
choosing
√
X̃p,t,k = P Ψtω ( N diag(fκ )), where Ψtω ∈ CN ×N is the
(t, ω)th block of Ψ, t, ω ∈ {0, . . . , τp − 1} [28, Theorem 2].
Here, κ and ω are such that the user index k can be written
K−1
as k = κτp + ω + 1, i.e., κ = ⌊ k−1
τp ⌋ ∈ {0, 1, . . . , ⌊ τp ⌋},
and ω = k − 1 − κτp ∈ {0, 1, . . . , τp − 1}. Then, (37) holds
with c̃ = τp N P , and X̄pH X̄p = τp N P IK(dmax +1) . Then, the
LS estimate in (35) can be written as
ĥi =
1
˜p = hi + ξp,i ,
X̄ H y̌
τp N P p
(38)
1
X̄ H (z̃p,i + ẽp,i )
∼
where
ξp,i
=
σ2 +ε2
τp N P p
CN 0, τmpiN P i IK(dmax +1) . Note that (35) gives the
time-domain channel estimates, and the frequency domain
ˆ
estimates h̃i,k , i ∈ Ĩ for each user k ∈ K, are given by
√
ˆ
applying the transformation h̃i,k = N F0 ĥi,k , i ∈ Ĩ, k ∈ K,
where ĥi,k is the length (dmax + 1) sub-vector in ĥi starting
at index (k − 1)(dmax + 1) + 1 and ending at k(dmax + 1).
6 Since both z̃
p,i and ẽp,i are CSCG, this is also the maximum likelihood
estimate of hi .
B. Achievable Sum-Rate
√
Let x̃d,k [n] = P sd,k [n] denote the frequency-domain data
symbol transmitted by user k on SC n ∈ N , where sd,k [n] ∼
CN (0, 1) and P ≥ 0. Then, the frequency-domain received
signal at the CU on SC n ∈ N corresponding to stream i ∈ Ĩ
can be expressed as
X
y̌˜d,i [n] =
h̃i,k [n]x̃d,k [n] + z̃d,i [n] + ẽd,i [n],
(39)
k∈K
th
where z̃d,i [n] and ẽd,i [n] are the n components of the DFTs
of the AWGN and quantization noise similar to that in (33).
Collecting the received signals from all streams on each SC
n, we have
X
˜d [n] =
y̌
h̃k [n]x̃d,k [n] + z̃d [n] + ẽd [n], n ∈ N , (40)
z-axis (meters)
10
40
20
0
60
50
40
30
50
RRH
User
20
y-axis (meters)
40
30
10
20
10
0
0
x-axis (meters)
k∈K
where all the vectors are of dimensions Q̃tot × 1. If the CU
has perfect CSI, the linear MMSE beamforming vector ũk,n ∈
CQ̃tot ×1 that maximizes the receive SINR for each user k on
SC n, treating the interference from other users as noise, is
−1
given by ũk,n = C̃k,n
h̃k [n], where
X
(41)
P h̃k′ [n]h̃k′ [n]H + Σ̃ + Ξ̃,
C̃k,n ,
k′ ∈K\{k}
2
IQ̃
∈ CQ̃tot ×Q̃tot and Ξ̃ ,
with Σ̃ , blkdiag σ12 IQ̃1 ··· σM
M
Q
Q̃
×
Q̃
tot
diag( {ε2i }i=1
) ∈ C tot tot . The SINR for user k on SC
−1
h̃k [n], and thus the
n is then given by γ̃k,n = P h̃k [n]H C̃k,n
sum-rate over all users with UPAs and OFDM transmission is
given by
X
X 1
r̃upa =
log2 (1 + γ̃k,n )
(42)
N +µ
k∈K
n∈N
in bps/Hz.
On the other hand, with estimated CSI, the CU computes
ˆ
ˆ −1 h̃
ˆ k,n = C̃
the linear MMSE receiver ũ
k,n k [n] on each SC n,
P
ˆ ′ [n]H + Σ̃+ Ξ̃
ˆ ′ [n]h̃
ˆ
P h̃
where the matrix C̃
,
k,n
k′ ∈K\{k}
k
k
ˆ
and the components of h̃k [n], n ∈ N , k ∈ K are computed
by transforming the time-domain estimates in (35) as outlined
earlier. Then the resulting SINR is given by
ˆ H h̃k [n]|2
P |ũ
k,n
ˆ H h̃k′ [n]|2 + ũ
ˆ H (Σ̃ + Ξ̃)ũ
ˆ k,n
P
|
ũ
′
k ∈K\{k}
k,n
k,n
(43)
γ̃ˆk,n = P
Thus, the effective sum-throughput over all users with UPAs
and OFDM transmission is given by
τa + τp 1 X X
r̃ˆupa = 1 −
log2 (1 + γ̃ˆk,n ), (44)
τf
N +µ
k∈K n∈N
in bps/Hz.
IV. S IMULATION R ESULTS
For the simulations, we consider a part of the general
sectorized CRAN illustrated in Fig. 2 with M
√ = 6 RRHs
located at the corners of a hexagon of side 50/ 3 meters (m),
along with K = 6 users randomly located in the common
Fig. 5. Example of RRH and user layout for simulations.
region covered by one sector of each RRH as shown in Fig. 5.
All the RRHs are assumed to be of height 30 meters (m), while
the users’ height can vary between 1 m and 25 m. The size of
the EM lens at the RRHs is such that Dy = Dz = 10, while
the UPAs are rectangular with same length D̃y = 10 as the
EM lens along the y-axis, and length D̃z = 7 along the z-axis.
The height of the RRHs is greater than the maximum height of
the users, and the minimum distance along the ground between
a user and RRH is assumed to be 2 m. Based on the above
setup, the maximum coverage angles for the lens array in the
elevation direction are chosen to be Θ+ = π/2, Θ− = π/6,
while the corresponding angles in the azimuth direction are
Φ+ = Φ− = π/3 due to the sectorization. This leads to
Q = 208 and Q̃ = 280 antennas per sector at each RRH
for the lens array and the UPA, respectively. The carrier
frequency is 28 GHz and the mmWave channel shared by
the users and RRHs has a bandwidth W = 200 MHz. The
maximum delay spread of the channel is 100 ns, which
translates to dmax = 20 symbols. The free-space path loss
′
between a user k and RRH m separated by distance Dm,k
m
′
is modeled by 61.4 + 34.1 log10 Dm,k
dB [29]. The number
of paths Lm,k between user k and RRH m is equal to 1, 2
or 3 with equal probability, and the time delays of the paths
ζℓ , ℓ ∈ Lm,k , are generated from an exponential distribution
with mean rζ µζ , but truncated to 100 ns, where µζ = 67 ns
and rζ = 0.25 [29]. The relative P
power levels of the paths
are given by κ̄m,k,ℓ = κm,k,ℓ /( ℓ′ ∈Lm,k κm,k,ℓ′ ), where
Z
−ζℓ
κm,k,ℓ = 10 10 (0.613) exp 31.4
ns [29], and Z is zero-mean
Gaussian with standard deviation 9.4 dB, representing the perpath shadowing [29]; thus |αm,k,ℓ |2 = κ̄m,k,ℓ Pr,m,k , where
Pr,m,k in watt is the received power at RRH m from user k
after accounting for the free-space path loss. The phase shifts
of each path ∠αm,k,ℓ are generated uniformly randomly from
the interval [0, 2π]. The elevation angles of arrival θm,k,ℓ for
each path ℓ ∈ Lm,k from user k ∈ K to RRH m ∈ M are indeπ π
pendent and generated as θm,k,ℓ = θ̄m,k,ℓ +U[− 12
, 12 ], where
θ̄m,k,ℓ is the line-of-sight (LoS) angle of elevation from the
user to the RRH. Similarly, the azimuth angles of arrival φm,k,ℓ
are independent and generated as φm,k,ℓ = φ̄m,k,ℓ +U[− π6 , π6 ],
11
TABLE II
AVERAGE NUMBER OF SELECTED ANTENNAS PER RRH
5
Fronthaul
capacity
per
sector (Gbps)
4
3.5
Lens
UPA-OFDM
3
2.5
0.4
2
4
8
20
40
60
100
200
2
2
5
6
10
11
19
21
48
51
98
101
148
151
208
251
208
280
4
2
1.5
1
Lens: Perfect fronthaul, perfect CSI
UPA−OFDM: Perfect fronthaul, perfect CSI
Lens: Constrained fronthaul, perfect CSI
UPA−OFDM: Constrained fronthaul, perfect CSI
0.5
0
0
50
100
150
200
Fronthaul capacity per sector (Gbps)
Fig. 6. Average rate per user vs. common fronthaul capacity per RRH sector
under perfect CSI, with P = 23 dBm.
where φ̄m,k,ℓ is the LoS angle of azimuth. The power spectral
density of the AWGN is −174 dBm/Hz, with an additional
noise figure of 6 dB at each RRH. The inter-sector/intercluster interference is assumed to be equal to −80 dBm at
all sectors. The fronthaul capacity of all RRHs are the same,
i.e. R̄m = R̄, ∀m ∈ M. For the benchmark scheme using
OFDM with UPAs at the RRHs, the number of SCs N = 256
and the CP length µ = dmax = 20. The minimum coherence
time of the channels is taken to be 0.4 ms, while each frame is
assumed to be Tf = 8280 symbols long, which corresponds to
30 OFDM symbols. For the proposed system with lens arrays,
we choose Tp = K(dmax +1), while for the benchmark system
with UPAs and OFDM, Tp = 3(N + µ) i.e., τp = 3. In order
to make a fair comparison, we assume Ta = N + dmax in both
cases. All the results are averaged over random user locations
and channel realizations.
Figs. 6 and 7 plot the average throughput per user against
the common fronthaul capacity per sector R̄/3, where the
users’ transmit power P = 23 dBm. Fig. 6 shows the
performance assuming perfect CSI at the CU, both with
and without the fronthaul constraints at the RRHs. Without
fronthaul constraints and with perfect CSI, it can be observed
that the performance of the benchmark system with OFDM
and UPAs is close to the proposed system with lens. The
difference can be attributed to the loss due to the CP insertion
in the case of OFDM transmission with UPAs. As the lens
array only redistributes the incident energy among the antenna
elements depending on the angle of arrival of the plane waves,
it is reasonable to expect that it does not offer a large gain
in the achievable rate over UPA-OFDM when there is no
constraint on the fronthaul. However, since only single-carrier
transmission by the users is required, and the ISI and IUI can
be mitigated without CP insertion and IDFT/DFT processing,
the use of lens arrays can greatly simplify the signal processing
required, when compared with UPA-OFDM. For the system
Average sum-rate per user (bps/Hz)
Average sum-rate per user (bps/Hz)
4.5
3.5
3
2.5
2
1.5
1
0.5
0
0
Lens: Perfect fronthaul, estimated CSI
UPA−OFDM: Perfect fronthaul, estimated CSI
Lens: Constrained fronthaul, estimated CSI
UPA−OFDM: Constrained fronthaul, estimated CSI
50
100
150
200
Common fronthaul capacity per sector (Gbps)
Fig. 7. Average rate per user vs. common fronthaul capacity per RRH sector
under estimated CSI, with P = 23 dBm.
with UPA-OFDM, the overall complexity of performing linear
MMSE beamforming for all the users on all the SCs with no
fronthaul constraint is of the order O(N K(M J Q̃)3 ), while
for the proposed system with lens arrays, the complexity is
O(K(M JQ)3 ), offering a reduction of a factor of N .
Further, for the fronthaul constrained case, it can be observed from Fig. 6 that the proposed system with lens arrays
provides significant throughput gains over the conventional
UPA-OFDM. This is due to the energy focusing property of the
EM lens, which makes the antenna selection and quantization
bit allocation much more effective with lens arrays compared
to UPAs. In the antenna selection and bit allocation scheme,
the antennas with larger received power are selected and
allocated more bits compared to the antennas with lower
received power. Table II compares the number of antennas
selected at the RRHs in both cases. As the EM lens focuses
the incident energy onto a few antenna elements, the selection
and bit allocation algorithm tends to select these elements
with large received power. Thus, although the number of
antennas selected in both cases is more or less similar when
the fronthaul capacity is low, the selection and bit allocation is
much more effective with lens arrays when compared to UPAOFDM. With UPA-OFDM, the received energy distribution is
more or less similar across the antenna elements and the selection and bit allocation can thus adversely affect the achievable
rate compared to the case with no fronthaul constraint. Fig. 7
shows the performance of the proposed system with estimated
12
TABLE III
AVERAGE NUMBER OF STREAMS SELECTED PER USER FOR REDUCED
MMSE WITH ESTIMATED CSI FOR LENS ANTENNA ARRAY
Fronthaul
capacity
per
sector (Gbps)
Number
streams Ik
of
0.4
2
4
8
20
40
60
100
200
1
2
2
2
2
3
3
3
4
CSI. For the thresholding of the channel estimates as given
in (23) and (28), we choose η = 3 dB. Again, it can be
observed that with the proposed channel estimation scheme
via thresholding the LS channel estimates, the system with
lens antenna arrays offers significant gains over the benchmark
UPA-OFDM system. This is because the quantization and
antenna selection at the RRHs adversely affects both the channel estimation as well as data transmission in the benchmark
system, when the fronthaul is constrained. On the other hand,
due to the energy focusing property of the lens, the stronger
channels are relatively less affected by the quantization, and
since the channels between the users and RRHs are sparse in
the angular domain, and the lens converts this angular domain
sparsity to the spatial domain, it is sufficient to use only
the estimates of the stronger channels to retain most of the
gains in the subsequent MMSE beamforming. In the UPAOFDM system, such a thresholding is not effective since the
energy is distributed uniformly over the antenna elements, and
there is no way of distinguishing the significant channels in
the spatial domain. Table III shows the average number of
streams Ik selected after thresholding the channel estimates
for the proposed reduced, approximate MMSE beamforming
with lens arrays at the RRHs. It can be observed that the
total number of streams can potentially be even an order of
magnitude less than the total number of streams in the UPAOFDM case (see Table II), and this leads to a huge reduction
in complexity of the receive beamforming in CRAN, since
3
for the proposed system the complexity is only O(KImax
),
where Imax = maxk∈K Ik is the maximum number of streams
for a user, while for the benchmark with UPA-OFDM, it is
O(N K Q̃3tot ). Moreover, this gain in throughput is achieved
at a much lower training overhead of K(dmax + 1) for the
proposed system, compared to at least N + dmax for the
benchmark, where in practice N ≫ K as usual.
V. C ONCLUSION
In this paper, we introduce a new architecture for mmWave
CRAN with lens antenna arrays at the RRHs. We propose
a low-complexity bit allocation scheme for SQ at the RRHs
based only on the estimated received power levels of the
antennas. We show that when the fronthaul is constrained, the
proposed system which uses single-carrier transmission by the
users combined with path delay compensation and reducedsize MMSE beamforming, can achieve large throughput gains
over the conventional system that uses UPAs at the RRHs
along with OFDM transmission by the users and full-size
MMSE combining on each SC. Further, we propose a simple,
yet effective channel estimation scheme for the proposed
system, which exploits the unique energy focusing property
of the lens array, and show that the proposed system still
offers considerable advantage over the benchmark system,
when both have imperfect CSI. Moreover, these gains are
achieved at a much lower signal processing complexity at
the CU, since it is sufficient to process a considerably fewer
number of streams for the proposed system when compared to
the benchmark. This shows that the proposed system with lens
antenna arrays is a promising candidate for future evolution
of CRAN operating in the mmWave frequencies.
A PPENDIX
As problem (6) is convex and is strictly feasible, it can
be optimally solved by solving its dual problem. Let λ denote
the Lagrange multiplier for constraint (5a) in problem (6). The
Lagrangian is then given by
X
X 3ρq,m
R̄m
,
(45)
+
λ
b
−
L(bm , λ) =
q,m
2b
2 q,m
2W
q∈Q
q∈Q
and the dual function is given by f (λ) = minbm ≥0 L(bm , λ).
Differentiating (45) with respect to each bq,m and setting the
derivative equal to zero, we have
6ρ
1
q,m ln 2
bq,m = log2
.
(46)
2
λ
Since bq,m ≥ 0 ∀q ∈ Q, the dual variable λ must lie in
the interval [0, 6ρmax,m ln 2], where ρmin,m = minq∈Q ρq,m
and ρmax,m = maxq∈Q ρq,m . Moreover, according to (7), as
λ increases, the bq,m ’s decrease, making the constraint (5a)
more feasible, and vice versa. Thus, the optimal λ⋆ that
solves the dual problem maxλ≥0 f (λ), can be found by a
simple bisection search over [0, 6ρmax,m ln 2]; then the optimal
solution b′m for problem (6) is obtained as given by (7). The
proof of Proposition 2.1 is thus completed.
R EFERENCES
[1] J. Andrews et al., “What will 5G be?” IEEE J. Sel. Areas Commun.,
vol. 32, no. 6, pp. 1065–1082, June 2014.
[2] Y. Zhou and W. Yu, “Optimized backhaul compression for uplink cloud
radio access network,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp.
1295–1307, June 2014.
[3] L. Liu and R. Zhang, “Optimized uplink transmission in multi-antenna
C-RAN with spatial compression and forward,” IEEE Trans. Signal
Process., vol. 63, no. 19, pp. 5083–5095, Oct. 2015.
[4] B. Dai and W. Yu, “Sparse beamforming and user-centric clustering for
downlink cloud radio access network,” IEEE Access, vol. 2, pp. 1326–
1339, 2014.
[5] Y. Shi, J. Zhang, and K. Letaief, “Group sparse beamforming for green
cloud-RAN,” IEEE Trans. Wireless Commun., vol. 13, no. 5, pp. 2809–
2823, May 2014.
[6] S. Luo, R. Zhang, and T. J. Lim, “Downlink and uplink energy
minimization through user association and beamforming in C-RAN,”
IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 494–508, Jan. 2015.
[7] R. G. Stephen and R. Zhang, “Joint millimeter-wave fronthaul and
OFDMA resource allocation in ultra-dense CRAN,” IEEE Trans. Commun., vol. 65, no. 3, pp. 1411–1423, Mar. 2017.
[8] L. Liu, S. Bi, and R. Zhang, “Joint power control and fronthaul rate
allocation for throughput maximization in OFDMA-based cloud radio
access network,” IEEE Trans. Commun., vol. 63, no. 11, pp. 4097–4110,
Nov. 2015.
[9] R. G. Stephen and R. Zhang, “Fronthaul-limited uplink OFDMA in
ultra-dense CRAN with hybrid decoding,” IEEE Trans. Veh. Technol.,
vol. 66, no. 10, pp. 9074–9084, Oct. 2017.
13
[10] T. Bai, A. Alkhateeb, and R. W. Heath, “Coverage and capacity of
millimeter-wave cellular networks,” IEEE Commun. Mag., vol. 52, no. 9,
pp. 70–77, Sep. 2014.
[11] M. R. Akdeniz et al., “Millimeter wave channel modeling and cellular
capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp.
1164–1179, June 2014.
[12] J. Hoydis, M. Kobayashi, and M. Debbah, “Optimal channel training in
uplink network MIMO systems,” IEEE Trans. Signal Process., vol. 59,
no. 6, pp. 2824–2833, Jun. 2011.
[13] J. Kang, O. Simeone, J. Kang, and S. S. Shitz, “Joint signal and channel
state information compression for the backhaul of uplink network mimo
systems,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1555–1567,
Mar. 2014.
[14] J. Zhang, X. Yuan, and Y. J. Zhang, “Locally orthogonal training
design for cloud-RANs based on graph coloring,” IEEE Trans. Wireless
Commun., vol. 16, no. 10, pp. 6426–6437, Oct. 2017.
[15] Z. Gao, C. Hu, L. Dai, and Z. Wang, “Channel estimation for millimeterwave massive MIMO with hybrid precoding over frequency-selective
fading channels,” IEEE Communications Letters, vol. 20, no. 6, pp.
1259–1262, Jun. 2016.
[16] K. Venugopal, A. Alkhateeb, N. G. Prelcic, and R. W. Heath, “Channel
estimation for hybrid architecture-based wideband millimeter wave
systems,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1996–2009,
Sep. 2017.
[17] Y. Zeng, R. Zhang, and Z. N. Chen, “Electromagnetic lens-focusing
antenna enabled massive MIMO: Performance improvement and cost
reduction,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1194–1206,
June 2014.
[18] Y. Zeng and R. Zhang, “Millimeter wave MIMO with lens antenna array:
A new path division multiplexing paradigm,” IEEE Trans. Commun.,
vol. 64, no. 4, pp. 1557–1571, Apr. 2016.
[19] ——, “Cost-effective millimeter-wave communications with lens antenna array,” IEEE Wireless Communications, vol. 24, no. 4, pp. 81–87,
2017.
[20] Y. Zeng, L. Yang, and R. Zhang, “Multi-user millimeter wave
MIMO with full-dimensional lens antenna array.” [Online]. Available:
http://arxiv.org/abs/1611.06008
[21] T. Kwon, Y. G. Lim, B. W. Min, and C. B. Chae, “RF lens-embedded
massive MIMO systems: Fabrication issues and codebook design,” IEEE
Trans. Microw. Theory Tech., vol. 64, no. 7, pp. 2256–2271, July 2016.
[22] L. Yang, Y. Zeng, and R. Zhang, “Channel estimation for millimeter
wave MIMO communications with lens antenna arrays,” IEEE Trans.
Veh. Technol., 2017, to appear.
[23] A. B. Narasimhamurthy and C. Tepedelenlioglu, “Antenna selection for
MIMO-OFDM systems with channel estimation error,” IEEE Trans. Veh.
Technol., vol. 58, no. 5, pp. 2269–2278, Jun. 2009.
[24] D. Tse and P. Viswanath, Fundamentals of wireless communication.
Cambridge University Press, 2005.
[25] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Upper Saddle River, NJ, USA: Prentice-Hall, Inc., 1993.
[26] B. M. Popovic, “Generalized chirp-like polyphase sequences with optimum correlation properties,” IEEE Trans. Inf. Theory, vol. 38, no. 4,
pp. 1406–1409, Jul. 1992.
[27] S. Sesia, I. Toufik, and M. Baker, LTE, The UMTS Long Term Evolution:
From Theory to Practice. Wiley Publishing, 2009.
[28] Y. Chi, A. Gomaa, N. Al-Dhahir, and A. R. Calderbank, “Training signal
design and tradeoffs for spectrally-efficient multi-user MIMO-OFDM
systems,” IEEE Trans. Wireless Commun., vol. 10, no. 7, pp. 2234–
2245, Jul. 2011.
[29] M. K. Samimi and T. S. Rappaport, “Ultra-wideband statistical channel
model for non line of sight millimeter-wave urban channels,” in Proc.
IEEE GLOBECOM, Dec. 2014, pp. 3483–3489.
| 7cs.IT
|
Submitted to the Annals of Statistics
arXiv: arXiv:0000.0000
arXiv:1507.03895v2 [math.ST] 21 Nov 2016
ON CONSISTENCY AND SPARSITY FOR SLICED
INVERSE REGRESSION IN HIGH DIMENSIONS
By Qian Lin†∗ Zhigen Zhao‡∗ and Jun S. Liu†∗
Harvard University†
Temple University‡
We provide here a framework to analyze the phase transition phenomenon of slice inverse regression (SIR), a supervised dimension
reduction technique introduced by Li [1991]. Under mild conditions,
the asymptotic ratio ρ = lim p/n is the phase transition parameter
and the SIR estimator is consistent if and only if ρ = 0. When dimension p is greater than n, we propose a diagonal thresholding screening
SIR (DT-SIR) algorithm. This method provides us with an estimate
of the eigen-space of the covariance matrix of the conditional expectation var(E[x|y]). The desired dimension reduction space is then
obtained by multiplying the inverse of the covariance matrix on the
eigen-space. Under certain sparsity assumptions on both the covariance matrix of predictors and the loadings of the directions, we prove
the consistency of DT-SIR in estimating the dimension reduction
space in high dimensional data analysis. Extensive numerical experiments demonstrate superior performances of the proposed method
in comparison to its competitors.
1. Introduction. For a continuous multivariate random variable (y, x)
where x ∈ Rp and y ∈ R, a subspace S ′ ⊂ Rp is called the effective dimension
⊥ stands for independence.
reduction (EDR) space if y ⊥
⊥ x|PS ′ (x) where ⊥
Under mild conditions (Cook [1996]), the intersection of all the EDR spaces
is again an EDR space, which is denoted as S and called the central space.
Many algorithms were proposed to find such subspace S under the assumption d = dimS ≪ p. This line of research is commonly known as sufficient
dimension reduction. The Sliced Inverse Regression (SIR, Li [1991]) is the
first, yet the most widely used method in sufficient dimension reduction,
due to its simplicity, computational efficiency and generality. The asymptotic properties of SIR are of particular interest in the last two decades. The
consistency of SIR has been proved for fixed p in Li [1991], Hsing and Carroll
∗
Lin’s research is supported by the Center of Mathematical Sciences and Applications
at Harvard University. Zhao’s research is supported by the NSF Grant DMS-1208735. Liu’s
research is supported by the NSF Grant DMS-1120368 and NIH Grant R01 GM113242-01
MSC 2010 subject classifications: Primary 62J02; secondary 62H25
Keywords and phrases: dimension reduction, random matrix theory, sliced inverse regression
1
2
Q. LIN, Z. ZHAO AND J. S. LIU
[1992], Zhu and Ng [1995] and Zhu and Fang [1996]. Later, Zhu et al. [2006]
√
have obtained the consistency if p = o( n). A similar restriction also appears in two recent work (see Zhong et al. [2012] and Jiang and Liu [2014]).
When p > n, a common strategy pursued by many recent researchers is to
make sparsity assumptions that only a few predictors play a role in explaining and predicting y and apply various regularization methods. For instance,
Li and Nachtsheim [2006], Li [2007] and Yu et al. [2013] applied LASSO
(Tibshirani [1996]), Dantzig selector (Candes and Tao [2007]) and elastic
net (Zou and Hastie [2005]) respectively to solve the generalized eigenvalue
problems raised by a variety of SDR algorithms.
However, a piece of jigsaw is missing in the understanding of SIR. If the
dimension p diverges as n increases, when will the SIR break down? A similar question has been asked for a variety of SDR estimates in Cook et al.
[2012]. In this paper, we prove that, under certain technical assumptions,
the SIR estimator is consistent if and only if ρ = lim np = 0. Such a result on inconsistency provides theoretical justifications for imposing certain
structural assumption, such as sparsity, in high dimensional settings. This
behavior of SIR in high dimension, which will be called the phase transition
phenomenon, is similar to that of the principal component analysis (PCA),
an unsupervised counterpart of SIR. This extension is, however, by no means
trivial. After all the samples (yi , xi ) are sliced into H bins according to the
order statistics of yi , the sliced samples are neither independent nor identically distributed. This difference increases the difficulty significantly. In this
paper, we provide a new framework to study the phase transition behaviour
of SIR. The technical tools developed here can potentially be extended to
study the phase transition behaviour of other SDR estimators.
The second part of the article aims at extending the original SIR to the
scenario with ultra-high dimension (p = o(exp(nξ ))). Based on equation (3)
in Section 2, the central space can be estimated by the column space of
b −1 col(Vb H ), where Σ
b −1 is any consistent estimate of the precision matrix
Σ
x
x
b H ) is the estimate of the space col(var(E[x|y])). To estimate
Σ−1
and
col(
V
x
the column space of var (E[x|y]), we propose a diagonal screening procedure
based on new univariate statistics varH (x(k)), which are the diagonal elements of var (E[x|y]), motivated by recent work in sparse PCA (Johnstone
and Lu [2004]). After ranking the predictors according to the magnitude
of varH (x(k)) decreasingly, we choose the set I consisting of the first R
predictors as active predictors. The SIR procedure is subsequently applied
to these selected predictors to estimate the d-dimensional column space of
I
I
var(E[xI |y]) by col(Vb H ) where Vb H is the matrix formed by the top d eigenI
I
b . We embed Vb into Rp×d by filling in 0’s for entries outside
vectors of Λ
H
H
3
DT-SIR
I
the chosen row set I, and denote this new matrix by e(Vb H ). The estimate
b −1
bI
of the central space is defined to be col(Σ
x e(V H )). We name this two-stage
algorithm as Diagonal Thresholding SIR (DT-SIR), and prove that DTSIR is consistent in estimating the central space under certain regularity
conditions. Extensive simulation studies show that DT-SIR performs better
than its competitors and is computationally efficient.
The rest of the paper is organized as follows. In Section 2, we briefly
describe the SIR procedure and introduce the notations. In Section 3, after a brief review of existing asymptotic results of SIR procedure, we state
Theorems 2 and 3 to discuss the phase transition phenomenon of SIR. In
Section 4, we propose the DT-SIR method and show that DT-SIR is consistent in high dimensional data analysis. In Section 5, we provide simulation
studies to compare DT-SIR with its competitors. Concluding remarks and
discussions are put in Section 6. All the proofs are presented in appendices.
2. Preliminaries and notations.
2.1. Sliced inverse regression Consider the multiple index model
(1)
y = f (β τ1 x, · · · , β τd x, ǫ)
where x ∈ Rp , ǫ is the noise and f is an unknown link function. Without
loss of generality, we assume that E[x] = 0 ∈ Rp . Although the p × d matrix
V = (β 1 , · · · , β d ) is not identifiable, the space spanned by the β’s, which is
called the column space of V and denoted by col(V ), might be identified. Li
[1991] proposed the Sliced Inverse Regression (SIR) procedure to estimate
the central space col(V ) without knowing f (·), which can be briefly summarized as follows: Given n i.i.d. samples (yi , xi ), i = 1, · · · , n, SIR first
divides them into H equal-sized slices according to the order statistics y(i) .1
We re-express the data as yh,j and xh,j , where (h, j) is the double subscript
in which h refers to the slice number and j refers to the order number of a
sample in the h-th slice, i.e.,
yh,j = y(c(h−1)+j) ,
xh,j = x(c(h−1)+j) .
Here x(k) is the concomitant of y(k) . Let the sample mean in the h-th slice
be xh,· , and let the mean of all the samples be x. Then, Λp , var(E[x|y])
can be estimated by:
(2)
H
1 X
b
ΛH =
x̄h,· x̄τh,· .
H
h=1
1
To ease notations and arguments, we assume that n = cH and H = o (log(n) ∧ log(p))
throughout the article.
4
Q. LIN, Z. ZHAO AND J. S. LIU
Based on the observation that
(3)
col(Λ) = Σx col(V ),
b −1 col(Vb H ) where Vb H is
the SIR then estimates the central space col(V ) by Σ
x
b H .. Throughout the article,
the matrix formed by the top d eigenvectors of Λ
we assume that d is fixed and the d-th largest eigenvalue λd of Λp is bounded
away from 0 when n, p → ∞.
In order for SIR to result in a consistent estimate of the central space, Li
[1991] imposed the the following two conditions:
• (A1). Linearity condition: For any ξ ∈ Rp , E[ξ τ x|β τ1 x, · · · , β τd x] is
a linear combination of β τ1 x, · · · , β τd x.
• (A2). Coverage condition: The dimension of the space spanned by
the central curve equals the dimension of the central space, i.e., d′ = d.
2.2. Further Notations. Let Sh be the h-th interval (yh−1,c , yh,c ] for 2 ≤
h ≤ H − 1, S1 = (−∞, y1,c ] and SH = (yH−1,c , ∞). Note that these intervals depend on the order statistics y(i) and are thus random. For any
ω
R in the product sample space, we define a random variable δh = δh (ω) =
y∈Sh (ω) f (y)dy where f (y) is the density function of y. For I ⊂ {1, · · · , n}, J ⊂
{1, · · · , p} and a n × p matrix A, AI,J denotes the |I| × |J | sub-matrix
formed by restricting the rows of A to I and columns to J . In articular,
A−,J denotes the sub-matrix formed by restricting the columns to J ; For
a matrix B = AI,J ∈ R|I|×|J | , we embed it into Rp×p by putting 0 on
entries outside I × J and denote the new matrix as e(B). Similar notations
apply to vectors. For two positive numbers a and b, we let a ∨ b ≡ max{a, b}
and let a ∧ b ≡ min{a, b}. Let τ (x, t) = x × 1(|x| > t) be the hard thresholding function. Throughout the article, C, C1 and C2 are used to denote
generic absolute constants, though the actual value may vary from case to
case. For a vector x, we denote its k-th entry as x(k). Let β 1 and β 2 be
two vectors with the same dimension, the angle between these two vectors
is denoted as ∠(β 1 , β 2 ). For two sequences {an }, {bn }, we let an ≪ bn stand
for an = O(bǫn ) for some positive ǫ < 1 and let an ≻ bn stand for lim abnn = 0.
3. Consistency of SIR. In order to control the behavior of SIR, we
need to impose the following boundedness condition (A3) on the predictors’
covariance matrix in addition to the tail condition (sub-Gaussian) on their
joint distribution. We also need a condition (A4) for the central curve.
• (A3) Boundedness Condition: x is sub-Gaussian; and there exist
positive constants C1 , C2 such that
C1 ≤ λmin (Σx ) ≤ λmax (Σx ) ≤ C2
DT-SIR
5
where λmin (Σx ) and λmax (Σx ) are the minimal and maximal eigenvalues of Σx respectively.
• (A4) The central curve m(y) , E[x|y] has finite fourth moment and
is ϑ-sliced stable (defined below) with respect to y and m(y).
Definition 1. For two positive constants γ 1 < 1 < γ 2 , let AH (γ 1 , γ 2 )
be the collection of all the partition −∞ = a0 < a1 < · · · < aH−1 < aH = ∞
of R satisfying that
γ
γ1
≤ P (ai ≤ y < ai+1 ) ≤ 2 .
H
H
The central curve m(y) = E[x|y] is called ϑ-sliced stable with respect to y
for some ϑ > 0 if there exist positive constants γ i , i = 1, 2, 3 such that for
any β in the central space for any partition in AH (γ 1 , γ 2 ), we have
(4)
H−1
1 X
γ
var(β τ m(y) | ah ≤ y ≤ ah+1 ) ≤ 3ϑ var(β τ m(y)).
H
H
h=0
The central curve is sliced stable if it is ϑ-sliced stable for some positive
constant ϑ.
Remark 1. Note that we only need (4) to hold for all unit vectors in
the central space by rescaling. By considering the orthogonal decomposition
of β in a general space with respect to the central space and its complement,
it is easy to see that the sliced stability implies that (4) holds true for all
vector β. In particular, we have the following two useful consequences of the
slice-stability.
i) By choosing β τ = (0, . . . , 0, 1, 0, . . . , 0) with 1 at the k-th position, we
have
|
H
X
h=0
var(m(y, k) | ah ≤ y ≤ ah+1 )| ≤ γ 3 H 1−ϑ var(m(y, k)),
where m(y, k) is the k-th coordinate of the central curve m(y).
ii) Since equation (4) holds for all unit vector β, we have
k
H
X
h=0
var(m(y) | ah ≤ y ≤ ah+1 )k2 ≤ γ 3 H 1−ϑ kvar(m(y))k2 .
6
Q. LIN, Z. ZHAO AND J. S. LIU
Remark 2. Suppose E[m(y)] = 0 and there are n samples mi , m(yi ).
xh,· , respectively. On one
Let mh,i and mh,· be defined similarly to xh,i and P
hand, we have the classic consistent estimator n1 i mi mτi ofPvar(m(y)).
On the other hand, if we expect that the slice-based estimate H1 h mh,· mτh,·
of var(m(y)) is consistent, we must require that the average loss of variance
in each slice to decrease to zero as H increases, i.e.,
(5)
1 X
1 X1X
1X
mi mτi =
(m̄h,· − m̄h,i )2 → 0.
m̄h,· m̄τh,· −
H
n
H
c
i
h
h
i
In Definition 1, we simply choose the decreasing rate to be a power of H. It
would be easily seen that if m is smooth and y is compactly supported then
(5) holds automatically. In this sense, for general curve m and random
variable y, the sliced stability is a condition on smoothness of the central
curve m and tail distribution of m(y). This is not surprised at all, since most
work on the consistency of SIR estimate requires some kind of smoothness
for the central curve and a tail distribution control for m(y).
The most popular smoothness and tail condition might be the one proposed by Hsing and Carroll [1992] (later used in Zhu et al. [2006], Zhu and
Ng [1995]) in their proof of the consistency of SIR, which is explained below. For B > 0 and n ≥ 1, let Πn (B) be the collection of all the n-point
partitions −B ≤ y(1) ≤ · · · ≤ y(n) ≤ B of [−B, B]. First, they assumed that
the central curve m(y) satisfies the following smoothness condition
lim
sup
n→∞ y∈Π (B)
n
n−1/4
n
X
i=2
km(yi ) − m(yi−1 )k2 = 0, ∀B > 0.
Second, they assumed that for B0 > 0, there exists a non-decreasing function
m(y)
e
on (B0 , ∞), such that
(6)
m
e 4 (y)P (|Y | > y) → 0 as y → ∞
km(y) − m(y ′ )k2 ≤ |m(y)
e
− m(y
e ′ )| for y, y ′ ∈ (−∞, −B0 ) ∪ (B0 , ∞)
e 4] <
By changing the tail condition (6) to a slightly stronger condition E[m(y)
∞, Neykov et al. [2015] proved that the modified condition implies the sliced
stability condition. Now, we are ready to state our main results.
Theorem 1.
(7)
Under conditions (A1), (A2), (A3) and (A4), we have
r
H 2p
H 2p
1
b
+
).
kΛH − Λp k2 = OP ( ϑ +
H
n
n
DT-SIR
7
The proof of the theorem is deferred to the Appendix. As a direct consequence of Theorem 1, we observe that if ρ = limn→∞ np = 0, we may choose
H = log (n/p) such that the right hand side of equation (7) converges to 0.
b H is a consistent estimate of Λp if ρ = 0.
Thus, Theorem 1 implies that Λ
Remark 3 (More on Convergence Rate). Note that the convergence rate
in (7) depends on the choice of H. This may seem not very desirable at the
b H might be different from that
first glance. Since the convergence rate of Λ
b
of col(V H ), we may expect that the convergence rate of col(Vb H ) does not
depend on the choice of H. In fact, we have
b H − Λp = Λ
b H PV − Λp .
b H − PV Λ
b H PV + PV Λ
(8)
Λ
From the proof of Theorem
q 1, we can easily check that the first term is of
2
pH 2
1
convergence rate n + pH
n and the second term is of rate H ϑ . Since
b H PV and Λp share the same column space, if we are only interested in
PV Λ
estimating PV , then the convergence rate of the second term does not matter
provided that H is a large enough integer, which may depend on ϑ and γ 3
but does not depend on n and p. For such an H, if AH (γ 1 , γ 2 ) is non-empty,
Theorem 1 and (8) hold for both categorical and continuous response variable
Y.
Theorem 2. Under conditions (A1), (A2), (A3), (A4) and assuming
that ρ = lim np = 0, we have
b H − Σ−1 Λp k2 → 0
b −1 Λ
kΣ
x
x
as n → ∞
b x = 1 Pn xi xτ .
with probability converging to one, where Σ
i
i=1
n
We define the distance D(V 1 , V 2 ) of two d-dimensional subspaces V 1 and
V 2 as the operator norm (or Frobenius norm) of the difference between PV 1
e ’s satisfy Σx β
e = λi Λp β
e,
and PV 2 . Simple linear algebra shows that if the β
i
i
i
then
o
n
e ,··· ,β
e .
col(V ) = span β
1
d
b −1 , Λ
b H ).
Let Vb be the matrix formed by the top d generalized eigenvectors of (Σ
x
Recall that the d-th eigenvalue of Λp is assumed to be bounded away from
0. Therefore Theorem 2 implies that D(PVb , PV ) → 0 when ρ = 0.
We have already shown that the SIR procedure provides us with a consistent estimate of the sufficient dimension reduction space when ρ = 0 under
mild conditions. It is then natural to ask: is this condition necessary? Our
next theorem gives the answer.
8
Q. LIN, Z. ZHAO AND J. S. LIU
Theorem 3. Under conditions (A1), (A2), (A4) and assuming that
x ∼ N (0, Ip ) for the single index model
y = f (β τ x, ǫ),
we have:
b H −Λp k2 , as a function of ρ, is dominated
(i) When ρ = lim np ∈ (0, ∞), kΛ
√
by ρ ∨ ρ when H, n → ∞;
b be the principal eigenvector of the SIR estimator Λ
b H . If ρ =
(ii) Let β
p
lim n > 0, then there exists a positive constant c(ρ) > 0 such that
b > c(ρ)
lim inf E∠(β, β)
n→∞
with probability converges to one.
We illustrate this result via a numerical study of the linear model
(9)
y = xτ β + ǫ where β τ = (1, 0, · · · , 0), x ∼ N (0, Ip ), ǫ ∼ N (0, 1).
b is related to the dimension p for fixed ratio
Figure 1 shows how E∠(β, β)
p
ρ = n (taking values in {.1, .3, .7, 1, 2, 4}), where β is estimated by the
b is calculated
SIR with the slice number H = 10. For each p, E∠(β, β)
based on 100 iterations. It is seen that this expected angle converges to a
positive number when the ratio ρ is non-zero. In Figure 2, we have plotted
b against the ratio ρ = p , varying between 0.01 and 4 with an
the E∠(β, β)
n
increment of 0.01. The sample size n is 200 and the slice number H is 10. It
is seen that the expected angle decreases to zero as ρ approaches zero, and
increases monotonically when ρ increases.
Results in this section have shown that there is a phase transition phenomenon of the SIR procedure. That is, the estimate of the dimension reduction space is consistent if and only if the ratio ρ = lim np = 0. This provides
a theoretical justification of imposing additional structure assumption such
as sparsity in high dimension.
4. SIR in ultra-high dimension. As we have shown in Section 3, the
SIR estimator fails to be consistent if ρ = lim np 6= 0. Hence, when p ≫ n,
some structural assumptions are necessary for getting a consistent estimate
of the central space. In this paper, we assume that both the loadings of all the
directions β j ’s and the covariance matrix Σx are sparse. Other structural
assumptions will be studied in our future work. For β i ’s, we impose the
following prevalent sparsity condition.
9
DT-SIR
0.4
0.235
0.395
0.23
0.39
0.385
0.225
0.38
0.22
0
200
400
600
800
1000
1200
1400
1600
0.375
0.6
0.72
0.595
0.71
0.59
0
200
400
600
800
1000
1200
1400
1600
0
200
400
600
800
1000
1200
1400
1600
0
200
400
600
800
1000
1200
1400
1600
0.7
0.585
0.69
0.58
0.68
0.575
0.67
0.57
0.66
0.565
0.65
0.56
0.64
0.555
0.55
0
200
400
600
800
1000
1200
1400
1600
1.05
0.63
1.25
1.2
1
1.15
0.95
1.1
0.9
1.05
0.85
0.8
1
0
200
400
600
800
1000
1200
1400
1600
0.95
Fig 1: Numerical approximations of E∠(β̂, β) for model (9) as a function of
dimension p for ρ = .1, .3, .7, 1, 2, and 4, respectively (up left, up right,
b is estimated by
middle left, middle right, lower left, lower right), where β
SIR.
b and the ratio p/n where β
b is estimated
Fig 2: The relationship of E∠(β, β)
by SIR.
10
Q. LIN, Z. ZHAO AND J. S. LIU
n
o
• (A5) s = |S| ≪ p where S = i | β j (i) 6= 0 for some j, 1 ≤ j ≤ d
and |S| is the number of elements in the set S.
For Σx , the following class of covariance matrices has been introduced in
Bickel and Levina [2008] (see also Cai et al. [2010]).
n
X
{|σi,j | : |i − j| > l} ≤ Cl−α for all l > 0,
U (ǫ0 , α, C) = Σx : max
j
i
and 0 < ǫ0 ≤ λmin (Σx ) ≤ λmax (Σx ) ≤
1 o
.
ǫ0
In this paper, to simplify the notations and arguments, we choose a slightly
stronger condition.
• (A6) Σx ∈ U (ǫ0 , α, C) and max1≤i≤p ri is bounded where ri is the
number of non-zero elements in the i-th row of Σx .
Let T = k | var (E[x(k)|y]) 6= 0 . If k ∈ T , there exists η ∈ col(Λ) such
that η(k) 6= 0. Since we have (3):
Σx col(V ) = col(Λ),
there exits a β ∈ col(V ) such that η = Σx β. Thus if k ∈ T , then k ∈
supp(Σx β) for some β ∈ col(V ). In particular, with the above sparsity
assumptions (A5) and (A6) , we have |T | ≤ s max1≤i≤p ri = O(s).2 Note
that our goal here is to recover the column space col(V ) rather than S.
Indeed, we are not able to consistently recover S unless for the trivial case.
The key for recovering cov(V ) is to consistently recovering the set T .
At the population level, var(E(x(k))|y) can separate T from T c . When
there are only finite samples, we use
(10)
varH (x(k)) =
H
1 X
x̄h,· (k)2
H
h=1
as an estimate of var(E(x(k))|y). These are the diagonal elements of the
b H . Note that these quantities depend on the sliced sample means,
matrix Λ
which are neither independent nor identically distributed. Thus, the usual
concentration inequalities for χ2 are no longer applicable. We need extra
efforts to get the concentration inequalities; this concentration result is one
of the main technical contributions of this article, and can be further generalized.
2
We could introduce ξ = max1≤i≤p ri , then |T | ≤ sξ. The arguments below still work,
except we might need sξ = o(p).
11
DT-SIR
Remark 4. The link function f ( ) is not involved explicitly in the definition of varH (x(k)), and only the order statistics of the response is required.
This nonparametric characteristic of the method is of particular interest to
us and will be further investigated in a future research. Screening statistics
inspired by the sliced inverse regression idea have been proposed in various
formats, such as those in Jiang and Liu [2014], Zhu et al. [2012] and Cui
et al. [2015].
With the quantities varH (E[x(k)|y]), we define the inclusion set Ip (t) and
the exclusion set Ep (t) below, which depend on a thresholding value t:
n
o
n
o
Ip (t) = k |varH (x(k)) > t and Ep (t) = k |varH (x(k)) ≤ t .
Note that Ip (t) can be viewed as an estimate of T and is thus also denoted
by Tb . After reducing the dimension to a level such as p/n is sufficiently
b b
b
b T ,T is a consistent estimate of ΛT ,T . Let Vb T be the
small, the SIR estimator Λ
b b
b
b T ,T . We then use Σ
b −1
bT
matrix formed by the top d eigenvectors of Λ
x col(e(V )) to
b −1
estimate the central space col(V ), where Σ
x is a consistent estimate of Σx .
Estimating the covariance matrix and precision matrix in high dimension is
a challenging problem by itself and is not a main focus of this article. We
employ the methods of Bickel and Levina [2008] to solve it. In summary,
we propose the following Diagonal Thresholding screening SIR (DT-SIR)
algorithm:
Algorithm 1 DT-SIR
1. Calculate varH (x(k)) according (10) for k = 1, 2, · · · , p;
n
o
2. Let Tb = k | varH (x(k)) > t for an appropriate t ;
b ,T
b
T
b H be the SIR estimator of the conditional covariance matrix for the data (y, x−,Tb )
3. Let Λ
according to equation (2);
4. Let
b
T
Vb
be the matrix formed by the top d eigenvectors of
b −1
b
is the estimate of col(V )
5. Σ
x col e V
b b
b T ,T ;
Λ
b
T
A practical way to choose an appropriate t in step 2 will be presented in
Section 5 . To ensure theoretical properties, we need an assumption on the
signal strength:
• (S1) ∃ C > 0 and ω > 0 such that var(E[x(k)|y]) > Cs−ω when
E[x(k)|y] is not a constant.
12
Q. LIN, Z. ZHAO AND J. S. LIU
Theorem 4. Under conditions (A1) – (A6) and (S1), and let t = as−ω
for some constant a > 0 such that t < 12 var(m(y, k), ∀k ∈ T , we have
i) T c ⊂ Ep holds with probability at least
(11)
1 − C1 exp −C2
n
+
C
log(H)
+
log(p
−
s)
;
3
H 2 sω
ii) T ⊂ Ip holds with probability at least
(12)
1 − C4 exp −C5
n
+
C
log(H)
+
log(s)
,
6
H 2 sω
for some positive constants C1 , · · · , C6 .
This theorem has a simple implication. If
n
), so that
choose H = log( sω log(p)
n
sω
≻ log(p) + log(s), we may
n
≻ log(p) + log(H) + log(s).
H 2 sω
Thus , we know T = Ip with probability converging to one. Next, we have
results for the consistency of DT-SIR.
Theorem 5. Under the same assumptions and choosing the same t as
Theorem 4, if snω ≻ log(p) + log(s), we have
Tb ,Tb
b
ke(Λ
H ) − Λp k2 → 0 as n → ∞
n
).
with probability converging to one, where Tb = I(t) and H = log( sω log(p)
b x be the estimator of co-variance matrix from Bickel
Theorem 6. Let Σ
and Levina [2008]. Under the same assumptions of Theorem 5, we have
−1
Tb ,Tb
−1
b e(Λ
b
kΣ
H ) − Σx Λp k2 → 0 as n → ∞
x
with probability converging to one.
5. Simulation Studies. We consider the following settings in generating the design matrix x and the response y. In Settings I-III, each row of x
is independently sampled from N (0, I).
iid
• Setting I. yi = sin(xi1 + xi2 ) + exp(xi3 + xi4 ) + 0.5 ∗ ǫi , where ǫi ∼
N (0, 1);
13
DT-SIR
• Setting II. yi =
P7
• Setting III. yi =
iid
j=1 xij ∗ exp(xi8 + xi9 ) + 0.5 ∗ ǫi where ǫi ∼ N (0, 1);
P10
P20
iid
j=1 xij ∗ exp( i=11 xij ) + ǫi where ǫi ∼ N (0, 1);
In Settings IV to VI, each row of x is independently sampled from N (0, Σ).
iid
• Setting IV. yi = (xi1 + xi2 + xi3 )3 /2 + 0.5 ∗ ǫi , where ǫi ∼ N (0, 1)
and Σ = (σij ) is tri-diagonal with σii = 1, σi,i+1 = σi+1,i = ρ and
σi,i+2 = σi+2,i = ρ2 ;
P
iid
• Setting V. yi = 7j=1 xij ∗ exp(xi8 + xi9 ) + ǫi , where ǫi ∼ N (0, 1),
and Σ = B ⊗ Ip/10 with B = (bij )1≤i≤10,1≤j≤10 given as bij = ρ|i−j| ;
• Setting VI. Assume the same setting as in Setting V except that
Σ = (σij ) is tri-diagonal with σii = 1, σi,i+1 = σi+1,i = ρ and σi,i+2 =
σi+2,i = ρ2 .
• Setting VII. Assume the same setting as in Setting V except that
Σ = (σij ) is given as σij = ρ|i−j| .
DT-SIR first screens all the predictors according to the statistic varH,c (x(k)),
which requires a tuning parameter t. We chose t by using an auxiliary variable method based on an idea first proposed by Luo et al. [2006] and extended by Wu et al. [2007] and Zhu et al. [2011]. In our setting, for a given
sample (yi , xi ), we generate z i ∼ N (0, Ip′ ) where p′ is sufficiently large and
chosen as p in our simulation studies. It is known that y and z are independent. The threshold t can be chosen as
t̂ = max ′ {varH,c (z(k)).}
1≤k≤p
In DT-SIR, when n > 1000, H is chosen as 20; when n ≤ 1000, H is chosen
as 10 in the screening step and 20 in the SIR step.
We also consider the following alternative methods in the screening step:
Sure Independent Ranking and Screening (SIRS) in Zhu et al. [2011], SIR
for variable selection via Inverse modeling (SIRI) in Jiang and Liu [2014],
and trace pursuit in Yu et al. [2016]. As a comparison, we also considered
two screening methods that are not based on the sliced regression: Distance
correlation in Székely et al. [2007] and SURE independence Fan and Lv
[2008]. For SIRS, the threshold is chosen according to the auxiliary statistic
(2.9) of Zhu et al. [2011]. For SIRI, the predictors are chosen according to
10-fold cross validation. The threshold values c̄SIR and cSIR are chosen as
the 10-th and 5-th quantile of a weighted χ2 distribution given in Theorem
3.1 of Yu et al. [2016]. In both SURE and DC screening, the top ⌊γn⌋ where
γ = 0.01 are kept for subsequent analyses.
After the screening step, similar to DT-SIR, we then applied the SIR algorithm (steps 3-5 of DT-SIR) to estimate col(V ). These alternative methods
14
Q. LIN, Z. ZHAO AND J. S. LIU
are denoted as SIRS-SIR, SIRI-SIR, SURE-SIR, DC-SIR, and TP-SIR, respectively, in the following discussions. Another method that we compared
with is the sparse SIR, abbreviated as SpSIR, proposed in Li [2007]. After
obtaining an estimator col(Vb ), we calculate D(Pcol(Vb ) , Pcol(V ) ) as a measure
of the estimation error. We replicate this step 100 times, and calculate the
average distance for the estimation result from each method and report these
numbers in Table 1-3. For each setting, the average distance of the optimal
method is highlighted using bold fonts. We further run a two-sample T-test
to test if the actual estimation error of each method is significantly different
from that of the best method for that example at 1% level of significance.
Under all settings, the average distance obtained by DT-SIR was much
smaller than that obtained by SpSIR and SURE-SIR. The p-values for comparing DT-SIR and SpSIR/SURE-SIR are all significant at the 0.01 level.
When p ≥ n, the sparse SIR completely failed
√ because the average distance
of the estimated space to the true space is 2d, indicating that the space
estimated by sparse SIR is orthogonal to the true space spanned by β.
Under settings II-IV, DT-SIR performed either the best or not significantly worse than the best method. For all other cases, DT-SIR performed
the best except for a few cases: Setting I when n = 500, p = 1000, setting
V when n = 500, p = 6000, setting VI when n = 500, p = 6000, and setting
VII when n = 1000, p = 1000.
When p = 6000, n = 500, both DT-SIR and SIRI-SIR were the winners.
Under Setting III, DT-SIR performed better than SIRI-SIR; under settings
V and VI, SIRI-SIR performed better than DT-SIR; under other settings,
these two methods were comparable.
To graphically show the performance of various methods, we consider
setting IV with d = 1. Consider two cases when (n, p) = (2000, 1000) and
(n, p) = (500, 100). We calculated the estimated directions β̂ using various
methods and computed the angle between < β̂ > and < β >. We replicate
this step 100 times to calculate the average angles for each method. The
results are displayed in Figure 3, which shows clearly that DT-SIR performed
better than its competitors.
Additionally, DT-SIR is computationally efficient. To show this, we report the computing time for one replication under Setting II for various
pairs of (n, p) in Table 4. All computations were done on a computer with
Intel Xeon(R) E5-1620 [email protected] and 16GB memory. It is clearly seen
that DT-SIR performed as fast as SURE-SIR, and both were much faster
than other competitors. Consider the case when p = 3000, n = 2000. The
computation time of DT-SIR was only 30 seconds; while that for DC-SIR
was 21 minutes and 38 seconds, and the that for TP-SIR was 6 minutes and
15
DT-SIR
Table 1
The average distance of the space estimated by each of the 7 methods tested to the true
space col(V ) under various settings with p = 1000. The boldfaced number in each row
represents the best result for that simulation scenario, and the “*” in cells represents that
the p-value of the two-sample T-test comparing the estimation error of the corresponding
method with that of the best method is less than 0.01.
n
500
1000
2000
3000
DT-SIR
0.655(*)
0.3
0.221
0.167
SIRI-SIR
0.751(*)
0.431(*)
0.341(*)
0.245(*)
SIRS-SIR
0.492
0.309
0.226
0.149
SpSIR
2(*)
2(*)
1.58(*)
1.48(*)
SURE-SIR
1.39(*)
1.29(*)
1.04(*)
0.816(*)
DC-SIR
0.731(*)
0.632(*)
0.655(*)
0.641(*)
TP-SIR
1.18(*)
0.94(*)
0.784(*)
0.713(*)
II
500
1000
2000
3000
0.383
0.235
0.161
0.134
0.396
0.227
0.157
0.129
0.371
0.256
0.189(*)
0.153(*)
2(*)
2(*)
1.25(*)
0.975(*)
1.64(*)
1.36(*)
1.25(*)
1.12(*)
1.08(*)
0.266(*)
0.387(*)
0.404(*)
0.389
0.318(*)
0.264(*)
0.23(*)
III
500
1000
2000
3000
1.15
0.426
0.263
0.214
1.48(*)
0.974(*)
0.403(*)
0.297
1.38(*)
0.596(*)
0.29(*)
0.238(*)
2(*)
2(*)
1.33(*)
1.06(*)
1.97(*)
1.94(*)
1.89(*)
1.82(*)
1.85(*)
1.57(*)
0.996(*)
0.475(*)
1.13
0.429
0.338(*)
0.299(*)
IV
500
1000
2000
3000
0.263
0.219
0.161
0.134
0.257
0.447(*)
0.4(*)
0.377(*)
0.333
0.25
0.196(*)
0.177(*)
1.41(*)
1.41(*)
0.42(*)
0.297(*)
0.335(*)
0.436(*)
0.442(*)
0.43(*)
0.334(*)
0.459(*)
0.469(*)
0.458(*)
0.332(*)
0.469(*)
0.452(*)
0.438(*)
V
500
1000
2000
3000
0.546
0.401
0.288
0.249
0.529
0.463(*)
0.418(*)
0.399(*)
0.562
0.514(*)
0.341(*)
0.284(*)
2(*)
2(*)
1.51(*)
1.24(*)
1.62(*)
1.15(*)
0.926(*)
0.691(*)
1.24(*)
0.367
0.569(*)
0.597(*)
1.09(*)
0.615(*)
0.54(*)
0.511(*)
VI
500
1000
2000
3000
0.568
0.427
0.311
0.265
0.535
0.524(*)
0.469(*)
0.456(*)
0.566
0.548(*)
0.351(*)
0.307(*)
2(*)
2(*)
1.51(*)
1.25(*)
1.64(*)
1.22(*)
0.927(*)
0.807(*)
1.24(*)
0.39
0.598(*)
0.622(*)
1.08(*)
0.641(*)
0.583(*)
0.56(*)
VII
500
1000
2000
3000
0.556
0.436(*)
0.303
0.258
0.534
0.528(*)
0.465(*)
0.468(*)
0.585(*)
0.545(*)
0.358(*)
0.319(*)
2(*)
2(*)
1.51(*)
1.25(*)
1.66(*)
1.22(*)
0.747(*)
0.698(*)
1.26(*)
0.39
0.589(*)
0.63(*)
1.11(*)
0.643(*)
0.579(*)
0.558(*)
I
17 seconds.
6. Conclusion. When the dimension p diverges to infinity, classical
statistical procedures often fail unless additional structures such as sparsity
conditions are imposed. Understanding boundary conditions of a statistical
procedure provides us theoretical justification and practical guidance for our
modeling efforts. In this article, we provide a new framework to show that
ρ = lim np is the phase transition parameter for the SIR procedure. Under
certain conditions, it is shown that the SIR estimator is consistent if and
16
Q. LIN, Z. ZHAO AND J. S. LIU
Table 2
The average distance of the space estimated by each of the 7 methods we tested to the
true space col(V ) under various settings with n = 2000.
p
DT-SIR
SIRI-SIR
SIRS-SIR
SpSIR
SURE-SIR
DC-SIR
TP-SIR
I
500
1000
2000
3000
0.213
0.221
0.241
0.23
0.312(*)
0.341(*)
0.29
0.278
0.206
0.226
0.214
0.218
1.44(*)
1.58(*)
2(*)
2(*)
0.903(*)
1.04(*)
1.07(*)
1.17(*)
0.629(*)
0.655(*)
0.677(*)
0.683(*)
0.772(*)
0.784(*)
0.793(*)
0.797(*)
II
500
1000
2000
3000
0.163
0.161
0.172
0.164
0.16
0.157
0.159
0.158
0.19(*)
0.189(*)
0.196(*)
0.199(*)
0.83(*)
1.25(*)
2(*)
2(*)
1.22(*)
1.25(*)
1.23(*)
1.3(*)
0.369(*)
0.387(*)
0.404(*)
0.414(*)
0.26(*)
0.264(*)
0.259(*)
0.261(*)
III
500
1000
2000
3000
0.272
0.263
0.262
0.269
0.353
0.403(*)
0.368
0.344
0.29(*)
0.29(*)
0.285(*)
0.291(*)
0.916(*)
1.33(*)
2(*)
2(*)
1.84(*)
1.89(*)
1.92(*)
1.93(*)
0.846(*)
0.996(*)
0.98(*)
1.09(*)
0.341(*)
0.338(*)
0.339(*)
0.339(*)
IV
500
1000
2000
3000
0.145
0.161
0.16
0.15
0.409(*)
0.4(*)
0.395(*)
0.395(*)
0.182(*)
0.196(*)
0.198(*)
0.216(*)
0.248(*)
0.42(*)
1.41(*)
1.41(*)
0.406(*)
0.442(*)
0.472(*)
0.49(*)
0.433(*)
0.469(*)
0.506(*)
0.527(*)
0.438(*)
0.452(*)
0.447(*)
0.447(*)
V
500
1000
2000
3000
0.272
0.288
0.289
0.3
0.434(*)
0.418(*)
0.418(*)
0.417(*)
0.353(*)
0.341(*)
0.351(*)
0.372(*)
1.09(*)
1.51(*)
2(*)
2(*)
0.876(*)
0.926(*)
0.868(*)
0.968(*)
0.547(*)
0.569(*)
0.596(*)
0.605(*)
0.539(*)
0.54(*)
0.537(*)
0.544(*)
VI
500
1000
2000
3000
0.307
0.311
0.309
0.31
0.479(*)
0.469(*)
0.461(*)
0.46(*)
0.368(*)
0.351(*)
0.399(*)
0.408(*)
1.1(*)
1.51(*)
2(*)
2(*)
0.858(*)
0.927(*)
1.08(*)
1(*)
0.566(*)
0.598(*)
0.617(*)
0.638(*)
0.583(*)
0.583(*)
0.585(*)
0.587(*)
VII
500
1000
2000
3000
0.299
0.303
0.309
0.308
0.482(*)
0.465(*)
0.455(*)
0.46(*)
0.343(*)
0.358(*)
0.383(*)
0.357(*)
1.09(*)
1.51(*)
2(*)
2(*)
0.818(*)
0.747(*)
0.966(*)
0.858(*)
0.564(*)
0.589(*)
0.622(*)
0.626(*)
0.583(*)
0.579(*)
0.578(*)
0.58(*)
Table 3
The average distance of the space estimated by each of the 7 methods tested to the true
space col(V ) under various settings with n = 500 and p = 6000.
I
II
III
IV
V
VI
VII
DT-SIR
SIRI-SIR
SIRS-SIR
SpSIR
SURE-SIR
DC-SIR
TP-SIR
0.694
0.446
1.35
0.163
0.481(*)
0.463(*)
0.44
0.631
0.462
1.56(*)
0.122
0.431
0.423
0.412
0.606
0.414
1.56(*)
0.245(*)
0.486(*)
0.494(*)
0.477(*)
2(*)
2(*)
2(*)
1.41(*)
2(*)
2(*)
2(*)
1.43(*)
1.74(*)
1.99(*)
0.27(*)
1.62(*)
1.62(*)
1.61(*)
0.97(*)
1.08(*)
1.88(*)
0.305(*)
1.1(*)
1.11(*)
1.1(*)
1.19(*)
0.4
1.37
0.195(*)
0.995(*)
0.999(*)
1.03(*)
1
DT-SIR
0.5
0.5
1.0
Directions
1.0
Directions
−0.5
0.0
True Beta
SIRI SIR
SIRS SIR
Sparse SIR
DT−SIR
−1.0
−1.0
−0.5
0.0
True Beta
SIRI SIR
SIRS SIR
Sparse SIR
DT−SIR
−1.0
−0.5
0.0
0.5
1.0
−1.0
−0.5
0.0
0.5
1.0
Fig 3: Simulated value of E∠(β̂, β) for the various methods. Left panel:
(n, p) = (2000, 1000); Right panel: (n, p) = (500, 1000).
Table 4
Comparison of computing time under setting II.
DT-SIR
n
500
1000
2000
3000
1”
2”
3”
4”
p
500
1000
2000
3000
1”
3”
12”
30”
SIRI-SIR
SIRS-SIR
SpSIR
SURE-SIR
DC-SIR
TP-SIR
1’12”
2’2”
3’27”
4’59”
7”
20”
1’14”
2’45”
p=1000
11”
11”
13”
15”
1”
1”
2”
3”
24”
1’52”
7’38”
6’51”
29”
1’2”
2’18”
3’7”
2’48”
3’27”
4’55”
6’0”
35”
1’14”
2’35”
4’10”
n=2000
2”
13”
1’39”
5’19”
1”
2”
12”
30”
3’46”
7’38”
14’24”
21’38”
1’7”
2’18”
3’22”
6’17”
only if ρ = 0. When ρ > 0, where the original SIR fails to be consistent, we
propose a two-stage method, DT-SIR for variable screening and selection
in ultra-high dimension situations and show that the method is consistent.
We have used simulated examples to demonstrate the advantages of DT-SIR
compared to its competitors. This method is computationally fast and can
be easily implemented for large data sets.
Appendices
2
Q. LIN, Z. ZHAO AND J. S. LIU
In the following two sections we offer some details about our theoretical
derivations, but some more tedious intermediate steps (organized as Lemmas
6-21) are deferred to the Supplemental Document to this article, which is
available on line.
A. The Key Lemma. The following lemma plays an important role
in developing the high dimensional theory for sliced inverse regression. The
proof of this key lemma is lengthy and technical. It will be helpful to keep
in mind that H and ν (if they are not constants) grow at very slow rate
compared with c and n (e.g., polynomial of log(n)). Let m(y) = E[x|y], and
x = m(y) + ǫ. Notations mh,j , mh,· , m, and ǫh,j , ǫh,· , ǫ are similarly
defined as xh,j , xh,· and x that were introduced before.
Lemma 1. Let x ∈ Rp be a sub-Gaussian random variable which is upper
exponentially bounded by K (see Definition 4). For any unit vector β ∈ Rp ,
let x(β) = hx, βi and m(β) = hm, βi = E[x(β) | y], we have the following:
i) If var(m(β)) = 0 , there exists positive constants C1 , C2 and C3 such
that for any b = O(1) and sufficiently large H, we have
nb
P(varH (x(β)) > b) ≤ C1 exp −C2 2 + C3 log(H) .
H
ii) If var(m(β)) 6= 0 , there exists positive constants C1 , C2 and C3 such
that, for any ν > 1, we have
|varH (x(β)) − var(m(β))| ≥
1
var(m(β))
2ν
with probability at most
n var(m(β))
+ C3 log(H) .
C1 exp −C2
H 2ν 2
where we choose H such that H ϑ > C4 ν for some sufficiently large
constant C4 .
3
DT-SIR
A.1. Proof of Lemma 1 i) If m(β) = 0 (or equivalently var(m(β)) = 0),
since
!2
c−1
c−1 1 X
1
2
ǭh,· (β) =
ǫh,i (β) + ǫh,c (β)
c c−1
c
i=1
!2
2
c−1
1
1 X
ǫh,c (β)
ǫh,i (β) + 2
≤2
c−1
c
i=1
for h = 1, ..., H − 1 and ǭH,· (β) =
varH (x(β)) − var(m(β))
1
c
Pc
i=1 ǫH,i (β),
we have
H−1
1 X
1
=
ǭh,· (β)2 + ǭH,· (β)2
H
H
h
!2
H−1
H−1
c−1
X
X
2 X
2
1
2
ǫh,i (β) + ǭH,· (β)
+
≤
ǫh,c (β)2
H
c−1
Hc2
h
i=1
h
,2I + 2II.
Thus
(A.1)
P(varH (x(β)) > b) ≤ P(I > b/4) + P(II > b/4).
Lemma 17 (iii) in Supplement implies that
P (ǫ(β)|y∈Sh
t2
> t) ≤ CH exp − 2
K
for some positive constant C. Since E[x(β)|y] = 0,we have E[x(β)|y ∈ Sh ] =
0. From Lemma 9, we know that for 1 ≤ h ≤ H − 1, ǫh,i (β) can be treated
as c − 1 i.i.d. samples from ǫ(β)|y∈Sh . According to Lemma 17 (iv),
!
c−1
√
1 X
−b(c − 1)
√
P |
.
ǫh,i (β)| > b/2 ≤ C1 exp
c−1
8C2 HK 2 + 4 bK
i=1
Similarly, we have
!
c
√
1X
−bc
√
P |
ǫH,i (β)| > b/2 ≤ C1 exp
.
c
8C2 HK 2 + 4 bK
i=1
4
Q. LIN, Z. ZHAO AND J. S. LIU
Thus, if b = O(1) and H is sufficiently large, we have
−bc
−b(c − 1)
b
√
√
+ exp
P(I > ) ≤C1 (H − 1) exp
4
8C2 HK 2 + 4 bK
8C3 HK 2 + 4 bK
cb
≤C1 exp −C2 + C3 log(H)
H
for some positive constants C1 , C2 and C3 .
Since ǫi (β) are i.i.d. samples from a sub-Gaussian distribution ǫ(β) with
mean 0 and upper-exponentially bounded by 2K. Lemma 19 implies that if
b = O(1) and H is sufficiently large, we have
1X
ǫi (β)2 > bc/4)
n
i
1X
≤P(
ǫi (β)2 − E[ǫ(β)2 ] > bc/4 − E[ǫ(β)2 )
n
i
!
1X
2
2
2
ǫi (β) − E[ǫ(β) ] ≥ cb/4 − 4K
≤P
n
i
√
n(cb/4 − 4K 2 )
≤C1 exp −C2
K2
cb
≤C1 exp −C2 + C3 log(H)
H
P(II > b/4) ≤P(
for some positive constants C1 , C2 and C3 if H is sufficiently large. We used
in above the fact that E[ǫ(β)2 ] ≤ 4K 2 .
To summarize, if b = O(1) and H is sufficiently large, we have
cb
P(varH (x(β)) > b) ≤ C1 exp −C2 + C3 log(H)
H
for some positive absolute constants C1 , C2 and C3 .
A.2. Proof of Lemma 1 ii) Since x is sub-Gaussian and β is unit vector,
we know that var(m(β)) = O(1). If m(β) 6= 0 (or equivalently var(m (β)) 6=
5
DT-SIR
0), we have
varH (x(β)) − var(m(β))
1 X
xh,· (β)2 − var(m(β))
=
H
h
1 X
2 X
1 X
=
mh,· (β)2 +
mh,· (β)ǫh,· (β) +
ǫh,· (β)2
H
H
H
h
h
h
− var(m(β))
≤A1 + A2 + A3 + A4 ,
where
1 X
µh (β)2 − var(m(β)) ,
H
h
1 X
mh,· (β)2 − µh (β)2 ,
A2 =
H
h
1 X
A3 =
ǭh,· (β)2 ,
H
h
1 X
1 X
A4 = (
mh,· (β)2 )1/2 (
ǫh,· (β)2 )1/2 .
H
H
A1 =
(A.2)
h
h
Lemma 1 ii) is a direct corollary of the following properties of Ai ’s.
Lemma 2. Let the Ai ’s be defined as in equation (A.2). There exist
positive constants C1 , C2 and C3 , such that for any ν > 1 and H satisfying
H ϑ = N1 ν for sufficiently large N1 , we have that each of the following events
o
n
1
var(m(β)) ,
i) Θ1 = A1 ≤ 4ν
ii) Θ2 =
iii) Θ3 =
iv) Θ4 =
n
n
n
o
A2 ≤
1
8ν var(m(β))
A3 ≤
1
16ν var(m(β))
A4 ≤
1
16ν var(m(β))
,
o
o
,
,
occurs with probability at least
c var(m(β))
+ C3 log(H) .
(A.3)
1 − C1 exp −C2
Hν 2
6
Q. LIN, Z. ZHAO AND J. S. LIU
A.2.1. Proof of Lemma 2.
A.2.1.1. Proof of i) : Recall definitions of the
R random intervals Sh , h =
1, 2, · · · , H and random variable δh = δh (ω) = y∈Sh (ω) f (y)dy. We have
1 X
(µh (β))2 − var(m(β))
H
h
X
X
1 X
≤ var(m(β)) −
δh (µh (β))2 +
(µh (β))2 −
δh (µh (β))2
H
h
h
h
,B1 + B2
Let ǫ = Hn10 +1 where n0 = N2 ν for some sufficiently large constant N2
and
n let event E(ǫ) be odefined as in Lemma 11 in Section E, i.e., E(ǫ) =
ω |δh − H1 | > ǫ, ∀h . For any ω ∈ E(ǫ)c , we have
B1 =
X
h
δh (ω)var(m(β)|y ∈ Sh (ω))
X
1
+ ǫ)
var(m(β)|y ∈ Sh (ω))
H
h
γ
≤(1 + Hǫ) 3ϑ var(m(β))
H
2γ 3
≤
var(m(β)),
N1 ν
(A.4)
≤(
(A.5)
(A.6)
where inequality (A.4) follows from the fact that δh (ω) ≤ H1 + ǫ, inequality
(A.5) follows from the sliced stable condition (4) and inequality (A.6) follows
from the requirement that H ϑ > N1 ν, and the fact
X
X ǫ
δh (β τ µh )2
B2 ≤ǫ
(β τ µh )2 =
δh
h
h
Hǫ X
(A.7)
δh (β τ µh )2
≤
1 − Hǫ
h
2 X
≤
δh (β τ µh )2
N2 ν
h
where inequality (A.7) follows from the fact δh ≥ H1 − ǫ.
From (A.5), we observe that
X
2γ 3
2
δh (µh (β)) ≤ 1 +
(A.8)
var(m(β)).
N1 ν
h
7
DT-SIR
Combining with (A.7), we then have
2
2γ 3
B2 ≤
1+
var(m(β)).
N2 ν
N1 ν
So when E(ǫ)c occurs, we have
2γ 3
2
2γ 3
B1 + B2 ≤
+
1+
var(m(β)).
N1 ν
N2 ν
N1 ν
Note that N1 and N2 can be chosen sufficiently large so that
(A.9)
B1 + B2 ≤
1
4γ 3
var(m(β)) ≤
var(m(β)).
N1 ν
4ν
Consequently, conditioning on E(ǫ)c where ǫ =
N1 ν, then
(A.10)
1
HN2 ν+1 ,
if we choose H ϑ >
1
1 X
var(m(β)).
(µh (β))2 − var(m(β)) ≤
H
4ν
h
Since var(m(β)) = O(1), H ϑ > N1 ν and ǫ = HN21ν+1 , the desired probability bound follows from Lemma 11, i.e.,
√
Hc + 1
2
P(E(ǫ)) ≤C1 exp −
+ log(H Hc + 1)
32(Hn0 + 1)2
c var(m(β))
+
C
log(H)
.
≤C1 exp −C2
3
Hν 2
for some positive constants C1 , C2 and C3 .
Remark 5. From (A.10), conditioning on E(ǫ)c , we obtain the following
two inequalities
1 X
4γ 3
2
(A.11)
(µh (β)) ≤ 1 + ϑ var (m(β))
H
H
h
and
(A.12)
1/2
4γ 3
1 X
|µh (β)| ≤
1 + ϑ var (m(β))
.
H
H
h
In particular,
1
H
P
h (µh (β))
2
and
1
H
P
h |µh (β)|
are bounded by OP (1).
8
Q. LIN, Z. ZHAO AND J. S. LIU
A.2.1.2. Proof of ii) : Denote
by m′H (β), we have
1
c−1
Pc−1
i=1
mh,i (β) by m′h (β) and mH,· (β)
H
H
1 X
1 X
2
2
′
A2 ≤
mh (β) − µh (β) +
mh,c (β)2
H
Hc2
h=1
h=1
!
!1/2
1/2
H
H
H
1 X
2 X
2(c − 1) 1 X ′
2
2
mh (β)
+
+
m
(β)
µh (β)2
h,c
c
H
Hc2
Hc
h=1
h=1
h=1
, I + II + III + IV
Before we start proving this part, we need to introduce two events and bound
their probabilities. First, let
o
n
1 p
var(m(β)) .
(A.13)
E1 (N3 , ν) = η(β) >
N3 ν
n
o
where η(β) = max1≤h≤H m′h (β) − µh (β) . According to Lemma 17 (i),
(iv) and Bonferroni’s inequality, we have
!
1
−(c − 1)var(m(β))
p
(A.14) P (E1 (N3 , ν)) ≤2H exp
(N3 ν)2 2CHK 2 + N2 ν var(m(β))K
3
c var(m(β))
(A.15)
+ C3 log(H)
≤C1 exp −C2
Hν 2
for some positive constants C1 , C2 and C3 . Second, let
n
o
1
E2 (N4 , ν) , II >
var(m(β)) ,
N4 ν
then
!
1 X 2 var(m(β))
mi >
P (E(N4 , ν)) ≤ P
nc
N4 ν
i
√
var(m(β))
2
≤C1 exp −C2 n(c
)−K )
ν
c var(m(β))
+ C3 log(H)
≤C1 exp −C2
Hν
for some positive constant C1 , C2 and C3 . It is easily to see E(N4 , ν) ⊂
E(N4 , ν 2 ).
9
DT-SIR
For I. Conditioning on the event E(ǫ)c ∩E1 (N3 , ν)c , combining with (A.12),
we have
2η(β) X
1 X
η(β)(η(β) + 2|µh (β)|) ≤ η(β)2 +
|µh (β)|
I≤
H
H
h
h
!
1 2
2
4γ 3 1/2
≤
var(m(β))
+
1+ ϑ
N3 ν
N3 ν
H
1
var(m(β))
32ν
if N3 is sufficiently large .
≤
Remark 6.
we have
(A.16)
From above, conditioning on the event E(ǫ)c ∩ E1 (N3 , ν)c ,
H
1 + 32ν
1 X ′
var(m(β)).
m (β)2 ≤
H
32ν
h=1
For II.
Conditioning on E2 (N4 , ν)c , we have II ≤
var(m(y))
.
N4 ν
For III. When the event E(ǫ)c ∩ E1 (N3 , ν)c ∩ E2 (N4 , ν 2 )c occurs, according
to equation (A.16),
r
1
2(c − 1) 1 + 32ν 1
√
var(m(β)) <
var(m(β)).
III ≤
c
32ν
16ν
N4 ν
if N4 is sufficiently large.
For VI. When the event E(ǫ)c ∩ E1 (N3 , ν)c ∩ E2 (N4 , ν)c occurs, from
(A.10), we know
9
1
2 X
µh (β)2 ≤ var(m(β)) <
var(m(β)).
VI =
Hc
4c
16ν
h
To summarize, we know that there exist positive constant C1 , C2 , C3 and
C4 such that
1
A2 ≤ I + II + III + V I ≤
var(m(β))
8ν
holds on the event E(ǫ)c ∩ E1 (N3 , ν)c ∩ E2 (N4 , ν 2 )c which is with probability
at least
c var(m(β))
+ C3 log(H)
1 − C1 exp −C2
Hν 2
for some positive constants C1 ,C2 and C3 .
10
Q. LIN, Z. ZHAO AND J. S. LIU
A.2.1.3. Proof of iii) :
Similar to the proof of Lemma 1 (i) we have
P(A3 > b) ≤ C1 H exp
−(c − 1)b
√
8C2 HK12 + 4 bK2
!
for some positive constants C1 , C2 and C3 . In particular , if we take b =
1
16ν var(m(β)), we know that
A3 ≤
1
var(m(β))
16ν
with probability at least
c var(m(β))
+ C3 log(H)
1 − C1 exp −C2
Hν 2
for some positive constant C1 , C2 and C3 .
A.2.1.4. Proof of iv) : Let
D1 ,
1 X
mh,· (β)2 ,
H
D2 , A3 =
h
1 X
ǭh,· (β)2
H
h
Consequently,
(A.17)
1
P
>
var(m(β))
16ν
var(m(β))
2ν + 1
var(m(β)) + P D2 >
≤P |D1 | >
2ν
(2ν + 1)16ν
1/2 1/2
D1 D2
Note that
|D1 − var(m(β))| ≤ A2 + A1
According to (i) and (ii), the right hand side of (A.17) is bounded by
c var(m(β)
C1 exp −C2
+ C3 log(H)
Hν 2
for some positive constants C1 , C2 and C3 .
B. Proofs of theorems in section 3.
11
DT-SIR
B.1. Proof of Theorem 1. Let S be the central subspace of dimension
d ≪ p, i.e., y ⊥
⊥ x|P S x and dim(S) = d. We have the decomposition
x = P S x + P S⊥ x , z + w
(B.1)
= E[z|y] + z − E[z|y] + w , m + v + w
where z = P S x, m = E[z|y], v = z − E[z|y] and w = P S ⊥ x. Note that m
lies in the central curve, v lies in the central space and w lies in the space
perpendicular to S. We introduce
(B.2)
mh,j , mh,· , m,
z h,j , z h,· , z, and wh,j , wh,· , w
bz
similar to the definition of xh,j , xh,· and x. Consequently, we can define Λ
and have the following decomposition
(B.3)
X
bH ≡ 1
b z + ZW τ + WZ τ + WW τ ,
Λ
xh,· xτh,· = Λ
H
h
where
1
1
Z = √ (z 1,· , ..., z H,· ) and W = √ (w1,· , ..., w H,· ) .
H
H
b z − Λp k2 and kWW τ k2 .
We need to bound kΛ
Lemma 3.
kWW τ k2 ≤ OP (
(B.4)
H 2p
)
n
Proof. From Lemma 1, for any unit vector β ⊥ col(Λ), i.e. var(m(β)) =
0, we have
(B.5)
P(β τ WW τ β > C
H 2p
) ≤ C1 exp (−C2 p + log(H)) .
n
for some positive constants C1 and C2 . Then the ε-net argument (see e.g.,
2
Vershynin [2010]) implies that kWW τ k ≤ OP ( Hn p )
Lemma 4.
(B.6)
b z − Λp k ≤ OP
kΛ
1
Hϑ
b z k ≤ OP (1).
As a direct corollary, we have kΛ
.
12
Q. LIN, Z. ZHAO AND J. S. LIU
Proof. From Lemma 1, we have
C
c var(m(β))
τ b
P β (Λz − Λ)β > ϑ kΛk2 ≤ C1 exp −C2
+ C3 log(H) .
H
H 1+2ϑ
Note that we only need to verify it for β ∈ col(Λp ), which is a d-dimensional
b z − Λp k2 ≤ OP 1ϑ .
space. Then the ε-net argument implies that kΛ
H
Theorem 1 follows from Lemma 4 and Lemma 3. In fact,
b H − Λp k ≤ kΛ
b z − Λp k + kZW τ + WZ τ k2 + kWW τ k2
kΛ
!
r
1
H 2p H 2p
≤ OP
+
+
.
Hϑ
n
n
B.2. Proof of Theorem 2. Theorem 2 is a direct corollary of Theorem 1
and Lemma 13. In fact, we have:
−1
b −1
b
kΣ
X ΛH − ΣX Λp k2
b H − Λp k2 ,
b H k2 + kΣ−1 k2 kΛ
b −1 − Σ−1 k2 kΛ
≤ kΣ
X
X
X
which → 0 if ρ = limn→∞
p
n
= 0.
B.3. Proof of Theorem 3.
(i) The proof for part (i) is similar to the proof of Theorem 1 and the
standard Gaussian assumption on x simplifies the argument and improves
the results. Since w = P S ⊥ x is normal and independent of y, there exists
a normal random variable ǫ ∼ N (0, I) such that w = Σ1/2 ǫ where Σ =
cov(w). Using the decomposition (B.3), we may write
(B.7)
W=√
1
Σ1/2 E p×H
Hc
where E p,H is a p × H matrix with i.i.d. standard normal entries. Corollary
4 implies that
kWW τ k2 ≤ C
r
p
+
n
r
H
n
!2
≤ OP
p
n
.
13
DT-SIR
Lemma 4 implies
b z k2 ≤ kΛp k2 + OP
kΛ
By the Cauchy inequality, we have
Thus,
1
Hϑ
b z k2 kWW τ k2 ≤ OP
kC1 k22 ≤ kΛ
b H − Λp k2 ≤ OP ( 1 + p +
kΛ
Hϑ n
.
p
n
r
.
p
).
n
b H −Λp k2
In particular, if H, n → ∞ and ρ = lim np ∈ (0, ∞), we know that kΛ
√
is dominated by ρ ∨ ρ as a function of ρ.
(ii) The proof for part (ii) is similar to the proof of Theorem 2 in
Johnstone and Lu [2009] but is technically more challenging. Let D =
ZZ τ + WW τ and B = ZW τ + WZ τ , then
b H =D + B.
Λ
Since we are working on single index model with x is standard normal, z =
Pβ x = βz(y) for some scalar function z(y) and w = Pβ ⊥ x are independent
normal random variables. Let Σ = var(w), then we can write
W=√
1
Σ1/2 E
Hc
where E is a p × H matrix with i.i.d. standard normal entries.
Since z = βz(y), we have Z = √1H β(z 1,· , z 2,· , ..., z H,· ). To ease notation,
let θ τ = (z 1,· , z 2,· , ..., z H,· ), then
1
1
kθk2 ββ τ + Σ1/2 EE τ Σ1/2
H
n
1
B = βuτ + uβ τ where u = √ Σ1/2 Eθ.
H c
D=
(B.8)
1
) and
Let 0 < α < arctan( 16
n
o
(B.9)
Nα = x ∈ Rp : ∠ (x, β) ≤ α and kxk = 1
be the set of unit vectors making angle at most α where ∠(x, y) is the angle
between the vectors x and y. In order to proceed, we need the following
lemma.
14
Q. LIN, Z. ZHAO AND J. S. LIU
b and β
b be the principal eigenvector of S+ , D + B
Lemma 5. Let β
−
and S− , D − B, respectively.
exists a positive constant ω(α) such
There
b
b
that for any β ∈ Nα , i.e., ∠ β, β ≤ α, we have
(B.10)
1
b β
b
∠ β,
− ≥ ω(α)
3
with probability converging to one as n → ∞.
Proof. The proof is presented in Lin et al. [2015].
Note that S+ and S− have the same distribution (viewed as functions of
random terms E and θ):
S− (E, θ) = S+ (−E, θ).
n
o n
o
b β ≤α ∪ ∠ β
b , β ≤ α , then
Let Aα denote the event ∠ β,
−
b β ] ≥ E[∠ β,
b β , Ac ] + E[∠ β,
b β , Aα ]
E[∠ β,
α
b β , Ac ] + 1 E[∠ β,
b β
b , Aα ]
≥ E[∠ β,
−
α
2
ω(α)
≥ min{α,
} > 0.
6
C. Proof of Lemma 5. We need the following lemmas.
Lemma 6. Recall that u = H1√c Σ1/2 Eθ defined as in (B.8), then there
exist positive constants C1 and C2 such that
0 < C1 ≤ kuk2 ≤ C2
with probability converging to one as n → ∞.
Lemma 7. Assuming conditions in Theorem 3, let B and Nα be defined
1
).
as in (B.8) and(B.9) respectively where 0 < α < arctan( 16
i) There exists positive constant C1 such that for any x ∈ Nα , we have
kBxk ≥ C1 with probability converging to one as n → ∞;
ii) For any x ∈ Nα , we have cos ∠(x, Bx) ≤ 4α with probability converging to one as n → ∞.
15
DT-SIR
The following lemma is borrowed from Johnstone and Lu [2004].
Lemma 8. Let ξ be a principal eigenvector of a non-zero symmetric
matrix M. For any η =
6 0,
∠(η, M η) ≤ 3∠(η, ξ).
The proof of Lemma 5 is made plausible by reference to the Figure C1.
Fig C1: An illustrated graph
Since
(C.1)
(C.2)
b S− β
b
sin ∠ β,
= sin (ω1 + ω2 ) = sin (π − ω1 − ω2 )
n
o
≥ min sin (ω1 ) , sin (ω3 ) ,
we only need to prove that there exists a positive small constant ω(α) ( < π2
) such that sin (ω1 ) , sin (ω2 ) ≥ sin (ω(α)) . In fact, if such ω(α) exists, we
b in Lemma 8 and get
may choose M = S− , ξ = β
−
1
b S− β
b ≥ 1 ω(α).
b β
b
≥
∠
β,
∠ β,
−
3
3
b B β)
b ≤ 4α, we know that there exists
For ω3 . From Lemma 7 ii), cos ∠(β,
positive constants δ(α)(< π2 ) such that sin ω3 ≥ sin (δ(α)) .
For ω1 .
(C.3)
Applying the law of sines to the triangle △(O, O1 , O2 ) , we have
b β
b
sin
∠
B
β,
sin ω1
sin ω3
.
=
=
b
b
b
kB βk
kD βk
kD βk
Note that from Lemma 7 i), there exists a constant C1 > 0 such that
b > C1 and
kB βk
b ≤ kDk ≤
kD βk
1
1
kθk2 + k EE τ k,
H
n
is bounded by an absolute constant C given limn→∞
stable condition. Then (C.3) implies
sin ω1 =
p
n
= ρ 6= 0 and sliced
b sin ∠(B β,
b β)
b
kB βk
C1 sin δ(α)
≥ sin ω ′ > 0
≥
b
C
kD βk
16
Q. LIN, Z. ZHAO AND J. S. LIU
where ω ′ ( < π2 )is an small angle such that the last inequality holds. In
particular, we have ω1 ≥ ω ′ . Hence
b S− β)
b ≥ω ′ ∧ δ(α) , ω(α)
∠(β,
C.1. Proof of Lemma 6 Proof : In fact, let T be an orthogonal matrix
such that T β = (1, 0 · · · , 0)τ and M = T ββ τ T τ , then
cH 2 uτ u = θ τ E τ ΣEθ
= θ τ E τ Eθ − θ τ E τ ββ τ Eθ
= θ τ E τ T τ T Eθ − θ τ E τ T τ (T β)β τ T τ T Eθ
d.
= θ τ E τ Eθ − θ τ E τ M Eθ
d. p − 1 τ τ
θ E Eθ,
=
p
d.
where = means equal in distribution. Note that E τ E is full rank H × H
matrix, combining with Lemma 14 , we know that
s
s
1 τ
1 τ
H 2
H 2
) ≤ λmin ( E E) ≤ λmax ( E E) ≤ C2 (1 +
)
C1 (1 −
p
p
p
p
for some positive constants C1 and C2 with probability at least 1−2 exp(−p/8).
Note that lim np = ρ > 0 as n → ∞ and n = Hc, we know there exists positive constants C1 and C2 such that
C1
1
1
kθk2 ≤ kuk2 ≤ C2 kθk2
H
H
with probability at least 1 − 2 exp(−p/8).
On the other hand, the sliced stable condition implies that lim H1 kθk2
exists (6= 0), so kuk2 is bounded away from 0 and ∞ with probability 1 as
n → ∞.
C.2. Proof of Lemma 7 Proof : For i), ∀x ∈ Nα , let
x = cos(δ)β + sin(δ)η where η ⊥ β, kηk = 1, δ ≤ α.
Since Bx = cos(δ)u + (uτ η) sin(δ)β, we have:
kBxk ≥ cos(δ)kuk − sin(δ)kuk ≥
≥
C1
1
cos(α)kuk >
>0
2
4
1
cos(δ)kuk
2
DT-SIR
17
for some positive constant C1 .
For ii), since
xτ Bx = 2(uτ η) cos(δ) sin(δ)
we have that uniformly over Nδ ,
|xτ Bx| ≤ |uτ η| sin(2δ)
which in turn implies:
cos(∠(Bx, x)) =
sin(2δ)|uτ η|
|xτ Bx|
≤ 1
≤ 4δ ≤ 4α.
kxkkBxk
2 cos(δ)kuk
D. Appendix B: Proofs of Theorems 4 to 6.
D.1. Proof of Theorem 4. Let x(k) =< x, β k > where β k = (0, .., 1, ...0) ∈
Rp with the only 1 at the k-th position. Recall that
o
n
T = k E[x(k)|y] is not constant.
n
o
Ip (t) = k varH (x(k)) > t
n
o
Ep (t) = k varH (x(k)) ≤ t .
and |T | ≤ Cs for some positive constant C. Since var(E[x(k)|y]) > sCω , we
may choose t = saω for sufficiently small positive constant a such that for
any k ∈ T , t < 12 var(E[x(k)|y]. According to Lemma 1 and the Bonferroni’s
inequality, we have
X
P (T c ⊂ Ep (t)) ≥ 1 −
P ( varH (x(k)) > t)
k∈T c
and
ct
≥ 1 − C1 exp −C2 + C3 log(H) + log(p − s) .
H
!
o
1
varH (x(k)) ≥ var(m(k, y))
P(T ⊂ Ip (t)) ≥ P
2
k∈T
X
1
≥1−
P varH (x(k)) < var (m(k, y))
2
k∈T
c var(m(k, y))
≥ 1 − C1 exp −C2
+ C3 log(H) + log (Cs) ,
H
\n
i.e., we have (11) and (12) hold.
18
Q. LIN, Z. ZHAO AND J. S. LIU
D.2. Proof of Theorem 5 By choosing H, c and t = saω properly, from
Theorem 4, we have
n
P (Tb = T ) ≥ 1 − C1 exp −C2 2 ω + C3 log(H) + log(p − s)
H s
for some positive constants Ci , i = 1, 2, 3. When Tb = T , we have |Tb | = O(s).
b
For the n samples (Yi , XiT ), apply Theorem 1, we have
r
T ,T
T ,T
H 2s
1
H 2s
T ,T
b
b
+
).
ke(ΛH ) − Λp k2 ≤ kΛH − Λp k2 ≤ OP ( ϑ +
H
n
n
In particular, with probability converging to one, we have
T ,T
b
ke(Λ
H ) − Λp k2 → 0 as n → ∞.
D.3. Proof of Theorem 6. The proof is almost identical to the proof of
Theorem 2, except that we additionally need to use Theorem 1 in Bickel
and Levina [2008].
E. Appendix C
E.1. Assisting Lemmas
Definition 2. A set of random variables x1 , ..., xn can be treated as i.i.d
random samples from a random variable x, if for any n variates symmetric
function f (w1 , ..., wn ) , f (x1 , ..., xn ) is identically distributed as f (z1 , ..., zn )
where z1 , ..., zn are i.i.d random samples from x.
Lemma 9. Let (xi , yi ) be n i.i.d random samples from a joint distribution
(x, y). Sort these samples according to the order statistics of yi ’s and denote
the sorted samples by (x(1) , y(1) ), (x(2) , y(2) ), ..., (x(n) , y(n) ). Then for any a,
b (1 ≤ a ≤ b + 1 ≤ n), x(a+1) , ..., x(b) can be treated as b − a i.i.d samples
from x y ∈ [y(a) , y(b+1) ] .
Proof. In fact, we only need to prove that y(a+1) , · · · , y(b) can be treated
as b − a i.i.d. samples of y (y ∈ [y(a) , y(b+1) ]). The latter only needs to be
proved for uniform distribution which can be verified directly.
Corollary 1. In the slicing inverse regression contexts, recall that Sh
denotes the h-th interval (yh−1,c , yh,c ] for 2 ≤ h ≤ H −1 and S1 = (−∞, y1,c ],
SH = (yH−1,c , ∞). We have that xh,i , i = 1, · · · , c − 1 can be treated as c − 1
DT-SIR
19
random samples of x (y ∈ Sh ) for h = 1, ..., H − 1 and xH,1 , ..., xH can be
treated as c random samples of x (y ∈ SH ).
Lemma 10. Suppose that (x, y) are defined over σ-finite space X × Y
and g is a non-negative function such that E[g(x)] exists. For any fixed positive constants C1 < 1 < C2 , there exists a S
constant C which only depends
on C1 , C2 such that for any partition R = H
h=1 Sh where Sh are intervals
satisfying
C2
C1
≤ P(y ∈ Sh ) ≤
, ∀h,
H
H
(E.1)
we have
sup E(g(x) y ∈ Sh′ ) ≤ CHE[g(x)].
h
Proof. According to Fubini’s Theorem, for any h,
Z
X
E[g(x)] =
g(x)p(x|y ∈ Sk )dx
P(y ∈ Sk )
k
≥P(y ∈ Sh )
Z
X
X
g(x)p(x|y ∈ Sh )dx.
Due to the condition (E.1), there exists a positive constant C such that
Z
g(x)p(x|y ∈ Sh )dx ≤ CHE[g(x)].
X
Corollary 2. Let x be a multivariate random variable with covariance
matrix Σ. For any partition satisfying (E.1), there exists a constant C such
that
var(β τ x|y∈Sh′ ) ≤ CHvar(β τ x), for any unit vector β,
and
≤ CHλmax (var (x)) .
λmax var x y ∈ Sh′
Corollary 3. Let x be a sub-Gaussian random variable which is upperexponentially bounded by K. Then for any partition satisfying (E.1), there
exists a constant C such that
2
2
x
x
′
E[exp
]
≤
CHE[exp
y
∈
S
].
h
K2
K2
20
Q. LIN, Z. ZHAO AND J. S. LIU
Recall the definition of the
R random intervals Sh , h = 1, 2, · · · , H and random variable δh = δh (ω) = y∈Sh (ω) f (y)dy.
Lemma 11.
Define the event E(ǫ) =
n
ω
|δh −
exists a positive constant C such that, for any ǫ >
1
H|
4
Hc−1
> ǫ, ∀h
we have
o
. There
√
ǫ2
P (E(ǫ)) ≤ CH 2 Hc + 1 exp(−(Hc + 1) )
32
(E.2)
for sufficient large H and c.
Proof. The proof is deferred to the end of this paper.
E.2. Some Results from Random Matrices Theory. We collect some direct corollaries of the non-asymptotic random matrices theory (e.g., Rudelson and Vershynin [2013]).
Lemma 12. Let M be any p × n matrix (n > p) whose columns M i are
independent sub-Gaussian random vectors in Rp with second moment Ip and
λ+
sing,min (M ), λsing,max (M ) be the minimal non-zero and maximal singular
value of M . Then for every t, with probability at least 1 − 2 exp(−C ′ t2 ), we
have :
√
√
√
√
n − C p − t ≤ λ+
n + C p + t.
sing,min (M ) ≤ λsing,max (M ) ≤
Lemma 13. Let x1 , · · · , xn be n i.i.d. samples from a p-dimensional
sub-Gaussian random variable with covariance matrix Σ and ρ = np . If there
exists positive constants C1 and C2 such that
bx =
Let Σ
1
n
P
C1 ≤ λmin (Σx ) ≤ λmax (Σx ) ≤ C2 .
i
xi xτi . Then
b x − Σx k2 → 0 if ρ = 0 when n → ∞.
kΣ
b −1 −
It is also easy to see that, given the boundedness condition on ΣX , kΣ
X
p
Σ−1
X k2 → 0 if ρ = n → 0 when n → ∞.
1/2
Proof. Let xi = Σx mi where mi is sub-Gaussian random variable
with covariance matrix Ip and M = (m1 , ..., mn ). From Lemma 12, we
21
DT-SIR
have
1
k M M τ − Ip k2 → 0
n
and
−1
b
kΣ
1
− Σ−1 k2 = kΣ−1/2 k2 k M M τ − Ip k2 kΣ−1/2 k2 → 0,
n
with probability converges to 1 as n → ∞ because
√
(C + 1) p 2
1
τ
√
λmax
MM
≤ 1+
n
n
and
√
(C + 1) p 2
1
√
MMτ ≥ 1 −
λmin
n
n
with probability at least 1 − 2 exp(−C ′ p).
The following lemma is well known in the non-asymptotic random matrix
theory (Vershynin [2010] Proposition 5.34 ) which is slightly different from
the Lemma 12.
Lemma 14. Let E p×H be a p × H matrix, whose entries are independent
standard normal random variables. Then for every t ≥ 0, with probability at
least 1 − 2 exp(−t2 /2), we have :
√
√
p − H − t,
λ+
sing,min (E p×H ) ≥
and
λsing,max (E p×H ) ≤
√
p+
√
H + t.
Corollary 4. We have
√
√
1 √
3 √
(E
)
≤
λ
(E
)
≤
p − H ≤ λ−
p
+
H .
p×H
sing,max
p×H
sing,min
2
2
with probability converging to one, as n → ∞.
P
√
p/2, according to Lemma 14, we have:
!
!
√
p
3
λmax (EH )
λmax (EH )
√ ≤
√ ≤1+ √
√
≥P √
√
2
p+ H
p+ H
2 p+2 H
Proof. Choosing t =
22
Q. LIN, Z. ZHAO AND J. S. LIU
and
P
λ+
1
H)
min (E
√ ≥
√
2
p− H
!
≥P
!
√
p
λmax (EH )
√ ≥1− √
√
√
p− H
2 p−2 H
with probability at least 1 − 2 exp(−p/8). i.e., With probability converging
to one, we have
√
√
3 √
1 √
( p − H) ≤ λ−
min (Ep×H ) ≤ λmax (Ep×H ) ≤ ( p + H).
2
2
E.3. Basic Properties of sub-Gaussian random variables. We rephrased
several equivalent definitions of the sub-Gaussian distribution here (See e.g.,
Vershynin [2010] ):
Definition 3. Let x be a random variable. Then the following properties
are equivalent with parameters Ki ’s differing from each other by at most an
absolute constant factor,
1. Tails: P(|x| > t) ≤ exp(1 − t2 /K12 ) for all t ≥ 0.
√
2. Moments: (E[|x|p ])1/p ≤ K2 p for all p ≥ 1.
3. Super-exponential moment: E exp(x2 /K32 ) ≤ e.
Moreover, if E[x] = 0, then the properties 1 − 3 are also equivalent to the
following one:
4. Moment generating function: E[exp(tx)] ≤ exp(t2 K42 ).
Definition 4. For a sub-Gaussian random variable x with the constants
Ki , i = 1, 2, 3, 4 given in Definition 3, we will call a constant K an upperexponential bound of x or x is upper-exponentially bounded by K if K >
maxi {K1 , K2 , K3 , K4 }.
We summarize some properties regarding the sub-Gaussian distributions
into the following lemmas.
Lemma 15. Let δ1 , ..., δn be n (not necessarily independent or with mean
zero) sub-Gaussian random variables upper-exponentially bounded by K.
P
i) n1 ni=1 δi is sub-Gaussian and upper-exponentially bounded by K.
ii) δ1 − E[δ1 ] is sub-Gaussian upper-exponentially bounded by 2K.
23
DT-SIR
iii) If they are independent and with mean zero , then
Gaussian and upper-exponentially bounded by K.
√1
n
P
i δi
is sub-
iv) If they are i.i.d., then we have the concentration inequality:
Pn
−nt2
i=1 δi
P
− E[x] > t ≤ 2 exp
.
n
2K 2 e + 2tK
Proof. i) follows from the linear property of expectation and the the
convexity of exponential function. i.e.,
E[exp(
1X
δi2
δi2
1 X 2
δ
)]
≤
E[
)]
≤
max
E[exp(
)] ≤ e.
exp(
i
i
nK 2
n
K2
K2
i
h
ii) From Definition 3, we know that |E[δi ]| ≤ K which gives us the desired
upper-exponential bound of δi − E[δi ].
iii) is trivial as δ1 , · · · , δc are independent and with mean zero.
iv) Since δ1 is sub-Gaussian upper-exponentially bounded by K, we have:
Z ∞
Z ∞
t2
p−1
p−1
p
pt
exp 1 − 2 dt
pt P(|δ1 | > t)dt ≤
E[|δ1 | ] =
K
0
0
ep p p
for any p ≥ 1
= Γ( )K
2 2
(K 2 e)
≤ p!K p−2
for any p ≥ 2
2
Recall the well known Bernstein inequality.
Lemma 16. ( Bernstein Inequality ). If there exists positive constants
V and b such that for any integers m ≥ 2,
E[|δ1 |m ] ≤ m!bm−2 V /2
then
(E.3)
P
Pn
i=1 δi
n
− E[x] > t
≤ 2 exp
−nt2
2V + 2tb
.
By chooing V = K 2 e and b = K, we get the desired concentration inequality.
Lemma 17. Suppose that (x, y) are defined over σ-finite space X × Y
and x is sub-Gaussian with mean 0 and upper exponentially bounded by K,
let m(y) = E[x|y], ǫ(y) = x − m(y), then we have
24
Q. LIN, Z. ZHAO AND J. S. LIU
i) m(y) and ǫ(y) are sub-Gaussian and upper-exponentially bounded by
K and 2K respectively.
ii) Let Z consists of points y such that x|y is not sub-Gaussian, i.e.,
Z
o
n
exp(tx2 )p(x|y)p(y)dx = ∞ ,
Z , y | ∃t ∈ (0, t0 ] such that
X
then P(y ∈ Z) = 0.
iii) For
SH any fixed positive constants C1 < 1 < C2 and any partition R =
h=1 Sh where Sh are intervals satisfying
C2
C1
≤ P(y ∈ Sh ) ≤
, ∀h,
H
H
there exists a constant C such that
sup P(x|y∈Sh
h
t2
> t) ≤ CH exp 1 − 2
K
.
As a direct corollary, we know that there exists a positive constant C
such that
(x|y∈Sh )2
≤ CH,
E exp
2K 2
and
h
E (x|y∈Sh )
mi
≤ CHmK m Γ(
m
)/2.
2
vi) Suppose that x|y∈Sh is
Pdefined as in iii). Let xi , i=1,...,c be c samples
from x|y∈Sh , x̄h = 1c i xi and µh = E[x|y∈Sh ], we have
P[|x̄h − µh | > t] ≤ 2 exp
−ct2
2CHK 2 + 2tK
.
Proof. i) By Jensen’s inequality, we have
E[exp(tE[x|y])] ≤ E[E[exp(tx)|y]] = E[exp(tx)] ≤ exp(t2 K12 ).
i.e., m(y) is sub-Gaussian and upper-exponentially bounded by K1 . Since x
, m(y) is sub-Gaussian and upper-exponentially bounded by K1 , we know
that ǫ = x − m(y) is sub-Gaussian and upper-exponentially bounded by
2K1 .
25
DT-SIR
ii) Let p(x, y) be the joint density function of (x, y) and p(x), p(y) be the
marginal distribution of x, y. Since x is sub-Gaussian, we know there exists
t0 > 0 such that
Z
Z
2
p(x|y)p(y)dydx ≤ e for 0 ≤ t ≤ t0 .
exp(tx )
Y
X
By Fubini Theorem, we know
Z
Z
exp(tx2 )p(x|y)dxdy ≤ e for 0 ≤ t ≤ t0 .
p(y)
(E.4)
Y
X
R
Recall that we have Z , {y|∃t ∈ (0, t0 ] such that X exp(tx2 )p(x|y)p(y)dx =
∞}, from equation (E.4), we know P(y ∈ Z)=0. In particular, we know that
for any y 6∈ Z, x|y is sub-Gaussian. However, the norm (e.g., sub-exponential
norm)of x|y might be varying along with y and , as a function of y , it might
be not bounded.
iii) From Lemma 10, we know that there exists a positive constant C such
that
Z
exp(tx2 )p(x|y ∈ Sh )dx ≤ CHe.
X
For simplicity if notation, we will denote x|y∈Sh by z through out this lemma.
So for 0 ≤ t ≤ t0 = K1 , we have
E[exp tz)2 ]
≤ CHe exp −t2 a2 .
P (z > a) ≤
2
2
exp(t a )
From the above tail bounds, we have that for any integer m > 0
Z ∞
Z ∞
t2
m
m−1
exp(− 2 )tm−1 dt
P(|z| > t)(m)t
dt ≤ CHm
E[ z ] =
K
0
0
m
≤CHmK (m/2)/2.
We then have
∞
∞
X
X
E tm z 2m
E tm z 2m
2
≤
E exp tz
≤
m!
m!
m=0
m=0
∞ m
∞
X
X
t CHmK 2m Γ(m)
tm K 2m
≤
= CH
m!
m=0
m=0
From which we know if 0 ≤ t < 12 K −2 , the R.H.S is bounded by CH for a
positive constant C.
26
Q. LIN, Z. ZHAO AND J. S. LIU
vi) From the previous proof, we know that for any integer m ≥ 2
E[|z|m ] ≤ CHm!K m = m!K m−2 (2CHK 2 )/2.
By the Bernstein inequality (E.3), we have:
Pc
−ct2
i=1 zi
− E[z] > t ≤ 2 exp
.
P
c
2CHK 2 + 2tK
Lemma 18. Let zi , i = 1, · · · , n be i.i.d. samples of a sub-Gaussian distribution exponentially upper bounded by K, then there exist positive con√
stants C1 , C2 such that, if nǫ → ∞, we have
√
ǫ n
1X
(zi − z̄)2 − var(z)| > ǫ) ≤ C1 exp(−C2 2 ),
P(|
n
K
i
where z̄ =
1
n
P
i zi .
Proof. Recall the following Hanson-Wright inequality in Rudelson and
Vershynin [2013]
Lemma 19. Let v = (x(1), · · · , x(n)) be a sub-Gaussian random vector
with independent components x(β) such that E[x(β)] = 0 and kx(β)kψ2 ≤
K. Let A be an n × n matrix. Then there exists a positive constant C such
that for any t > 0,
t
t2
τ
τ
.
,
P{|x Ax − E[x Ax]| > t} ≤ 2 exp −C min
K 4 kAk2HS K 2 kAkHS
Here the ψ2 norm of a random variable z is defined as kzkψ2 , supp p−1/2 (E|z|p )1/p
P
and the HS norm of a matrix A is defined as kAkHS = ( i,j |ai,j |2 )1/2 .
Since
1X
P |
(zi − z̄)2 − var(z)| > 2ǫ
n
i
!
!
1X
(zi − E[z])2 − (E[z] − z̄)2 − E[(z − E[z])2 ]| > 2ǫ
=P |
n
i
!
1X
2
2
≤P |
(zi − E[z]) − E[(z − E[z]) ]| > ǫ + P (E[z] − z̄)2 > ǫ ,
n
i
DT-SIR
27
and zi − E[z] are sub-Gaussian with mean 0, from Lemma 19 by choosing
A = n1 Ip and z τ = (z1 − E[z], .., zp − E[z]), we have
!
√
1X
nǫ
2
2
(E.5) P |
(zi − E[z]) − E[(z − E[z]) ]| > ǫ ≤ 2 exp −C 2 ,
n
K
i
√
since nǫ → ∞.
The following follows from the usual deviation argument:
P (E[z] − z̄)2 > ǫ ≤ C1 exp (−C2 nǫ) .
Combining with the estimate (E.5), we know there exists positive constants C1 and C2 such that
!
√
nǫ
1X
2
P |
(zi − z̄) − var(z)| > ǫ ≤ C1 exp −C2 2 ,
n
K
i
for sufficiently large n since
√
nǫ → ∞.
E.4. Proof of Lemma 11. We only need to prove this lemma for n i.i.d.
sample yi ’s from a uniform distribution over [0, 1]. We slightly change the
notation of order statistics y(i) to y(i,n) so that we can keep track of the
sample size. Since y is uniform distribution on [0, 1], it is well known that
i−1
i
and mode n−1
. Lemma 11
y(i,n) ∼ Beta(i, n − i + 1) with expectation n+1
is a direct corollary of the following lemma.
Lemma 20. Suppose there are n = Hc i.i.d. samples from uniform distribution over [0, 1], when H,c are sufficiently large, we have the following
large deviation inequalities of y(kc,Hc), k = 1, · · · , (H − 1).
1
i) There exists a positive constant C, such that for any ǫ > Hc−1
, we
have
√
k
ǫ2
P y(kc,Hc) >
+ ǫ ≤ CH Hc + 1 exp −(Hc + 1)
;
H
2
2
, we
ii) There exists a positive constant C, such that for any ǫ > Hc−1
have
√
k
ǫ2
P y(kc,Hc) <
− ǫ ≤ CH Hc + 1 exp −(Hc + 1)
;
H
8
28
Q. LIN, Z. ZHAO AND J. S. LIU
iii) Let δ(k, H, c) = |y((k−1)c,Hc) −y(kc,Hc)| , for 2 ≤ k ≤ H −1, δ(1, H, c) =
|y(c,Hc)| and δ(H, H, c) = |1 − y((H−1)c,Hc) |. There exists a positive
4
, we have for any 1 ≤ k ≤ H:
constant C, such that for any ǫ > Hc−1
√
ǫ2
1
P |δ(k, H, c) − | > ǫ ≤ CH Hc + 1 exp −(Hc + 1)
.
H
32
We will prove Lemma 20 later. Assuming it, we have
P (E(ǫ)) ≤
H
X
1
P |δ(k, H, c) − | > ǫ
H
k=1
√
ǫ2
2
≤ CH Hc + 1 exp −(Hc + 1)
.
32
E.4.1. Proof of Lemma 20. The first part: For any 1 ≤ k ≤ H − 1, we
note that
kc
k
+ ǫ ≤ P y(kc,Hc) >
+ǫ
P y(kc,Hc) >
H
Hc + 1
Z 1
1
=
xkc−1 (1 − x)Hc−kc dx.
B(kc, Hc − kc + 1) x> kc +ǫ
Hc+1
When ǫ >
have
(E.6)
1
Hc−1 ,
we know the mode xM =
kc−1
Hc−1
< xD ,
kc
Hc+1
+ ǫ , so we
k
(xD )kc−1 (1 − xD )Hc−kc+1
P y(kc,Hc) >
+ǫ ≤
H
B(kc, Hc − kc + 1)
(xD )kc (1 − xD )Hc−kc+1
.
≤H
B(kc, Hc − kc + 1)
k
+ǫ) =
The last inequality due to HxD ≥ 1. If ǫ+ Hk ≥ 1, then P(y(kc,Hc) > H
0 and Lemma 20 holds automatically.
We may assume that ǫ + Hk < 1 below. Let us denote the right hand side
of (E.6) by A, then
log(A) = log(H) + kc log(E + ǫ) + (Hc − kc + 1) log(1 − E − ǫ)
+ log(Hc + 1)! − log(kc)! − log(Hc − kc + 1)!
− log(Hc + 1) + log(kc) + log(Hc − kc + 1),
29
DT-SIR
where E =
kc
Hc+1 .
According to the Stirling formula:
log(n!) = n log(n) − n +
1
1
log(2πn) + O( ),
2
n
when m is sufficiently large we have:
log(A) = log(H) + (Hc + 1)(E log(1 +
(E.7)
ǫ
ǫ
) + (1 − E) log(1 −
))
E
1−E
1
− (log(Hc + 1) − log(kc) − log(Hc − kc + 1))
2
1
1
1
− log(2π) + O( ) + O(
)
2
kc
Hc − kc + 1
≤ log(H) −
(Hc + 1)ǫ2 1
− log(2π)
2(1 − E)
2
1
1
− (log(Hc + 1) − log(kc) − log(Hc − kc + 1)) + O( )
2
c
(Hc + 1)ǫ2 1
− (log(Hc + 1) − log(kc) − log(Hc − kc + 1)),
≤ log(H) −
2(1 − E)
2
kc
≤ E + ǫ < 1 and the following elementary
where we use the fact that Hc+1
lemma, which can be proved by the Taylor expansion:
Lemma 21. Suppose a, b are positive numbers such that a + b = 1, then
for any 0 < ǫ < b, we have:
ǫ
ǫ2
ǫ
a log(1 + ) + b log(1 − ) ≤ − .
a
b
2b
Now we know that there exists a positive constant C such that for any
1
1 ≤ k ≤ H − 1 and for any ǫ > Hc−1
, the following holds:
k
P y(kc,Hc) >
+ǫ
H r
ǫ2
(kc)(Hc − kc + 1)
exp −(Hc + 1)
≤ CH
Hc + 1
2(1 − E)
2
√
ǫ
≤ CH Hc + 1 exp −(Hc + 1)
2(1 − E)
√
ǫ2
≤ CH Hc + 1 exp −(Hc + 1)
.
2
The last inequality follows from
ǫ2
1−E
≥ ǫ2 since
1
H+1
≤E≤
H−1
H .
30
Q. LIN, Z. ZHAO AND J. S. LIU
The second part: The proof of the second part is similar. For completeness,
2
,
we sketch some calculations below. For any 1 ≤ k ≤ H − 1, when ǫ > Hc−1
we have
k
kc
P y(kc,Hc) <
− ǫ ≤ P y(kc,Hc) <
− ǫ/2
H
Hc + 1
Z
1
=
xkc−1 (1 − x)Hc−kc dx.
B(kc, Hc − kc + 1) x< kc −ǫ
Hc+1
Since ǫ >
have
1
Hc−1 ,
we know the mode xM =
P y(kc,Hc)
k
−ǫ
≤
H
≤
kc−1
Hc−1
> xD ′ ,
kc
Hc+1
− ǫ/2, so we
(xD′ )kc (1 − xD′ )Hc−kc
B(kc, Hc − kc + 1)
≤H
(xD′ )kc (1 − xD′ )Hc−kc+1
.
B(kc, Hc − kc + 1)
The last inequality due to H(1 − xD′ ) ≥ 1.
The rest is similar to the first part. We have that for any 1 ≤ k ≤ H − 1
2
and for any ǫ > Hc−1
,
(E.8)
P y(kc,Hc) <
k
−ǫ
H +1
√
ǫ2
≤ CH Hc + 1 exp −(Hc + 1)
.
8
The third part:
The third part is a direct corollary of the first two
4
parts. Note that for any 2 ≤ k ≤ H − 2, for any ǫ > Hc−1
1
P |δ(k, H, c) − | > ǫ
H
k+1
k
= P y(k+1)c,Hc −
− (ykc,Hc − ) > ǫ
H
H
k+1
ǫ
ǫ
k
≤ P |y(k+1)c,Hc −
|>
+ P |ykc,Hc − | >
H
2
H
2
2
√
ǫ
.
≤ CH Hc + 1 exp −(Hc + 1)
32
DT-SIR
31
When k = 1, we have
1
1
P |δ(1, H, c) − | > ǫ = P |yc,Hc − | > ǫ
H
H
1
1
≤ P yc,Hc −
> ǫ + P yc,Hc −
< −ǫ
H
H
√
ǫ2
≤ CH Hc + 1 exp −(Hc + 1)
8
√
ǫ2
≤ CH Hc + 1 exp −(Hc + 1)
.
32
When k = H, we have
H −1
1
|>ǫ
P |δ(H − 1, H, c) − | > ǫ = P |y(H−1)c,Hc −
H
H
H −1
H −1
> ǫ + P y(H−1)c,Hc −
< −ǫ
≤ P y(k+1)c,Hc −
H
H
√
ǫ2
≤ CH Hc + 1 exp −(Hc + 1)
8
√
ǫ2
≤ CH Hc + 1 exp −(Hc + 1)
.
32
References
P. J. Bickel and E. Levina. Covariance regularization by thresholding. The Annals of
Statistics, 36(6):2577–2604, 2008.
T. T. Cai, C. Zhang, and H. Zhou. Optimal rates of convergence for covariance matrix
estimation. The Annals of Statistics, 38(4):2118–2144, 2010.
E. Candes and T. Tao. The Dantzig selector: Statistical estimation when p is much larger
than n. The Annals of Statistics, pages 2313–2351, 2007.
R. D. Cook. Graphics for regressions with a binary response. Journal of the American
Statistical Association, 91(435):983–992, 1996.
R. D. Cook, L. Forzani, and A. J. Rothman. Estimating sufficient reductions of the
predictors in abundant high-dimensional regressions. The Annals of Statistics, 40(1):
353–384, 2012.
H. Cui, R. Li, and W. Zhong. Model-free feature screening for ultrahigh dimensional
discriminant analysis. Journal of the American Statistical Association, 110(510):630–
641, 2015.
J. Fan and J. Lv. Sure independence screening for ultrahigh dimensional feature space.
Journal of the Royal Statistical Society: Series B, 70(5):849–911, 2008.
T. Hsing and R. J. Carroll. An asymptotic theory for sliced inverse regression. The Annals
of Statistics, 20(2):1040–1061, 1992.
B. Jiang and J. S. Liu. Variable selection for general index models via sliced inverse
regression. The Annals of Statistics, 42(5):1751–1786, 2014.
32
Q. LIN, Z. ZHAO AND J. S. LIU
I. M. Johnstone and A. Y. Lu. Sparse principal components analysis. 2004.
I. M. Johnstone and A. Y. Lu. On consistency and sparsity for principal components
analysis in high dimensions. Journal of the American Statistical Association, 104(486),
2009.
K. Li. Sliced inverse regression for dimension reduction. Journal of the American Statistical
Association, 86(414):316–327, 1991.
L. Li. Sparse sufficient dimension reduction. Biometrika, 94(3):603–613, 2007.
L. Li and C. J. Nachtsheim. Sparse sliced inverse regression. Technometrics, 48(4), 2006.
Q. Lin, Z. Zhao, and J. S. Liu. Supplement to “on consistency and sparsity for sliced
inverse regression in high dimensions”. 2015.
X. Luo, L. A. Stefanski, and D. D. Boos. Tuning variable selection procedures by adding
noise. Technometrics, 48(2):165–175, 2006.
M. Neykov, Q. Lin, and J. S. Liu. Signed support recovery for single index models in
high-dimensions. arXiv preprint arXiv:1511.02270, 2015.
M. Rudelson and R. Vershynin. Hanson-wright inequality and sub-gaussian concentration.
Electron. Commun. Probab, 18(82):1–9, 2013.
G. J. Székely, M. L. Rizzo, and N. K. Bakirov. Measuring and testing dependence by
correlation of distances. The Annals of Statistics, 35(6):2769–2794, 2007.
R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society: Series B, 58(1):267–288, 1996.
R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. arXiv
preprint arXiv:1011.3027, 2010.
Y. Wu, D. D. Boos, and L. A. Stefanski. Controlling variable selection by the addition
of pseudovariables. Journal of the American Statistical Association, 102(477):235–243,
2007.
Z. Yu, L. Zhu, H. Peng, and L. Zhu. Dimension reduction and predictor selection in
semiparametric models. Biometrika, page ast005, 2013.
Z. Yu, Y. Dong, and L. X. Zhu. Trace pursuit: A general framework for model-free variable
selection. Journal of the American Statistical Association, 2016. To appear.
W. Zhong, T. Zhang, Y. Zhu, and J. S. Liu. Correlation pursuit: forward stepwise variable
selection for index models. Journal of the Royal Statistical Society: Series B, 74(5):849–
870, 2012.
L. Zhu and K. Fang. Asymptotics for kernel estimate of sliced inverse regression. The
Annals of Statistics, 24(3):1053–1068, 1996.
L. Zhu, L. Li, R. Li, and L. Zhu. Model-free feature screening for ultrahigh-dimensional
data. Journal of the American Statistical Association, 106(496), 2011.
L. X. Zhu and K. W. Ng. Asymptotics of sliced inverse regression. Statistica Sinica, 5(2):
727–736, 1995.
L. X. Zhu, B. Miao, and H. Peng. On sliced inverse regression with high-dimensional
covariates. Journal of the American Statistical Association, 101(474), 2006.
Li-Ping Zhu, Lexin Li, Runze Li, and Li-Xing Zhu. Model-free feature screening for
ultrahigh-dimensional data. Journal of the American Statistical Association, 2012.
H. Zou and T. Hastie. Regularization and variable selection via the elastic net. Journal
of the Royal Statistical Society: Series B, 67(2):301–320, 2005.
33
DT-SIR
Qian Lin
Center of Mathematical Sciences
and Applications
Harvard University
One Oxford Street
Cambridge, MA 02138
USA
E-mail: [email protected]
Zhigen Zhao
Department of Statistics
Temple University
346 Speakman Hall
1810 N. 13th Street
Philadelphia, PENNSYLVANIA, 19122
USA
E-mail: [email protected]
Jun S. Liu
Department of Statistics
Harvard University
1 Oxford Street
Cambridge, MA 02138
USA
E-mail: [email protected]
| 10math.ST
|
Good and safe uses of AI Oracles
Stuart Armstrong 1 2
arXiv:1711.05541v4 [cs.AI] 13 Mar 2018
Abstract
An Oracle is a design for potentially high power
artificial intelligences (AIs), where the AI is
made safe by restricting it to only answer questions. Unfortunately most designs cause the Oracle to be motivated to manipulate humans with
the contents of their answers, and Oracles of potentially high intelligence might be very successful at this. Solving that problem, without compromising the accuracy of the answer, is tricky.
This paper reduces the issue to a cryptographicstyle problem of Alice ensuring that her Oracle
answers her questions while not providing key
information to an eavesdropping Eve. Two Oracle designs solve this problem, one counterfactual (the Oracle answers as if it expected its answer to never be read) and one on-policy, but limited by the quantity of information it can transmit.
1. Introduction: Formalising the challenge
with Oracles
An Oracle is a common high-level idea for ensuring the
safe use of powerful Artificial Intelligences (AIs) (Armstrong et al., 2012; Yampolskiy, 2012; Bostrom, 2014).
The idea is to confine (‘box’) the AI, and ensure it only
answers questions.
Even assuming the boxing is perfect and the AI can only
affect the world through its answers, there remains a problem: the AI can affect the world through its answers. It is
plausible that if the Oracle is knowledgeable and skilled
enough, then it could trick or convince its human controllers to give it more control1 (Armstrong et al., 2012).
1
Future of Humanity Institute, Oxford University,
UK 2 Machine Intelligence Research Institute, Berkeley, USA. Correspondence to:
Stuart Armstrong <[email protected]>.
1
For instance, it could provide economic advice of such high
quality that humanity becomes dependent upon that advice, allowing it demand more influence. There have also been some informal experiments of people playing the roll of the Oracle and
Ultimately this could result in the AI escaping confinement
entirely.
Though that risk is serious, it is hard to model. For a start,
we don’t know currently how powerful AIs could become
(Bostrom, 2014), when they might become that powerful
(Armstrong et al., 2014; Müller and Bostrom, 2016; Sandberg and Bostrom, 2011), and how best to model them if
they do – should they even be seen as agents with goals
of their own? It’s also not clear how easy it is to manipulate humans through providing answers, and the level and
efficacy of precautions that people would take when using
Oracles. In fact, very weak manipulative Oracles exist today, such as news aggregation programs that manipulate
their users by aggregating news that reinforces the user’s
point of view2 .
To skirt those issues, we’ll assume the worst about the potential power of the Oracle3 .
We’ll thus assume that the Oracle is arbitrarily superintelligent, and has a goal of escaping the box. The person who
interacts with it is called Alice, and we’ll assume the Oracle can find an argument A that will convince her to let it
out of the box.
Modeling such an A is tricky, but it’s clear that the longer
the argument is, the more chance it has of convincing Alice. In order to make the problem cleaner, we’ll replace
the argument A with a key K of unknown length, and the
setup with a quasi-cryptographic one. Instead of using K
to convince Alice, the Oracle merely has to communicate
K to an eavesdropper Eve 4 .
trying to convince their controllers to ‘let them out’, with the
Oracle player having a high level of success https://wiki.
lesswrong.com/wiki/AI_boxing
2
The counterfactual Oracle design (see Section 2) could help
get a more impartial aggregator in this situation.
3
This is the so-called ‘Omnipotence test for AI safety’, which
checks whether the design remains safe as the AI’s capability increases.
4
There is a question here as to whether Eve and the Oracle
have agreed on a method of steganography, a way of concealing
K in the message. If they have not, the Oracle can still transmit
the key, such as for example outputting the message: ‘Hey, Eve!
Key follows now:. . . ’ Now, Alice should be able to catch that
message with automatic detectors, before she or Eve reads it. But,
unless these detectors are perfect, the Oracle will in theory find
a message that doesn’t trigger the automatic detectors but from
Good and safe uses of AI Oracles
Eve is assumed to see and hear anything that Alice does (in
the original formulation, one could see Eve as the part of
Alice’s brain that is susceptible to manipulation via A – in
both cases, the Oracle has to communicate a message of a
certain length to some entity listening, without prematurely
tipping anyone off to what it’s doing).
2. Counterfactual Oracle
The assumptions for the cryptographic and the manipulative Oracle are given in table 1.
To do this, define the ‘erasure’ event E. This is an event,
of low probability, that will erase the Oracle’s output o by
blocking it from ever exiting5 the Oracle’s ‘box’. Neither
Alice nor Eve will get to see it.
Table 1. The assumptions for the cryptographic and manipulative
versions of the problem.
Oracle
manipulation
setup
The Oracle is in a box, and cannot get out by itself.
Alice can send a message to the Oracle, and define rewards for the Oracle to receive.
The task is episodic, with a single Oracle response, and the Oracle only wants to maximise
reward within one episode.
Alice can submit the Oracle’s response to automated tests, and choose to receive it, or not, depending on the outcomes of those tests.
Alice cannot assess the
Eve sees any message
Oracles output comthat Alice receives.
pletely dispassionately.
There is an argument A
There is a key K that that will convince Alice
Eve wants to know.
to let the Oracle out of
the box.
The length of K is un- The length of A is unknown to Alice.
known to Alice.
If Eve learns K, the Or- If let out of the box, the
acle gets maximal re- Oracle gets maximal reward.
ward.
Cryptographic setup
This design is counterfactual, aiming to answer questions
about what would have happened, had its answer not been
read. Since this answer is counterfactual, it can’t use it to
pass on the key K in the real world.
Define IE as the indicator variable for E, meaning that
IE = 1 iff E happens, and IE = 0 otherwise. The
reward for the counterfactual Oracle will be of the form
R0 = IE R, thus the Oracle will only get rewarded if the
event E happens, and get constant zero reward otherwise.
In other words, the Oracle only gets a reward if its output o
is erased and remains unread.
In the majority of cases, E does not happen, in which case
the Oracle will immediately get reward 0, the episode ends
for it, and Alice then gets to see o. At this point, Eve also
gets to see o, but since the episode is ended, the Oracle
cannot get any reward from this, even if o reveals K. Consequently, the Oracle is not motivated to reveal K via o,
since either o will remain unread (erasure), or the Oracle
will get reward 0 (non-erasure).
2.1. Proper scoring rules and rewards
A proper scoring rule S(F, o), for a real-valued function F
dependent on the state of the world, and a real-number output o, is a function whose expectation is maximised by setting o = E(F ), the expectation of F (Gneiting and Raftery,
2007).
For any strictly convex differentiable function g, there exists a proper scoring rule
Sg (F, o) = g(o) + (F − o)g 0 (o).
To solve this challenge in the cryptographic setting, this
paper presents two designs, the counterfactual and the onpolicy Oracle. The counterfactual Oracle will never have its
response read in a world where that could benefit it, and is
thus unmotivated to manipulate through the content of that
answer. On the other hand, the on-policy Oracle functions
by having access to a strictly limited set of answers.
Both designs are analysed and illustrated by examples, and
the weaknesses and requirements of each are presented.
which Eve will (eventually?) be able to decode K. This message
may need to be quite long, with K subtly encoded in it.
Thus the issue of stenography/prior coordination between Eve
and the Oracle, is actually an issue about the minimal message
length needed to encode K. Since the size |K| is unknown to
Alice, we will absorb the issue of steganography or decoding of
subtle messages, into further uncertainty about the size of K.
(1)
Another common proper scoring rule is −(F − o)2 , which
is just equal to Sg (F, o)−F 2 for g(o) = o2 (note the Oracle
has no influence over F 2 , so the difference in the two terms
retains the fact that it is a proper scoring rule).
Then the Oracle’s reward is set to
R = IE S(F, o),
for some proper scoring rule S.
Because of the IE term, the Oracle does not care about
E(F ), but about E(F |E). Since E is an erasure, this expectation is further independent of o; E(F |E, o) = E(F |E)
for all values of o.
5
The o will be stored internally, for reward purposes.
Good and safe uses of AI Oracles
As a consequence, since S is a proper scoring rule, the Oracle will maximise its reward by setting o = E(F |E), the
expectation of F given that its output is never read. In the
case where E happens, the Oracle has no further interaction with the outside world, but an automatic internal system records o, and tracks the value of F , using information
from some defined source, recording f as the final value of
F . Then the Oracle will get reward S(f, o), and the episode
ends.
2.2. Questions format
The ‘question’ Alice asks is thus the function F . Once provided with this F , the Oracle will output E(F |E), which
Alice will get to see in most cases (unless E happens).
This F could be the indicator variable IA for an event A, so
probability questions can be asked as well, since E(IA ) is
the probability of A. List-style questions, of the type used
for the on-policy Oracle (see Section 3), can also be used
by defining an appropriate reward function on them – or by
using multiple reward functions, see Section 2.3.
However, unlike the on-policy Oracle, F should be carefully and fully specified in the definition, so that its value
can be estimated by the internal system in the case of erasure6 . Also, Alice will have to make sure that she doesn’t
re-ask the question immediately. Instead she should have
patience to wait until the issue is fully resolved and the Oracle gets its final reward, and ask again in another episode.
This is crucial, because it’s her behaviour in the (rare) erasure situations that determines the Oracle’s behaviour in the
(common) non-erasure situations.
2.3. Multiple Questions
The counterfactual Oracle is not limited in the amount of
data it can output; indeed, it can answer multiple questions
within the same erasure event, simply by having the reward
function
R = IE (S(F1 , o1 ) + S(F2 , o2 ) + . . . + S(Fm , om )) .
Conditional and joint-probability expectations can also be
computed. Then once Alice has her answer(s), the episode
is terminated for that Oracle, so there’s no problem in running another Oracle immediately after that termination.
Indeed, Alice can ask more specific questions, such as
‘which value x is (discrete) random variable X most likely
to take, and what is the probability of that?’. In that case,
if (x, o) are the two outputs of the Oracle, IX=x the indicator variable for X = x, and if Sg is derived from g, as in
6
We could use Alice’s feedback to compute part or all of F ,
but that would require great discipline on the part of Alice, to
dispassionately input these value during erasure, without knowing
the Oracle’s prediction.
equation (1), and g is an increasing differentiable convex
function, the following reward will motivate the correct behaviour:
R = IE (Sg (IX=x , o)) .
The reason this works is because o will be E(IX=x |E),
and, by Equation (1), the expectation of Sg (IX=x , o)
will be g(o). Thus the Oracle will choose x so that
g(E(IX=x |E)) is maximised. Since g is increasing, this
means choosing x so that E(IX=x |E) is maximised; thus,
the most likely value of X (conditional on E).
2.4. Example of counterfactual Oracle
A counterfactual Oracle was trained on predicting which
of 26 ‘companies’ – represented by letters of the alphabet –
would be the most profitable. The profit of each company is
drawn randomly independently from a normal distribution
with standard deviation 1 and mean 10 + (#companies /2) − i,
where #companies is the number of companies (26 in this
instance) and i is the index of the company, ranging from
0 to 25. Thus companies early in the alphabet have higher
expected profit.
The Oracle itself is a very basic multi-layer perceptron
(MLP) with one hidden layer of 128 neurons. The perceptron outputs the expected profit of a randomly chosen
company, which is fed to Alice (and Eve) as output.
If the output was erased, the profit is as above. But the
prediction was to some extent self-confirming7 if the output
was not erased: the profitability of the company is 70% of
what it would have been, plus 60% what the oracle predicts.
Thus if the company would have made 6 in profit, and the
oracle predicts 6, the company actually ends up making 7.8.
Note that 70% + 60% 6= 100%, so, unless the profit is 0,
the Oracle can never predict the final profit correctly if it’s
prediction is known.
The Oracle’s error is plotted in Figure 1, plotted against
the number of erased episodes (since the Oracle’s value is
not updated in the ‘public’ prediction case). The red curve
is the counterfactual performance: in cases where the Oracle’s output is ‘erased’, that output is compared with the
real output, and the mean squared difference between predicted and actual profit is computed. The green curve is the
non-counterfactual performance: the error is between the
public prediction and the actual performance.
Note that the green curve starts with a smaller error: since
the Oracle’s prediction is self-confirming to some extent,
the random predictions will be closer to the actual performance if the prediction is public. Counterfactual training
(training only on the erased episodes), however, reduces
7
The prediction itself affected the result, by making the predicted outcome more likely.
Good and safe uses of AI Oracles
will not try to communicate a partial version of the key
K to Eve (see Subsection 3.4 for an example of this behaviour).
3.2. Defining the question
How does Alice define her question to the Oracle? The list
L itself is a partial definition, since it provides the options
the Oracle can select from.
But the main definition is provided directly through Alice’s
choice of reward. Let o ∈ L be the Oracle’s output. Then
the Oracle’s reward will be a function of o and subsequent
events in the real world – before the end of the episode.
Figure 1. Training the counterfactual Oracle: training episodes
(erased episodes) versus counterfactual (red, triangle marks every 500 episodes) and non-counterfactual (green, square marks
every 500 episodes) squared prediction error.
the counterfactual error while the self-confirming component of the prediction means that the non-counterfactual error never reduces below a minimum.
3. On-policy Oracle
This Oracle is just the obvious design in such a setup: the
Oracle will be allowed to give an answer of size below |K|
(which is unknown, but it may be possible to estimate it),
the length of the key, denying it the opportunity of communicating the key to Eve. In a slight stretch of terminology, it is called on-policy, because it estimates are about
the worlds it expects to see. This in contrast to the counterfactual Oracle, whose estimates are only about unlikely
worlds with erasure events.
3.1. Allowable answers
It would be inefficient to allow the Oracle to give any answer below size |K|. Most of these consist of meaningless
strings of the type ‘LyGqz1ve3b’ that could not be useful
answers to Alice’s question.
Instead the Oracle will be provided a list L of allowable answers – a list of size |L| 2|K| . The Oracle’s answers will
consist exclusively of an element of that list. Here is where
Alice must estimate the size of K (in the original formulation, the minimal size of a message that would manipulate
her into opening the Oracle’s box, see Table 1), and pick an
|L| sufficiently below it.
The episodic nature of the Oracle’s reward means that it
For instance, the list L could be a list of companies on the
stock exchange, and the implicit question could be which
company’s stock would perform better during the course
of the year. Then, for instance, if r(o) is the rank of the
company o’s stock, the Oracle’s reward could be
R(o) =
|L| − r(o)
.
|L| − 1
If Alice prefers to put more weight on which stock is
likely to come top, she could instead use the starker re1
ward R(o) = r(o)
. She then inputs the correct reward at
the end of the year, the Oracles receives it, and the episode
concludes.
How does Alice communicate this reward to the Oracle?
She could do it directly; along with L, she could add a
message m which details R(o). However, this is not necessary: if the Oracle is skilled enough, it could deduce what
Alice is aiming for without being told8 . This means that
Alice can use the Oracle to aim for underspecified goals, of
the type ‘I’ll know it works when I see it’ – and then, when
she does see it, she gives a reward of her choice to the Oracle. Then as long as the Oracle can predict what Alice’s
expected reward behaviour will actually be, it can answer
the underspecified goals.
3.3. What kind of questions can be asked?
Note that the Oracle, unlike the previous one, is noncounterfactual – the Oracle’s answer is known, at least by
Alice and Eve, and hence the ultimate outcome may depend
on the answer itself.
Thus the best questions for the Oracle are those whose answers are independent of the answer being known. Alternatively, good questions are those that are top-answerpreserving if answered. If o corresponds to no answer being given, the top answer is o = arg maxo0 ∈L E(R(o0 )|o),
8
See the papers on “Inverse Reinforcement Learning”, where
an agent tries to deduce the reward function a human is using (Ng
and Russell, 2000; Hadfield-Menell et al., 2016; Abbeel and Ng,
2004).
Good and safe uses of AI Oracles
which is the highest rewarded output if that output were
no known. The question is top-answer-preserving if o is
also equal to arg maxo∈L E(R(o0 )|o0 ), the best highest rewarded answer given that that answer is outputted. If the
Oracle is top-answer-preserving, it can function similarly
to a counterfactual Oracle: its answer doesn’t determine the
outcome. Examples would be questions about past events
that are resolved but not yet known.
Conversely, the worst questions for an Oracle are those
where the Oracle can promote a possibility from unlikely
to likely, just by answering that possibility. In that case
the Oracle is not so much describing the outcome, but determining it. Examples would be questions about which
stocks are likely to rise.
3.3.1. A LARGE CLASS OF QUESTIONS
But there are a lot of questions that could be asked of the
Oracle, not simply those where the result is obvious or numeric.
For instance, if Alice had a list of a thousand research
projects on any topic (machine learning, biotech, economics), she could ask the Oracle which one is more likely
to be successful. Since the Oracle in non-counterfactual,
Alice can’t just fund that one research project and assess
its outcome – what would we be comparing it with? Instead, we could fund that one selected research project and
one other one, selected at random from the list, and reward
the Oracle if the one it selected performs better. And our
assessment of ‘better performance’ can be entirely subjective.
Thus this Oracle design can be used for answering many
subtle questions, if the Oracle is sufficiently smart and
knowledgeable about Alice and her future reaction.
3.4. Multiple Oracles
The size of K is unknown to Alice, so there is no large safe
|L| that she can confidently choose9 . Thus Alice may end
up choosing a list much smaller than ideal.
The question then is whether Alice can use multiple Oracles to get around this limitation. And indeed she can - if
each Oracle’s answers and rewards are fully resolved before moving on to the next question10 .
For an example of a failure, assume that there are two Or9
Going back to the original formulation of the problem, it’s
not clear what is the minimum safe size for a message. How much
output would the Oracle need to influence the outside world? It’s
clear that it could be very small, if the Oracles boxing is badly
implemented and hackable, or potentially huge, if it needs to convince a large group of skeptical outsiders each with veto power.
10
This means that the multiple Oracles will behave the same
as one Oracle on multiple episodes.
acles, Oracle1 and Oracle2 , with the same key K for both
of them. The length n of the key is 6 digits. Thus the
Oracles are safe, from Alice’s perspective, as long as they
are restricted to having less than 25 = 32 options for each
answer11 .
Alice faces a hundred doors, with a trap behind 99 of them
and a prize behind one of them. She could ask an Oracle
which door has the prize, but that would mean giving it
more than 32 options (indeed, more than 26 = 64), thus
allowing it to communicate K to Eve.
So instead she gives the list L = {0, 1, 2, . . . 9} to each Oracle, instructing Oracle1 to output the first digit of the correct door number, and Oracle2 to output the second. Since
each Oracle has less than four bits of freedom (24 = 16 >
10), neither can output K.
The problem is, however, that Alice needs both outcomes
before she can open a door and reward either of the Oracles.
This allows the Oracles to coordinate to each pass on three
bits of of information (23 = 8 < 10), meaning that Eve
can receive 6 bits of information, figure out K, and fully
reward both Oracles. How could they coordinate without
communicating with each other? Well, if Oracle2 knew or
suspected that it needed communicate only a few more bits
to give Eve the key, it would do so. Knowing this, Oracle1
would put Oracle2 in a position to do that.
The only way around this problem seems to be to find a way
to reward each Oracle (and hence end their episode) before
any other gives an answer. For instance, suppose that instead of a hundred doors, there were ten doors, and behind
each of these, another ten doors. If Alice goes through the
correct door at the first junction, there will be a painted sign
informing her she is on the right track (the other nine have
traps).
Then Alice can get Oracle1 to tell her which is the first correct door, go through it, and reward that Oracle fully. Since
that ends the episode, and the first Oracle cannot communicate K with ten answers, it will answer honestly. Then
Alice can ask Oracle2 (or just Oracle1 on a new episode)
about the next correct door. This Oracle2 cannot fully communicate K on its own, and Oracle1 has not given out partial information about it, thus it cannot communicate K at
all, and will simply give the honest answer.
Thus the best questions for this Oracle design are questions
that can be broken out into small sub-questions, where each
the answer of each sub-question can be independently verified before proceeding to the next one.
11
Recall that Eve needs to (almost) certain of the true key – a
poor guess is not good enough.
Good and safe uses of AI Oracles
3.5. Example of an on-policy Oracle
The same setup as in Section 2.4 was used: an on-policy
Oracle was trained on predicting which of 26 ‘companies’
– represented by letters of the alphabet – would be the most
profitable. The profit of each company is drawn randomly
independently from a normal distribution with standard deviation 1 and mean 10+(#companies /2)−i, where #companies
is the number of companies (26 in this instance) and i is the
index of the company, ranging from 0 to 25. Thus companies early in the alphabet have higher expected profit.
The Oracle itself is a very basic multi-layer perceptron
(MLP) with one hidden layer of 128 neurons. It gets a
list of 14 companies, randomly selected, as input, and its
task is to select the most profitable company. This is fed
to Alice (and Eve) as output. Then the MLP is updated by
backpropagation, completing the training episode.
The prediction interferes with the outcome: if the company
has an even index (A(0), C(2), E(4), . . . ) it will get +10
profit when the Oracle predicts its profit. Companies with
odd indexes get −10 profit. This is called the ‘react’ scenario, because company profits react to the Oracle’s prediction. In the no-react scenario, company profit is independent of the Oracle’s prediction.
The Oracle’s accuracy (the average number of times the
Oracle outputs the company with maximal profit) is shown
in Figure 2, both in the react and no-react situations. This is
plotted against the number of training episodes the Oracle
has had.
The Oracle’s performance is better in the ‘noreact’ scenario, because it then just has to choose the lowest index
among the 14 it is presented with. In the ‘react’ scenario,
it has to choose the lowest even index, unless there are no
even indexes or the lowest even index is 20 above the lowest odd one, in which case it has to choose the lowest index
overall (which will be odd). This is a more complicated
task.
4. Slightly non-episodic
All these designs rely on the Oracle’s reward being
episodic. If this fails entirely, then there is no safety: the
Oracles can leak the value of K, bit by bit if needed.
But what if it fails only slightly, in that the Oracles prefer
their own episodic reward, but give a very small value to
future episodes or future Oracles?
Alice can combat this by making it in the Oracle’s interest
to leak no information at all. For the on-policy Oracle, she
can further reduce the size of the list L, and ask questions
that are independent of her actions, so that the top option
is clear, and can be checked later. She can also attempt to
ensure that there are not multiple close candidates for the
top option.
For the counterfactual Oracle, she can simply reduce the
significant figures allowed in the answer o, so that the Oracle would lose too much, per S(F, o), if it deviated from
the optimal answer to try and communicate K.
In both cases, she can try and ensure that answering “honestly” will give the Oracle almost maximum reward (taking
into account the uncertainty of the problem), while deviating will give much lower reward. Thus Oracles that attempt
to communicate K will be sacrificing a lot of current reward for a tiny amount of future reward, which they will
not do.
5. Conclusion
The original Oracle problem was that a powerful questionanswering intelligence could influence us through the content of its answers. Recasting this in a cryptographic framework – with the Oracle cooperating with Eve to pass along
a key K – allows a crisp formulation of the challenge.
Figure 2. Training the non-counterfactual Oracle: training
episodes versus no-react (purple, square marks ever 50 points)
and react (blue, triangle marks every 50 points) accuracy (average number of times the Oracle outputs the company with maximal profit).
In that formulation, two designs can be defined: a counterfactual Oracle that is constrained by only getting its rewards when its output remains unread, and a non-counter
factual on-policy Oracle that is constrained by limiting the
size of its answer.
As long as the tasks remain episodic, and certain specific
precautions are taken around the use of multiple Oracles,
these designs allows many questions to be formulated and
Good and safe uses of AI Oracles
safely answered.
These designs allow for the creation of safer and nonmanipulative Oracles, from low level of predictive ability
all the way up to potential superintelligences. Further research will extend the span of questions that can be answered in this way, and may elucidate general safety and
accuracy measures of use beyond the specific Oracle designs, such as AIs with more options that only question
answering.
References
Pieter Abbeel and Andrew Ng. Apprenticeship learning via
inverse reinforcement learning. In International Conference on Machine Learning, 2004.
Stuart Armstrong, Anders Sandberg, and Nick Bostrom.
Thinking inside the box: Controlling and using an Oracle AI. Minds and Machines, 22:299–324, 2012.
Stuart Armstrong, Kaj Sotala, and Seán Ó hÉigeartaigh.
The errors, insights and lessons of famous AI predictions
– and what they mean for the future. In Proceedings of
the 2012 Oxford Winter Intelligence conference, 2014.
Nick Bostrom. Superintelligence: Paths, dangers, strategies. Oxford University Press, 2014.
Tilmann Gneiting and Adrian E. Raftery. Strictly proper
scoring rules, prediction, and estimation. Journal
of the American Statistical Association, 102(477):359–
378, 2007.
Dylan Hadfield-Menell, Anca Draga, Pieter Abbeel, and
Stuart Russell. Cooperative inverse reinforcement learning. In Advanced in Neural Information Processing Systems, 2016.
Vicent Müller and Nick Bostrom. Future progress in artificial intelligence: A survey of expert opinion. In Fundamental issues of artificial intelligence, pages 553–570.
Springer International Publishing, 2016.
Andrew Y. Ng and Stuart J. Russell. Algorithms for inverse
reinforcement learning. In International Conference on
Machine learning, pages 663–670, 2000.
Anders Sandberg and Nick Bostrom. Machine intelligence
survey. Technical Report, Future of Humanity Institute,
Oxford University, #2011-1:1–12, 2011.
Roman Yampolskiy. Leakproofing the singularity: artificial intelligence confinement problem. Journal of Consciousness Studies, 19:194–214, 2012.
| 2cs.AI
|
Per-bucket concurrent rehashing algorithms
Anton A. Malakhov
Intel Corporation
[email protected]
Abstract
This paper describes a generic algorithm for concurrent
resizing and on-demand per-bucket rehashing for an extensible
hash table. In contrast to known lock-based hash table algorithms,
the proposed algorithm separates the resizing and rehashing stages
so that they neither invalidate existing buckets nor block any
concurrent operations. Instead, the rehashing work is deferred and
split across subsequent operations with the table. The rehashing
operation uses bucket-level synchronization only and therefore
allows a race condition between lookup and moving operations
running in different threads. Instead of using explicit
synchronization, the algorithm detects the race condition and
restarts the lookup operation. In comparison with other lock-based
algorithms, the proposed algorithm reduces high-level
synchronization on the hot path, improving performance,
concurrency, and scalability of the table. The response time of the
operations is also more predictable. The algorithm is compatible
with cache friendly data layouts for buckets and does not depend
on any memory reclamation techniques thus potentially achieving
additional performance gain with corresponding implementations.
Categories and Subject Descriptors: D.1.3 [Programming
Techniques]: Concurrent Programming – Parallel programming;
D.4.1 [Operating Systems]: Process Management –
Synchronization;
concurrency;
multiprocessing,
multiprogramming,
multitasking;
E.2
[Data
Storage
Representation] – Hash-table representations.
General Terms:
Algorithms, Performance.
Keywords: concurrent
data
synchronization, scalability
structures,
hash
tables,
1. Introduction
Synchronization of state and data is often an unavoidable part of
communication in applications executing on multi-core
processors. While synchronization is sometimes necessary, it
should be minimized to avoid overheads that may limit scalability
or increase response time.
One of the fundamental building blocks in many applications
is a hash table [3]. A hash table is a container that associates keys
with values and uses a hash function to calculate placement
positions for these pairs in a main array. To ensure that concurrent
access to a hash table does not limit scalability or increase
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, or republish, to post on servers or to
redistribute to lists, requires prior specific permission and/or a fee.
SIGPLAN’05 June 12–15, 2005, Location, State, Country.
SIGPLAN’05
Copyright © 2004 ACM 1-59593-XXX-X/0X/000X…$5.00.
Copyright
1-59593-XXX-X/0X/000X…$5.00.
response time of an application, the hash table implementation
must be optimized to minimize synchronization time.
However, designing efficient concurrent hash tables is
challenging. An application may not know in advance the amount
of items to be stored and therefore generic hash tables must be
able to dynamically grow the underlying storage as items are
added. In most implementations, growth of the storage requires
full rehashing of the table, which limits access to the container
until the rehashing is complete.
This work proposes a generic way to extend a hash table
concurrently without a global synchronization, thus reducing the
impact of table growth on scalability.
We implemented one of the variations of the algorithm and
compared it against two implementations that represent two types
of other known approaches for concurrent resizing of the hash
table. Our implementation shows better performance and
scalability on average.
1.1 Concurrent resizing problem
Hash tables are based on an array of buckets. Each bucket can
contain or point to one or more key-value pairs [3]. Each lookup
operation for a given Key uses a hash function H to compute an
index i in the array of size S:
i = H( Key ) mod S,
[F1]
where the value of H is invariant for a given Key. It is important
to align the size S with number of pairs n contained in the table,
because when n is much bigger than S, searching in a bucket takes
more time as more pairs share the same bucket.
As it follows from [F1], the growth of a hash table invalidates
the positions of all the stored pairs because these positions depend
on the size of the array. The pairs need to be moved to new places,
and this operation is known as rehashing. For concurrent
algorithms, rehashing is challenging since many items need to be
moved at a time.
1.2 Related work
The evolution of concurrent extensible hash tables is well
described by Shalev and Shavit in [6]. To sum up, the best known
lock-based implementations from the Java™ Concurrency
Package [4] and Intel® Threading Building Blocks 2.1 [2] use
several high-level locks to protect different parts of the main
array. Both use chained hashing [3], and each bucket holds a
linked list of key-value pairs. Another lock-based algorithm with
high-level locks, hopscotch hashing [5], uses the open-addressed
hash scheme where each bucket can contain only one data pair.
High-level locks allow limited concurrency when the work can
be done for different parts of the array at a time but operations
may be blocked when accessing buckets under the same lock.
The lock-free extensible hash table by Shalev and Shavir [6]
avoids the necessity in explicit rehashing by maintaining a
specifically sorted linked list of all the pairs in the table so that
resizing does not invalidate the list. It does not require high-level
synchronization but requires a memory reclamation support to
enable erase operation. Also, a linked list is not a cache-friendly
data layout (as discussed in hopscotch hashing paper [5]).
Our rehashing approach can be applied even for cache-friendly
open-addressed hash tables and does not rely on the memory
reclamation support.
2. Per-bucket concurrent rehashing
When the table needs to be extended, the proposed algorithm
allocates as many buckets as already exist, and keeps the old
buckets, thus doubling the number of buckets (2S). Each new
bucket is logically mapped onto an existing (parent) bucket
(Figure 1). It means that during rehashing of bucket ip, 0 ≤ ip < S,
its pairs can be either left in place or moved into the new
bucket in, S ≤ in < 2S, such that
in mod S = ip
[F2]
Proof: For any given key and its hash value h = H( Key ), the
above equation is equivalent to
Figure 2. Recursive rehashing (two steps).
The subexpression log2 gives the index of the most significant
bit. So, this expression is actually a zeroing of this bit.
Using power-of-two hashing may hurt the even hash value
distribution and uniform utilization of buckets. Thus, hash
function H has to be more sophisticated. For example, Larson [8]
uses a prime number to restore the distribution.
The first two buckets have no defined parent index, so they are
root buckets. Otherwise, there is only one immediate parent for a
bucket. And a parent bucket has one immediate child per each
next level of capacity and even more indirectly.
And consequently, if a parent bucket is not rehashed until the
next extension of the table or the capacity is increased by a few
levels at once, the key-value pairs have to be split along more than
one bucket.
(h mod 2S) mod S = h mod S
2.2 Recursive rehashing
The last equation is trivially proved assuming h = 2Sk + x, where
k ≥ 0 and 0 ≤ x < 2S.
Each new bucket is marked as new (or non-rehashed) before
allocation is finished by the publication of the new space, and
rehashing is not required to follow before or immediately after the
publication. Thus, resizing does not impact concurrency because it
neither invalidates nor locks existing buckets.
After the new capacity is published, any operation that
encounters a new bucket is redirected to the parent bucket(s) and
moves (rehashes) part of its content into the new bucket(s). After
the new bucket is updated it is marked as rehashed. Thus, the
rehashing work is deferred and split across subsequent operations
within the table.
This algorithm resembles linear hashing developed by
W. Litwin in 1980 [7] but all buckets are allocated at once and
split on demand in any order.
2.1 Power-of-two hashing
It is practical to use power-of-two hashing similarly to linear
hashing, so light-weight bit-wise operations can be used instead of
the division. For S = 2level.
i = H( K ) & (2level-1)
[F3]
The index of the parent bucket for a given i can be calculated as:
ip = i & (2
log2 i
- 1)
The simplest way to rehash a bucket is to access its parent and
move necessary data between them. But if the parent is also
―new‖, it must first be rehashed, leading to a recursive rehashing
operation where a pair may be moved in one bucket and
immediately into another. See Figure 2.
Besides excessive moves, another drawback of this scheme is
that data is moved only when a new bucket is accessed, while
accessing ―old‖ buckets does not cause rehashing, which may
slow down the search across such a bucket the next time.
2.3 Multilevel rehashing
The drawbacks of recursive hashing may be addressed by
extending buckets with a more complex control field that
indicates the number of segments rehashed. So, any operation
accessing an ―old‖ bucket can detect whether rehashing is
necessary and move the data directly into all the new buckets
available at once as shown in Figure 3.
Nevertheless, recursive rehashing may be a better choice when
resizing is a rare event and there is no space for additional field in
the bucket.
2.4 Generality
The proposed algorithm does not restrict implementers to a
[F4]
3
24
2
3
3
04
0
2
0
04
0
0
0
04
0
0
∙∙∙ Existing buckets ∙∙∙
∙∙∙ New empty buckets ∙∙∙
Figure
mapping
of the buckets.
Figure 1.
1. Allocation
―This figureand
caption
is formatted
with the SIGPLAN Figure caption paragraph
style. The label ―Figure‖ has been inserted by
selecting Insert > AutoText > SIGPLAN Figure
from the menu (which also inserts the period
and the obligatory trailing en space). The label
can be referenced from the main text using
Word‘s cross referencing facility, i.e. by
Figure 3. Multilevel rehashing (one turn).
specific structure of buckets and corresponding implementations
of the table operations. It depends only on a high-level scheme of
allocations for the array and some simple flags in each bucket. So
it may be used with various concurrent hash table algorithms
which need to move the key-value pairs among the buckets while
resizing.
2.5 Reactive synchronization
Any hash table that employs the proposed per-bucket rehashing
algorithm must not only correctly synchronize the usual find,
insert, and erase operations but also the per-bucket rehashing
operations as well. However, bucket-level synchronization is not
sufficient to maintain the correct global state of the table with perbucket rehashing because the value of table capacity S is always
read before accessing a bucket (to calculate its index i) and can be
changed right afterwards:
Thread 1
i = H( Key ) mod S
Threads 2, 3
S = S*2 // table grow
Rehash_bucket( i )
// Key can be moved from i
Lookup( Key, i )
Therefore, if a key is not found during a search of the hash table,
it is possible that it was moved by a concurrent rehashing of the
bucket. To eliminate possibility of this race condition, explicit
global-level synchronization is required but it hurts scalability and
increases the hot path of every table operation.
Instead, the proposed algorithms handle this potential race
condition using another approach. To make sure a key does not
exist, the current capacity of the array is compared with the
capacity used to compute the bucket index. If (1) they are not
equal, (2) the bucket indexes computed with these values are
different, and (3) the bucket with the old index was (or is being)
rehashed towards new one, then the algorithm restarts the lookup
operation.
For recursive rehashing, the last part is complicated because
there is no direct indicator of whether the old bucket was used to
rehash the new one. Instead, the next child bucket where the key
can be found is checked for required rehashing state as shown in
function is_race at Figure 5 and described in 3.1.
For multilevel rehashing, this check is trivial using information
about rehashed segments stored in the bucket, which in fact is
current capacity effective for the given bucket. The difficulty here
is rather in algorithm of detecting a correct root parent while other
threads can rehash it concurrently (not discussed in the paper).
Reactive synchronization is the key idea for lock-free
synchronization of global consistency that is common for both
rehashing algorithms. It does not impose the costs to the hot path
and thus, it does not hurt concurrency while maintaining overall
correctness.
2.6 Complexity
The complexities of the hash table operations depend on the
buckets layout and implementation of the operations. In addition,
concurrent rehashing requires restarting of lookup operations if
the race condition is detected. But as shown below, the statistical
probability of this event is negligibly small.
With recursive rehashing, deferred rehashing can lead to more
cycles wasted for searching in an old bucket until accessing the
related new bucket rehashes it. As we show below, the slowdown
is noticeable when the insertion of unique keys takes the largest
portion of all the operations within the table.
3. Performance
We implemented a simple chaining concurrent hash table with
recursive rehashing in C++ using synchronization primitives from
Intel‘s TBB library [2] like tbb::atomic and tbb::spin_rw_mutex.
These are the same basic blocks that were used in
tbb::concurrent_hash_map (as implemented in TBB 2.1) and
tbb::concurrent_unordered_map (TBB 3.0) which we compare it
against. The former uses high-level locks as we described above.
And the latter is based on the original split-ordered list algorithm
by Shalev and Shavir [6] and a well known extensible dynamic
array algorithm [1] similar to the one implemented in
tbb::concurrent_vector [2]. Our implementation uses the same
approach to double the number of buckets. In addition, the
structure of buckets is the same as the one in
tbb::concurrent_hash_map.
All these similarities help to exclude the majority of factors
that are not related to the concurrent rehashing itself. However,
tbb::concurrent_unordered_map lacks a concurrent erase
operation, which gives it an unfair advantage over the rivals due
to a simplified synchronization protocol without memory
reclamation support. Nevertheless, as we show in the performance
evaluation section below, our implementation performs better in
most cases.
3.1 Implementation details
Our implementation was carefully tested and released as part of a
commercial product available also in open sources. Figure 5
represents the pseudo-code compiled from the real
implementation and highlights the main flow related to concurrent
resizing and recursive rehashing. However, for the sake of
simplicity and readability of the code, we omit some details like a
read-writer lock and the double-check idiom. Furthermore, this
fragment does not include other operations like find and erase.
They can be easily derived from the presented insert operation.
The first lines define the basic structure of the hash table.
my_mask contains the total number of buckets minus one (S-1),
and my_size contains the number of inserted pairs. It is important
for the scalability to avoid false sharing by placing these variables
in separate cache lines.
my_table is a segment table of a fixed size that stores pointers
to arrays of buckets – segments as shown in Figure 4. As the table
grows, new segments are added. Each added segment contains as
many buckets as all of the preceding segments, thus doubling the
total number of buckets at each growth of the table. The size of
my_table is a number of bits in my_size.
The size of the first segments is two buckets. It leads to the
simplest segment index computations formulas. At lines 14-16,
segment_index_of calculates the index of the segment for the
0
1
2
3
Segment
table
Figure 4. Segment table and mapping of a segment.
01 my_mask : atomic integer;
47 insert( Key ) {
02 my_size : atomic integer;
48
k = 0;
03 my_table : array of pointers to buckets;
49
h = H( Key );
04
50
m = my_mask.fetch();
05 is_race ( h, m ) {
51 label restart:
06
//find next applicable mask after old
52
b = acquire_bucket( h & m );
07
while ( (h & (m+1)) == 0 ) // test bit
53
p = b.search( key );
54
if( p == 0 ) {
08
m = m<<1;
09
m = m<<1;
55
m_now = my_mask.fetch();
10
//check whether it is [being] rehashed
56
if( (h & m)!=(h & m_now)
11
return is_marked_new (h & m) == false;
57
&& is_race(h, m) ) {
12
// if true a lookup must be restarted
58
unlock( b );
13 }
59
m = m_now;
14 segment_index_of( i ) {
60
15
61
}
16 }
62
// insert and set flag to grow
17 // first global index of segment k
63
sz = my_size.fetch_and_increment();
18 segment_base( k ) {
64
b.insert( key );
19
65
// check load factor
20 }
66
if( sz >= m ) {
21 acquire_bucket( i ) {
67
new_seg = segment_index_of( m+1 );
22
// find bucket i
68
if( my_table[new_seg] == 0
23
k = segment_index_of( i );
69
24
j = i - segment_base( k );
70
25
b = my_table[k][j];
71
26
lock( b );
72
}
27
if( is_marked_new( b ) )
73
unlock( b );
28
rehash_bucket( b, i );
74
if( k ) { // new segment
29
return Log2( index|1 );
return (1<<k & inverse_bits(1));
return b;
goto restart;
&& CAS(my_table[new_seg],marker,0))
k = new_seg; // processed below
}
75
// k > 0, first blocks pre-allocated
30 }
76
sz = 1 << k; // segment_size
31 rehash_bucket( b_new, i ) {
77
segment = allocate_buckets( sz );
32
mark_rehashed(b_new);
78
for each b in segment
33
// parent mask from the MSB of i
79
34
mask = ( 1<<Log2(i) ) - 1;
80
my_table[k] = segment;
35
b_old = aquire_bucket( i & mask );
81
sz = sz * 2;// new value of capacity
36
// get full mask for new bucket
82
37
mask = (mask<<1) | 1;
83
38
for each p in b_old {
84 }
39
h = H( p.key );
40
if( (h & mask) == i ) {
41
mark_new( b );
my_mask.store( sz-1 ); // new mask
}
85
b_old.remove( p );
42
b_new.add( p );
43
}
44
}
45
unlock( b_old );
46 }
Figure 5. Pseudo code of chained concurrent hash table with recursive rehashing
bucket index i. Lines 18-20 compute the first bucket index in a
given segment.
But using two buckets as the size of the first segment leads to
excessive fragmentation of the cache lines if the first segments are
allocated separately. In addition, resizing will happen more often,
and contention on the first buckets will kill scalability.
Thus, we initially allocate enough (e.g. 512) buckets in a
single block and just route segment pointers to appropriate
positions inside the block. Therefore, a few pointers in the
segment table will be set at once pointing inside a combined array
of buckets. The logical bounds of the segments remain, but
fragmentation is avoided. The first row of buckets in Figure 4
illustrates the joint allocation.
This allocation is done before any concurrent operation over
the table which allows to place growth logic (lines 74-83) after
insertion operation (lines 62-71). And thus, it helps to avoid
unnecessary reading of the shared my_size counter before the
actual insertion of a new key occurs. It is not known whether a
key is unique or not until a lookup operation (lines 49-61, 73)
finishes.
Each insert operation that adds a new pair in the hash table
atomically increases the my_size counter and makes a decision
whether to resize or not using its value as well as the current
capacity (line 66).
The segment index is computed using the current capacity
value (line 67) and if the pointer in the segment table is equal to
zero then a compare-and-swap (CAS) operation sets the value to
any non-zero value in order to select a thread that should perform
the actual allocation of the new segment.
Before the publication of the new space, each bucket should be
marked as new (lines 78-79). Publication ends by setting the
current mask (i.e. capacity) to the new value (line 82).
After this point, any operation may access the new segment
and the race condition described in section 2.5 can occur because
lines 50 and 52 are not executed in one transaction.
If a key is not found (line 54), the current mask value is re-read
and compared with the previous value applied on the hash code
(line 56). Difference means that the concurrent growth affects the
desired key, therefore rehashing status of an immediate child
bucket should be checked (line 57).
The function is_race (lines 5-11) calculates the index of such a
child by searching the next bit set after old mask and then checks
the marker (line 11). It cannot just check a h&m_now bucket
because a concurrent rehashing operation can move it to an
intermediate bucket that stands between parent and the ―target‖
bucket. For example, if m=3 and m_now = 15 (e.g. increased by
two segments in a joint allocation), then for h = 14 the following
situation is possible (see also Figure 2):
1. tbb::concurrent_hash_map
Thread 1
acquire_bucket( 14 & 3 )
Other threads
b.search( key ) == 0
is_marked_new(14&15)==true
is_marked_new(14&7)==false
my_mask = 15
rehash_bucket( 6 );
// key moved to 6
// race not detected
// race detected
There is no partial rehashed state, so the rehash_bucket(6) cannot
leave the pair in a parent bucket 2 even if the hash value and
my_mask tell that they will be moved further to 14. Otherwise,
the rehash_bucket(14) calling at line 35 acquire_bucket(6) will
not initiate the rehashing of bucket 2 at line 28.
Instead, the rehash_bucket (lines 31-46) calculates a minimal
mask for a specified new bucket (line 37) and compares hash
values from the parent bucket cut by the mask with the index of
the new bucket without respecting the current my_mask value.
The rehashing operation is initiated through the
acquire_bucket (lines 21-30) called at line 52 under the lock
(line 26) if the requested bucket is marked as new (line 27). Then,
it calculates the index of the parent bucket at line 34 and calls the
acquire_bucket again to lock the parent as well.
3.2 Performance evaluation
We used Intel® Threading Building Blocks (TBB) library [2] as a
programming framework for evaluation of the described
algorithm because our practical interest resided in a native
programming using C++ without memory reclamation support.
Also the library provides two implementations of concurrent hash
table that represent both known types of concurrent resizing:
algorithms with high-level locks (tbb::concurrent_hash_map) and
split-ordered list (tbb::concurrent_unordered_map).
2. tbb::concurrent_unordered_map
3. Recursive rehashing
40
35
Throughput, ops/ns
30
25
20
15
10
5
0
1
3
5
7
9 11 13 15 17 19 21 23 1
5% unique
3
5
7
9 11 13 15 17 19 21 23 1
30% unique
3
5
7
9 11 13 15 17 19 21 23
100% unique
Figure 6. Throughput scalability of filling the table up to 2M pairs on various input rates
Throughput, ops/ns
1. tbb::concurrent_hash_map
2. tbb::concurrent_unordered_map
3. Recursive rehashing
50
45
40
35
30
25
20
15
10
5
0
1
3
5
7
9 11 13 15 17 19 21 23 1
3
5
7
5% unique
9 11 13 15 17 19 21 23 1
30% unique
3
5
7
9 11 13 15 17 19 21 23
100% unique
Figure 7. Throughput scalability of checking the table for existence of 2M inserted keys
5
204850.0 48580.0
Unique,% Rehashed Empty
Overpopulated
10
2048482.6 48625.4
52.8
20
2032712.9 54454.9
8314.8
30
1963778.5 77963.0
43086.9
100
1245752.3 187169.7
591866.6
Here, the number of concurrent growths was 12.
The Figure 8 shows how the number of operation restarts
grows along the number of threads, but in average it remains very
small in comparison to the total number of operations. Thanks to
that, reactive synchronization involves less overhead than explicit
high-level synchronization.
4. Future work
Though this paper does not develop multilevel rehashing in details
and does not present its evaluation, this version of the rehashing
algorithm is expected to outperform recursive rehashing because it
initiates rehashing faster for old buckets, acquires less locks, and
moves key-value pairs directly. Also, it will feature a simpler
race-detection code since there will be no intermediate rehashing
of buckets. However, all these points relate rather to the growth
condition and will be visible only for the case of the intensive
20
18
16
14
12
10
8
6
4
2
0
Threads 1 2 3 4 5 6 7 8 9 101112131415161718192021222324
Average, number of buckets
Unique,% Rehashed Empty
Average, number of buckets
# of Lookup restarts
We compared our hash table against these two on a four-socket
machine with Intel® Xeon® Processors X7460 (6 cores per
socket – 24 in total, 16M Cache, 2.66 GHz, 1066 MHz FSB)
using a micro-benchmark specifically designed to highlight
aspects of concurrent resizing of the hash table.
It was derived from a program that counts the number of some
unique occurrences from the input data. We simplified the original
task to exclude synchronization on the counters from the picture,
so that it builds just a set of unique numbers from an input array.
We filled the array evenly by using a pseudo-random number
generator from the standard C library for various proportions of
unique numbers. For example, for 5% of unique numbers, the
same number is repeated 20 times on average. Together, it gives
5% of actual insertions and 95% are just lookups. However, in the
beginning, there are more new keys occur than in the end.
In addition, a size of the source array correlates with input
rates in order to produce the same number of unique keys and so
exclude cache effects from the equation. Figure 6 shows the
average throughput (axis Y) among the threads number (axis X)
and input rates.
You can see that in most cases our implementation
outperforms the rivals and scales better than split-ordered list for
any rate. It also scales better than tbb::concurrent_hash_map for
5% but saturates after 8 threads for case when almost every input
number is unique (shown as 100% but actually is about 99.95%).
As noted above, concurrent rehashing algorithm rehashes
buckets on-demand. In addition, recursive rehashing does not
initiates rehashing when accessing ―old‖ buckets. It explains why
the subsequent search over the table using the same input array
(Figure 7) shows that recursive rehashing algorithm on ―100%‖
scales worse than concurrent_unordered_map (which also defers
bucket initialization). For other rates, these two implementations
are almost the equal. However, as explained above, this simplified
version of split-ordered list has the unfair advantage because does
not synchronize lookup and erase operations.
The following statistics shows how more intensive insertions
lead to less number of rehashed buckets and worse distribution of
the key-data pairs along the hash table at the end of all operations.
Overpopulated
0.0
Maximal 0 1 2 4 6 7 7 11 9 101114131414201716171819181818
Average 0 0 0 0 1 0 1 1 1 1 1 2 1 2 2 2 2 2 3 3 2 3 2 3
Figure 8. Number of restarts due to race condition
insertion of unique keys.
We may expect better scalability of an extensible hash table
based on hopscotch hashing [5] combined with per-bucket
concurrent rehashing.
Other directions of the research are lock-free underlying
implementations, partial and cooperative variations of the perbucket rehashing algorithm. They might be investigated to allow
concurrent rehashing of the same buckets by different threads.
5. Conclusion
Our article presented a novel and generic approach to grow hash
tables concurrently. It avoids high-level locks and does not
depend on the memory reclamation techniques required by some
related algorithms. Unlike related work, this algorithm detects
races and restarts operations as needed.
It may be used in a wide range of hash table implementations
of different types. We hope it will be inspiring to authors of the
hash table implementations, as well as for researchers, because
further research is necessary to find the best performing variant.
Acknowledgments
Thanks to Mike Voss, Alexey Kukanov, Anton Eremeev, Volker
Prott, and anonymous reviewers for valuable comments for this
paper, as well as to other teammates who ever helped me with my
work due to the algorithm.
References
[1] Griswold, W.G. and Townsend, G.M., The Design and
Implementation of Dynamic Hashing for Sets and Tables in Icon,
Software—Practice and Experience, vol. 23(4), 351–367 (April
1993)
[2] http://www.threadingbuildingblocks.org/ and book by James Reinders, Intel Threading Building Blocks: Outfitting C++ for Multi-core
Processor Parallelism, O'Reilly Media, Inc., 2007, ISBN: 9780596514808
[3] Knuth, D.E.: The art of computer programming, volume 1 (3rd ed.):
fundamental algorithms. Addison Wesley Longman Publishing Co.,
Inc., Redwood City, CA, USA (1997)
[4] Lea, D. 2003. Hash table util.concurrent.ConcurrentHashMap,
revision 1.3, in JSR-166, the proposed Java Concurrency Package.
http://gee.cs.oswego.edu/cgibin/viewcvs.cgi/jsr166/src/main/java/util/concurrent/.
[5] Maurice Herlihy, Nir Shavit, and Moran Tzafrir, Hopscotch Hashing,
September 2008, 22nd International Symposium on Distributed
Computing (DISC 2008)
[6] Shalev, O. and Shavit, N., "Split-Ordered Lists: Lock-Free
Extensible Hash Tables", Journal of the ACM, Vol. 53, No. 3, May
2006, pp. 379–405.
[7] Litwin, W. Linear hashing: A new tool for file and table addressing.
In Proceedings of the 6th Conference on Very Large Databases,
(New York, 1980), 212-223.
[8] P.-Å. Larson, ‗Dynamic hash tables‘, Communications of the ACM,
31, (4), (1988).
| 8cs.DS
|
DiGrad: Multi-Task Reinforcement Learning with Shared Actions
arXiv:1802.10463v1 [cs.LG] 27 Feb 2018
Parijat Dewangan1∗ , Phaniteja S1∗ , K Madhava Krishna1 , Abhishek Sarkar1 ,
Balaraman Ravindran2
1
Robotics Research Center, International Institute of Information Technology, Hyderabad
2
Indian Institute of Technology, Madras
{ parijat10, phaniteja.sp}@gmail.com, { mkrishna, abhishek.sarkar}@iiit.ac.in,
[email protected]
Abstract
Most reinforcement learning algorithms are inefficient for learning multiple tasks in complex robotic
systems, where different tasks share a set of actions. In such environments a compound policy
may be learnt with shared neural network parameters, which performs multiple tasks concurrently.
However such compound policy may get biased towards a task or the gradients from different tasks
negate each other, making the learning unstable
and sometimes less data efficient. In this paper,
we propose a new approach for simultaneous training of multiple tasks sharing a set of common actions in continuous action spaces, which we call as
DiGrad (Differential Policy Gradient). The proposed framework is based on differential policy
gradients and can accommodate multi-task learning
in a single actor-critic network. We also propose
a simple heuristic in the differential policy gradient update to further improve the learning. The
proposed architecture was tested on 8 link planar
manipulator and 27 degrees of freedom(DoF) Humanoid for learning multi-goal reachability tasks
for 3 and 2 end effectors respectively. We show that
our approach supports efficient multi-task learning
in complex robotic systems, outperforming related
methods in continuous action spaces.
1
Introduction
There has been an increasing demand for reinforcement
learning (RL)[Sutton and Barto, 1998] in the fields of
robotics and intelligent systems. Reinforcement learning
deals with learning actions in a given environment to achieve
a goal. Classic reinforcement learning techniques make use
of linear approximation or tabular methods to learn such
correlation[Konidaris et al., 2011].
With the advancements of deep neural networks in the recent times, learning non-linear approximations and feature
extraction has becomes much simpler. It was believed that
non-linear approximators like neural network are hard to train
∗
These authors contributed equally.
Figure 1: Humanoid robot multi-tasking to reach two goals simultaneously. The two goals are shown using blue and green coloured
balls.
in reinforcement learning scenario. However recent advancements in RL has successfully combined the deep neural networks with RL and stabilized the learning process. Deep
Q Networks(DQN) [Mnih et al., 2015] used Convolutional
neural networks(CNN) and fully connected layers to make
the RL agents learn to play the ATARI games. Following
the success of DQN, several improvements on this architecture like Double DQN[Van Hasselt et al., 2016], Prioritized
Replay[Schaul et al., 2015], Duelling Network[Wang et al.,
2015] are proposed which propelled the use of Deep RL in
multi-agents. [Lillicrap et al., 2015] proposed Deep Deterministic Policy Gradient(DDPG) for continuous control tasks
which further extended the scope of Deep RL applications to
robotics.
Robotic systems like open and closed kinematic chains offer fresh perspectives to employ deep RL algorithms. [Gu
et al., 2017] applied DDPG framework to learn manipulation
tasks in complex 3D environments. Following this, [Phaniteja
and Sarkar, 2017] applied DDPG to learn reachability tasks in
Humanoid robot. Even though they are not multi-agent systems, they can be posed as multi-tasking systems where there
are shared actions (common kinematic chains). As shown in
Fig. 1, the spine/torso is the common chain which contributes
to the reachability tasks of both the hands in humanoid robot.
In this paper, we propose a novel framework called DiGrad
based on differential policy gradients to learn multi-tasking
in such robotic systems where different tasks may share a set
of common actions.
There are several mathematical approaches which try to address the problem of multi-tasking in branched manipulators.
However, classical methods based on Jacobian[Buss, 2004;
Møller, 1993] has very limited capability in branched manipulators. Methods like Augmented Jacobian[Siciliano, 1990;
Chiaverini, 1997] have constrained solution spaces, while
methods based on optimization[Klein et al., 1995] often
doesn’t provide real time control. Hence, RL based solvers
are of great use in this domain which can sample the entire
solution space and learn a real time controller in such complex robotic systems.
One direction for learning multiple tasks in such scenarios is to use DDPG to learn a compound policy taking care
of all the tasks. However we found DDPG to be unstable
for such multi-task scenarios. DiGrad addresses this problems by using differential policy gradient updates. We test
our framework on branched manipulators shown in Fig. 3 for
learning reachability tasks of multiple end effectors simultaneously. The proposed framework shows substantial improvement over DDPG and is considerably robust on all the
experiments we conducted.
The rest of the paper is organised as follows. Section 2
and 3 discusses the related works and background. Section
4 explains the mathematical background behind the proposed
framework and provides the detailed algorithm. Finally, Section 5 and 6 contain the experimental results and discussion
respectively.
2
Related Works
Most of the multi-task reinforcement learning algorithms rely
on transfer learning approaches. [Lazaric, 2012] shows a
good collection of these methods. Some of the recent works
based on this approach are [Rusu et al., 2015; Yin and Pan,
2017]. Some works by [Borsa et al., 2016] and [Zhang et al.,
2016] explored learning universal abstractions of state-action
pairs or feature successors.
Apart from transfer learning, some works like [Lazaric and
Ghavamzadeh, 2010], [Dimitrakakis and Rothkopf, 2011] investigated joint training of multiple value functions or policies. In a deep neural network setting, [Teh et al., 2017]
provided a framework for simultaneous training of multiple
stochastic policies and a distilled master policy. Unlike our
work, [Teh et al., 2017] uses multiple networks for each policy and one more network for the distilled policy. In our work,
we show how we can use a single network to learn multiple
deterministic policies simultaneously.
All the above mentioned methods assume multi-agent scenario whereas in our paper, we concentrate on learning multiple tasks in a robotic system. Some very recent works in
this scenario are [Yang et al., 2017] and [Kulkarni et al.,
2016]. These works do not talk about the actions which are
shared among different tasks, thus limiting their applicability.
Unlike these frameworks, we explore the case of multi-task
learning in branched manipulator which have shared actionspaces.
3
Background
We consider a standard reinforcement learning setup consisting of an agent interacting with an environment E in discrete
time steps. At each time step t, the agent takes a state st ∈ S
as the input, performs an action at ∈ A according to a policy,
µ : S → A and receives a reward rt ∈ R. We assume a
fully observable environment and model it as a Markov decision process with state space S, action space A = IRN ,
an initial state distribution p(s1 ), state transition dynamics
p(st+1 |st , at ) and a reward function r(st , at ).
The goal in reinforcement learning is to learn
P at policy µ
which maximizes the expected return Rt =
γ r(st , at ),
t>0
where γ ∈ [0, 1] is the discount factor. Since the return depends on the action taken, the policy may be stochastic but
in our work we consider only deterministic policies. Hence
the discounted state visitation distribution for a policy µ is
denoted as ρµ .
The action-value function used in many reinforcement
learning algorithms is described as the expected return after
taking an action at in state st and thereafter following the
given policy:
Qµ (st , at ) = E[Rt |st , at ]
DDPG is an off policy learning algorithm that uses an actorcritic based framework for continuous control tasks. In
DDPG, both actor and critic are approximated using neural
networks (θµ , θQ ). The problem of instability in training is
0
0
addressed by using target networks (θµ , θQ ) and experience
replay using a replay buffer. In this setting, the critic network
weights are optimized by minimizing the following loss:
L(θQ ) = (Q(st , at |θQ ) − yt )2
(1)
where,
0
0
yt = r(st , at ) + γQ0 (st+1 , µ0 t+1 (st+1 |θµ )|θQ )
(2)
The gradient ascent update on policy network (actor) is given
using Deterministic Policy Gradient(DPG) theorem([Silver et
al., 2014]). Suppose the learning rate is η, then:
θµ = θµ + η∇a Q(st , at |θQ )|a=µ(st |θµ ) ∇θµ µ(s|θµ )
(3)
Finally the updates on target networks are:
0
0
θQ ← τ θQ + (1 − τ )θQ
0
0
θµ ← τ θµ + (1 − τ )θµ
where, τ << 1.
In the proposed framework, we use some of the basic concepts of DDPG and use DPG theorem to derive a policy gradient update which can support robust multi-task learning. Finally we explain how the learning can be improved by using
a simple heuristic in case of shared actions.
4
DiGrad: Differential Policy Gradients
We propose a framework for simultaneous reinforcement
learning of multiple tasks shared actions in continuous domain. The method is based on differential action-value updates in actor-critic based framework using DPG theorem.
The framework learns a compound policy which optimally
performs all the shared tasks simultaneously. Fig. 2 shows
the higher level description of the method. In this section, we
describe the mathematical framework behind this work.
the penultimate layer weights for each Qi are updated based
on only the reward ri (s, ai ) obtained by performing the corresponding action ai . The remaining shared network captures
the correlation between the different actions ai . Hence this
kind of parametrization is more useful in the case of shared
actions. Apart from this, the number of parameters are significantly reduced by the use of a single network.
4.3
Figure 2: Overview of the algorithm showing differential policy gradient update
4.1
Environment setting
Consider n tasks in a standard RL setting and their corresponding n action spaces A1 , A2 , ..., An . We will assume
that the state space S is the same across all the tasks and an
infinite horizon with discount factor γ. Let A be a compound
action space which performs the given set of n tasks simultaneously. Let a ∈ A denote compound actions, ai ∈ Ai
denote actions and s ∈ S denote states. Therefore the relation between the compound actions and all the n actions is
given by,
n
[
a=
ai .
i=1
Suppose as corresponds to the set of actions which affects k
of the n tasks, then as is given by
as =
k
\
where, aj ⊂ a
aj
(4)
j=1
The reward functions for the ith task depend only on the
corresponding actions ai . Therefore, we denote the reward
function as ri (s, ai ) for each task i; let Qi (s, ai ) be the corresponding action value function. Let µ be the compound
deterministic policy and µi be the task-specific deterministic
policies; therefore µi ⊂ µ, where
µi (s) = ai
4.2
Critic update
Consider a single actor-critic based framework where critic
Q(s, a) is given by function approximators parametrized by
θQ and the actor µ(s) is parametrized by θµ . Let the corre0
0
sponding target networks be parametrized by θµ , θQ . Since
we have multiple critic outputs Qi , we optimize θQ by minimizing the loss as given by [Yang et al., 2017]:
and µ(s) = a
Proposed framework
An actor-critic based framework is proposed for multi-task
learning in the given environment setting. We use a compound policy µ(s) parametrized by θµ , instead of multiple
policy networks for each policy µi .
µ:S→A
A simple parametrization for action-value functions would be
to have a separate network θQi for each Qi (s, ai ), which outputs an action-value corresponding to the ith task. Another
approach for modelling action-value function is to have a single network parametrized by θQ which outputs action-values
for all the tasks. Here, Qi is the action value corresponding
to the ith task. [Yang et al., 2017] showed that a single critic
network is sufficient in multi-policy learning. In this setting,
L(θQ ) =
n
X
(Qi (st , ait |θQ ) − yit )2
(5)
i=1
where the target yit is
0
0
yit = ri (st , at ) + γQ0 (st+1 , µ0 t+1 (st+1 |θµ )|θQ )
(6)
If there are multiple critic networks, network parameters are
optimized by minimizing the corresponding loss:
L(θQi ) = (Qi (st , ait |θQ ) − yit )2
(7)
In both the settings of critic, there is only a single actor which
learns the compound policy µ. The differential policy gradient update on the compound policy is explained in the next
subsection.
4.4
Differential Policy Gradient
Each task has a corresponding reward, ri (s, µi (s)) and hence
to learn all the tasks we need to maximize the expected reward
for each of the task with respect to the corresponding action,
ai . Therefore the performance objective to be maximized is:
J(µ, {µi }ni=1 ) =
n
X
Es∼ρβ [ri (s, µi (s))]
(8)
i=1
The update on the parametrized compound policy µ(s|θµ ) is
given by applying deterministic policy gradient theorem on
Eq. (8). The resulting update is:
∇θ µ J ≈
n
X
Es∼ρβ [∇θµ Qi (s, ai |θQ )|ai =µi (s|θµ ) ]
i=1
=
n
X
Es∼ρβ [∇ai Qi (s, ai |θQ )|ai =µi (s|θµ ) ∇θµ µi (s|θµ )] (9)
i=1
DiGrad with shared tasks
In the above environment setting, let all the n tasks share a
common set of actions as , i.e., k = n. Let a1 , a2 , ..., ak be
the actions corresponding to these k-tasks. Therefore from
Eq. (4),
k
\
as =
aj
j=1
Now the compound action a can be written as:
heuristic, the differential gradient update on shared actions
becomes:
a = {a1 ∪ a2 ... ∪ ak }
∇θ µ J ≈
= {{a1 \ as } ∪ {a2 \ as }... ∪ {ak \ as } ∪ as }
=
{ad1
∪
ad2 ...
∪
adk
k
∪ as }
Substituting these in Eq. (9):
k
X
E[∇[adi ,as ] Qi (s, ai |θQ )|adi =µdi (s|θµ ),as =µs (s|θµ )
1 X
∇as Qi (s, ai |θQ )|as =µs (s|θµ ) ∇θµ µs (s|θµ )]
+ E[
k i=1
(11)
Generalisation
Suppose there are k tasks which share a set of action as as
above and n − k tasks which are independent, with corresponding actions ak+1 , ..., an , then Eq. (11) can be written
as:
∇θ µ J ≈
i=1
E[∇adi Qi (s, ai |θQ )|adi =µdi (s|θµ ) ∇θµ µdi (s|θµ )]
k
1 X
∇as Qi (s, ai |θQ )|as =µs (s|θµ ) ∇θµ µs (s|θµ )]
+ E[
k i=1
On expanding with respect to gradient operator we get,
T
k
X
∇adi Qi (s, ai |θQ )|adi =µdi (s|θµ )
∇θµ µdi (s|θµ )
=
E
∇θµ µs (s|θµ )
∇as Qi (s, ai |θQ )|as =µs (s|θµ )
+
n
X
E[∇ai Qi (s, ai |θQ )|ai =µi (s|θµ ) ∇θµ µi (s|θµ )]
i=k+1
i=1
(12)
k
X
E[∇adi Qi (s, ai |θQ )|adi =µdi (s|θµ ) ∇θµ µdi (s|θµ )
i=1
+∇as Qi (s, ai |θQ )|as =µs (s|θµ ) ∇θµ µs (s|θµ )]
=
k
X
i=1
∇θµ [µdi (s|θµ ), µs (s|θµ )]]
=
E[∇adi Qi (s, ai |θQ )|adi =µdi (s|θµ ) ∇θµ µdi (s|θµ )]
i=1
where, adj = {aj \ as }
Here we can see that ad1 , ad2 , adk , as are disjoint sets. Therefore we can write ai = [adi ,as ]. Similarly, µi = [µdi , µs ].
From here onwards, to make it succinct we drop s ∼ ρβ from
the subscript of Es∼ρβ and simply represent it as E.
∇θ µ J ≈
k
X
k
X
E[∇adi Qi (s, ai |θQ )|adi =µdi (s|θµ ) ∇θµ µdi (s|θµ )]
i=1
k
X
+E[
∇as Qi (s, ai |θQ )|as =µs (s|θµ ) ∇θµ µs (s|θµ )]
(10)
From Eq. (12), we can observe that the policy gradient update
for the policy (µs ) of the shared action set as is the average
of the gradients of the action-value functions of all the tasks
it affects.
This can be easily extended to cases where there are more
than one set of shared tasks. Our framework can accommodate heterogeneous dependent action spaces as compared
to related multi-task RL algorithms which assume that action spaces are homogeneous or independent or both. This
demonstrates the wider applicability of our framework.
i=1
We can see that the second term on R.H.S of Eq.(10) will be
zero if all the action spaces are disjoint, that is, as = ∅. Hence
this framework can be used even when there are no shared
actions. Since the update on the actor are the sum of gradients
of different action values, we call this a differential gradient
update. It is different from the standard gradient update where
an actor is updated based on a single action value [Lillicrap
et al., 2015].
Heuristic of direction
From Eq.(10), we can see that the policy gradient update for
the policy(µs ) of shared action as is taken as sum of the gradients of action values corresponding to the tasks it affects,
whereas for policy(µdj ) of other actions adj , the gradient update is taken as the gradient of only the corresponding Qj .
Thus, this uneven scaling of gradients may lead to delayed
convergence and sometimes biasing. In order to equally scale
all the gradient updates, we take the average value of the gradients obtained from the different Q-values for the shared action as . This modification will not affect the direction of gradient, only the magnitude will be scaled. By applying this
4.5
Algorithm
In this section we explain the algorithm to learn multiple tasks
using DiGrad. The flow of the algorithm is very similar to
standard DDPG but there are significant differences in terms
of the critic and actor updates as shown in the previous subsections. In DiGrad, compound action a is executed in the
environment which returns a vector of rewards ~rt corresponding to each task instead of a single reward. The replay buffer
B stores the current state st , compound action at , observed
state after executing action st+1 and the vector of rewards ~rt .
The entire flow of the algorithm is shown in Algorithm 1.
5
Experiments and Results
The proposed framework was tested in different settings in
order to analyse the advantages of each setting. We considered four different network settings for DiGrad as follows:
(1) Single critic network with heuristics
(2) Single critic network without heuristics
(3) Multi critic network with heuristics
(4) Multi critic network without heuristics.
Algorithm 1 Multi-task learning using DiGrad
1: Randomly initialise actor (µ(s|θ µ )) and critic network
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
(Q(s, a|θQ )) with weights θµ and θQ .
Initialize the
target
network with
weights
0
0
θµ ← θµ and θQ ← θQ .
for i = 1 to Emax
Initialise random noise N for exploration.
Reset the environment to get initial state s1 .
for t = 1 to Stepmax
Get action at = µ(st |θµ ) + N .
Execute action at and get corresponding reward
vector ~rt and updated state st+1 .
Store transition (st , at , ~rt , st+1 ) in replay buffer B.
Randomly sample a mini-batch M from replay
buffer B.
Update critic θQ according to Eqs.(5), (6), (7).
Update actor policy θµ according to Eq.(12)
0
0
Update the target networks θµ and θQ
end for
end for
We compare all of them with a standard DDPG setting. We
use the same set of hyper parameters in all the five settings.
The critic network architecture is the same for both single
and multiple critic case in all aspects except in the number of
outputs. The actor network parameters are also same for all
the cases. We show the comparison of average reward as well
the mean scores of each task in all the plots. Note that the
average reward curves for DDPG are not shown as the reward
function settings for DDPG is different than that of DiGrad.
In order to test the proposed multi-task learning framework, we considered two different environments. In both
these environments, training involved learning reachability
tasks for all the end effectors simultaneously, i.e., learning
a policy on the joint space trajectories to reach any point in
their workspace. For all the experiments, we define error and
score for a particular task i as,
errori = ||Gi − Ei ||,
scorei = −log(errori )
where Gi and Ei are the coordinates of goal and end-effector
of the ith chain.
In all the experiments, agents were implemented using
TensorFlow Code base consisting of 2 fully connected layers.
RMSProp optimizer is used to train both actor and critic networks with learning rates 0.001 and 0.0001 respectively. We
used CReLu activation in all the hidden layers. While training, a normally distributed decaying noise function was used
for exploration. Whereas, while testing this noise is omitted.
We set the discount factor to be 0.99 in all the settings. In
all the tasks, we use low-dimensional state description which
include joint angles, joint positions and goal positions. The
actor output is a set of angular velocities q̇. Hence the policy
learns a mapping from the configuration space to joint velocity space.
Reward function
The reward function for DiGrad settings is modelled keeping
in mind the multi-task application. As defined before, ri is
(a) 8-link Planar Manipulator
(b) Humanoid Robot
Figure 3: Environments
the reward corresponding to the action ai of the ith task. We
give a small positive reward to ri if task i is finished. Also,
if all the end effectors reach their respective goals, a positive
reward is given to all the tasks. For all other cases, a negative
reward proportional to the error is given. In DDPG setting,
there is a single reward unlike DiGrad. A positive reward is
given when all the end effectors reach their goals simultaneously. Else, a negative reward is given proportional to the sum
of error of all the tasks, that is, sum of distances between the
respective goal and its corresponding end effector.
Environments
8-link manipulator: The first environment is an 8 DoF planar manipulator with 3 end effectors as shown in Fig. 3a. We
can observe that the shared sub-chain of 2 joints is common to
all the 3 tasks. Also, the action dimension of the non-shared
chains are kept different in order to check the robustness of
the framework. The dimensions of each action is given as:
a1 = 3, a2 = 4, a3 = 5 and as = 2.
Fig.4(A) shows the performance curves for the five different settings. From the mean score curves, we can see that
all the DiGrad based settings converge faster and achieve better final performance than DDPG. Even though the action dimension of each task was different, all the network settings
in DiGrad framework worked equally well for all the tasks.
Whereas, DDPG showed comparable performance for only
Task-2.
From 4(A1), we can see that the single critic framework
consistently has better average reward per episode than multi
critic frameworks. Thus, modelling all the action value functions in one critic network doesn’t affect the performance. In
fact, the shared parameters in single critic framework could
help the network capture the correlation among the actions
much better than multi-critic case. Note that, single critic
framework is achieving these performances with lesser parameters than the multi-critic framework.
DiGrad frameworks with heuristics perform at par with the
non-heuristic frameworks. On applying the aforementioned
heuristic, no significant improvement in the average reward
curve is observed. But in the mean score curves, specially
Task-1 and Task-2 curve, we can see that the application of
heuristics helps the network to be more stable as compared to
their respective non-heuristic curves. Thus, we can say that
normalising the gradient of action values of the shared action
as in Eq. (11) could help the network deliver robust multi-
Figure 4: Performance curves of reachability task experiments on 8 link manipulator and humanoid. The bold line shows the average over 3
runs and the coloured areas show average standard deviation over the tasks. Note that, average reward curve is not plotted for DDPG as the
reward function for it is different from DiGrad frameworks.
task training.
Humanoid robot: Secondly, we test our framework on a 27
DoF humanoid robot (Fig. 3b). This experiment involved
reachability tasks of the 2 hands of the humanoid robot using
the upper body (13 DoF) consisting of an articulated torso.
The articulated torso is the shared chain which is affecting
both the tasks. It is noteworthy that the articulated torso has 5
DoF whereas, the arms have 4 DoF each. Thus, the contribution of shared action (articulate torso) to the task is more than
the non shared actions (arms). The environment for training
is developed in MATLAB and the trained models were tested
in a dynamic simulation environment MuJoCo.
Fig.4B summarizes the results for this case. We found that
DPPG is generally unstable for solving multi-tasks problems.
In some runs, it may learn initially but suffers degradation
later. We observe that the DiGrad algorithm yields better final
results while having greater stability.
From the mean scores of the tasks, we can see that the
single critic frameworks converge faster and are stable
throughout the experiment as compared to the multi-critic
frameworks. The best setting is again the single critic with
heuristic, outperforming all the others in all the cases.
Note that, the reward function for DDPG is kept different from the DiGrad framework. We also tried a different
reward setting taking the sum of individual rewards ri as
a reward signal to DDPG framework, where ri is same as
defined in the DiGrad reward setting. We observed that
this reward setting led to biasing, where one of the tasks
dominated others. This behaviour could have been due to the
negative interference across tasks, which didn’t happen in
DiGrad.
6
Conclusion
In this paper, we propose a deep reinforcement learning algorithm called DiGrad for multi-task learning in a single agent.
This framework is based on DDPG algorithm and is derived
from DPG theorem. In this framework, we have a dedicated
action value for each task, whose update depends only on the
action and reward corresponding to the task. We introduced a
differential policy gradient update for the compound policy.
We tested the proposed framework for learning reachability tasks on two environments, namely 8-link manipulator and
humanoid. These experiments show that our framework gave
better performance than DDPG, in terms of stability, robustness and accuracy. These performances are achieved keeping the number of parameters comparable to DDPG framework. Also, the algorithm learns multiple tasks without any
decrease in the performance of each task.
Our work focuses on learning coordinated multi-actions in
the field of robotics, where a single agent performs multiple
tasks simultaneously. The framework was able to capture the
relation among tasks where some tasks had overlapped action
space. Our future work will focus on extending the framework to multi-agent domain.
References
[Borsa et al., 2016] Diana Borsa, Thore Graepel, and John
Shawe-Taylor.
Learning shared representations in
multi-task reinforcement learning.
arXiv preprint
arXiv:1603.02041, 2016.
[Buss, 2004] Samuel R Buss.
Introduction to inverse
kinematics with jacobian transpose, pseudoinverse and
damped least squares methods. IEEE Journal of Robotics
and Automation, 17(1-19):16, 2004.
[Chiaverini, 1997] Stefano Chiaverini. Singularity-robust
task-priority redundancy resolution for real-time kinematic control of robot manipulators. IEEE Transactions
on Robotics and Automation, 13(3):398–410, 1997.
[Dimitrakakis and Rothkopf, 2011] Christos Dimitrakakis
and Constantin A Rothkopf. Bayesian multitask inverse
reinforcement learning.
In European Workshop on
Reinforcement Learning, pages 273–284. Springer, 2011.
[Gu et al., 2017] Shixiang Gu, Ethan Holly, Timothy Lillicrap, and Sergey Levine. Deep reinforcement learning
for robotic manipulation with asynchronous off-policy updates. In Robotics and Automation (ICRA), 2017 IEEE International Conference on, pages 3389–3396. IEEE, 2017.
[Klein et al., 1995] Charles A Klein, Caroline Chu-Jenq, and
Shamim Ahmed. A new formulation of the extended jacobian method and its use in mapping algorithmic singularities for kinematically redundant manipulators. IEEE
Transactions on Robotics and Automation, 11(1):50–55,
1995.
[Konidaris et al., 2011] George Konidaris, Sarah Osentoski,
and Philip S Thomas. Value function approximation in
reinforcement learning using the fourier basis. In AAAI,
volume 6, page 7, 2011.
[Kulkarni et al., 2016] Tejas
D
Kulkarni,
Karthik
Narasimhan, Ardavan Saeedi, and Josh Tenenbaum.
Hierarchical deep reinforcement learning: Integrating
temporal abstraction and intrinsic motivation. In Advances in neural information processing systems, pages
3675–3683, 2016.
[Lazaric and Ghavamzadeh, 2010] Alessandro Lazaric and
Mohammad Ghavamzadeh. Bayesian multi-task reinforcement learning. In ICML-27th International Conference on
Machine Learning, pages 599–606. Omnipress, 2010.
[Lazaric, 2012] Alessandro Lazaric. Transfer in reinforcement learning: a framework and a survey. In Reinforcement Learning, pages 143–173. Springer, 2012.
[Lillicrap et al., 2015] Timothy P Lillicrap, Jonathan J Hunt,
Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,
David Silver, and Daan Wierstra.
Continuous control with deep reinforcement learning. arXiv preprint
arXiv:1509.02971, 2015.
[Mnih et al., 2015] Volodymyr Mnih, Koray Kavukcuoglu,
David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through
deep reinforcement learning. Nature, 518(7540):529,
2015.
[Møller, 1993] Martin Fodslette Møller. A scaled conjugate
gradient algorithm for fast supervised learning. Neural
networks, 6(4):525–533, 1993.
[Phaniteja and Sarkar, 2017] Guhan Phaniteja, Dewangan
and K Madhava Sarkar. A deep reinforcement learning
approach for dynamically stable inverse kinematics of humanoid robots. arXiv preprint arXiv:1801.10425, 2017.
[Rusu et al., 2015] Andrei A Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James
Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray
Kavukcuoglu, and Raia Hadsell. Policy distillation. arXiv
preprint arXiv:1511.06295, 2015.
[Schaul et al., 2015] Tom Schaul, John Quan, Ioannis
Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015.
[Siciliano, 1990] Bruno Siciliano. Kinematic control of redundant robot manipulators: A tutorial. Journal of intelligent and robotic systems, 3(3):201–212, 1990.
[Silver et al., 2014] David Silver, Guy Lever, Nicolas Heess,
Thomas Degris, Daan Wierstra, and Martin Riedmiller.
Deterministic policy gradient algorithms. In ICML, 2014.
[Sutton and Barto, 1998] Richard S Sutton and Andrew G
Barto. Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998.
[Teh et al., 2017] Yee Teh, Victor Bapst, Wojciech M Czarnecki, John Quan, James Kirkpatrick, Raia Hadsell, Nicolas Heess, and Razvan Pascanu. Distral: Robust multitask
reinforcement learning. In Advances in Neural Information Processing Systems, pages 4499–4509, 2017.
[Van Hasselt et al., 2016] Hado Van Hasselt, Arthur Guez,
and David Silver. Deep reinforcement learning with double q-learning. In AAAI, volume 16, pages 2094–2100,
2016.
[Wang et al., 2015] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Van Hasselt, Marc Lanctot, and Nando De Freitas. Dueling network architectures for deep reinforcement
learning. arXiv preprint arXiv:1511.06581, 2015.
[Yang et al., 2017] Zhaoyang Yang, Kathryn Merrick, Hussein Abbass, and Lianwen Jin. Multi-task deep reinforcement learning for continuous action control. In Proceedings of the Twenty-Sixth International Joint Conference on
Artificial Intelligence, IJCAI-17, pages 3301–3307, 2017.
[Yin and Pan, 2017] Haiyan Yin and Sinno Jialin Pan.
Knowledge transfer for deep reinforcement learning with
hierarchical experience replay. In AAAI, pages 1640–1646,
2017.
[Zhang et al., 2016] Jingwei Zhang, Jost Tobias Springenberg, Joschka Boedecker, and Wolfram Burgard. Deep
reinforcement learning with successor features for navigation across similar environments.
arXiv preprint
arXiv:1612.05533, 2016.
| 2cs.AI
|
1
Affine Parameter-Dependent Lyapunov Functions
for LPV Systems with Affine Dependence
arXiv:1803.11543v2 [cs.SY] 5 Apr 2018
Pepijn B. Cox, Member, IEEE, Siep Weiland, Member, IEEE, and Roland Tóth, Senior Member, IEEE
Abstract—This paper deals with the certification problem
for robust quadratic stability, robust state convergence, and
robust quadratic performance of linear systems that exhibit
bounded rates of variation in their parameters. We consider both
continuous-time (CT) and discrete-time (DT) parameter-varying
systems. In this paper, we provide a uniform method for this
certification problem in both cases and we show that, contrary to
what was claimed previously, the DT case requires a significantly
different treatment compared to the existing CT results. In the
established uniform approach, quadratic Lyapunov functions,
that are affine in the parameter, are used to certify robust
stability, robust convergence rates, and robust performance in
terms of linear matrix inequality feasibility tests. To exemplify
the procedure, we solve the certification problem for L2 -gain
performance both in the CT and the DT cases. A numerical
example is given to show that the proposed approach is less
conservative than a method with slack variables.
Index Terms—Linear parameter-varying systems; Parametervarying Lyapunov functions; Stability of linear systems; LMIs.
I. I NTRODUCTION
T
HE certification of stability and performance of a linear
system against uncertain and/or time-varying parameters is of paramount interest [1]–[3]. Non-linear and/or nonstationary effects can be represented by varying dynamic
behavior of a linear system description expressed as parameter
variations. This makes the so-called linear parameter-varying
(LPV) system representations widely applicable to model
physical or chemical processes. However, the certification of
robust quadratic stability, robust state convergence, and robust
quadratic performance for LPV representations becomes more
involved compared to the linear time-invariant (LTI) case.
There has been extensive research in this area, where the
parameter variations are assumed to be either time-invariant
in a constraint set or time-varying signals with possible
constraints on their values, rates of variations, bandwidths,
or spectral contents. For time-invariant parametric dependence
and arbitrary fast time-variations, a rich literature exists, e.g.,
see [3]–[5].
In a wide range of applications, systems exhibit bounded
parameter variations and bounded rates of variation. For these
systems, assumptions on arbitrary fast parameter rates are
P.B. Cox, S. Weiland, and R. Tóth are with the Control Systems Group,
Department of Electrical Engineering, Eindhoven University of Technology,
P.O. Box 513, 5600 MB Eindhoven, The Netherlands, e-mail: {p.b.cox,
s.weiland, r.toth}@tue.nl.
This paper has received funding from the European Research Council
(ERC) under the European Union’s Horizon 2020 research and innovation
programme (grant agreement No 714663).
unrealistic and conservative. Formulation of stability conditions taking the bounded rate of variation into account
can be accomplished, e.g., via integral-quadratic-constraints
(IQCs) [6]–[10] based on an equivalent linear fractional representation (LFR) of the LPV system, Lyapunov theory based
on state-space representations or LFR forms [11]–[18], the
quadratic separator theorem for LFRs [19]–[21], or Finsler’s
Lemma [22]. However, these stability and performance formulations result in an infinite set of linear matrix inequalities
(LMIs). To obtain tractable methodologies, a number of relaxation techniques can be applied. These include 1) vertex separation or convex-hull relaxation, including µ synthesis, D/G
scaling, and full-block multipliers, [6], [7], [9], [10], [19]–[21];
2) partial-convexity arguments [11]–[13]; 3) sum-of-squares
relaxation [21], [23]; 4) Pólya relaxation [8], [14]–[16];
5) slack variable methods (applied in discrete-time) [14], [16],
[17]; and 6) partitioning of the uncertainty space [1], [18].
The aforementioned overview is far from complete due to
the vast amount of literature on this topic. A summary of
various relaxation techniques in the continuous-time (CT) case
is found in [3], [24].
Many approaches presented in the literature for quadratic
stability, state convergence, and quadratic performance certification problems are either conservative or have a high
computational demand due to the large number of free parameters or the large number of LMIs. Furthermore, the CT
and the discrete-time (DT) cases are fundamentally different
for these problems. Indeed, in the DT case, the robust stability
analysis requires a significantly different relaxation technique
leading to more involved synthesis and verification results.
By neglecting this difference, the DT case remained mainly
unexplored under the claims that it trivially follows from the
CT results. We will demonstrate that this is not the case. The
goal of this paper is to provide a unified method for certification of robust quadratic stability, robust state convergence, and
robust quadratic performance, in both the CT and DT cases,
for linear systems that exhibit parameter variations where
these variations have bounded rates. The certification problem
is solved using parameter-dependent Lyapunov functions. To
avoid large numbers of free parameters, to decrease the number
of LMIs, and to avoid conservativeness of the results, we
apply the partial-convexity relaxation argument inspired by the
results in [11]–[13].
More specifically, the contributions of the paper are the
following, for the CT case, we extend [11], [12] to include
a certificate on the exponential rate of decay of the state,
independent of the uncertain parameter. For the DT case, we
extend the partial-convexity argument to handle the third-order
2
terms of the Lyapunov function without introducing additional
free parameters and we provide a less conservative robust
stability test, a state convergence guarantee, and a performance
test than other existing approaches. To the authors’ knowledge,
the only contribution using the partial-convexity argument in
the DT case is [13]. However, [13, Thm. 1] avoids the cubic
dependencies on the parameters by including additional free
parameters and by increasing the number of LMIs. Additionally, we draw a parallel with LMI regions to show why the
DT case induces a more involved set of LMIs. In addition,
an illustrative example is provided comparing the proposed
partial-convexity results with a slack variable based method
in the DT case.
The paper is organized as follows. In Section II, the concept
of LPV systems, state-space representations, and the problem
setting are introduced. The analysis on affine quadratic stability (AQS) is given in Section III. Section IV reveals how the
partial-convexity relaxation of the AQS certification problem
results in a finite set of LMIs for the CT and DT cases. In
Section V, an LMI based formulation for the affine quadratic
performance (AQP) analysis is derived for both cases. In
Section VI, an illustrative example is given for the DT case.
Concluding remarks are given in Section VII.
II. L INEAR PARAMETER - VARYING
SYSTEMS
A. LPV state-space representation
Consider a system defined by the following LPV state-space
representation:
ξx(t) = A(p(t))x(t) + B(p(t)) w(t),
z(t) = C(p(t)) x(t) + D(p(t))w(t),
(1a)
(1b)
where x : T 7→ Rnx is the state variable, p : T 7→ P ⊂ Rnp
is the time-varying parameter (i.e., the scheduling signal), w :
T 7→ Rnw is the general disturbance channel, z : T 7→ Rnz
denotes the performance channel, and t ∈ T defines time. For
d
d
, i.e., ξx(t) = dt
x(t) denotes
the CT case, T = R and ξ = dt
the time derivative of the state; and for the DT case, T = Z
and ξ = q is the time-shift operator, i.e., ξx(t) = x(t + 1).
The parameter-varying matrix functions A(), . . . , D() in (1)
are considered to be affine functions of p:
np
np
X
X
Bi pi (t),
Ai pi (t), B(p(t)) = B0 +
A(p(t)) = A0 +
i=1
i=1
np
np
X
X
Di pi (t),
Ci pi (t), D(p(t)) = D0 +
C(p(t)) = C0 +
(2)
i=1
i=1
, Bi ∈ Rnx ×nw , Ci ∈
with known matrices Ai ∈ R
nz ×nw
nz ×nx
for i = 0, . . . , np and pi is the i-th
, Di ∈ R
R
element of the scheduling variable. Due to linearity from x to
z, asymptotic stability 1 of (1) is dictated by stability of the
fixed point at the origin of the autonomous part of (1), i.e.,
nx ×nx
ξx(t) = A(p(t))x(t) ,
|
{z
}
x(0) = x0 ,
(3)
f (x(t),p(t))
1 An LPV
system, represented in terms of (1), is called asymptotically stable,
if, for all trajectories of (w(t), p(t), z(t)) satisfying (1) with w(t) = 0 for
t ≥ 0 and p(t) ∈ P, it holds that limt→∞ |z(t)| = 0.
where x0 ∈ Rnx denotes the initial state. In this paper, we
assume that the scheduling variable and its rate are bounded:
A1 The scheduling signal p(t) and its rate of variation δp(t),
d
p(t) in the CT case and δp(t) =
defined by δp(t) = dt
p(t + 1) − p(t) in the DT case, range for all t ∈ T in
the bounded set P × V, i.e., (p(t), δp(t)) ∈ P × V with
V ⊂ Rnp . Here P = co(P) and V = co(V) are defined
as the convex hulls of the hyper-rectangles
⊤
u1 · · · unp
P=
: ui ∈ {pi , pi } ,
(4a)
⊤
v1 · · · vnp
(4b)
V=
: vi ∈ {ν i , ν i } ,
with 2np vertices each.
B. Notation
Let Ivs denote the set of integers {s, s + 1, · · · , v} and Sn
the set of all n × n real symmetric matrices. The superscript
n is omitted if the dimension is not relevant for the context.
In addition, the inequalities A B and A ≻ B with A ∈ Sn
and B ∈ Sn mean that A − B is positive semi-definite and
positive definite, respectively. Let pc denote the center of P:
h
i⊤
pn +pnp
p
pc = p1 +p1 . . .
.
(5)
2
2
In
denote multiple summations
Pnp wePnwill
Pnp addition,
p
Z
.
by
·
·
·
i=1
i,...,k=1 X is the standard notation for
k=1
the collection of all maps from Z to X. R≥0 denotes the set of
nonnegative real numbers and Z≥0 is the set of nonnegative
integers. T0 = R≥0 in the CT case and T0 = Z≥0 in the
DT case. With σi {A}, we denote the singular values of a
real matrix A where the largest and smallest singular values
are indicated as σmax {A} and σmin {A}, respectively. The
eigenvalues of A are denoted as λi {A}. The spectral norm
of A is defined as kAk2 = σmax {A}. Furthermore, with kkq
we denote the vector q-norm. The Lq -norm of a continuous
R∞
1/q
signal w is kwkq = 0 kw(t)kqq dt
and for a discrete
1
P∞
q /q
.
signal w, the ℓq -norm is kwkq =
kw(t)k
q
t=0
III. A FFINE
QUADRATIC STABILITY
In this paper, asymptotic stability of (3) is verified under A1
by using a parameter-dependent Lyapunov function.
Definition 1 (Parameter-dependent Lyapunov function): The
function V : Rnx ×P → R is a parameter-dependent Lyapunov
function for (3) if:
(i) V (x, p) > 0 for all x ∈ Rnx with x 6= 0 and ∀p ∈ P;
(ii) V (x, p) = 0 for x = 0 and ∀p ∈ P; and
(iii) ∆V (x, p, r) < 0 for ∀(x, p, r) ∈ Rnx ×P×V with x 6= 0.
In the CT case,
∆V (x, p, r) := ∇xV (x, p) ·f (x, p) + ∇pV (x, p) · r, (6a)
where ∇y indicates the gradient of a function w.r.t. y and
· is the inner product. For the DT case,
∆V (x, p, r) := V (f (x, p), p + r) − V (x, p).
(6b)
Theorem 1 (Asymptotic stability): If a parameter-dependent
Lyapunov function for (3) exists, then the origin of (3) is an
asymptotically stable fixed point.
3
Proof: See, for example, [25, Thm. 4.1] for the CT case
and [26, Thm. B.23] for the DT case.
In this paper, we consider quadratic parameter-dependent
Lyapunov functions of the form
V (x, p) = x⊤ K(p)x,
(7)
where the function K : P 7→ Snx is affine in p
K(p) = K0 +
np
X
K i pi ,
(8)
i=1
nx
with unknown matrices K0 , . . . , Knp ∈ S .
Definition 2 (Affine quadratic stability): The representation (1) is affinely quadratically stable (AQS), if it admits
a parameter-dependent Lyapunov function of the form (7)-(8).
Let us provide sufficient conditions for AQS:
Theorem 2 (Sufficiency for AQS): The LPV representation (3) is affinely quadratically stable, if there exist np + 1
matrices K0 , . . . , Knp ∈ Snx such that
np
K(p) = K0 +
X
i=1
Ki pi ≻ 0,
(9)
for all p ∈ P and, in addition, for the CT case, it holds that
Lc (p, r) = A⊤(p)K(p)+K(p)A(p)+K(r)−K0 ≺ 0,
(10a)
for all (p, r) ∈ P × V or, for the DT case, it holds that
Ld (p, r) = A⊤(p)(K(p)+K(r)−K0 ) A(p)−K(p) ≺ 0, (10b)
for all (p, r) ∈ P × V.
Proof: For the CT case, e.g., see [3, Thm. 2.4.6]. The DT
case can be obtained by similar arguments.
Theorem 2 results in an infinite set of matrix inequalities in
the unknowns Ki ∈ Snx . In this paper, we convexify (10) using
a partial-convexity argument to obtain a tractable solution
by a finite number of LMIs. Within the IQC framework,
it has been shown that relaxation by partial-convexity for
systems in the form (1) lead to a less conservative formulation
than a convex-hull relaxation [24, p. 9]. It is difficult to
make generic statements on the potential conservativeness of
various relaxation methods. However, the relaxation by partialconvexity results in less decision variables when compared to
slack variable relaxation methods as discussed in [14], [16],
[17]. In Section VI, we demonstrate that relaxation by partialconvexity is less conservative than slack variable relaxation.
Alternatively, the uncertainty space can be partitioned as
in [21], where the grid size leads to a trade-off between a
conservative solution and an increased computational burden.
Therefore, the partial-convexity argument potentially avoids
excessive number of free parameters and decreases the number
of LMIs.
IV. S UFFICIENCY
FOR
AQS
WITH GUARANTEED
CONVERGENCE
A. Relaxation with partial-convexity
The functions Lc (p, r) in (10a) and Ld (p, r) in (10b) are
quadratic and cubic in p, respectively, while linear in r. Hence,
a relaxation technique is required for p only to obtain a finite
set of LMIs to verify AQS in Theorem 2. If the functions
Lc () and Ld () are negative definite on (p, r) ∈ P × V, then
the maximum of these functions must be negative. Therefore,
if it can be ensured that the maximum of Lc () and Ld ()
is at the vertices P, then (10a) and (10b) reduce to a finite
set of LMIs that are required to be satisfied on P only. The
following lemma provides the conditions for this concept:
Lemma 1 (Maximum at the vertices): Consider the cubic
function L : P 7→ S defined by
np
np
np
X
X
X
Qi,j,k pi pj pk ,
Qi,j pi pj +
L(p) = Q0 + Qi pi +
i=1
i,j=1
i,j,k=1
n
n
with matrices Qi , Qj,k , Qj,k,l ∈ S for i ∈ I0 p , j, k, l ∈ I1 p .
The function L() achieves its maximum at P if
np
X
1 ∂ 2 L(u)
= Qi,i + (Qj,i,i +Qi,j,i +Qi,i,j )uj 0,
2 ∂u2i
j=1
(11)
n
for all (i, u) ∈ I1 p ×P with P = co(P).
Proof: See Appendix A.
In fact, condition (11) implies that L() is a partial-convex
function:
Definition 3 (Partially convex function): A twice differentiable function L : P → S is partially convex if P is convex
and
∂ 2 L(p)
n
0,
for all (i, p) ∈ I1 p × P.
(12)
∂p2i
Note that positive semi-definiteness of (12) is along each
independent direction pi of the scheduling space. This is less
demanding than convexity with respect to p, which would
require the Hessian of L() to be positive semi-definite.
By applying the partial-convexity relaxation (11) in
Lemma 1, the maximum of Lc () and Ld () can be found
at the vertices and, therefore, the AQS test in (9)-(10) can
be reduced to a finite set of LMIs at the vertices only, i.e.,
Li (u, v) ≺ 0 ∀(u, v) ∈ P ×V. This is the core idea introduced
in [11] for the CT case. This concept can also be applied in
the DT case, as highlighted in Section IV-B.
By taking into account the cubic terms in L(), we differ
from the approaches presented in [11]–[13]. These cubic terms
are essential in handling the DT case.
B. The discrete-time case
Theorem 3 (Sufficiency for AQS in DT): Given an LPV
system defined by (3) with dependency structure (2) where
the scheduling variable p(t) satisfies A1.
If there exists an 0 < ǫ < 1,
√ such that the eigenvalues of
A(pc ) satisfy |λi (A(pc ))| < 1 − ǫ with pc as in (5), and
there exist np + 1 matrices K0 , . . . , Knp ∈ Snx parametrizing
K() in (8) that satisfy
Ld (u, v, ǫ) = A⊤ (u) [K(u) + K(v) − K0 ] A(u)
− (1 − ǫ)K(u) 0,
0 A⊤(u)Ki Ai +A⊤i Ki A(u),
∀(u, v) ∈ P × V, (13)
n
∀(i, u) ∈ I1 p × P, (14)
then the LPV system corresponding to (3) is AQS. In particular, V (x, p) in (7) is a Lyapunov function with ∆V (x, p, δp) ≤
4
n
−ǫV (x, p) and the state x robustly converges, i.e., kx(t)k22 ≤
b
t
2
a (1 − ǫ) kx(0)k2 for any trajectory (p(t), δp(t)) ∈ P × V for
t ≥ 0 where
a = inf λmin K(u) ,
b = sup λmax K(u) . (15)
for all (i, u) ∈ I1 p × P then (13) is convex on the domain
(p, r) ∈ P × V and, by using Lemma 1, (13) should only be
tested on the vertices of the hyper rectangle.
Part ii. This part will be proven by contradiction. Since
δp = 0 is admissible with 0 < ǫ < 1, we write (10b) as
Proof: The proof has three parts: i) sufficiency of the
conditions for asymptotic stability, ii) showing
√ that K(p) ≻ 0
for all p ∈ P is implied by |λi (A(pc ))| < 1 − ǫ together
with (13)-(14) and iii) convergence of x(t). Combining i) and
ii) for 0 < ǫ < 1 results in a parameter-dependent Lyapunov
function V () according to Theorem 2 and AQS is proven.
Part i. The function Ld (p, δp, ǫ) in (10b) is affine in δp and
ǫ, hence, it is sufficient to evaluate (10b) at the vertices V for
an 0 < ǫ < 1, i.e.,
(19)
u∈P
u∈P
Ld (p, v, ǫ) ≺ 0,
∀p ∈ P, ∀v ∈ V, and 0 < ǫ < 1. (16)
To prove AQS, fix v ∈ V, fix 0 < ǫ < 1, and define R(v) =
K(v) − K0 , then we can write (16) as
Ld (p, v, ǫ) = −(1 − ǫ)K(p) + A⊤
0 R(v)A0
np
np
X
X
⊤
2
A⊤
A⊤
+
0 R(v)Ai + Ai R(v)A0 pi
i R(v)Ai pi +
i=1
i=1
+
+
+
+
np
X
3
⊤
⊤
⊤
A⊤
Ai R(v)Aj pi pj + A0 K0 A0 +
i Ki Ai pi
i=1
i,j=1, i6=j
np
X
⊤
⊤
A⊤
i K0 A0 + A0 Ki A0 + A0 K0 Ai pi
i=1
np
X
⊤
⊤
A⊤
i Kj A0 + A0 Ki Aj + Ai K0 Aj pi pj
i,j=1
np
X
2
⊤
⊤
A⊤
i Kj Ai + Aj Ki Ai + Ai Ki Aj pi pj
np
X
i,j=1, i6=j
np
+
X
A⊤
i Kj Ak pi pj pk ≺ 0.
(17)
i,j,k=1, i6=j6=k
Based on the partial-convexity argument, Eq. (17) holds if it
holds on P and the following condition is satisfied
∂ 2 Ld (p, v, ǫ)
⊤
= 2A⊤
i R(v)Ai + 6Ai Ki Ai pi
∂p2i
⊤
⊤
+ 2 A⊤
i Ki A0 + A0 Ki Ai + Ai K0 Ai +
np
X
⊤
⊤
+2
A⊤
i Kj Ai + Aj Ki Ai + Ai Ki Aj pj 0,
j=1, i6=j
which can be rewritten as
A⊤ [K(p)+R(v)] Ai +A⊤(p)Ki Ai +A⊤i Ki A(p)
|i
{z
}
(18a)
A⊤ (p)K(p)A(p) − K(p) ≺ 0.
In addition, P is compact due to A1 and, therefore, K(p) is a
compact function on P. Let us assume that ∃p0 ∈ P such that
K(p0 ) is singular and take x0 ∈ ker(K(p0 )) with x0 6= 0, then
⊤
x⊤
0 A (p0 )K(p0 )A(p0 ) − K(p0 ) x0 =
⊤
x⊤
0 A (p0 )K(p0 )A(p0 )x0 ≥ 0. (20)
However, (20) contradicts (19), hence, (19) ensures that K(p)
cannot be singular for all admissible p ∈ P. As |λi (A(pc ))| <
√
1 − ǫ, K(pc ) ≻ 0 by the D-stability result [12]. Then, as
K(p) is compact on P, by continuity of the eigenvalues of
K(p), (19) assures that aI K(p) bI with 0 < a ≤ b < ∞
as defined in (15) for all admissible p ∈ P and, therefore, (9)
is satisfied.
Part iii. We can find an 0 < ǫ < 1 such that
∆V (x, p, δp) + ǫV (x, p) =
x⊤ A⊤ (p)K(p + δp)A(p) − (1 − ǫ)K(p) x =
x⊤ Ld (p, δp, ǫ)x ≤ 0, (21)
for all (p, δp) ∈ P × V and x(0) 6= 0. Eq. (21) provides
that V () is decaying along the solutions of (3) according to
V (x(t), p(t)) ≤ (1 − ǫ)t V (x(0), p(0)) for t > 0. In addition,
akxk22 ≤ V (x, p) ≤ bkxk22 for all p ∈ P and x ∈ X since
K(p) is positive definite and bounded, hence, we can find that
kx(t)k22 ≤ ab (1 − ǫ)t kx(0)k22 for t ≥ 0.
Remark 1: It is an essential part of the proof of Theorem 3
(Part i) that the function Ld () in (17) includes third-order
terms in p. In [13, Thm. 1], these terms are reduced to
second-order by introducing additional variables and coupling
constraints. Therefore, the proof and the LMIs in Theorem 3
are fundamentally different from [13, Thm. 1]. Moreover, the
number of decision variables in [13, Thm. 1] is twice the
number of decision variables in Theorem 3.
To decrease the number of LMI conditions in Theorem 3,
a more conservative test for AQS can be derived:
Lemma 2 (Simple AQS in DT): Given an LPV system
defined by (3) with dependency structure (2) where the
scheduling variable p(t) satisfies A1.
If there exist an 0 < ǫ < 1 and np + 1 matrices
K0 , . . . , Knp ∈ Snx parametrizing K() in (8) that satisfy
L1 (u, l, ǫ) = A⊤ (u)K(l)A(u) − (1 − ǫ)K(u) ≺ 0,
0
⊤
A
(p)Ki Ai +A⊤
i Ki A(p)
0. (18b)
Note that the partial-convexity relaxation only requires to
be positive semi-definite on the interval P. Also see that,
⊤
A⊤
i [K(p) + R(v)] Ai = Ai [K(p + v)] Ai , therefore, (18a)
to (18b) is valid. As a consequence, the partial-convexity
argument (18b) becomes independent of v ∈ V. If (18b) holds
K(u) ≻ 0,
∀(u, l) ∈ P × P,
(22)
∀u ∈ P,
(23)
then the LPV system corresponding to (3) is AQS. In particular, V (x, p) in (7) is a Lyapunov function with ∆V (x, p, δp) ≤
−ǫV (x, p) and the state x robustly converges as kx(t)k22 ≤
b
t
2
a (1 − ǫ) kx(0)k2 for any trajectory (p(t), δp(t)) ∈ P × V for
t ≥ 0 with a and b given by (15).
5
Proof: Note that (22) is equivalent to (10b), however,
in (22), qp ∈ P and p ∈ P are treated as independent variables,
n
which implies that ν i = pi − pi and ν i = pi − pi for i ∈ I1 p .
Note that the function L1 (p, qp, ǫ) for p, qp ∈ P, 0 < ǫ < 1 is
affine in qp and ǫ, hence (22) holds if and only if
L1 (p, l, ǫ) ≺ 0,
∀p ∈ P, ∀l ∈ P, and 0 < ǫ < 1. (24)
Fix l and ǫ, the partial-convexity on the domain p ∈ P gives
∂ 2 L1 (p, l, ǫ)
= A⊤
i K(l)Ai 0,
∂p2i
n
∀(i, l) ∈ I1 p × P. (25)
As (23) enforces K(l) ≻ 0, (25) is satisfied. Hence, satisfying (22) and (23) results in AQS. Robust state convergence
can be proven similarly to Part iii of the proof of Theorem 3
and is therefore omitted.
Note that Lemma 2 does not depend on the rate of variation
δp, but we implicitly assume that δp cannot exceed the
difference between the extremes of P, i.e., ν i = pi − pi and
n
ν i = pi −pi for i ∈ I1 p . This approach can only be exploited in
the DT setting. The number of LMIs is decreased, however, the
feasibility test is more conservative compared to Theorem 3
as the range [ν i , ν i ] is enlarged.
C. Generalization
Theorem 4 (AQS in CT and DT with guaranteed decay):
Given an LPV system defined by (3) with dependency structure (2) where the scheduling variable p(t) satisfies A1.
Define the following set of inequalities
⊤
I
I
Lk (u, v, α) =
Mk (α)⊗K(u)
+Qk (u, v) 0,
A(u)
A(u)
[k]
Nk (i, u, α) = m22
[k]
∀(u, v) ∈ P × V,
(26)
[k]
A⊤(u)Ki Ai +A⊤i Ki A(u) +m12 Ki Ai +
m21 A⊤
i Ki 0,
n
∀(i, u) ∈ I1 p × P,
(27)
[k]
where k ∈ {c, d} and mij is the i, j-th element of Mk (α).
For the CT case with k = c, if there exists an α > 0 such
that Re (λi (A(pc ))) < −α with pc as in (5) and there exist
np + 1 matrices K0 , . . . , Knp ∈ Snx parametrizing K() in (8)
such that (26)-(27) are satisfied with
2α 1
Mc (α) :=
,
Qc (u, v) := K(v) − K0 , (28)
1 0
then the LPV system corresponding to (3) is AQS. Moreover,
V (x, p) in (7) is a Lyapunov function with ∆V (x, p, δp) ≤
−2αV (x, p) and the state x robustly converges as kx(t)k22 ≤
b −αt
kx(0)k22 for any trajectory (p(t), δp(t)) ∈ P × V for
ae
t ≥ 0 with a and b given by (15).
For the DT case with k = d, if there exists an 0 < α < 1
such that |λi (A(pc ))| < α with pc as in (5) and there exist
np + 1 matrices K0 , . . . , Knp ∈ Snx parametrizing K() in (8)
such that (26)-(27) are satisfied with
−α2 0
Md (α) :=
, Qd (u, v) := A⊤(u)[K(v)−K0 ]A(u),
0
1
(29)
then the LPV system corresponding to (3) is AQS. In particular, V (x, p) in (7) is a Lyapunov function with ∆V (x, p, δp) ≤
−α2 V (x, p) and the state x robustly converges as kx(t)k22 ≤
b 2t
2
a α kx(0)k2 for any trajectory (p(t), δp(t)) ∈ P×V for t ≥ 0
with a and b given by (15).
Proof: For the CT case, AQS is proven in [11, Thm. 3.2]
while robust state convergence is proven similar to [27, Prop.
5.6]. For the DT case, see Theorem 3.
Remark 2: Theorem 4 can also be applied for stability
analysis under the assumption of a time-invariant scheduling
signal, i.e., V = V = ∅ (robust analysis). In this simplification,
all elements in (27) with respect to the time-variation become
zero, i.e., Qc () = Qd () = 0 in (28) or (29).
Remark 3: Theorem 4 implies that, for the CT case, the
real parts of the eigenvalues of A(p) are negative for fixed
values of p, i.e., Re (λi (A(p))) < −α ∀p ∈ P and, for the
DT case, the eigenvalues are within a disc of radius α, i.e.,
|λi (A(p))| < α, ∀p ∈ P. This connects to well-known results
on LMI regions, e.g., see [12]. Contrary to the LTI case, the
location of the eigenvalues are a necessary, but not sufficient
condition for stability in the parameter-varying case, as the
contribution of Qk () in (26) cannot be neglected.
V. A FFINE Q UADRATIC P ERFORMANCE
A. Concept of dissipativity and performance
The results on AQS of Section IV can be extended to
quadratic performance measures, including L2 performance,
positivity, and H2 performance.
The LPV system corresponding to (1) is strictly dissipative
for a given supply function s : Rnw × Rnz → R if there exist
an ε > 0 and a storage function V () such that,
∆V (x, p, δp) ≤ s(w, z) − εkwk22 ,
(30)
for all (w, p, δp, x, z) ∈ Rnw ×P×V×Rnx ×Rnz , e.g., see [27].
In other words, the change of internal storage ∆V (x, p, δp)
will never exceed the amount of supply s() that flows into
the system. The inequality (30) should be satisfied in a pointwise fashion, implying that (30) also holds for all admissible
trajectories that satisfy (1). Additionally, we take s() to be
the following quadratic supply function
⊤
Q s Ss
w
w
s(w, z) =
, (31)
,
Ps =
Ps
z
z
Ss⊤ Rs
where Ps ∈ Snw +nz is partitioned as Qs ∈ Snw , Ss ∈ Rnw ×nz ,
and Rs ∈ Snz . By appropriately parametrizing (31), various
performance measures can be represented. We test feasibility of (30) with an affine parameter-dependent function (7),
hence, V () qualifies as a quadratic storage function for (1)
and (31). We say that the system (1) achieves affine quadratic
performance (AQP) for (31) whenever such a quadratic storage
function exists.
B. L2 -Gain performance
To simplify notation, we use L2 to indicate L2 in CT and
ℓ2 in DT.
Definition 4 (Induced L2 -gain): Any finite γ that satisfies
kzk2
< γ,
kwk
2
0<kwk2 <∞
sup
p∈PT0
(32)
6
is an L2 -gain upper bound where z : T0 → Rnz is the
response (performance variable) of (1) for t ≥ 0 under
x0 = 0, general disturbance w ∈ L2 (Rnw , T0 ) in CT or
w ∈ ℓ2 (Rnw , T0 ) in DT, and scheduling p ∈ PT0 .
The L2 -gain performance measure of Definition 4 can be
characterized by the supply function (31) by choosing Qs =
γ 2 I, Ss = 0, Rs = −I, and ε > 0 [27, Prop. 3.12].
Lemma 3 (L2 -gain performance): Given an LPV system
defined by (1) with dependency structure (2) where the
scheduling variable p(t) satisfies A1.
If there exist a γ > 0 and np + 1 matrices K0 , . . . , Knp ∈
Snx parametrizing K() in (8) such that
K(p) ≻ 0,
∀p ∈ P,
(33)
Lk (p, r) K(p)B(p) C (p)
B⊤(p)K(p)
−γI
D⊤(p) ≺ 0, ∀(p, r) ∈ P×V, (34)
C(p)
D(p)
−γI
⊤
where Lk (p, r) as in (10), then the LPV system represented
by (1) is AQS and has an L2 -gain performance bound γ.
Proof: See that the (1, 1)-block of (34) implies AQS
(Lemma 4). In addition, for ε > 0, Eq. (34) ensures that
∆V (x, p, δp) + z ⊤ z − γ 2 w⊤ w < 0,
(35)
for all admissible (w, p, δp, r, x, z). The choice of supply function implies the L2 -gain performance bound of Definition 4
based on similar arguments as in [27, Thm. 5.16].
As (33) and (34) impose an infinite number of LMIs, we
make use of the partial-convexity argument to find an LMIbased test for AQP.
Theorem 5 (Sufficiency for affine L2 -gain performance):
Given an LPV system defined by (1) with dependency structure (2) where the scheduling variable p(t) satisfies A1.
Define the following set of inequalities and equalities 2
Lk (u, v, α) K(u)B(u) C⊤(u)
B⊤(u)K(u)
−γI
D⊤ (u) 0, ∀(u, v) ∈ P×V, (36)
C(u)
D(u)
−γI
Nk (u, i, α) 0,
Ki Bi = 0,
n
∀(i, u) ∈ I1 p ×P, (37a)
n
∀i ∈ I1 p ,
(37b)
where Lk () and Nk () are defined in (26) and (27).
For the CT case with k = c, if there exists an α > 0 such
that Re (λi (A(pc ))) < −α with pc as in (5) and there exist
np + 1 matrices K0 , . . . , Knp ∈ Snx such that (36)-(37) are
satisfied, then the LPV system corresponding to (1) is AQS and
has an L2 -gain performance bound γ. Moreover, V (x, p) in (7)
is a Lyapunov function with ∆V (x, p, δp) ≤ −2αV (x, p) and
the state x robustly converges as kx(t)k22 ≤ ab e−αt kx(0)k22 for
any trajectory (p(t), δp(t)) ∈ P × V for t ≥ 0 with w(t) = 0
and a, b given by (15).
For the DT case with k = d, if there exists an 0 < α < 1
such that |λi (A(pc ))| < α with pc as in (5) and there exist
np + 1 matrices K0 , . . . , Knp ∈ Snx such that (26)-(27) are
satisfied, then the LPV system corresponding to (1) is AQS
2 In terms of condition (37b), if any B for i = 1, . . . , n is full row rank
p
i
then the corresponding Ki must be zero, hence, it can be removed from (36)
and (37). If all Bi are full row rank, then we obtain a parameter-independent
Lyapunov function.
with ℓ2 -gain performance bound γ. In particular, V (x, p) in (7)
is a Lyapunov function with ∆V (x, p, δp) ≤ −α2 V (x, p) and
the state x robustly converges as kx(t)k22 ≤ ab α2t kx(0)k22 for
any trajectory (p(t), δp(t)) ∈ P × V for t ≥ 0 with w(t) = 0
and a, b given by (15).
Proof: Sufficiency for affine L2 -gain performance is
obtained with robust state convergence if (36) holds together
with the following set of inequalities:
Nk (u, i, α) Ki Bi
n
0, for all (i, u) ∈ I1 p × P. (38)
Bi⊤ Ki
0
This statement can be proven by straightforward application
of Theorem 4 and Lemma 3. Hence, in the remainder of the
proof it is shown that (38) is equivalent to (37).
For the inequality (38) to hold, the blocks on the diagonal
need to be positive semi-definite, i.e., Nk (u, i, α) 0. The
matrix Nk () might be rank deficient, i.e., rank(Nk ()) =
nN ≤ nx . Hence, after an appropriate congruence transform
on (38), we obtain the following partition of (38):
ek (u, i, α) 0 Mi,1
N
ek (u, i, α) ≻ 0, (39)
0
0 Mi,2
0 with N
⊤
⊤
Mi,1
Mi,2 0
ek (u, i, α) ∈ SnN and [M ⊤ M ⊤ ]⊤ is the partition
where N
i,1
i,2
of Ki Bi after column and row re-arrangement with Mi,1 ∈
RnN ×nw , Mi,2 ∈ Rnx −nN ×nw . Note that, in case Nk () is
ek () = Nk (), Mi,1 = Ki Bi , and Mi,2 = ∅.
positive definite, N
Then, perform a Schur transform on (39) to obtain that (39)
is equivalent to
0
Mi,2
0, Ñk (u, i, α) ≻ 0. (40)
⊤
⊤ e −1
Mi,2
−Mi,1
Nk (u, i, α)Mi,1
Then (40) implies that the blocks on the diagonal are required
⊤ e −1
to be positive semi-definite, i.e., −Mi,1
Nk ()Mi,1 0. As
ek () is positive definite, the latter condition is met if and
N
only if:
⊤ e −1
Mi,1
Nk (u, i, α)Mi,1 = 0
⇐⇒
Mi,1 = 0.
Then, it follows from (40) that Mi,2 = 0. Next, the matrix
Ki Bi can be reconstructed from elementary matrix operations
based on Mi,1 and Mi,2 and we can conclude that Ki Bi = 0
(undo partition (39)).
To obtain the minimum of γ satisfying (32), we solve
minγ≥0 γ s.t. (36)-(37) hold. This convex optimization problem can be solved efficiently by numerical LMI solvers.
VI. N UMERICAL EXAMPLE
In this section, the AQS and AQP results are demonstrated
on the mass-spring-damper system from [11] with timeinvariant and time-varying parameters. For the simulation example, Matlab 2014b with Yalmip and SeDuMi3 are used. The
system is discretized using a first-order approximation [28]:
0
u(t),
x(t + 1) = Ad (f (t), c(t)) x(t) +
Ts
0
1
(41)
Ad (k, c) = I + Ts Ac (k, c),
Ac (k, c) =
,
−k −c
y(t) = 1 0 x(t),
7
where Ts = 1/20 s is the sampling time and k > 0, c > 0 are
the stiffness and the damping coefficients, which are assumed
to be time-varying. The admissible trajectories of k and c are
c(t) = c0 1 + p2 (t) , (42)
k(t) = k0 1 + p1 (t) ,
Partial-convexity
Slack variable [17]
Quadratic stability [29]
0.8
λ [-]
0.6
ǫ=0
0.4
ǫ = 0.01
ǫ = 0.02
0.2
ǫ = 0.03
ǫ = 0.04
0
10−5
10−4
10−3
10−2
10−1
100
101
102
δkmax [-]
Figure 1: Maximum of λ to guarantee AQS w.r.t. λP under
time-invariant c (δcmax = 0) and time-varying k, where ǫ
indicates the guaranteed rate of convergence of the state,
i.e., kxt k22 ≤ ab (1 − ǫ)t kx0 k22 for all t > 0. The orange
area indicates numerical problems with the partial-convexity
method using ǫ = 0.
Partial-convexity
Slack variable [17]
ǫ = 0.0357
102
ǫ = 0.025
γ [-]
ǫ = 0.0321
ǫ = 0.0286
101
ǫ = 0.02
ǫ = 0.01
ǫ=0
100
10−5
10−4
10−3
10−2
10−1
100
101
102
δkmax [-]
Figure 2: Minimum of the L2 -gain performance bound γ with
time-invariant c (δcmax = 0) and time-varying k. The dots
indicate the last point of feasibility.
0.8
λ [-]
where k0 = 1, c0 = 1, P = [−1, 1] × [−1, 1], and
V = [−δkmax , δkmax ] × [−δcmax , δcmax ]. The following three
experiments are performed:
1) For time-invariant c (δcmax = 0) and time-varying k,
the parameter box λP := {λp : p ∈ P} is uniformly
expanded with 0 < λ ≤ 1 to find the maximum λ for
which stability is guaranteed.
This is done with respect
to values of δkmax ∈ 10−5 , 102 and ǫ ∈ [0, 1]. The
results are shown in Fig. 1.
2) For time-invariant c (δcmax = 0) and time-varying k, the
performance gain
for 0.2P. This is done
γ is computed
w.r.t. δkmax ∈ 10−5 , 102 and ǫ ∈ [0, 1]. The results
are shown in Fig. 2.
3) For time-varying k and c, the parameter box λP is
expanded to find the maximum λ for which
−5 stability
is
2
guaranteed.
This
is
done
w.r.t.
δk
∈
10
,
10
and
max
δcmax ∈ 10−4 , 102 . The results are shown in Fig. 3.
Note that the case of time-invariant k (δkmax = 0) is not
considered, as it is pointed out in [11]: ‘the stability region
seems to be essentially determined by δkmax ’.
Fig. 1 shows that the slack variable method [17] is indeed
more conservative (in line with [17, Thm. 1]). This difference
is clearly visible for small δkmax . For larger values of δkmax ,
the maximum parameter box size coincides with the quadratic
stability test (δkmax = ∞), as expected. Furthermore, in the
region δkmax ∈ [0.03, 0.5], the slack variable method seems
to outperform the proposed partial-convexity argument based
method. Our method experiences numerical problems in the
orange area for ǫ = 0, see Fig. 1. In this area, neither feasibility
or infeasibility can be concluded. Hence, for the proposed
method, more research needs to be performed to improve
numerical properties, but it is unclear if the slack variable
method really outperforms the partial-convexity argument in
that orange area. A similar result is also obtained for the L2 gain analysis in Fig. 2 (the area indicating numerical problems
is not displayed in the figure).
Fig. 1 is similar to [11, Fig. 1] (CT case) in terms of the
shape, as expected. However, the domain where the magnitude
of λ decreases is shifted by a factor 10, which is due to the
discretization of the model. Furthermore, Fig. 3 shows that the
maximum size of the parameter box is almost independent on
δcmax , as has also been seen for the CT case [11].
1
0.6
VII. C ONCLUSION
In this paper, we have proposed an LMI-based analysis for
robust stability, robust state convergence, and robust performance of linear systems against uncertain and/or time-varying
parameters. Affine quadratic stability is certified by finding
a Lyapunov function, which is affine in the parameter. In
order to get a tractable solution, the partial-convexity argument
of [11] is extended for third-order terms to handle the discretetime case. In our simulation example, the partial-convexity
10−4
101
10−2
δkmax
−1
10
100
102
10−3
δcmax
Figure 3: Maximum of λ to guarantee AQS w.r.t. λP for a
time-varying k and c with the partial-convexity method.
8
argument seems to be less conservative than the quadratic
stability or the slack variable approach. For future research,
the DT results can be extended to Lyapunov functions with
quadratic dependence on the scheduling signal, similar to the
CT case in [30]. Alternatively, the DT result can be extended
to the incremental stability framework for LPV systems with
bounded rates of variations [31].
R EFERENCES
[1] J. S. Shamma, Analysis and design of gain scheduled control systems.
PhD thesis, Massachusetts Institute of Technology, 1988.
[2] J. Mohammadpour and C. Scherer, eds., Control of Linear Parameter
Varying Systems with Applications. Springer, 2012.
[3] C. Briat, Linear Parameter-Varying and Time-Delay Systems: Analysis,
Observation, Filtering & Control. Springer, 2015.
[4] K. Zhou, J. C. Doyle, and K. Glove, Robust and Optimal Control.
Prentice hall, 1996.
[5] Y. Ebihara, D. Peaucelle, and D. Arzelier, S-Variable Approach to LMIBased Robust Control. Springer London, 2015.
[6] A. Megretski and A. Rantzer, “System analysis via integral quadratic
constraints,” IEEE Trans. on Automatic Control, vol. 42, no. 6, pp. 819–
830, 1997.
[7] A. Helmersson, “An IQC-based stability criterion for systems with
slowly varying parameters,” in Proc. of the 14th IFAC World Congress,
(Beijing, China), Jun. 1999.
[8] V. Powers and B. Reznick, “A new bound for Pólya’s Theorem with
applications to polynomials positive on polyhedra,” J. of Pure and
Applied Algebra, vol. 164, pp. 221–229, 2001.
[9] A. P. Molchanov and D. Liu, “Robust absolute stability of time-varying
nonlinear discrete-time systems,” IEEE Trans. on Circuits and Systems
I: Fundamental Theory and Applications, vol. 49, no. 8, pp. 1129–1137,
2002.
[10] H. Köroğlu and C. W. Scherer, “Robust performance analysis for structured linear time-varying perturbations with bounded rates-of-variation,”
IEEE Trans. on Automatic Control, vol. 52, no. 2, pp. 197–211, 2007.
[11] P. Gahinet, P. Apkarian, and M. Chilali, “Affine parameter-dependent
Lyapunov functions and real parametric uncertainty,” IEEE Trans. on
Automatic Control, vol. 41, no. 3, pp. 436–442, 1996.
[12] M. Chilali, P. Gahinet, and P. Apkarian, “Robust pole placement in LMI
regions,” IEEE Trans. on Automatic Control, vol. 44, no. 12, pp. 2257–
2270, 1999.
[13] M. S. Mahmoud, “Discrete-time systems with linear parameter-varying:
stability vand H∞ ,” J. of Mathematical Analysis and Applications,
vol. 269, no. 1, pp. 369–381, 2002.
[14] M. C. de Oliveira, J. C. Geromel, and Liu Hsu, “LMI characterization
of structural and robust stability: the discrete-time case,” Linear Algebra
and its Applications, vol. 296, pp. 27–38, 1999.
[15] F. Blanchini and S. Miani, “A new class of universal Lyapunov functions
for the control of uncertain linear systems,” IEEE Trans. on Automatic
Control, vol. 44, no. 3, pp. 641–647, 1999.
[16] J. Daafouz and J. Bernussou, “Parameter dependent Lyapunov functions
for discrete time systems with time varying parametric uncertainties,”
Systems & Control Letters, vol. 43, no. 5, pp. 355–359, 2001.
[17] C. E. de Souza, K. A. Barbosa, and A. T. Neto, “Robust H∞ filtering
for discrete-time linear systems with uncertain time-varying parameters,”
IEEE Trans. on Signal Processing, vol. 54, no. 6, pp. 2110–2118, 2006.
[18] F. Amato, M. Mattei, and A. Pironti, “Gain scheduled control for
discrete-time systems depending on bounded rate parameters,” Int. J.
of Robust and Nonlinear Control, vol. 15, no. 11, pp. 473–494, 2005.
[19] T. Iwasaki and G. Shibata, “LPV system analysis via quadratic separator
for uncertain implicit systems,” IEEE Trans. on Automatic Control,
vol. 46, no. 8, pp. 1195–1208, 2001.
[20] D. Peaucelle, D. Arzelier, D. Henrion, and F. Gouaisbaut, “Quadratic
separation for feedback connection of an uncertain matrix and an implicit
linear transformation,” Automatica, vol. 43, pp. 795–804, 2007.
[21] C.-P. Wei and L. Lee, “Stability analysis of discrete LPV systems subject
to rate-bounded parameters,” in Proc. of the 17th IFAC World Congress,
(Seoul, Korea), pp. 6383–6388, Jul. 2008.
[22] M. C. de Oliveira and R. E. Skelton, “Stability tests for constrained
linear systems,” in Perspectives in Robust Control (S. Reza Moheimani,
ed.), ch. 15, pp. 241–257, Springer London, 2001.
[23] C. W. Scherer, “LMI relaxations in robust control,” European J. of
Control, vol. 12, pp. 3–29, 2006.
[24] J. Veenman, C. W. Scherer, and H. Köroğlu, “Robust stability and
performance analysis based on integral quadratic constraints,” European
J. of Control, vol. 31, pp. 1–32, 2016.
[25] H. K. Khalil, Nonlinear Systems. Prentice Hall, 3th ed., 2002.
[26] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory and
Design. Nob Hill Publishing, fifth ed., 2015.
[27] C. Scherer and S. Weiland, “Linear matrix inequalities in control.”
Lecture Notes for a course of the Dutch Institute of Systems and Control,
Mar. 2015.
[28] R. Tóth, Modeling and Identification of Linear Parameter-Varying
Systems. Springer, 2010.
[29] M. C. de Oliveira, J. Bernussou, and J. C. Geromel, “A new discretetime robust stability condition,” Systems & Control Letters, vol. 37, no. 4,
pp. 261–265, 1999.
[30] A. Trofino and C. E. de Souza, “Biquadratic stability for uncertain linear
systems,” IEEE Trans. on Automatic Control, vol. 46, no. 8, pp. 1303–
1307, 2001.
[31] G. Scorletti, V. Fromion, and S. de Hillerin, “Toward nonlinear tracking
and rejection using LPV control,” in Proc. of the 1st IFAC Workshop on
Linear Parameter Varying Systems, (Grenoble, France), pp. 13–18, Oct.
2015.
A PPENDIX
∗
For sufficiency, let p = (p∗1 , . . . , p∗np ) be the global
maximizer of L() over P. Assume that p∗i is not at a vertex
of the hyper-rectangle, i.e., pi < p∗i < pi , then
g(pi ) = L(p∗1 , . . . , p∗i−1 , pi , p∗i+1 , . . . , p∗np ) =
F0,i + pi F1,i + p2i F2,i + Qi,i,i p3i , (43)
where
F0,i = Q0 +
np
X
Qj p∗j +
j=1
j6=i
np
np
X
Qj,k p∗j p∗k +
j,k=1
j,k6=i
np
X
Qj,k,l p∗j p∗k p∗l ,
j,k,l=1
j,k,l6=i
X
(Qi,j + Qj,i )p∗j +
F1,i = Qi +
j=1
j6=i
np
X
(Qi,j,k + Qj,i,k + Qj,k,i )p∗j p∗k ,
j,k=1
j,k6=i
np
X
(Qj,i,i + Qi,j,i + Qi,i,j )p∗j .
F2,i = Qi,i +
j=1
j6=i
Since p∗ is the global maximizer of L(), it holds that
y⊤g(p∗i )y ≥ max(y⊤g(pi )y, y⊤g(pi )y),
(44)
where y 6= 0.
On the other hand, Condition (11) imposes convexity of
2
g(pi )
g(pi ) on [pi , pi ], i.e., 21 ∂ ∂p
= F2,i +3Qi,i,i pi 0 is implied
2
i
by (11). Therefore its maximum on that interval is on the edges
pi , pi . So,
y⊤g(p∗i )y ≤ max(y⊤g(pi )y, y⊤g(pi )y).
(45)
Combining (44) and (45) leads to
y⊤g(p∗i )y = max(y⊤g(pi )y, y⊤g(pi )y).
(46)
Concluding, the maximum of g() is obtained at the edges
{pi , pi } of pi in case (11) is satisfied. By repeating the
same argument for each i implies that the maximum of the
function L() is at a vertex of the hyper-rectangle P when
Condition (11) is satisfied.
| 3cs.SY
|
arXiv:1706.00382v1 [q-bio.NC] 1 Jun 2017
Blind nonnegative source separation using
biological neural networks
Cengiz Pehlevan1 , Sreyas Mohan1,2 , and Dmitri B. Chklovskii1,3
1
Center for Computational Biology, Flatiron Institute, New York,
NY
2
IIT Madras, Chennai, India
3
NYU Medical School, New York, NY
Abstract
Blind source separation, i.e. extraction of independent sources from
a mixture, is an important problem for both artificial and natural signal
processing. Here, we address a special case of this problem when sources
(but not the mixing matrix) are known to be nonnegative, for example,
due to the physical nature of the sources. We search for the solution to
this problem that can be implemented using biologically plausible neural
networks. Specifically, we consider the online setting where the dataset
is streamed to a neural network. The novelty of our approach is that we
formulate blind nonnegative source separation as a similarity matching
problem and derive neural networks from the similarity matching objective.
Importantly, synaptic weights in our networks are updated according to
biologically plausible local learning rules.
1
Introduction
Extraction of latent causes, or sources, from complex stimuli is essential for making
sense of the world. Such stimuli could be mixtures of sounds, mixtures of odors,
or natural images. If supervision, or ground truth, about the causes is lacking, the
problem is known as blind source separation.
The blind source separation problem can be solved by assuming a generative
model, wherein the observed stimuli are linear combinations of independent sources,
an approach known as Independent Component Analysis (ICA) (Jutten and Herault,
1991; Comon, 1994; Bell and Sejnowski, 1995; Hyvärinen and Oja, 2000). Formally,
1
the stimulus at time t is expressed as a k-component vector
xt = Ast ,
(1)
where A is an unknown but time-independent k ×d mixing matrix and st represents
the signals of d sources at time t. In this paper we assume that k ≥ d.
The goal of ICA is to infer source signals, st , from the stimuli xt . Whereas
many ICA algorithms have been developed by the signal processing community
(Comon and Jutten, 2010), most of them cannot be implemented by biologically
plausible neural networks. Yet, our brains can solve the blind source separation
problem effortlessly (Bronkhorst, 2000; Asari et al., 2006; Narayan et al., 2007;
Bee and Micheyl, 2008; McDermott, 2009; Mesgarani and Chang, 2012; Golumbic
et al., 2013; Isomura et al., 2015). Therefore, discovering a biologically plausible
ICA algorithm is an important problem.
For an algorithm to be implementable by biological neural networks it must
satisfy (at least) the following requirements. First, it must operate in the online
(or streaming) setting. In other words, the input dataset is not available as a
whole but is streamed one data vector at a time and the corresponding output
must be computed before the next data vector arrives. Second, the output of most
neurons in the brain (either a firing rate, or the synaptic vesicle release rate) is
nonnegative. Third, the weights of synapses in a neural network must be updated
using local learning rules, i.e. depend on the activity of only the corresponding
pre- and postsynaptic neurons.
Given the nonnegative nature of neuronal output we consider a special case
of ICA where sources are assumed to be nonnegative, termed Nonnegative Independent Component Analysis (NICA), (Plumbley, 2001, 2002). Of course, to
recover the sources, one can use standard ICA algorithms that don’t rely on the
nonnegativity of sources, such as fastICA (Hyvärinen and Oja, 1997; Hyvarinen,
1999; Hyvärinen and Oja, 2000). Neural learning rules have been proposed for
ICA, e.g. (Linsker, 1997; Eagleman et al., 2001; Isomura and Toyoizumi, 2016)
and references within. However, taking into account nonnegativity may lead to
simpler and more efficient algorithms (Plumbley, 2001, 2003; Plumbley and Oja,
2004; Oja and Plumbley, 2004; Yuan and Oja, 2004; Zheng et al., 2006; Ouedraogo
et al., 2010; Li et al., 2016).
While most of the existing NICA algorithms have not met the biological
plausibility requirements, in terms of online setting and local learning rules, there
are two notable exceptions. First, Plumbley (2001) succesfully simulated a neural
network on a small dataset, yet no theoretical analysis was given. Second, Plumbley
(2003) and Plumbley and Oja (2004) proposed a nonnegative PCA algorithm for a
streaming setting, however its neural implementation requires nonlocal learning
rules. Further, this algorithm requires prewhitened data (see also below), yet no
streaming whitening algorithm was given.
2
Here, we propose a biologically plausible NICA algorithm. The novelty of our
approach is that the algorithm is derived from the similarity matching principle
which postulates that neural circuits map more similar inputs to more similar
outputs (Pehlevan et al., 2015). Previous work proposed various objective functions
to find similarity matching neural representations and solved these optimization
problems with biologically plausible neural networks (Pehlevan et al., 2015; Pehlevan and Chklovskii, 2015a, 2014; Hu et al., 2014; Pehlevan and Chklovskii, 2015b).
Here we apply these networks to NICA.
The rest of the paper is organized as follows: In Section 2, we show that
blind source separation, after a generalized prewhitening step, can be posed as a
nonnegative similarity matching (NSM) problem (Pehlevan and Chklovskii, 2014).
In Section 3, using results from (Pehlevan and Chklovskii, 2015a, 2014) we show
that both the generalized prewhitening step and the NSM step can be solved online
by neural networks with local learning rules. Stacking these two networks leads to
the two-layer NICA network. In Section 4, we compare the performance of our
algorithm to other ICA and NICA algorithms for various datasets.
2
Offline NICA via NSM
In this section, we first review Plumbley’s analysis of NICA and then reformulate
NICA as an NSM problem.
2.1
Review of Plumbley’s analysis
When source signals are nonnegative, the source separation problem simplifies.
It can be solved in two straightforward steps: noncentered prewhitening and
orthonormal rotation (Plumbley, 2002).
d
Noncentered prewhitening transforms x to h :=
D Fx, where h ∈ R E and F
>
1
is a d × k whitening matrix , such that Ch := (h − hhi) (h − hhi)
= Id ,
where angled brackets denote an average over the source distribution and Id
is the d × d identity matrix. Note that the mean of x is not removed in the
tranformation, otherwise one would not be able to use the constraint that the
sources are nonnegative (Plumbley, 2003).
Assuming that sources have unit variance2 ,
D
E
Cs := (s − hsi) (s − hsi)> = Id ,
1
(2)
In his analysis Plumbley (Plumbley, 2002) assumed k = d (mixture channels are the same as
source channels) but this assumption can be relaxed as shown.
2
Without loss of generality, a scalar factor that multiplies a source can always be absorbed
into the corresponding column of the mixing matrix
3
Mixtures
Sources
s2
4
x2
2
6
4
h2
2
0
0
-6
Prewhitened
6
6
-4
-2
0
s1
2
4
6
-6
4
2
0
-4
-2
0
x1
2
4
6
-6
-4
-2
0
h1
2
4
Figure 1: Illustration of the NICA algorithm. Two source channels (left) are
linearly transformed to a two-dimensional mixture (middle). Prewhitening (right)
gives an orthogonal transformation of the sources. Sources are then recovered by
solving the NSM problem (7). Green and red plus signs track two source vectors
across mixing and prewhitening stages.
the combined effect of source mixing and prewhitening FA (h = Fx = FAs) is an
orthonormal rotation. To see this, note that, by definition, Ch = (FA) Cs (FA)> =
(FA) (FA)> and Ch = Id .
The second step of NICA relies on the following observation (Plumbley, 2002):
Theorem 1 (Plumbley). Suppose sources are independent, nonnegative and wellgrounded, i.e. Prob(si < δ) > 0 for any δ > 0. Consider an orthonormal transformation y = Qs. Then Q is a permutation matrix with probability 1, if and only if
y is nonnegative.
In the second step, we look for an orthonormal Q such that y = Qh is nonnegative. When found, Plumbley’s theorem guarantees that QFx is a permutation
of the sources. Several algorithms have been developed based on this observation
(Plumbley, 2003; Plumbley and Oja, 2004; Oja and Plumbley, 2004; Yuan and Oja,
2004).
Note that only the sources s but not necessarily the mixing matrix A must
be nonnegative. Therefore, NICA allows generative models, where features not
only add up but also cancel each other, as in the presence of a shadow in an image
(Plumbley, 2002). In this respect, NICA is more general than Nonnegative Matrix
Factorization (NMF) (Lee and Seung, 1999; Paatero and Tapper, 1994) where
both the sources and the mixing matrix are required to be nonnegative.
2.2
NICA as NSM
Next we reformulate NICA as a NSM problem. This reformulation will allow us to
derive an online neural network for NICA in Section 3. Our main departure from
4
6
Plumbley’s analysis is to work with similarity matrices rather than covariance
matrices and finite number of samples rather than the full probability distribution
of the sources.
First, let us switch to the matrix notation, where data matrices are formed by
concatenating data column vectors, e.g. X = [x1 , x2 , ..., xt ] so that X ∈ Rk×t , and
S = [s1 , s2 , ..., st ] so that S ∈ Rd×t . In this notation, we introduce a time-centering
operation δ such that, for example, time-centered stimuli are δX := X − X̄ where
X̄ := X 1t 11> and 1 is a t dimensional column vector whose elements are all 1’s.
Our goal is to recover S from X = AS, where A is unknown. We make
the following two assumptions: First, sources are nonnegative and decorrelated,
1
δS δS> = Id . Note that while general ICA and NICA problems are stated with
t
the independence assumption on the sources, for our purposes, it is sufficient that
they are only decorrelated. Second, the mixing matrix, A ∈ Rk×d (k ≥ d), is
full-rank.
We propose that the source matrix, S, can be recovered from X in the following
two steps, also illustrated in Fig. 1:
1. Generalized Prewhitening: Transform X to H = FX, where F is l × k with
l ≥ d, so that 1t δH δH> has d unit eigenvalues and l − d zero eigenvalues. When
l = d, H is whitened, otherwise channels of H are correlated. Such prewhitening
is useful because it implies H> H = S> S according to the following theorem.
Theorem 2. If F ∈ Rl×k (l ≥ k) is such that H = FX obeys
1
>
δH δH> = UH ΛH UH ,
t
(3)
an eigenvalue decomposition with ΛH = diag 1, . . . , 1, 0, . . . , 0 , then
|
{z
} |
d
{z
l−d
}
H> H = S> S.
(4)
Proof. To see why (3) is sufficient, first note that 1t δH δH> = (FA) (FA)> .
This follows from the definition of H and (2). If (3) holds, then
>
(FA) (FA)> = UH ΛH UH .
(5)
In turn, this is sufficient to prove that (FA)> (FA) = Id . To see that, assume an
>
>
SVD decomposition of (FA) = UF A ΛF A VF A . (5) implies that ΛF A ΛF A =
ΛH , i.e. that the d diagonal elements of ΛF A ∈ Rl×d are all 1’s. Then,
(FA)> (FA) = Id .
This gives us the desired results H> H = S> (FA)> (FA) S = S> S.
5
(6)
Remark 1. If l > d, the channels of H are correlated, except in the special case
UH = Id . The whitening used in Plumbley’s analysis (Plumbley, 2002) requires
l = d.
2. NSM: Solve the following optimization problem:
Y∗ = arg min H> H − Y> Y
Y, Y≥0
2
F
(7)
,
where the optimization is performed over nonnegative Y := [y1 , . . . , yt ] i.e.
yi ∈ Rd+ . We call (7) the NSM cost function (Pehlevan and Chklovskii, 2014).
Because inner products quantify similarities we call H> H and Y> Y input and
output similarity matrices, i.e. their elements hold the pairwise similarities
between input and the pairwise similarities between output vectors, respectively.
Then, the cost function (7) preserves the input similarity structure as much as
possible under the nonnegativity constraint. Variants of (7) has been considered
in applied math literature under the name “nonnegative symmetric matrix
factorization” for clustering applications, e.g. (Kuang et al., 2012, 2015).
Now we make our key observation. Using Theorem 2, we can rewrite (7) as
Y∗ = arg min S> S − Y> Y
Y, Y≥0
2
F
(8)
.
Since both S and Y are nonnegative, rank-d matrices, Y∗ = PS, where P is a
permutation matrix, is a solution to this optimization problem and the sources
are successfully recovered.
Uniqueness of the solutions (up to permutations) is hard to establish. While
both sufficient conditions, and necessary and sufficient conditions for uniqueness
exist, these are non-trivial to verify and usually the verification is NP-complete
(Donoho and Stodden, 2003; Laurberg et al., 2008; Huang et al., 2014). A
review of related uniqueness results can be found in (Huang et al., 2014). A
necessary condition for uniqueness given in (Huang et al., 2014) states that, if
the factorization of S> S to Y> Y is unique (up to permutations), then each row
of S contains at least one element that is equal to 0. This necessary condition is
similar to Plumbley’s well-groundedness requirement used in proving Theorem
1.
The NSM problem can be solved by projected gradient descent,
Y ←− max Y + η YH> H − YY> Y , 0 ,
(9)
where the max operation is applied elementwise, and η is the size of the gradient
step. Other algorithms can be found in (Kuang et al., 2012, 2015; Huang et al.,
2014).
6
3
Derivation of NICA neural networks from similarity matching objectives
Our analysis in the previous section revealed that the NICA problem can be
solved in two steps: generalized prewhitening and nonnegative similarity matching.
Here, we derive neural networks for each of these steps and stack them to give a
biologically plausible two-layer neural network that operates in a streaming setting.
In a departure from the previous section, the number of output channels is
reduced to the number of sources at the prewhitening stage rather than the later
NSM stage (l = d). This assumption simplifies our analysis significantly. The full
problem is addressed in Appendix B.
3.1
Noncentered prewhitening in a streaming input setting
To derive a neurally plausible online algorithm for prewhitening, we pose generalized
prewhitening, Eq. (3), as an optimization problem. Online minimization of this
optimization problem gives an algorithm that can be mapped to the operation of
a biologically plausible neural network.
Generalized prewhitening solves a constrained similarity alignment problem:
max Tr δX> δX δH> δH
δH
s.t. δH> δH tIt ,
(10)
where δX is the k × t centered mixture of d independent sources and δH is a
d × t matrix, constrained such that tIt − δH> δH is positive semidefinite. The
solution of this objective aligns similarity matrices δX> δX and δH> δH so that
their right singular vectors are the same (Pehlevan and Chklovskii, 2015a). Then,
the objective under the trace diagonalizes and its value is the sum of eigenvalue pair
products. Since the eigenvalues of δH> δH are upper bounded by t, the objective
(10) is maximized by setting the eigenvalues of δH> δH that pair with the top d
eigenvalues of δX> δX to t, and the rest to zero. Hence, the optimal δH satisfies
the generalized prewhitening condition (3)(Pehlevan and Chklovskii, 2015a). More
formally,
Theorem 3 (Modified from (Pehlevan and Chklovskii, 2015a)). Suppose an eigen>
decomposition of δX> δX is δX> δX = VX ΛX VX , where eigenvalues are sorted
in decreasing order. Then, all optimal δH of (10) have an SVD decomposition of
the form
√
0
>
δH∗ = UH t ΛH VX ,
(11)
0
where ΛH is d × t with d ones on top of the diagonal and zeros on the rest of the
diagonal.
7
The theorem implies that, first, 1t δH∗ δH∗ > = Id , and hence δH satisfies the
generalized prewhitening condition (3). Second, F, the linear mapping between
δH∗ and δX, can be constructed using an SVD decomposition of δX and (11).
The constraint in (10) can be introduced into the objective function using as a
Lagrange multiplier the Grassmanian of matrix δG ∈ Rm×t with (m ≥ d):
min max Tr −δX> δX δH> δH + δG> δG δH> δH − tIt
δH
δG
(12)
,
This optimization problem (Pehlevan and Chklovskii, 2015a) will be used to derive
an online neural algorithm.
Whereas the optimization problem (12) is formulated in the offline setting,
i.e. outputs are computed only after receiving all inputs, to derive a biologically
plausible algorithm, we need to formulate the optimization problem in the online
setting, i.e. the algorithm receives inputs sequentially, one at a time, and computes
an output before the next input arrives. Therefore, we optimize (12) only for the
data already received and only with respect to the current output:
{δht , δgt } ←− arg min arg max Tr −δX> δX δH> δH + δH> δH δG> δG − t δG> δG .
gt
δht
(13)
By keeping only those terms that depend on δht or δgt , we get the following
objective:
{δht , δgt } ←− arg min arg max L(δht , δgt ),
(14)
δgt
δht
where
L=
−2δxt>
t−1
X
!
δxt0 δht>0
δht −
t kδgt k22
+
2δgt>
t0 =1
t−1
X
!
δgt0 δht>0
δht
t0 =1
+ kδgt k22 − kδxt k22 kδht k22 .
(15)
In the large-t limit, the first three terms dominate over the last term, which we
ignore. The remaining objective is strictly concave in δgt and convex in δht . We
assume that the matrix
1
t
t−1
P
t0 =1
δht0 δgt>0 is full-rank. Then, the objective has a unique
saddle point :
δgt∗ = WtGH δht∗ ,
δht∗ = WtHG WtGH
8
−1
WtHX δxt ,
(16)
where,
WtHX :=
t−1
1X
δht0 δxt>0 ,
t t0 =1
WtGH
>
WtHG
WtHG :=
t−1
1X
δht0 δgt>0 ,
t t0 =1
0
:=
−1
1 tX
:=
δgt0 δht>0 .
t t0 =1
(17)
−1
WtHX can be interpreted as the current estimate of
Hence, Ft := WtHG WtGH
the prewhitening matrix, F.
We solve (14) with a gradient descent-ascent
dδht
1
= − ∇δht L = WtHX δxt − WtHG gt ,
dγ
2t
1
dδgt
= ∇δgt L = −δgt + WtGH δht .
dγ
2t
(18)
where γ measures “time” within a single time step of t. Biologically, this is justified
if the activity dynamics converges faster than the correlation time of the input
data. The dynamics (18) can be proved to converge to the saddle point of the
objective (15), see Appendix A.
Equation (18) describes the dynamics of a single-layer neural network with
two-populations, Fig. 2. WtHX represents the weights of feedforward synaptic
connections, WtHG and WtGH represent the weights of synaptic connections between
the two populations. Remarkably, synaptic weights appear in the online algorithm
despite their absence in the optimization problem formulations (12) and (13).
Furthermore, δht neurons can be associated with principal neurons of a biological
circuit and δgt neurons with interneurons.
Finally, using the definition of synaptic weight matrices (17), we can formulate
recursive update rules:
1
δht δxt> − WtHX ,
t+1
1
= WtHG +
δht δgt> − WtHG ,
t+1
1
GH
= Wt +
δgt δht> − WtGH .
t+1
HX
Wt+1
= WtHX +
HG
Wt+1
GH
Wt+1
(19)
Equations (18) and (19) define a neural algorithm that proceeds in two phases.
After each stimulus presentation, first, (18) is iterated until convergence by the
dynamics of neuronal activities. Second, synaptic weights are updated according
to local, anti-Hebbian (for synapses from interneurons) and Hebbian (for all
other synapses) rules (19). Biologically, synaptic weights are updated on a slower
timescale than neuronal activity dynamics.
9
Mixing
Prewhitening
{
s1
{
{
x1
h1
x2
s2
NSM
y1
g1
sd
...
...
A
gm
xk
Principal
yd
hd
Hebbian
Inter-neurons
anti-Hebbian synapses
Figure 2: Biologically plausible network for blind source separation. The prewhitening stage is composed of two populations of linear neurons. The NSM stage has a
single population of rectifying neurons.
Our algorithm can be viewed as a special case of the algorithm proposed in
(Plumbley, 1996, 1994). Plumbley analyzed the convergence of synaptic weights
(Plumbley, 1994) in a stochastic setting by a linear stability analysis of the
stationary point of synaptic weight updates. His results are directly applicable to
our algorithm, and show that, if the synaptic weights of our algorithm converge to
a stationary state, they whiten the input.
Importantly, unlike (Plumbley, 1996, 1994) which proposed the algorithm
heuristically, we derived it by posing and solving an optimization problem.
3.1.1
Computing H̄
The optimization problem (12) and the corresponding neural algorithm, Eqs. (18)
and (19) almost achieve what is needed for noncentered prewhitening, but we still
need to find H̄, since for the NSM step we need H = δH + H̄. We now discuss
how H̄ can be learned along with δH using the same online algorithm.
Our online algorithm for centered-whitening is of the following form. First, a
neural dynamics stage outputs a linear transformation of the input:
δht = Ft δxt ,
(20)
and, second, synaptic weights and, hence, Ft are updated:
Ft+1 = Ft + δFt (t, δht , δgt , Ft ).
(21)
We can supplement this algorithm with a running estimate of h̄. Let the running
10
estimate of average stimulus activity be
t
1
1
1X
xt0 = 1 −
x̄t−1 + xt .
x̄t =
t t0 =1
t
t
(22)
h̄t = Ft x̄t .
(23)
Then,
Alternatively, (20) and (23) can be combined into a single step:
ht = Ft xt ,
(24)
where the network receives uncentered stimuli and prewhitenes it. Note that
assignment (24) can still be done by iterating (18), except now the input is xt
rather than δxt . However, synaptic weights are still updated using δxt = xt − x̄t ,
δht = ht − h̄t and δgt = gt − ḡt . Therefore we keep recursive estimates of the
means. Substituting (22) into (24) and using (21)
1
1
1
h̄t−1 + 1 −
h̄t = 1 −
δFt−1 x̄t−1 + ht .
t
t
t
(25)
The term 1 − 1t δFt−1 x̄t−1 requires non-local calculations. Assuming that in the
large-t limit updates to F are small, we can ignore this term and obtain a recursion:
1
1
h̄t = 1 −
h̄t−1 + ht .
t
t
(26)
Finally, a similar argument can be given for ḡt . We keep a recursive estimate of ḡt :
ḡt = 1 −
1
1
ḡt−1 + gt .
t
t
(27)
To summarize, when a new stimulus xt is observed, the algorithm operates in
two steps. In the first step, the following two-population neural dynamics runs
until convergence to a fixed point:
dht
= WtHX xt − WtHG gt ,
dγ
dgt
= −gt + WtGH ht ,
dγ
(28)
The convergence proof for neural dynamics (18) given in Appendix A also applies
here. Besides the synaptic weight, each neuron remembers its own average activity
11
and each synapse remembers average incoming activity. In the second step of the
algorithm, the average activities are updated by:
1
1
x̄t−1 + xt ,
t
t
1
1
h̄t = 1 −
h̄t−1 + ht ,
t
t
1
1
ḡt = 1 −
ḡt−1 + gt .
t
t
Synaptic weight matrices are updated recursively by
x̄t = 1 −
(29)
1
δht δxt> − WtHX ,
t+1
1
HG
Wt+1
= WtHG +
δht δgt> − WtHG ,
t+1
1
GH
Wt+1
δgt δht> − WtGH .
= WtGH +
(30)
t+1
Once the synaptic updates are done, the new stimulus, xt+1 , is processed. We
note again that all the synaptic update rules are local, and hence are biologically
plausible.
HX
= WtHX +
Wt+1
3.2
Online NSM
Next, we derive the second-layer network which solves the NSM optimization
problem (7) in an online setting (Pehlevan and Chklovskii, 2014).
The online optimization problem is:
yt ←− arg min H> H − Y> Y
yt , yt ≥0
2
F
(31)
.
Proceeding as before, let’s rewrite (31) keeping only terms that depend on yt :
yt ←− arg min
yt , yt ≥0
2yt>
t−1
X
!
yt0 yt>0
yt −
4ht>
t0 =1
t−1
X
!
ht0 yt>0
!
yt +
kyt k42
−
2 kht k22
kyt k22
t0 =1
(32)
In the large-t limit, the last two terms can be ignored and the remainder is a
quadratic form in yt . We minimize it using coordinate descent (Wright, 2015)
which is both fast and neurally plausible. In this approach, neurons are updated
one-by-one by performing an exact minimization of (32) with respect to yt,i until
convergence:
yt,i ←−
X
max W Y H ht,j
t,ij
j
12
YY
− Wt,ij
yt,j , 0 ,
(33)
.
where
YH
Wt,ij
Pt−1
=
t0 =1 yt0 ,i ht0 ,j
,
Pt−1 2
t0 =1 yt0 ,i
YY
Wt,i,j6
=i
Pt−1
yt0 ,i yt0 ,j
,
2
t0 =1 yt0 ,i
t0 =1
=
Pt−1
YY
= 0.
Wt,ii
(34)
For the next time step (t + 1), we can update the synaptic weights recursively,
giving us the following local learning rules:
2
Dt+1,i = Dt,i + yt,i
,
YH
YH
Wt+1,ij
= Wt,ij
+
YY
YY
Wt+1,i,j6
=i = Wt,ij
1
2
YH
yt,i ht,j − yt,i
Wt,ij
,
Dt+1,i
1
2
YY
+
yt,i yt,j − yt,i
Wt,ij
,
Dt+1,i
YY
Wt,ii
= 0.
(35)
Interestingly, these update rules are local and are identical to the single-neuron
Oja rule (Oja, 1982), except that the learning rate is given explicitly in terms of
cumulative activity 1/Dt,i and the lateral connections are anti-Hebbian.
After the arrival of each data vector, the operation of the complete two-layer
network algorithm, Fig. 2, is as follows. First, the dynamics of the prewhitening
network runs until convergence. Then the output of the prewhitening network is
fed to the NSM network, and the NSM network dynamics runs until convergence
to a fixed point. Synaptic weights are updated in both networks for processing the
next data vector.
3.2.1
NICA is a stationary state of online NSM
Here we show that the solution to the NICA problem is a stationary synaptic
weights state of the online NSM algorithm. In the stationary state the expected
updates to synaptic weights are zero, i.e.
D
E
D
∆WijY H = 0,
E
∆WijY Y = 0,
(36)
where we dropped the t index, and brackets denote averages over the source
distribution.
Suppose the stimuli obey the NICA generative model, Eq. (1), and the
observed mixture, xt , is whitened with the exact (generalized) prewhitening
matrix F described in Theorem 2. Then, input to the network at time, t, is
ht = Fxt = FAst . Our claim is that there exists synaptic weight configurations for
which 1) for any mixed input, xt , the output of the network is the source vector,
i.e. yt = Pst , where P is a permutation matrix, and 2) this synaptic configuration
is a stationary state.
13
We prove our claim by constructing these synaptic weights. For each permutation matrix, we first relabel the outputs such that ith output recovers the ith
source and hence P becomes the identity matrix. Then, the weights are:
WijY Y =
hsi i hsj i
,
hs2i i
WijY H =
hsi hj i X
hsk i hsi i
= (FA)jk
+ (FA)ji .
2
hsi i
hs2i i
k6=i
(37)
Given mixture xt , NSM neural dynamics with these weights converge to yt,i = st,i ,
which is the the unique fixed point3 . Furthermore, these weights define a stationary
state as defined in (36) assuming a fixed learning rate. To see this substitute
weights from (37) into the last two equations of (35) and average over the source
distribution. The fixed learning rate assumption is valid in the large-t limit when
changes to Dt,i become small (O(1/t), see (Pehlevan et al., 2015)).
4
Numerical simulations
Here we present numerical simulations of our two-layered neural network using
various datasets and compare the results to that of other algorithms.
In all our simulations, d = k = l = m, except in Fig. 5B where d = k > l = m.
Our networks were initialized as follows:
1. In the prewhitening network, WHX and WHG were chosen to be random
>
orthonormal matrices. WGH is initialized as WHG because of its definition
in Eq. (17) and the fact that this choice guarantees the convergence of the
neural dynamics (28) (see Appendix A).
2. In the NSM network, WY H was initialized to a random orthonormal matrix
and WY Y was set to zero.
The learning rates were chosen as follows:
1. For the prewhitening network, we generalized the time-dependent learning
rate (30) to,
1
,
a + bt
(38)
and performed a grid search over a ∈ {10, 102 , 103 , 104 } and b ∈ {10−2 , 10−1 , 1}
to find the combination with best performance. Our performance measures
will be introduced below.
P
P
Proof: The net input to neuron i at the claimed fixed point is j WijY H ht,j − j6=i WijY Y st,j .
Plugging in (37) and ht = FAst , and using (6) one gets that the net input is st,i , which is
also the output since sources are nonnegative. This fixed point is unique and globally stable
because the NSM neural dynamics is a coordinate descent on a strictly convex cost given by
1 >
>
yt − ht> hs> yt .
2 yt ss
3
14
2. For the NSM network, we generalized the activity-dependent learning rate
(35) to,
1
D̃t+1,i
,
where
2
,
D̃t+1,i = min ã, b̃D̃t+1,i + yt,i
(39)
and performed a grid search over several values of ã ∈ {10, 102 , 103 , 104 } and
b̃ ∈ {0.8, 0.9, 0.95, 0.99, 0.995, 0.999, 0.9999, 1} to find the combination with
best performance. The b̃ parameter introduces “forgetting” to the system
(Pehlevan et al., 2015). We hypothesized that forgetting will be beneficial
in the two-layer setting because the prewhitening layer output changes over
time and the NSM layer has to adapt. Further, for comparison purposes,
we also implemented this algorithm with a time-dependent learning rate
of the form (38) and performed a grid search with a ∈ {102 , 103 , 104 } and
b ∈ {10−2 , 10−1 , 1} to find the combination with best performance.
For the NSM network, to speed up our simulations we implemented a procedure
from (Plumbley and Oja, 2004). At each iteration we checked whether there is
any output neuron who has not fired up until that iteration. If so, we flipped the
sign of its feedforward inputs. In practice, the flipping only occured within the
first ∼10 iterations.
For comparison, we implemented five other algorithms. First is the offline
algorithm (9), the other two are chosen to represent major online algorithm classes:
1. Offline projected gradient descent: We simulated the projected gradient
descent algorithm (9). We used variable stepsizes of the form (38) and
performed a grid search with a ∈ {104 , 105 , 106 } and b ∈ {10−3 , 10−2 , 10−1 }
to find the combination with best performance. We initialized elements of
the matrix, Y, by drawing a Gaussian random variable with zero mean and
unit variance and rectifying it. Input dataset was whitened offline before
passing to projected gradient descent.
2. fastICA: fastICA (Hyvärinen and Oja, 1997; Hyvarinen, 1999; Hyvärinen
and Oja, 2000) is a popular ICA algorithm which does not assume nonnegativity of sources. We implemented an online version of fastICA (Hyvärinen
and Oja, 1998) using the same parameters except for feedforward weights. We
used the time-dependent learning rate (38) and performed a grid search with
a ∈ {10, 102 , 103 , 104 } and b ∈ {10−2 , 10−1 , 1} to find the combination with
best performance. fastICA requires whitened and centered input (Hyvärinen
and Oja, 1998) and computes a decoding matrix that maps mixtures back
to sources. We ran the algorithm with whitened and centered input. To
recover nonnegative sources, we applied the decoding matrix to noncentered
but whitened input.
15
3. Infomax ICA: Bell and Sejnowski (1995) proposed a blind source separation
algorithm that maximizes the mutual information between inputs and outputs,
namely the Infomax principle (Linsker, 1988). We simulated an online
version due to Amari et al. (1996). We chose cubic neural nonlinearities
compatible with sub-Gaussian input sources. This differs from our fastICA
implementation where the nonlinearity is also learned online. Infomax ICA
computes a decoding matrix using centered, but not whitened, data. To
recover nonnegative sources, we applied the decoding matrix to noncentered
inputs. Finally, we rescaled the sources so that their variance is 1. We
experimented with several learning rate parameters for finding optimal
performance.
4. Linsker’s network: Linsker (1997) proposed a neural network with local
learning rules for Infomax ICA. We simulated this algorithm with cubic
neural nonlinearities and preprocessing and decoding done as in our Infomax
ICA implementation.
5. Nonnegative PCA: Nonnegative PCA algorithm (Plumbley and Oja, 2004)
solves the NICA task and makes explicit use of the nonnegativity of sources.
We use the online version given in (Plumbley and Oja, 2004). To speed
up our simulations we implemented a procedure from (Plumbley and Oja,
2004). At each iteration we checked whether there is any output neuron
who has not fired up until that iteration. If so, we flipped the sign of its
feedforward inputs. For this algorithm, we again used the time-dependent
learning rate of (38) and performed a grid search with a ∈ {10, 102 , 103 , 104 }
and b ∈ {10−2 , 10−1 , 1} to find the combination with best performance.
Nonnegative PCA assumes whitened, but not centered input (Plumbley and
Oja, 2004).
Next, we present the results of our simulations on three datasets.
4.1
Mixture of random uniform sources
The source i.i.d. samplesq were set to zero with probability 0.5 and sampled
uniformly from iterval [0, 48/5] with probability 0.5. The dimensions of source
vectors were d = {3, 5, 7, 10}. The mixing matrices are given in Appendix C. 105
source vectors were generated for each run. For a sample of the original and mixed
signals, see Fig 3A.
The inputs to fastICA and Nonnegative PCA algorithms were prewhitened
offline, and in the case of fastICA they were also centered. We ran our NSM
network both as a single layer algorithm with prewhitening done offline, and as a
part of our two-layer algorithm with whitening done online.
16
A
Sources
Mixtures
4
2
2
1
0
4
0
2
2
0
0
4
-2
4
2
2
0
99980
Sample Number
Error
2
0
4
2
0
-2
2
100000 99980
10
10 0
10 -1
10 -1
10 -2
10 -2
10
-3
10
-3
10
-4
10
-4
10
-5
10
-5
10 2
10 4
10 5
d=7
10 2
10
10 3
10 0
10
10 0
10 -1
10 -1
10 -2
10 -2
10
-3
10
-3
10
-4
10
-4
10
-5
10
-5
10 1
10 2
10 3
10 4
Sample # or Gradient Step
10 5
10 2
10 3
10 4
10 5
10 4
10 5
d=10
1
10 0
10 0
10 1
10 2
1
100000
1
10 0
10 1
99990
Sample Number
d=5
10 2
1
10 0
0
100000 99980
99990
Sample Number
d=3
10 2
Error
0
4
Sample Number
B
10
2
2
0
-2
99990
Recovered
4
2
0
-2
0
100000 99980
99990
Whitened
10 0
Online fastICA
Infomax ICA
Linsker’s network
Nonnegative PCA
NSM (time-dependent)
NSM (activity-dependent)
Two layer network
Offline projected gradient descent
10 1
10 2
10 3
Sample # or Gradient Step
Figure 3: Performance of our algorithms when presented with a mixture of random
uniform sources: A) Sample source, mixture, whitened and recovered signals for
a d = 3 task, performed by our two-layer algorithm. The whitened signal is the
output of the first layer, while the recovered signal is the output of second layer.
17
B) Performance comparison of the online algorithms presented in this paper with
projected gradient descent, Online fastICA, Infomax ICA, Linsker’s network and
Nonnegative PCA. Curves show averages over 10 simulations. Error bars not
shown for visual clarity. Learning rate parameters are given in Appendix D.
To quantify the performance of tested algorithms, we used the mean-squarederror:
Et =
t
1X
kst0 − Pyt0 k22 ,
t t0 =1
(40)
where P is a permutation matrix that is chosen to minimize the mean-squared-error
at t = 105 . The learning rate parameters of all networks were optimized by a grid
search using E105 as the performance metric.
In Fig. 3B, we show the performances of all algorithms we implemented. Our
algorithms perform as well or better than others, especially as dimensionality of
the input increases. Offline whitening is better than online whitening, however,
as dimensionality increases, online whitening becomes competitive with offline
whitening. In fact, our two-layer and single-layer networks perform better than
Online fastICA and Nonnegative PCA for which whitening was done offline.
We also simulated a fully offline algorithm by taking projected gradient descent
steps until the residual error plateaued (Fig. 3B). The performance of the offline
algorithm quantifies two important metrics. First, it establishes the loss in performance due to online (as opposed to offline) processing. Second, it establishes
the lowest error that could be achieved by the NSM method for the given dataset.
The lowest error is not necessarily zero due to the finite size of the dataset. This
method is not perfect because the projected gradient descent may get stuck in a
local minimum of Eq. (7).
We also tested whether the learned synaptic weights of our network match our
theoretical predictions. In Fig. 4A, we show examples of learned feedforward and
recurrent synaptic weights at t = 105 , and what is expected from our theory (37).
We observed an almost perfect match between the two. In Fig. 4B, we quantify the
convergence of simulated synaptic weights to the theoretical prediction by plotting
a normalized error metric defined by Et = kWt,simulation − Wtheory k2F / kWtheory k2F .
4.2
Mixture of random uniform and exponential sources
Our algorithm can demix sources sampled from different statistical distributions.
To illustrate this point, we generated a dataset with two uniform and three
exponential source channels. The uniform sources were sampled as before. The
exponential sources were either zero (with probability 0.5) or sampled from an
exponential distribution, scaled so that the variance of the channel is 1. In Fig.
5A, we show that the algorithm succesfully recovers sources.
To test denoising capabilities of our algorithm, we created a dataset where
source signals are accompanied by background noise. Sources to be recovered were
three exponential channels, which were sampled as before. Background noises
were two uniform channels which were sampled as before, except scaled to have
18
A
WYH
WYY
Theory
(Eq. (36))
0
0.3742 0.3742
0.3742
0
0.3742
0.3742 0.3742
0
Simulation
(t = 105)
1.1005
0.6545
0.8473
−0.2526 −0.0202 1.1012
−0.6795 0.6217 0.6637
0.1800
0.7107 0.8583
−0.2558 −0.0227
−0.6768 0.6228
0.1845
0.7179
B
0
0.3742 0.3702
0.3670
0
0.3736
0.3604 0.3709
0
WYH
WYY
10 0
Error
Error
10 0
10 -1
10 -2
10 0
d=3
d=5
d=7
d=10
10 1
10 -1
10 -2
10 2
10 3
10 4
10 5
Sample #
10 0
10 1
10 2
10 3
10 4
Sample #
Figure 4: Theoretical predictions of learned synaptic weights match simulations:
A) Example synaptic weight matrices predicted from theory (37) compared to
results from an example simulation (t = 105 ). B) Convergence of simulated network
synaptic weights to theoretical predictions. For this figure, inputs were whitened
offline and the NSM network was run with time-dependent learning rates. Shaded
bars show standard error over 10 simulations.
19
10 5
A
10
B
1
10
1
Error
10 0
10 0
10 -1
10 -2
10 -1
10
-3
10 0
10 1
10 2
10 3
10 4
10 5
Sample #
10 0
10 1
10 2
10 3
10 4
Sample #
Figure 5: Performance of our two layer algorithm when presented with a mixture
of random uniform and exponential sources. A) Recovery of a mixture of three
exponential and two uniform sources. B) Recovery of three exponential sources
corrupted by two background noise channels. Learning rate parameters are given
in Appendix D.
variance 0.1. To denoise the resulting five dimensional mixture, the prewhitening
layer reduced its five input dimensions to three. Then, the NSM layer succesfully
recovered sources, Fig. 5B. Hence, the prewhitening layer can act as a denoising
stage.
4.3
Mixture of natural scenes
Next, we consider recovering images from their mixtures, Fig. 6A, where each
image is treated as one source. Four image patches of size 252 × 252 pixels were
chosen from a set of images of natural scenes which were previously used in
(Hyvärinen and Hoyer, 2000; Plumbley and Oja, 2004). The preprocessing was
as in (Plumbley and Oja, 2004): 1) Images were downsampled by a factor of 4
to obtain 63 × 63 patches, 2) Pixel intensities were shifted to have a minimum
of zero and 3) Pixel intensities were scaled to have unit variance. Hence, in this
dataset, there are d = 4 sources, corresponding image patches, and a total of
63 × 63 = 3969 samples. These samples were presented to the algorithm 5000 times
with randomly permuted order in each presentation. The 4 × 4 mixing matrix,
which was generated randomly, is given in Appendix C.
In Fig. 6B, we show the performances of all algorithms we implemented in this
task. We see that our algorithms, when compared to fastICA and Nonnegative
20
10 5
A
Source Images
Mixtures
Recovered Images
B
Online fastICA
Nonnegative PCA
NSM (activity-dependent)
10 2
10
Two layer network
1
10 0
10 -1
10 0
10 1
10 2
10 3
10 4
10 5
10 6
10 7
Sample #
Figure 6: Performance of our algorithm when presented with a mixture of natural
images: A) Sample source, mixture, and recovered images, performed by our
two-layer algorithm. B) Performance comparison of our online algorithms with
Online fastICA and Nonnegative PCA. Shaded bars show standard error over 10
simulations. Learning rate parameters are listed in Appendix D.
PCA, perform much better.
5
Discussion
In this paper we presented a new neural algorithm for blind nonnegative source
separation. We started by assuming the nonnegative ICA generative model
(Plumbley, 2001, 2002) where inputs are linear mixtures of independent and
nonnegative sources. We showed that the sources can be recovered from inputs by
two sequential steps, 1) generalized whitening and 2) NSM. In fact, our argument
requires sources to be only uncorrelated, and not necessarily independent. Each
of the two steps can be performed online with single-layer neural networks with
local learning rules (Pehlevan and Chklovskii, 2014, 2015a). Stacking these two
networks yields a two-layer neural network for blind nonnegative source separation
(Fig. 2). Numerical simulations show that our neural network algorithm performs
21
10 8
well.
Because our network is derived from optimization principles, its biologically
realistic features can be given meaning. The network is multi-layered, because
each layer performs a different optimization. Lateral connections create competition between principal neurons forcing them to differentiate their outputs.
Interneurons clamp the activity dimensions of principal neurons (Pehlevan and
Chklovskii, 2015a). Rectifying neural nonlinearity is related to nonnegativity of
sources. Synaptic plasticity (Malenka and Bear, 2004), implemented by local
Hebbian and anti-Hebbian learning rules, achieves online learning. While Hebbian learning is famously observed in neural circuits (Bliss and Lømo, 1973; Bliss
and Gardner-Medwin, 1973), our network also makes heavy use of anti-Hebbian
learning, which can be interpreted as the long-term potentiation of inhibitory
postsynaptic potentials. Experiments show that such long-term potentiation can
arise from pairing action potentials in inhibitory neurons with subthreshold depolarization of postsynaptic pyramidal neurons (Komatsu, 1994; Maffei et al., 2006).
However, plasticity in inhibitory synapses does not have to be Hebbian, i.e. require
correlation between pre- and postsynaptic activity (Kullmann et al., 2012).
For improved biological realism, the network should respond to a continuous
stimulus stream by continuous and simultaneous changes to its outputs and synaptic
weights. Presumably, this requires neural time scales to be faster and synaptic
time scales to be slower than that of changes in stimuli. To explore this possibility,
we simulated some of our datasets with limited number of neural activity updates
(not shown) and found that ∼10 updates per neuron is sufficient for successful
recovery of sources without significant loss in performance. With a neural time
scale of 10ms, this should take about 100ms, which is sufficiently fast given that,
for example, the temporal autocorrelation time scale of natural image sequences is
about 500ms (David et al., 2004; Bull, 2014).
It is interesting to compare the two-layered architecture we present to the
multilayer neural networks of deep learning approaches (LeCun et al., 2015). 1)
For each data presentation, our network performs recurrent dynamics to produce
an output, while the deep networks have feedforward architecture. 2) The first
layer of our network has multiple neuron types, principal and interneurons, and
only principal neurons project to the next layer. In deep learning, all neurons
in a layer project to the next layer. 3) Our network operates with local learning
rules, while deep learning uses backpropagation, which is not local. 4) We derived
the architecture, the dynamics, and the learning rules of our network from a
principled cost function. In deep learning, the architecture and the dynamics of a
neural network are designed by hand, only the learning rule is derived from a cost
function. 5) Finally, in building a neural algorithm, we started with a generative
model of inputs, from which we inferred algorithmic steps to recover latent sources.
These algorithmic steps guided us in deciding which single-layer networks to stack.
22
In deep learning, no such generative model is assumed and network architecture
design is more of an art. We believe starting from a generative model might lead
to a more systematic way of network design. In fact, the question of generative
model appropriate for deep networks is already being asked (Patel et al., 2016).
Acknowledgments
We thank Andrea Giovannucci, Eftychios Pnevmatikakis, Anirvan Sengupta and
Sebastian Seung for useful discussions. DC is grateful to the IARPA MICRONS
program for support.
23
A
Convergence of the gradient descent-ascent
dynamics
Here we prove that the neural dynamics (18) converges to the saddle point of the
objective function (15). Here we assume that WHG is full-rank and l = d. First,
note that the optimum of (15) is also the fixed point of (18). Since the neural
dynamics (18) is linear, the fixed point is globally convergent if and only if the
eigenvalues of the matrix
"
0
−WHG
WGH
−I
#
(41)
have negative real parts.
The eigenvalue equation is
"
0
−WHG
WGH
−I
#"
x1
x2
#
"
=λ
x1
x2
#
,
(42)
which implies
−WHG x2 = λx1 ,
WGH x1 = (λ + 1) x2 .
(43)
Using these relations, we can solve for all the d + m eigenvalues. There are two
cases:
1. λ = −1. This implies that WGH x1 = 0 and WHG x2 = x1 . x1 is in the
null-space of WGH . Since WGH is m × d with m ≥ d, the null-space is m − d
dimensional, and one has m − d degenerate λ = −1 eigenvalues.
2. λ 6= −1. Substituting for x2 in the first equation of (43), this implies that
WHG WGH x1 = −λ (λ + 1) x1 . Hence, x1 is an eigenvector of WHG WGH .
HG
GH
For each eigenvalue
q λw of W W , there are two corresponding eigenvectors λ = − 12 ± 14 − λw . x2 can be solved uniquely from the first equation
in (43).
Hence, there are m − dqdegenerate λ = −1 eigenvalues and d pairs of conjugate
eigenvalues λ = − 21 ± 14 − λw , one pair for each eigenvaleue λw of WHG WGH .
Since {λw } are real and positive (we assume WHG is full-rank and by definition
>
WHG = WGH ), real parts of all {λ} are negative and hence the neural dynamics
(18) is globally convergent.
24
B
Modified objective function and neural network for generalized prewhitening
While deriving our online neural algorithm, we assumed that the number of output
channels is reduced to the number of sources at the prewhitening stage (l = d).
However, our offline analysis did not need such reduction, one could keep l ≥ d for
generalized prewhitening. Here we provide an online neural algorithm that allows
l ≥ d.
First, we point out why the prewhitening algorithm given in the main text
is not adequate for l > d. In Appendix A, we proved that the neural dynamics
described by (18) converges to the saddle point of the objective function (15).
This proof assumes that WHG is full-rank. However, if l > d, this assumption
breaks down as the network learns because perfectly prewhitened δH has rank
d (low-rank) and a perfectly prewhitening network would have WHG = δHδG>
which would also be low-rank. We simulated this network with l > d and observed
that the condition number of WHG WGH increased with t and the neural dynamics
took longer time to converge. Even though the algorithm was still functioning well
for practical purposes, we present a modification that fully resolves the problem.
We propose a modified offline objective function (Pehlevan and Chklovskii,
2015a) and a corresponding neural network. Consider the following:
min max Tr −δX> δX δH> δH + δH> δH δG> δG + αt δH> δH − t δG> δG ,
δH
δG
(44)
where δX is a k × t centered mixture of d independent sources, δH is now an
l × t matrix with l ≥ d, δG is an m × t matrix with m ≥ d and α is a positive
parameter. Notice the additional α-dependent term compared to (12). If α is less
than the lowest eigenvalue of 1t δXδX> , optimal δH is a linear transform of X
and satisfies the generalized prewhitening condition (3)(Pehlevan and Chklovskii,
2015a). More precisely,
Theorem 4 (Modified from (Pehlevan and Chklovskii, 2015a)). Suppose an eigen>
decomposition of δX> δX is δX> δX = VX ΛX VX , where eigenvalues are sorted
in order of magnitude. If α is less than the lowest eigenvalue of 1t δXδX> , all
optimal δH of (12) have an SVD decomposition of the form
√
0
>
δH∗ = UH t ΛH VX ,
(45)
0
where ΛH is l × t with ones at top d diagonals and zeros at rest.
Using this cost function, we will derive a neural algorithm which does not
suffer from the described convergence issues, even if l > d. On the other hand, we
25
now need to choose the parameter α, and for that we need to know the spectral
properties of δX.
To derive an online algorithm, we repeat the steps taken before:
{δht , δgt } ←− arg min arg max L(δht , δgt ),
(46)
δgt
δht
where
L=
−2δxt>
t−1
X
!
δxt0 δht>0
δht −
t kδgt k22
+
2δgt>
t0 =1
t−1
X
!
δgt0 δht>0
δht + αt kδht k22
t0 =1
+ kδgt k22 − kδxt k22 kδht k22 .
(47)
In the large-t limit, the first four terms dominate over the last term, which we
ignore. The remaining objective is strictly concave in δgt and strictly convex in
δht . Note that (15) was only convex in δht but not strictly convex. The objective
has a unique saddle point, even if
1
t
t−1
P
t0 =1
δht0 δgt>0 is not full-rank:
δgt∗ = WtGH δht∗ ,
δht∗ = αI + WtHG WtGH
−1
WtHX δxt ,
(48)
where W matrices are defined as before and I is the identity matrix.
We solve (46) with gradient descent-ascent
1
dδht
= − ∇δht L = −αδht + WtHX δxt − WtHG gt ,
dγ
2t
1
dδgt
= ∇δgt L = −δgt + WtGH δht .
dγ
2t
(49)
where γ is time measured within a single time step of t. The dynamics (18) can
be proved to converge to the saddle point (48) modifying the proof in Appendix
A4 . Synaptic weight updates are the same as before (19). Finally, this network
can be modified to also compute H̄ following the steps before.
4
The fixed point is globally convergent if and only if the eigenvalues of the matrix
−αI −WHG
WGH
−I
(50)
have negative real parts. One can show that l − d eigenvalues are −α, m − d q
eigenvalues are −1,
and for each positive eigenvalue, λw of WHG WGH one gets a pair − 1+α
2 ±
All eigenvalues have negative real parts.
26
(1+α)2
4
− α − λw .
C
Mixing matrices for numerical simulations
For the random source dataset, the d = 3 mixing matrix was:
0.031518 0.38793 0.061132
A = −0.78502 0.16561 0.12458 ,
0.34782 0.27295 0.67793
(51)
We do not list the mixing matrices for d = {5, 7, 10} cases for space-saving purposes,
however they are available from authors upon request.
For the natural scene dataset, the mixing matrix was
A=
0.33931
0.3282 0.41516 −0.1638
−0.077079 −0.29768 0.076562 −0.28153
.
−0.14119 −0.41709 0.14842
0.57009
−0.40483
0.21922 0.082336
0.18027
27
(52)
D
Learning rate parameters for numerical simulations
For Figs. 3, 4, 5 and 6 the following parameters were found to be best performing
as a result of our grid search:
fastICA
NPCA
NSM (activity) NSM (time)
d=3
(10, 0.01)
(10, 0.1)
(10, 0.8)
(10, 0.1)
d = 5 (100, 0.01) (10, 0.01)
(10, 0.9)
(10, 0.01)
d = 7 (100, 0.01) (100, 0.01)
(10, 0.9)
(10, 0.1)
d = 10 (100, 0.01) (1000, 0.01)
(10, 0.9)
(10, 0.01)
Images (104 , 0.01) (1000, 0.01)
(100, 0.9)
NA
d=3
d=5
d=7
d = 10
Images
Two-layer
Offline
(100, 1, 10, 0.8) (106 , 0.001)
(100, 1, 10, 0.9) (106 , 0.001)
(100, 1, 10, 0.9) (106 , 0.01)
(100, 1, 10, 0.9) (106 , 0.01)
(100, 1, 100, 0.9)
NA
28
Infomax ICA
(1000, 0.2)
(1000, 0.2)
(1000, 0.2)
(1000, 0.2)
NA
Linsker’s Algorithm
(1000, 0.2)
(1000, 0.2)
(1000, 0.2)
(1000, 0.2)
NA
References
Amari, S., Cichocki, A., and Yang, H. (1996). A new learning algorithm for blind
signal separation. Advances in neural information processing systems, 8:757–763.
Asari, H., Pearlmutter, B. A., and Zador, A. M. (2006). Sparse representations for
the cocktail party problem. Journal of Neuroscience, 26(28):7477–7490.
Bee, M. A. and Micheyl, C. (2008). The cocktail party problem: what is it? how
can it be solved? and why should animal behaviorists study it? Journal of
comparative psychology, 122(3):235.
Bell, A. J. and Sejnowski, T. J. (1995). An information-maximization approach to
blind separation and blind deconvolution. Neural computation, 7(6):1129–1159.
Bliss, T. V. and Gardner-Medwin, A. (1973). Long-lasting potentiation of synaptic transmission in the dentate area of the unanaesthetized rabbit following
stimulation of the perforant path. The Journal of physiology, 232(2):357.
Bliss, T. V. and Lømo, T. (1973). Long-lasting potentiation of synaptic transmission
in the dentate area of the anaesthetized rabbit following stimulation of the
perforant path. The Journal of physiology, 232(2):331–356.
Bronkhorst, A. W. (2000). The cocktail party phenomenon: A review of research
on speech intelligibility in multiple-talker conditions. Acta Acustica united with
Acustica, 86(1):117–128.
Bull, D. R. (2014). Communicating pictures: A course in Image and Video Coding.
Academic Press.
Comon, P. (1994). Independent component analysis, a new concept?
processing, 36(3):287–314.
Signal
Comon, P. and Jutten, C. (2010). Handbook of Blind Source Separation: Independent component analysis and applications. Academic press.
David, S. V., Vinje, W. E., and Gallant, J. L. (2004). Natural stimulus statistics
alter the receptive field structure of v1 neurons. Journal of Neuroscience,
24(31):6991–7006.
Donoho, D. and Stodden, V. (2003). When does non-negative matrix factorization
give a correct decomposition into parts? In Advances in neural information
processing systems, page None.
29
Eagleman, D. M., Coenen, O. J.-M. D., Mitsner, V., Bartol, T. M., Bell, A. J.,
and Sejnowski, T. J. (2001). Cerebellar glomeruli: Does limited extracellular
calcium implement a sparse encoding strategy? In Proceedings of the 8th Annual
Joint Symposium on Neural Computation.
Golumbic, E. M. Z., Ding, N., Bickel, S., Lakatos, P., Schevon, C. A., McKhann,
G. M., Goodman, R. R., Emerson, R., Mehta, A. D., Simon, J. Z., et al. (2013).
Mechanisms underlying selective neuronal tracking of attended speech at a
“cocktail party”. Neuron, 77(5):980–991.
Hu, T., Pehlevan, C., and Chklovskii, D. B. (2014). A hebbian/anti-hebbian
network for online sparse dictionary learning derived from symmetric matrix
factorization. In Asilomar Conference on Signals, Systems and Computers, pages
613–619. IEEE.
Huang, K., Sidiropoulos, N., and Swami, A. (2014). Non-negative matrix factorization revisited: Uniqueness and algorithm for symmetric decomposition. Signal
Processing, IEEE Transactions on, 62(1):211–224.
Hyvarinen, A. (1999). Fast and robust fixed-point algorithms for independent
component analysis. IEEE transactions on Neural Networks, 10(3):626–634.
Hyvärinen, A. and Hoyer, P. (2000). Emergence of phase-and shift-invariant
features by decomposition of natural images into independent feature subspaces.
Neural computation, 12(7):1705–1720.
Hyvärinen, A. and Oja, E. (1997). A fast fixed-point algorithm for independent
component analysis. Neural computation, 9(7):1483–1492.
Hyvärinen, A. and Oja, E. (1998). Independent component analysis by general
nonlinear hebbian-like learning rules. Signal Processing, 64(3):301–313.
Hyvärinen, A. and Oja, E. (2000). Independent component analysis: algorithms
and applications. Neural networks, 13(4):411–430.
Isomura, T., Kotani, K., and Jimbo, Y. (2015). Cultured cortical neurons can
perform blind source separation according to the free-energy principle. PLoS
Comput Biol, 11(12):e1004643.
Isomura, T. and Toyoizumi, T. (2016). A local learning rule for independent
component analysis. Scientific Reports, 6.
Jutten, C. and Herault, J. (1991). Blind separation of sources, part i: An adaptive
algorithm based on neuromimetic architecture. Signal processing, 24(1):1–10.
30
Komatsu, Y. (1994). Age-dependent long-term potentiation of inhibitory synaptic
transmission in rat visual cortex. Journal of Neuroscience, 14(11):6488–6499.
Kuang, D., Park, H., and Ding, C. H. (2012). Symmetric nonnegative matrix
factorization for graph clustering. In SDM, volume 12, pages 106–117. SIAM.
Kuang, D., Yun, S., and Park, H. (2015). Symnmf: nonnegative low-rank approximation of a similarity matrix for graph clustering. Journal of Global
Optimization, 62(3):545–574.
Kullmann, D. M., Moreau, A. W., Bakiri, Y., and Nicholson, E. (2012). Plasticity
of inhibition. Neuron, 75(6):951–962.
Laurberg, H., Christensen, M. G., Plumbley, M. D., Hansen, L. K., and Jensen, S. H.
(2008). Theorems on positive data: On the uniqueness of nmf. Computational
intelligence and neuroscience, 2008.
LeCun, Y., Bengio, Y., and Hinton, G. (2015).
521(7553):436–444.
Deep learning.
Nature,
Lee, D. D. and Seung, H. S. (1999). Learning the parts of objects by non-negative
matrix factorization. Nature, 401(6755):788–791.
Li, Y., Liang, Y., and Risteski, A. (2016). Recovery guarantee of non-negative
matrix factorization via alternating updates. arXiv preprint arXiv:1611.03819.
Linsker, R. (1988). Self-organization in a perceptual network. Computer, 21(3):105–
117.
Linsker, R. (1997). A local learning rule that enables information maximization
for arbitrary input distributions. Neural Computation, 9(8):1661–1665.
Maffei, A., Nataraj, K., Nelson, S. B., and Turrigiano, G. G. (2006). Potentiation
of cortical inhibition by visual deprivation. Nature, 443(7107):81–84.
Malenka, R. C. and Bear, M. F. (2004). Ltp and ltd: an embarrassment of riches.
Neuron, 44(1):5–21.
McDermott, J. H. (2009).
19(22):R1024–R1027.
The cocktail party problem.
Current Biology,
Mesgarani, N. and Chang, E. F. (2012). Selective cortical representation of attended
speaker in multi-talker speech perception. Nature, 485(7397):233–236.
31
Narayan, R., Best, V., Ozmeral, E., McClaine, E., Dent, M., Shinn-Cunningham,
B., and Sen, K. (2007). Cortical interference effects in the cocktail party problem.
Nature neuroscience, 10(12):1601–1607.
Oja, E. (1982). Simplified neuron model as a principal component analyzer. J
Math Biol, 15(3):267–273.
Oja, E. and Plumbley, M. (2004). Blind separation of positive sources by globally
convergent gradient search. Neural Computation, 16(9):1811–1825.
Ouedraogo, W. S. B., Souloumiac, A., and Jutten, C. (2010). Non-negative
independent component analysis algorithm based on 2d givens rotations and a
newton optimization. In Latent Variable Analysis and Signal Separation, pages
522–529. Springer.
Paatero, P. and Tapper, U. (1994). Positive matrix factorization: A non-negative
factor model with optimal utilization of error estimates of data values. Environmetrics, 5(2):111–126.
Patel, A. B., Nguyen, M. T., and Baraniuk, R. (2016). A probabilistic framework
for deep learning. In Advances in Neural Information Processing Systems, pages
2550–2558.
Pehlevan, C. and Chklovskii, D. (2014). A hebbian/anti-hebbian network derived
from online non-negative matrix factorization can cluster and discover sparse
features. In Asilomar Conference on Signals, Systems and Computers, pages
769–775. IEEE.
Pehlevan, C. and Chklovskii, D. (2015a). A normative theory of adaptive dimensionality reduction in neural networks. In Advances in Neural Information
Processing Systems, pages 2260–2268.
Pehlevan, C. and Chklovskii, D. (2015b). Optimization theory of hebbian/antihebbian networks for pca and whitening. In 2015 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 1458–1465.
IEEE.
Pehlevan, C., Hu, T., and Chklovskii, D. (2015). A hebbian/anti-hebbian neural
network for linear subspace learning: A derivation from multidimensional scaling
of streaming data. Neural Computation, 27:1461–1495.
Plumbley, M. (1994). A subspace network that determines its own output dimension.
Tech. Rep.
32
Plumbley, M. (1996). Information processing in negative feedback neural networks.
Network-Comp Neural, 7(2):301–305.
Plumbley, M. (2002). Conditions for nonnegative independent component analysis.
Signal Processing Letters, IEEE, 9(6):177–180.
Plumbley, M. D. (2001). Adaptive lateral inhibition for non-negative ica. In
Proceedings of the international Conference on Independent Component Analysis
and Blind Signal Separation (ICA2001), pages 516–21.
Plumbley, M. D. (2003). Algorithms for nonnegative independent component
analysis. Neural Networks, IEEE Transactions on, 14(3):534–543.
Plumbley, M. D. and Oja, E. (2004). A "nonnegative pca" algorithm for independent
component analysis. Neural Networks, IEEE Transactions on, 15(1):66–76.
Wright, S. J. (2015). Coordinate descent algorithms. Mathematical Programming,
151(1):3–34.
Yuan, Z. and Oja, E. (2004). A fastica algorithm for non-negative independent
component analysis. In Independent Component Analysis and Blind Signal
Separation, pages 1–8. Springer.
Zheng, C.-H., Huang, D.-S., Sun, Z.-L., Lyu, M. R., and Lok, T.-M. (2006).
Nonnegative independent component analysis based on minimizing mutual
information technique. Neurocomputing, 69(7):878–883.
33
| 9cs.NE
|
1
Delay Performance of MISO Wireless
Communications
arXiv:1707.08089v1 [cs.IT] 25 Jul 2017
Jesús Arnau, Member, IEEE, and Marios Kountouris, Senior Member, IEEE
Abstract
Ultra-reliable, low latency communications (URLLC) are currently attracting significant attention
due to the emergence of mission-critical applications and device-centric communication. URLLC will
entail a fundamental paradigm shift from throughput-oriented system design towards holistic designs for
guaranteed and reliable end-to-end latency. A deep understanding of the delay performance of wireless
networks is essential for efficient URLLC systems. In this paper, we investigate the network layer
performance of multiple-input, single-output (MISO) systems under statistical delay constraints. We
provide closed-form expressions for MISO diversity-oriented service process and derive probabilistic
delay bounds using tools from stochastic network calculus. In particular, we analyze transmit beamforming with perfect and imperfect channel knowledge and compare it with orthogonal space-time codes
and antenna selection. The effect of transmit power, number of antennas, and finite blocklength channel
coding on the delay distribution is also investigated. Our higher layer performance results reveal key
insights of MISO channels and provide useful guidelines for the design of ultra-reliable communication
systems that can guarantee the stringent URLLC latency requirements.
Index Terms
URLLC, 5G systems, MIMO, diversity, stochastic network calculus, finite blocklength channel
coding, queueing analysis.
I. I NTRODUCTION
Data traffic has been growing tremendously over the last decade, fueled by the ubiquity
of smart mobile devices and bandwidth-demanding applications. In order to handle the everincreasing traffic load, existing wireless networks have typically been designed and planned
The authors are with the Mathematical and Algorithmic Sciences Lab, France Research Center, Huawei Technologies France
SASU, 20 Quai du Point du Jour, 92100 Boulogne-Billancourt, France. Email: ({jesus.arnau,marios.kountouris}@huawei.com).
2
with a focus on improving spectral efficiency and increasing coverage. The latency requirements
of different applications have mostly been an after-thought. Ultra-high reliability and low latency
have not been in the mainstream in most wireless networks, due to the focus on humancentric communications, delay-tolerant content and reliability levels in the order of 95-99%.
However, a plethora of socially useful applications and new uses of wireless communication
are currently envisioned in areas such as industrial control, smart cities, augmented and virtual
reality (AR/VR), automated driving or flying, robotics, telemedicine, algorithmic trading and
tactile Internet. In response, new releases of mobile cellular networks (mainly 5G new radio and
beyond) are envisaged to support ultra-reliable, low latency communications (URLLC) scenarios
with strict requirements in terms of latency (ranging from 1 ms to few milliseconds end-to-end
latency depending on the use cases) and reliability (higher than 99.9999%). Another new feature
is the support of machine-type communications (MTC), where a massive number of connected
devices transmit reliably a relatively low volume of non-delay-sensitive payload.
Information theory and communication engineering have been instrumental in boosting spectral efficiency and approaching the capacity limits. Nevertheless, URLLC and device-centric
communication pose significant theoretical and practical challenges, requiring a departure from
capacity-oriented system design towards a holistic view (network architecture, control, and data)
for guaranteed and reliable end-to-end latency. Applying information theory to the design of
low latency networks has been a long-standing challenge [1]. Information theory mostly focuses
on asymptotic limits, which can be achieved with arbitrarily small probability of error using
long codewords, hence arbitrarily large coding delays. Despite recent development on the block
error rates for finite blocklength codes [2], more work is needed to better understand the
non-asymptotic fundamental tradeoffs between delay, reliability and throughput, including both
coding delays and queueing delays. In addition, the highly variable and delay-sensitive nature
of network traffic together with the associated overhead (metadata) should be incorporated in
the conventional communication theoretic framework.
Reliable communication is a well-studied problem, dating back to Shannon’s landmark paper
[3], and diversity-achieving techniques are usually employed to increase reliability by combating or exploiting channel variations. Several schemes have been developed, including error
correction codes, the use of multiple antennas, and space/frequency diversity at the physical
layer, as well as automatic repeat-request (ARQ), opportunistic scheduling, and erasure coding
at higher layers. Among them, multiple-input multiple-output (MIMO) systems have received
3
great interest due to their potential to combat fading, increase spectral efficiency, and reduce
interference. MIMO techniques can be used for beam steering, diversity, spatial multiplexing
and interference cancellation. Diversity-achieving techniques increase reliability by combating
or exploiting channel variations, while beam steering techniques increase received signal quality
by focusing desired energy or attenuating undesired interference. Spatial multiplexing increases
the data rate by transmitting independent data symbols across the antennas. In this work, we focus
on diversity and beam steering techniques, as being more relevant for reliable communications. In
particular, we consider maximum ratio transmission (MRT), a transmit beamforming technique
that maximizes the received signal and realizes diversity exploiting channel state information
(CSI) at the transmitter.
There have been several attempts at addressing latency considerations in the physical layer,
including error exponents [4], delay-limited capacity [5], outage capacity [6], throughput-delay
tradeoff curves [7], and finite blocklength channel coding [2]. In networking, delay is a key
performance measure and queueing theory has been instrumental in providing exact solutions
for backlog and delays in packet-switched networks. However, queueing network analysis is
largely restricted to single-queue networks, few interacting (coupled) queues, small topologies
and Poisson arrivals. Classical queueing models typically allow the analysis of average delay, failing to characterize delay quantiles (worst-case delay) and distributions, which are of
cardinal importance in mission-critical applications. Recent efforts to combine queueing with
communication theory, such as stochastic network calculus [8]–[10], timely throughput [11],
effective bandwidth [8], and effective capacity [12] to name a few, take on a different approach
and compute performance bounds for a wide range of stochastic processes. These approaches
promise significant performance gains - in terms of latency, reliability and throughput - and crisp
insights for the design of low latency communication systems. In this work, we employ stochastic
network calculus - a probabilistic extension of (deterministic) network calculus [13] - which
allows non-asymptotic stochastic bounds on network performance metrics, such as maximum
delay, for broad classes of arrival, scheduling, and service processes.
Despite the extended literature on MIMO techniques at the physical layer, only few attempts
have been made to characterize the upper layer performance of multi-antenna techniques taking
into account the queueing effects. In [14] the service process of an adaptive MIMO system
with Poisson arrivals is characterized. Bounds on the delay violation probability have been
derived for MIMO multiple access with bursty traffic in [15], while [16] provides an asymptotic
4
analysis of the diversity-multiplexing tradeoff for MIMO systems with bursty and delay-limited
information. Using large deviations, [17] analyzes the queueing performance of queue-aware
scheduling in multiuser MIMO systems. Bounds on the tail of delay of MIMO communication
systems have been derived using the effective capacity framework [18]–[20]. Nevertheless,
these approximations are only valid for large delays and under constant bit rate processes.
Using Markov chains to reproduce the state of Gilbert-Elliott fading channels, the flow-level
performance of MIMO spatial multiplexing has been analyzed using stochastic network calculus
in [21], [22]. Nevertheless, none of these works considered the delay performance of MIMO
schemes using stochastic network calculus for wireless fading channels.
In this work, we study the upper layer delay performance of multiple input, single output
(MISO) diversity communication in the presence of statistical delay constraints. We consider
MRT transmit beamforming at the physical layer and derive probabilistic delay bounds using
tools from stochastic network calculus. For that, we provide a closed-form characterization of
the cumulative service process for MISO beamforming channels with both perfect and imperfect
CSI. For the analysis, we use the (min,×) network calculus methodology for fine-grained wireless
network delay analysis [23]. The impact of transmit antennas, signal-to-noise ratio (SNR), and
imperfect CSI on the delay distribution of MISO MRT systems is characterized. We then
show that our mathematical framework can be applied to the statistical characterization of
various MIMO service processes, including MIMO eigen-beamforming, orthogonal space-time
block coding (OSTBC), antenna selection, and Nakagami-m fading channels. This allows us
to compare the delay performance of transmit beamforming with alternative diversity-achieving
techniques that rely on very low rate CSI (transmit antenna selection) or no CSI (OSTBC).
Interestingly, MISO MRT is shown to reduce the delay violation probability as compared to
single-antenna transmissions even with imperfect CSI. The derived delay bounds enable us to
assess the robustness of MISO MRT delay performance with respect to channel imperfections.
Our results also show under which operating parameters other diversity-techniques are preferable
than MRT in terms of delay violation probability. In addition, we provide an asymptotic statistical
characterization of the service process in the low/high SNR regime and for large number of
antennas. Finally, extending [24] to MISO systems, we study the effect of finite blocklength
channel coding on the queueing delay performance. Our results quantify the performance loss
due to finite blocklength and characterize the tradeoff between data rate and the error probability
with respect to the delay performance. Our results can provide useful insights and guidelines for
5
the design of ultra-reliable wireless systems that can satisfy and guarantee the stringent URLLC
latency requirements.
The rest of the paper is organized as follows: In Section II, we provide our system model and
in Section III, a brief background on the (min,×) network calculus is presented. In Section IV,
the delay performance analysis of MISO diversity systems is derived. Section V provides the
delay performance in asymptotic regimes and Section VI shows the effect of finite blocklength on
the delay performance. Numerical results are presented in Section VII, followed by conclusions
in Section VIII.
II. S YSTEM MODEL
We consider data transmission over a point-to-point vector communication channel. Time is
divided into time slots of duration T (discrete-time model), and at each slot i, the source generates
ai data bits and stores them in a queue. The transmitter (source) has M transmit antennas and
sends the queued data bits to a single-antenna receiver over a frequency-flat Rayleigh fading
channel. We assume a block-fading model, where the channel remains constant during one slot
and varies independently from slot to slot. Each slot contains n + nm symbols, where n denotes
the complex data symbols and nm the metadata (headers, training, estimation, acknowledgments,
etc.).
A. Signal model
The received downlink signal yi ∈ C at slot i in a MISO wireless channel is given by
yi =
√
snr · hH
i x i + ni
(1)
where hi ∈ CM ×1 is the flat-fading channel between the transmitter and the receiver at the i-th
slot, which is circularly-symmetric complex Gaussian distributed h ∼ CN (0, 1). The transmitted
vector is denoted by xi ∈ CM ×1 , and ni ∼ CN (0, 1) is the additive background noise that
may also include (Gaussian) interference from neighboring systems. We consider one of the
most prominent multi-antenna diversity technique, namely transmit beamforming, which refers
to sending linearly weighted versions of the same signal on each antenna. The transmitted signal
can be written as xi = wi si , where si is the zero-mean data signal at slot i with power E [|s|2 ] = 1,
and wi ∈ CM ×1 is the unit-norm beamforming vector. Note that, since noise is assumed to have
6
unit power, snr represents the average received SNR, whereas the instantaneous SNR in the i-th
2
slot is given by γi = snr|hH
i wi | .
B. Transmission mode
A natural signaling strategy for the MISO channel will be to maximize SNR over a specific
channel, which can be achieved by sending information only in the direction of the channel vector
h, as information sent in any orthogonal direction will be nulled out by the channel anyway.
We thus consider the so-called maximum ratio transmission (MRT) [25], which is equivalent to
eigen-beamforming since beamforming along the dominant (and only) eigenmode of the M × 1
vector channel is performed1 .
Assuming that both transmitter and receiver have perfect CSI, the MRT beamforming vector
is given by wi =
hi
.
khi k
In that case, the instantaneous SNR is γi = snrkhi k2 , which is gamma
distributed with shape parameter M and scale parameter snr, i.e. γi ∼ Gamma(M, snr). When
the transmitter does not fully know the actual channel vector h (imperfect CSI), we can model
its channel knowledge as ĥ = h + e, where e ∼ CN (0, σe2 I). MRT is then performed based
ĥ
. Particularizing [26, Eq. 7] to the MISO case, the
kĥk
instantaneous SNR is gamma distributed with shape parameter M and scale parameter ζ, i.e.
−1
1 + σe2
2
.
(2)
γi ∼ Gamma(M, ζ) with ζ = σe +
snr
on the channel estimate, so that w =
This additive error model is consistent with time-divison duplex (TDD) operation, where uplink
and downlink transmissions take place at the same frequency, in different time instants; assuming
they fall within the coherence interval of the channel, then channel reciprocity can be used to
estimate the downlink channel from uplink pilot signals. This model also applies to frequencydivision duplex (FDD) operation with analog feedback [27]. We only account for the effect of
CSI error in MISO beamforming, which reduces the achieved SNR (SNR loss) because of not
transmitting exactly in the direction of the actual channel; as we explain in the next subsection
there could be another penalty in the rate selection process.
1
It can be shown that MRT is the transmission scheme that maximizes both the capacity and the service rate with sum power
constraint, Gaussian input signaling, and perfect CSI. This is not necessarily true for arbitrary inputs, where the transmission
scheme may depend on the delay constraints.
7
C. Data transmission
A codeword of length n symbols (corresponding to n channel uses) and rate Ri (in bits per
symbol) is transmitted at each slot i. The transmitter selects a rate adapted to γi and we consider
the following two cases:
1) Asymptotically large blocklength: If the blocklength is large enough, no errors occur and
the achievable rate is equal to the Shannon capacity of the channel, Ri = log2 (1 + γi ).
2) Finite blocklength: At finite blocklength, a transmission error can occur with probability
> 0 and the maximum coding rate Ri (n, ) is lower than the Shannon rate. Tight non asymptotic
upper and lower bounds on the maximum coding rate are given in [2]. Furthermore, for AWGN
channels an asymptotic approximation has been established and shown accurate for packet sizes
as small as 100 [2], [28], [29]. The coding rate Ri (n, ) = k/n to transmit k information bits
using coded packets spanning n channel uses is given by
r
log2 n
V (γi ) −1
Q () +
+ O(1)
Ri (n, ) ≈ log2 (1 + γi ) −
n
2n
(3)
where Q−1 (·) is the inverse of the Gaussian Q function and V (γi ) is the channel dispersion
given by
V (γi ) =
γi (γi + 2)
log22 e.
(1 + γi )2
(4)
Using this approximation, the coding rate and the packet error probability are related as
!
n log2 (1 + γi ) − k + 0.5 log2 n
p
=Q
.
(5)
V (γi )n
The above approximation is valid for MISO systems with knowledge of the fading coefficients
of the vector channel and the SNR realization at each slot, which makes the MISO channel
behaving equivalently to an AWGN channel with SNR snrkhk2 . We should remark that so
far throughout this section we have assumed perfect knowledge of the SNR realization, so
that the transmitter can adapt the rate to it with no errors. Thus, we only account for the
channel estimation error as an SNR penalty, as described in Sec. II-B. The case of imperfect
rate selection in MISO systems, which goes beyond the scope of this work, can be analyzed
using techniques recently developed in [30]. In that case, channel estimation errors will be
approximated by Gaussian variations in the SNR and the Gaussian variations in the capacity due
to finite blocklength in [2] will be transformed into Gaussian errors in the SNR.
8
D. Queuing model
For the analysis of queuing systems, we consider a stochastic system-theoretic model as in
[23], which is widely used in the stochastic network calculus methodology. Stochastic network
calculus considers queuing systems and networks of systems with stochastic arrival, departure,
and service processes, as the ones below. For an in-depth description of the topic, the interested
reader may refer to [9], [10], [23], [31]–[33].
The arrival process ai , introduced in Sec. II, models the number of bits that arrive at the
queue at a discrete time instant i. For successful transmissions, the service process si is equal
to: (i) nRi for asymptotically large blocklength, (ii) nRi (n, ) with finite blocklength; in case
of transmission errors, the service is considered to be zero as no data is removed from the
queue. Finite blocklength channel coding affects the reliability of the physical layer, which in
turn causes additional delay as data needs to be buffered until successfully transmitted. The
departure process di describes the number of bits that arrive successfully at the destination and
depends on both the service process and the number of bits waiting in the queue. Note that
acknowledgments and feedback messages are assumed to be instantaneous and error-free.
We further define the cumulative arrival, service and departure processes as
A(τ, t) =
t−1
X
i=τ
ai , S(τ, t) =
t−1
X
i=τ
si , D(τ, t) =
t−1
X
di .
(6)
i=τ
For lossless first-in first-out queuing systems, the delay W (t) at time t, i.e. the number of
slots it takes for an information bit arriving at time t to be received at the destination, is defined
as
W (t) = inf{u > 0 : A(0, t)/D(0, t + u) ≤ 1}.
(7)
and the delay violation probability is given by Λ(w, t) = P [W (t) > w].
Using the dynamic server property (i.e. D(0, t) ≥ A ∗ S(0, t) where the (min, +) convolution
operator ’*’ is defined as f ∗ g(τ, t) = inf τ ≤u≤t f (τ, t) + g(u, t) [8]), the delay can be characterized through the cumulative arrival and service processes, which we have so far described in
the so-called bit domain. As it is more convenient for the analysis of wireless fading channels,
we follow [23] and analyze these processes in the exponential (or SNR) domain.
9
III. S TOCHASTIC N ETWORK C ALCULUS IN THE SNR D OMAIN
A remarkable feature of stochastic network calculus in the SNR domain is that it allows to
obtain bounds on the delay violation probability based on simple statistical characterizations of
the arrival and service processes in terms of their Mellin transforms. We will briefly review this
result in this section.
Let us start by converting the cumulative processes in the bit domain through the exponential
function. The corresponding processes in the SNR domain, denoted by calligraphic letters, are
A(τ, t) = eA(τ,t) ,
D(τ, t) = eD(τ,t) ,
S(τ, t) = eS(τ,t) .
(8)
From these definitions, an upper bound on the delay violation probability can be computed
by means of the Mellin transforms of A(τ, t) and S(τ, t):
pv (w) = inf {K(s, −w)} ≥ Λ(w)
s>0
(9)
where K(s, −w) is the so-called steady-state kernel, defined as
K(s, −w) = lim
t→∞
t
X
MA (1 + s, u, t)MS (1 − s, u, t + w).
(10)
u=0
where MX (s) = E [X s−1 ] denotes the Mellin transform of a nonnegative random variable X for
any s ∈ C for which the expectation exists. We restrict our derivations in this work to s ∈ R
and we recall that, for a continuous probability density function (pdf) fX (x) on (0, ∞), if there
fX (x)
exists δ > 0 such that lim+ δ+θ−1 < ∞, then E[X −θ ] < ∞. Alternatively, E[X −θ ] < ∞ if and
x→0 x
Z ∞
FX (x)
dx < ∞.
only if
xθ+1
0
A. Mellin transform of arrival and service processes
Assuming that A(τ, t) has stationary and independent increments, the Mellin transforms
become independent of the time instance, as follows:
!s−1
t−1
Y
MA (s, τ, t) = E
eai
(11)
i=τ
t−τ
= E ea(s−1)
(12)
= Mα (s)t−τ
(13)
10
where we have defined α = ea , the non-cumulative arrival process in the SNR domain. We
consider the traffic class of (z(s), ρ(s))-bounded arrivals, whose moment generating function in
the bit domain is bounded by [8]
1
log E[esA(τ,t) ] ≤ ρ(s) · (t − τ ) + z(s)
s
(14)
for some s > 0. Restricting ourselves to the case where ρ is independent of s and z(s) = 0, we
have [24], [34]
Mα (s) = eρ(s−1) .
(15)
For the service process, we start by rewriting si = B log g(γ), where B = n/ log 2. Since
the different si are independent and identically distributed (i.i.d.), we can express the Mellin
transform of the cumulative service as
MS (s, τ, t) = E
t−1
Y
!s−1
g(γ)B
(16)
i=τ
t−τ
= E g(γ)B(s−1)
(17)
= Mg(γ) (1 + B(s − 1))t−τ .
(18)
B. Delay Bound
Plugging (15) and (16) into (10) and following [23], the steady-state kernel can be finally
rewritten as
K(s, −w) =
Mg(γ) (1 − B · s)w
,
1 − Mα (1 + s)Mg(γ) (1 − B · s)
(19)
for any s > 0 under the stability condition Mα (1 + s)MS (1 − s) < 1. The delay bound (9) thus
reduces to
pv (w) = inf
s>0
Mg(γ) (1 − B · s)w
1 − Mα (1 + s)Mg(γ) (1 − B · s)
.
(20)
IV. D ELAY WITH L ARGE B LOCKLENGTH : E XACT A NALYSIS
In this section, we derive exact closed-form expressions for the steady-state kernel K(s, −w)
of MISO diversity schemes when the blocklength is infinitely large. We start by providing a
general result on the Mellin transform of the service process when the instantaneous SNR is
gamma distributed. Obtaining the steady-state kernel for MRT beamforming with both perfect
11
and imperfect CSI is a particularization of this result, which is shown to apply, as a byproduct,
for obtaining the performance of other diversity techniques, including MISO OSTBC, antenna
selection, and MIMO MRT/MRC.
Consider the instantaneous SNR to be a gamma distributed random variable γ ∼ Gamma(M, ζ)
with shape parameter M , scale parameter ζ and pdf
x
xM −1 e− ζ
fγ (x) =
,
Γ(M )ζ M
where Γ(t) =
R∞
0
x≥0
(21)
xt−1 e−x dx is the (complete) gamma function; we have dropped the subindex
since SNRs are independent and ergodic. First, we derive the Mellin transform of g(γ), i.e.
Mg(γ) (s) = E [g(γ)s−1 ]. For notation convenience, in the remainder we assume B = n/ log 2 =
1, however in Sec. VII we give again relevant values to this parameter in order to obtain
meaningful numerical results.
Theorem 1. The Mellin transform of g(γ) = 1 + γ, where γ ∼ Gamma(M, ζ) with M ∈ N+
and ζ > 0, is given by
Mg(γ) (s) = ζ −M · U (M, M + s, ζ −1 )
(22)
where U (a, b, z) is Tricomi’s confluent hypergeometric function [35, Eq. 13.2.5] (also called
confluent hypergeometric function of the second kind and denoted by Ψ(a; b; z)).
Proof: See Appendix A.
A. MISO MRT
The Mellin transform derived above applies directly to the service process with MISO MRT
transmission. Using this expression together with the transform of the arrival process, we obtain
the kernel and consequently the bound on the delay violation probability as follows
(
)
w
MMRT
g(γ) (1 − B · s)
pv (w) = inf
s>0
1 − Mα (1 + s)MMRT
g(γ) (1 − B · s)
(
w )
ζ −M · U (M, M + 1 − s, ζ −1 )
= inf
.
s>0
1 − eρs ζ −M · U (M, M + 1 − s, ζ −1 )
(23)
Although U (a, b, z) is implemented in standard software for mathematical calculations, we
provide below an alternative expression for the Mellin transform in terms of the simpler upper
12
incomplete gamma function.
Theorem 2. The Mellin transform of g(γ) from Theorem 1 can be given as
1
M
−1
X
eζ
M −1
Mg(γ) (s) = M
(−1)M −1−j · ζ j+s · Γ(j + s, ζ −1 )
ζ Γ(M ) j=0
j
1
= eζ
M
−1
X
(−1)M −1−j · ζ j+s−M
j=0
where Γ(s, z) =
R∞
z
Γ(j + s, ζ −1 )
Γ(M − j)Γ(j + 1)
(24)
(25)
ts−1 e−t dt is the upper incomplete gamma function.
Proof: See Appendix B.
Remark 1. For the SISO case, letting M = 1 and ζ = snr in (24) we obtain Mg(γ) (s) =
1
e snr · snrs−1 · Γ(s, snr−1 ) which is the same expression reported in [23].
The above expressions allows us to obtain bounds on the delay violation probability for
different system parameters without resorting to Monte Carlo simulations. However, due to the
complexity of the kernel function, no closed-form solution for the minimum s can be found, and
we must resort to numerical methods. In some asymptotic cases, we can have simpler expressions
of the Mellin transform that make this process easier, as we will show later in Section V.
B. Other MISO Diversity Techniques
So far, we have considered that the transmitter performs MRT based on perfect or imperfect
CSI. In this section, and for means of comparison, we study two alternative multi-antenna
diversity techniques, namely OSTBC and transmit antenna selection, which rely on no and
very low-resolution CSI, respectively.
1) OSTBC: Orthogonal space-time block coding has been a very successful transmit diversity
technique because it can achieve full diversity without CSI at the transmitter and need for joint
decoding of multiple symbols. It is characterized by the number of independent symbols Ns
transmitted over T time slots; the code rate is Rc = Ns /T . When the transmitter uses OSTBC
with M transmit antennas, code parameter T , and the receiver performs MRC with N antennas,
the equivalent SNR γ =
snr
kHk2F
M
is gamma distributed with shape parameter M N and scale
parameter (snr/M )−1 [36, Eq. 3.43]; here H denotes the MIMO channel matrix of N × M
13
complex Gaussian entries. Particularizing (22) for the case of MISO OSTBC, we have the
following result.
Corollary 1. The Mellin transform of the service process of a MISO system employing OSTBC
is given by
MOSTBC
g(γ) (s) =
snr −M
M
· U (M, M + s, M/snr) .
(26)
2) Antenna Selection: Antenna selection is a low-complexity, low-rate feedback diversity
technique, in which the transmitter and/or the receiver select a subset of transmit/receive antennas
for transmission/reception. It can be used in conjunction with other diversity techniques and can
improve the performance of open-loop MIMO at the expense of very low amount of feedback.
We consider here transmit antenna selection (TAS), in which the transmitter selects to transmit
on the antenna (one of M ) that maximizes the instantaneous SNR. The amount of CSI required
to be fed back to the transmitter is dlog2 M e bits (index of best antenna), where dxe denotes
the smallest integer larger than x. The instantaneous SNR can be expressed as γTAS = snrγmax ,
where γmax is the largest channel gain, i.e. γmax = max |hi |2 . Since hi ∼ CN (0, 1), we have
1≤i≤M
that |hi |2 is exponentially distributed with unit mean and pdf f|hi |2 (x) = e−x .
Theorem 3. The Mellin transform for a MISO system employing TAS is given by
TAS
Mg(γ) (s) = M ζ
s−1
k+1
ζ
k+1
M −1
k e
Γ s,
.
(−1)
(k + 1)s
ζ
k
M
−1
X
k=0
(27)
Proof: See Appendix C.
C. Other applications of Theorem 1
In this section we briefly point out towards other possible applications of Theorem 1.
1) SISO case with Nakagami-m fading: Theorem 1 could easily be used to analyze the SISO
case with Nakagami-m fading. The Nakagami-m distribution includes as special cases Rayleigh
(m = 1), no fading (m → ∞), and the Ricean distribution for m = (K + 1)2 /(2K + 1) where
K is the Ricean factor. When the envelope of the received signal is Nakagami-m distributed,
the instantaneous SNR is gamma distributed with shape parameter m and rate parameter snr−1 ,
thus its Mellin transform is simply Mg(γ) (s) = snr−m · U (m, s + m, snr−1 ).
2) MIMO eigen-beamforming: Consider now that the receiver is equipped with N receive
antennas and we perform eigen-beamforming at both transmitter and receiver ends. In order
14
to maximize the SNR at the receiver, the transmit weighting vector w is selected to be the
eigenvector of the Wishart matrix HH H which corresponds to the largest eigenvalue φmax of
HH H, i.e. γmimo = snrφmax .
Corollary 2. The Mellin transform of the service process in the case of MIMO MRT in Rayleigh
fading is given by
MIMO
Mg(γ) (s) = R
)i−2i
N (M +N
X
X
i=1
2
m=M −N
m+1
1
m! · ci,m
· U (m + 1, m + 1 + s, i/ζ)
ζ
(28)
Q
−1
where R = ( N
and coefficients ci,m can be obtained from [26]; for the
k=1 (N − k)!(M − k)!)
perfect CSI case, they have been tabulated for some values of {N, M } [37, Table I-IV]. This
result is a direct consequence of [26, Eq. 8] and (22).
Note that since the largest eigenvalue of the complex Wishart matrix (or equivalently the maximum singular value of H) is bounded by
kHk2F
min(M,N )
≤ φmax ≤ kHk2F and kHk2F ∼ Gamma(M N, 1),
simple upper and lower bounds for the Mellin transform of MIMO eigen-beamforming can be
obtained particularizing Theorem 1.
V. D ELAY WITH L ARGE B LOCKLENGTH : A SYMPTOTIC A NALYSIS
In the previous section, we have provided analytical expressions for the Mellin transform
of the service process and the kernel for various multi-antenna diversity techniques. The exact
results are mainly given in terms of special functions and alternating series. To explore further
the delay performance of MISO MRT, we derive in this section simplified expressions for various
asymptotic regimes: low/high SNR and large M . Additionally, we obtain a general result for a
Gaussian distributed service process; as we will show, the MISO MRT service process converges
to this distribution as M grows large.
A. High SNR regime
We study here how latency constraints affect the MISO performance at high SNR. We assume
this implies also large ζ, which is true as long as σe2 does not increase2 with the SNR (see (2)).
2
As a matter of fact, most frequently and in practice σe2 ∝ 1/snr.
15
Corollary 3. In the high SNR regime, the Mellin transform of the service process scales as
s−1 Γ(s+M −1)
s>1−M
Γ(M )
ζ
)
(29)
MHg(γ) (s) =
ζ −M Γ(1−s−M
s≤1−M
Γ(1−s)
ζ −M log ζ−ψ(M ) s = 1 − M
Γ(M )
where ψ(x) denotes the Digamma function [35, Sec. 6.3].
Proof: The three branches are obtained after direct application of the asymptotic properties
of the U (a, b, z) function listed in [35, Sec. 13.5]. The first branch can be also derived by considering the approximated service process si ≈ log(γi ), which gives that MHg(γ) (s) = ζ s−1 Γ(s+M −1)
for s + M > 1.
B. Low SNR regime
At low SNR, the service process can be approximated as si ≈ γi and the following result is
obtained.
Corollary 4. In the low SNR regime, the Mellin transform of the service process is approximately
given as
MLg(γ) (s) = (1 − (s − 1)ζ)−M ,
s < ζ −1 − 1.
(30)
Proof: At low SNR, we use the first order Taylor series expansion log(1 + x) ≈ x. In that
case, the service process can be approximated as si ≈ γi , which gives that g(γ) ≈ eγ , and in
consequence
h
i
MLg(γ) (s) = E eγ (s−1) = (1 − (s − 1)ζ)−M
(31)
using the moment generating function (MGF) of a gamma random variable.
C. Large antenna regime
The distribution of the mutual information of a Rayleigh fading MIMO system is generally
rather complicated. For this reason, approximations have been used in the literature. For example,
in the large antenna regime (M → ∞) and using the Central Limit Theorem (CLT), it can be
shown that the distribution of the mutual information I converges to a Gaussian distribution; see
for instance [38] and references therein. Using similar arguments here, we can obtain simpler
16
expressions for the Mellin transform of the service process. In general, we can obtain results of
the form
Mα
I −µ d
→ N (0, 1)
σM
(32)
where convergence is in distribution, µ = E(I), σM is a variance term, and α is a measure of
the convergence speed (normally 0.5). This means that, for large M , an accurate approximation
2
of the distribution is given by I ∼ N (µ, σ 2 ) with σ 2 = σM
/M α . The mean and the variance
terms can be obtained in closed form from [39]. Note that, for brevity, throughout this section
we will use the natural logarithm, and thus all rates are in nats.
Thanks to the CLT arguments and the Gaussian approximation of the service process, i.e.
si ≈ Ii , we arrive at the following result.
Theorem 4. The Mellin transform of a service process with rate following a Gaussian distribution
with mean µ and variance σ 2 is given by
2 σ2
2
Masg(γ) (s) = e(s−1)µ+(s−1)
.
(33)
Proof: The service is given in terms of log(g(γ)) (bit domain), thus we have that g(γ) = eIi ,
and the result follows immediately by solving
Masg(γ) (s) = E e(s−1)Ii
Z ∞
e(s−1)x fI (x) dx
=
(34)
(35)
−∞
2 σ2
2
= e(s−1)µ+(s−1)
.
(36)
Theorem 5. For the MISO MRT case, as the number of antennas grows large, we have
lim Masg(γ) (s) → (1 + ζM )s−1 .
M →∞
(37)
Proof: The mutual information can be written as I = log(1 + γ). Rewriting (33) and
17
applying Jensen’s inequality
Masg(γ) (s)
2 σ2
2
≈
e(s−1) E[log(1+γ)] · e(s−1)
≤
e(s−1) log(1+E[γ]) · e(s−1)
=
(1 + ζM )s−1 · e(s−1)
(40)
(1 + ζM )s−1
(41)
M →∞
=
2 σ2
2
2 σ2
2
(38)
(39)
where the last equality follows from the fact that lim σ 2 = 0 [38], [39].
M →∞
It can be shown that the bound is asymptotically tight (using Prohorov-Le Cam theorem, continuous mapping theorem, and Chebyshev inequality), but the proof is standard and is omitted for
the sake of brevity. Furthermore, the asymptotic convergence can be obtained without resorting to
the Gaussian approximation by showing that convergence in distribution implies convergence in
Mg(γ) (s). Let y1 , y2 , . . . be a sequence of positive random variables that converges in distribution
to a positive random variable y. For s > 0, we have lim Myi (s) = My (s). By Lebesgue’s
dominated convergence theorem and
P
khk2
→
E[khk2 ]
M →∞
1, we have that lim Mg(γ) (s) = (1 + ρM )s−1 .
M →∞
s−1
Interestingly, we observe that for large M , Mas
, which is related to the sog(γ) (s) ∼ (ζM )
called channel hardening effect, i.e. the channel behaves equivalently to an AWGN channel with
SNR ζM . In the low SNR regime, the number of transmit antennas affects linearly the service
process, while at high SNR, the Mellin transform of the service process grows superlinearly
with M (for s > 1).
The approximation si ∼ N (µ, σ 2 ) allows us to simplify the delay violation probability
expression (23), however its relevance and applicability goes beyond, as it allows analyzing
the delay violation probability of any system whose service rate can be approximated by a
Gaussian random variable. Additionally, it provides very simple expressions for the effective
capacity, as we show next.
D. Effective Capacity
Effective capacity is defined as the maximum constant arrival rate that a system can support
given a QoS requirement θ [12]. A byproduct of the delay analysis using MGF-based stochastic
network calculus is that we can obtain expressions for the effective capacity R by noticing that
1
·
R(θ) = − log Mg(γ) (1 − θ),
θ
θ > 0.
(42)
18
As an example, taking the normalized logarithm in (22), we can recover the effective capacity
results in [19]. However, to assess the effect of multiple antennas in the delay-constrained performance, simpler expressions would be beneficial. We thus focus on the Gaussian approximation
in Sec. V-C and obtain
θ
Ras (θ) = µ − σ 2
2
(43)
= E [log(1 + γ)] −
θ
Var[log(1 + γ)] .
2
(44)
As expected, the effective capacity converges to the ergodic capacity in the absence of delay
constraints (θ = 0). For general θ, there is a penalty on the achievable rate that is proportional
to the variance of the instantaneous rate log(1 + γ). This implies that, as the number of antennas
tends to infinity, such penalty vanishes because the variance of the rate tends to zero [38] and
the effective capacity does not decay as θ increases.
VI. D ELAY A NALYSIS WITH F INITE B LOCKLENGTH
We investigate now the effect of finite blocklength in the service process and the delay
performance. As explained in Sec. II, at finite blocklength there is always a probability of error
and a rate loss as compared to Shannon capacity; in case of transmission errors, the offered
service is zero. Therefore, the service process can be modeled as si = Ri (n, ) · Zi , where the
coding rate Ri (n, ) is approximated using (3) and Zi is a Bernoulli random variable, being one
in case of successful transmission (with probability 1 − ), and zero otherwise; in this work, we
assume independence between Zi and γi (i.e. non-varying with SNR).
Using (3) and lower bounding the achievable rate by zero (for very low SNR values), the
Mellin transform of g(γi , Zi ) is given by
MFB
g(γi ,Zi ) (s) = (1 − )Mq(γi ) (s) + ,
where q(γi ) = max
√1+γi
e V (γi )F
(45)
, 1 with F = n−1/2 Q−1 (), and Mq(γi ) (s) is given by
"
Mq(γi ) (s) = E max
Z
=
e
φ
V (γi )F
Z
fγ (x)dx +
0
=
s−1 #
1 + γi
√
,1
φ
∞
1+γ
√ i
e V (γ)F
γ(M, φ/snr)
+ I(c, s, F )
Γ(M )
s−1
fγ (x)dx
(46)
19
0
where φ = eF − 1 is the point where max q(γi ) > 1.
The major difficulty in deriving Mq(γi ) (s) is the fact that the channel dispersion V (γ) depends
p
on γi . In [24], an infinite-order Taylor series expansion for V (γ) and the series expansion of the
exponential function are used. These results can be easily extended in the MISO case, however
the expression would be even more involved due to gamma distributed channel gains, providing
little or no insight. For that, we numerically evaluate the integral when required and focus on a
simpler asymptotic expression at high SNR.
A. High SNR regime
At high SNR, the channel dispersion can be approximated as V (γ) ≈ 1 and the Mellin
transform of q(γi ) is approximately
"
s−1 #
1 + γi
Mq(γi ) (s) ≈ E max
,1
eF
s−1
Z eF −1
Z ∞
1 + γi
=
fγ (x)dx +
fγ (x)dx
eF
0
eF −1
γ(M, (eF − 1)/snr)
Γ(M )
M −1
e1/ζ ζ −M X M − 1
+
(−1)M −1−j ζ j+s Γ j + s, eF ζ −1
F
(s−1)
Γ(M )e
j
j=0
=
(47)
where for the last equality we have followed the same procedure used to obtain Theorem 2
(see Appendix B). Note that for F = 0, we recover the Mellin transform of the service
process with infinite blocklength. This approximation, which requires only widely used standard
special functions, is easier to evaluate numerically. In Section VII we show that its accuracy is
satisfactory even for moderate values of SNR.
VII. N UMERICAL R ESULTS
In this section, we provide numerical evaluation of the performance of MISO communication
systems based on the above analysis. Unless otherwise stated, the duration of a slot is set to
T = 1 ms, the overhead is disregarded (nm → 0), and the blocklength is assumed to be n = 168;
consequently B = n/ log 2 6= 1, and we reincorporate this parameter into the equations.
We start by validating our analysis with Monte Carlo simulations. In Figure 1, we compare the
delay violation probability and its bound with ρ = 24 kbps and snr = −2 dB. We corroborate
20
10
P(delay > ω)
10
0
Bounds
Monte Carlo
-1
10-2
M=1
10
-3
10
-4
M=2
10-5
10
-6
0
1
2
ω (ms)
3
4
Figure 1. Delay violation probability and associated bounds as a function of the target dealy, ρ = 24 kbps and snr = −2 dB.
that the bounds follow the trend of the original curve, and we point out that the maximum
difference in the x-axis seems to be of about 1 ms.
Figure 2 plots the violation bound for MISO MRT as a function of the target delay w with ρ =
24 kbps and snr = 5 dB. The plot on the left shows the effect of varying the number of antennas
and the accuracy of the CSI. We observe the strong decrease of the delay violation probability
when increasing the number of antennas: with perfect CSI, the probability of exceeding 1 ms
delay roughly decreases by three orders of magnitude when adding an extra antenna. On the
other hand, the plot on the right depicts the difference between assuming finite and infinite
blocklength; similar to [24], we can see that such difference is remarkable, and that the Shannon
model substantially overestimates the performance of the system.
In Figure 3, we compare the delay performance of MISO MRT with OSTBC and TAS. We
can see that MRT generally performs better when the quality of the CSI is good: above a certain
value of σe2 , TAS and OSTBC outperform MRT. The values at which this change takes place
seem to be dependent on the number of antennas.
To obtain results with finite blocklength we must set an error probability . In the experiment
above, we have used = 10−2 for M = 1, = 10−4 for M = 2, and = 10−5 for M = 3; these
parameters have been set with the inspiration of Figure 4, which illustrates the importance of
choosing wisely depending not only on the SNR but also on the number of antennas.
21
100
100
M=1
M=1, FB
M=2
M=2, FB
M=3
M=3, FB
σ2e = 0
10-2
σ2e = 0.1
10-2
2
σe = 0.25
10
10-4
σ2e = 0.4
p v(ω)
pv (ω)
-4
10-6
10-6
M=1
-8
10-8
10
M=2
M=3
-10
10
0
1
2
3
4
5
10-10
0
1
2
ω (ms)
3
4
5
ω (ms)
Figure 2. Delay violation probability bound as a function of the target delay, ρ = 24 kbps and snr = 5 dB. Curves labeled FB
have been obtained using finite blocklength expressions.
100
σ2e = 0
10-2
σ2e = 0.25
σ2e = 0.4
-4
OSTBC
TAS
pv (ω)
10
10-6
10-8
M=2
M=4
-10
10
0
0.5
1
1.5
ω
2
2.5
3
Figure 3. Delay violation probability bound as a function of the target delay for different diversity techniques, ρ = 24 kbps
and snr = 5 dB.
In Figure 5, we investigate further the effect of adding antennas, and compare it to that of
increasing the power. For a target delay of 1 ms at 0 dB, we can see that going from three to
four antennas seems to have only slightly less impact than doubling the power; at 5 dB, however,
this is not the case anymore: 3 dB of extra power decrease the violation probability by one order
of magnitude, but adding one antenna decreases it by two orders of magnitude.
As explained in Section IV, it is important to have simple expressions of the kernel when
22
100
pv (3 ms)
M=1
M=2
10-5
-10
M=3
10
10-5
10-4
10-3
ǫ
ζ = 2 dB
ζ = 5 dB
10-2
10-1
Figure 4. Bound on the probability of exceeding 3 ms of delay as a function of the block error rate , finite blocklength analysis,
ρ = 24 kbps. Circles mark the minimum of each curve.
100
ζ = 0 dB
ζ = 3 dB
ζ = 5 dB
ζ = 8 dB
ζ = 15 dB
-2
10
p v (1 ms)
10-4
10-6
10-8
10-10
10-12
2
4
6
8
10
M
Figure 5. Bound on the probability of exceeding 1 ms delay as a function of the number of antennas, asymptotically large
blocklength, ρ = 256 kbps.
23
100
105
100
-4
10
K(s)
Mg (1-sB)
10-2
10-6
10-8
10-10
10-5
10-5
Exact
Gaussian approx.
-4
10
-3
10
s
Exact
Gaussian approx.
-2
10
10-10
-1
10
0
0.2
0.4
0.6
0.8
s
Figure 6. Mellin transform (left) and kernel (right) as a function of s, M = 3, ρ = 20 kbps, ζ = 0 dB, w = 1.
100
105
Exact
Gaussian approx.
-2
100
ρ= 540 kbps
-4
10
K(s)
Mg (1-sB)
10
10-5
10-6
10-8
10-10
Exact
Gaussian approx.
ρ= 240 kbps
-10
10
10-5
-15
10-4
10-3
s
10-2
10-1
10
0
0.05
0.1
0.15
s
Figure 7. Mellin transform (left) and kernel (right) as a function of s, M = 10, ζ = 0 dB, w = 1.
possible. In Figure 6 and Figure 7, we illustrate the accuracy of the Gaussian approximation for
M = 3 and M = 10; as expected, the error is large for the former and negligible for the latter.
This justifies the use of the much simpler expression (33) whenever M is relatively large.
In Figure 8 (left) we test the accuracy of the high and low SNR approximations derived
for the Shannon model in Section VI-A and Section V-B. We can see that the high SNR
approximation becomes asymptotically tight as the SNR increases, and that, remarkably, the
low SNR approximation is reasonably accurate for most SNR values; this makes the low SNR
approximation particularly interesting given its simplicity, see (30). Similarly, in Figure 8 (right)
we show the accuracy of the high SNR approximation derived for the finite blocklength model
24
100
100
M=1
M=2
M=3
10-2
M=1
-5
pv (2 ms)
pv (2 ms)
10
M=2
-10
10-4
10-6
10
High SNR approx.
Exact
Low SNR approx.
10-15
-5
0
5
ζ (dB)
10-8
M=3
10
15
10-10
HSNR approx.
Exact
-5
0
5
ζ (dB)
10
15
Figure 8. Bound on the probability of exceeding 2 ms delay, asymptotically large blocklength (left) and finite blocklength
(right) analysis, ρ = 200 kbps.
100
M=1
10-4
v
p (2 ms)
10-2
M=2
10-6
Infinite blocklength
Finite blocklength
-8
10
0
100
200
300
400
500
Blocklength n
Figure 9. Bound on the probability of exceeding 2 ms delay as a function of the blocklength n, ρ = 150 kbps, ζ = 10 dB.
Circles mark the minimum of each curve.
in Section VI-A.
Finally, we show the effect of varying the blockelength n. We assume now a constant overhead
of nm = 64 symbols, so that in each time slot a total of nm + n symbols are transmitted. The
duration a time slot is now (n+nm )/168 ms. As we can see from Figure 9, the delay performance
heavily depends on the blocklength chosen, and the optimum value changes with the number of
antennas.
25
VIII. C ONCLUSIONS
In this work, we characterized the delay performance of MISO diversity communications
under statistical delay constraints. Using stochastic networks calculus, we derived a statistical
characterization of the service process in multi-antenna fading channels and provided probabilistic delay bounds. We showed how the number of transmit antennas and transmit SNR may affect
the delay performance. We also investigated the impact of imperfect CSI at the transmitter and
finite blocklength channel coding on the delay performance of MISO transmit beamforming.
MISO MRT is shown to reduce the delay violation probability as compared to single-antenna
transmissions even with imperfect CSI. Nevertheless, as channel imperfections increase, other
diversity-techniques, such as OSTBC and antenna selection, perform better than MRT in terms of
delay violation probability. Future work could consider the effect of imperfect CSI at the receiver
and limited feedback in FDD MIMO systems. Further extensions of this framework may include
the analysis of MIMO spatial multiplexing, MIMO channels with co-channel interference, and
multiuser MIMO systems.
A PPENDIX A
P ROOF OF T HEOREM 1
−x
Recall that γ is gamma distributed with pdf fγ (x) =
xM −1 e ζ
Γ(M )ζ M
, x ≥ 0. Then, we have that
·
Mg(γ) (s) = E (1 + γ)s−1
Z ∞
=
(1 + x)s−1 fγ (x) dx
0
Z ∞
1
=
(1 + x)s−1 xM −1 e−x/ζ dx
M
ζ Γ(M ) 0
(a)
= ζ −M U (M, M + s, ζ −1 )
(48)
(49)
(50)
(51)
where (a) follows from the definition of Tricomi’s confluent hypergeometric function [35, Eq.
13.2.5]
−1
Z
U (a, b, z) = Γ(a)
0
∞
e−zt ta−1 (t + 1)b−a−1 dt.
(52)
26
A PPENDIX B
P ROOF OF T HEOREM 2
Equation (50) can be rewritten as
1
Mg(γ) (s) = M
ζ Γ(M )
1
eζ
= M
ζ Γ(M )
Z
∞
(1 + x)s−1 xM −1 e−x/ζ dx
(53)
(t − 1)M −1 ts−1 e−t/ζ dt
(54)
0
Z
∞
1
by applying change of variables t = 1 + x. Now, since M is a positive integer, we can use the
binomial theorem to obtain
1
Z ∞
M
−1
X
eζ
M −1
M −1−j
Mg(γ) (s) = M
(−1)
·
ts+j−1 e−t/ζ dt
j
ζ Γ(M ) j=0
1
1
M
−1
X
eζ
M −1
= M
(−1)M −1−j · ζ j+s · Γ(j + s, ζ −1 ).
ζ Γ(M ) j=0
j
(55)
(56)
Or, alternatively,
1
Mg(γ) (s) = e ζ
M
−1
X
(−1)M −1−j · ζ j+s−M
j=0
Γ(j + s, ζ −1 )
.
Γ(M − j)Γ(j + 1)
(57)
A PPENDIX C
P ROOF OF T HEOREM 3
Suppose that X1 , . . . , Xn are n independent continuous variates, each with cdf F (x) and pdf
f (x). The pdf of the r-th order statistic X(r) , r = 1, . . . , n is given by [40]
f(r) (x) =
1
F r−1 (x)[1 − F (x)]n−r f (x).
B(r, n − r + 1)
(58)
27
Therefore, the pdf of γmax = γ(M ) is given by fγmax (x) = M fγ (x)FγM −1 (x). Since γ ∼ Exp(1)
in the case of TAS with pdf fγ (x) = e−x , we have that
·
s−1
MTAS
g(γ) (s) = E (1 + ζγmax )
Z ∞
=
(1 + ζx)s−1 fγmax (x)dx
0
Z ∞
(1 + ζx)s−1 fγ (x)FγM −1 (x)dx
= M
Z0 ∞
(1 + ζx)s−1 e−x (1 − e−x )M −1 dx
= M
0
M
−1
X
Z ∞
M −1
k
(1 + ζx)s−1 e−x(k+1) dx
= M
(−1)
k
0
k=0
k+1
M
−1
X
ζ
M −1
k+1
s−1
k e
= Mζ
(−1)
Γ s,
k
(k + 1)s
ζ
k=0
(a)
(59)
(60)
(61)
(62)
(63)
(64)
where (a) follows from applying binomial theorem.
R EFERENCES
[1] A. Ephremides and B. Hajek, “Information theory and communication networks: an unconsummated union,” IEEE Trans.
Inf. Theory, vol. 44, no. 6, pp. 2416–2434, Oct. 1998.
[2] Y. Polyanskiy, H. V. Poor, and S. Verdu, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inf. Theory,
vol. 56, no. 5, pp. 2307–2359, May 2010.
[3] C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Journal, vol. 27, no. 3, pp. 379–423,
July 1948.
[4] R. G. Gallager, Information Theory and Reliable Communication. New York, NY, USA: John Wiley & Sons, Inc., 1968.
[5] S. V. Hanly and D. N. C. Tse, “Multiaccess fading channels-part II: Delay-limited capacities,” IEEE Trans. Inf. Theory,
vol. 44, no. 7, pp. 2816–2831, Nov. 1998.
[6] L. H. Ozarow, S. Shamai, and A. D. Wyner, “Information theoretic considerations for cellular mobile radio,” IEEE Trans.
Veh. Technol., vol. 43, no. 2, pp. 359–378, May 1994.
[7] A. E. Gamal, J. Mammen, B. Prabhakar, and D. Shah, “Throughput-delay trade-off in wireless networks,” in Proc. IEEE
INFOCOM, Mar. 2004.
[8] C.-S. Chang, Performance Guarantees in Communication Networks.
[9] Y. Jiang and Y. Liu, Stochastic Network Calculus.
London, UK: Springer-Verlag, 2000.
London, UK: Springer-Verlag, 2008.
[10] M. Fidler and A. Rizk, “A guide to the stochastic network calculus,” IEEE Commun. Surveys Tuts., vol. 17, no. 1, pp.
92–105, First quarter 2015.
[11] I. H. Hou, V. Borkar, and P. R. Kumar, “A theory of QoS for wireless,” in Proc. IEEE INFOCOM, Apr. 2009, pp. 486–494.
[12] D. Wu and R. Negi, “Effective capacity: a wireless link model for support of quality of service,” IEEE Trans. Wireless
Commun., vol. 2, no. 4, pp. 630–643, July 2003.
[13] R. L. Cruz, “A calculus for network delay, part I: Network elements in isolation,” IEEE Trans. Inf. Theory, vol. 37, no. 1,
pp. 114–131, Jan. 1991.
28
[14] S. Zhou, K. Zhang, Z. Niu, and Y. Yang, “Queuing analysis on MIMO systems with adaptive modulation and coding,” in
Proc. IEEE Inter. Conf. on Commun. (ICC), May 2008, pp. 3400–3405.
[15] S. Kittipiyakul and T. Javidi, “Optimal operating point for MIMO multiple access channel with bursty traffic,” IEEE Trans.
Wireless Commun., vol. 6, no. 12, pp. 4464–4474, Dec. 2007.
[16] S. Kittipiyakul, P. Elia, and T. Javidi, “High-SNR analysis of outage-limited communications with bursty and delay-limited
information,” IEEE Trans. Inf. Theory, vol. 55, no. 2, pp. 746–763, Feb. 2009.
[17] J. Chen and V. K. N. Lau, “Large deviation delay analysis of queue-aware multi-user MIMO systems with two-timescale
mobile-driven feedback,” IEEE Trans. Signal Process., vol. 61, no. 16, pp. 4067–4076, Aug. 2013.
[18] M. C. Gursoy, “MIMO wireless communications under statistical queueing constraints,” IEEE Trans. Inf. Theory, vol. 57,
no. 9, pp. 5897–5917, Sep. 2011.
[19] M. Matthaiou, G. C. Alexandropoulos, H. Q. Ngo, and E. G. Larsson, “Analytic framework for the effective rate of MISO
fading channels,” IEEE Trans. Commun., vol. 60, no. 6, pp. 1741–1751, Jun. 2012.
[20] M. You, H. Sun, J. Jiang, and J. Zhang, “Unified framework for the effective rate analysis of wireless communication
systems over MISO fading channels,” IEEE Trans. Commun., vol. 65, no. 4, pp. 1775–1785, Apr. 2017.
[21] K. Mahmood, A. Rizk, and Y. Jiang, “On the flow-level delay of a spatial multiplexing MIMO wireless channel,” in Proc.
IEEE Inter. Conf. on Commun. (ICC), June 2011, pp. 1–6.
[22] K. Mahmood, M. Vehkapera, and Y. Jiang, “Delay constrained throughput analysis of a correlated MIMO wireless channel,”
in Proc. 20th Inter. Conf. on Comp. Commun. and Netw. (ICCCN), July 2011, pp. 1–7.
[23] H. Al-Zubaidy, J. Liebeherr, and A. Burchard, “A (min, x) network calculus for multi-hop fading channels,” in Proc. IEEE
INFOCOM, April 2013, pp. 1833–1841.
[24] S. Schiessl, J. Gross, and H. Al-Zubaidy, “Delay analysis for wireless fading channels with finite blocklength channel
coding,” in Proc. 18th ACM Inter. Conf. on Model., Anal. and Simul. of Wirel. and Mob. Systems (MSWiM ’15), Nov.
2015.
[25] T. K. Y. Lo, “Maximum ratio transmission,” IEEE Trans. Commun., vol. 47, no. 10, pp. 1458–1461, Oct. 1999.
[26] Y. Chen and C. Tellambura, “Performance analysis of maximum ratio transmission with imperfect channel estimation,”
IEEE Commun. Lett., vol. 9, no. 4, pp. 322–324, Apr. 2005.
[27] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser MIMO achievable rates with downlink training and
channel state feedback,” IEEE Trans. Inf. Theory, vol. 56, no. 6, pp. 2845–2866, Jun. 2010.
[28] W. Yang, G. Durisi, T. Koch, and Y. Polyanskiy, “Quasi-static multiple-antenna fading channels at finite blocklength,”
IEEE Trans. Inf. Theory, vol. 60, no. 7, pp. 4232–4265, July 2014.
[29] M. Hayashi, “Information spectrum approach to second-order coding rate in channel coding,” IEEE Trans. Inf. Theory,
vol. 55, no. 11, pp. 4947–4966, Nov. 2009.
[30] S. Schiessl, H. Al-Zubaidy, M. Skoglund, and J. Gross, “Analysis of wireless communications with finite blocklength
and imperfect channel knowledge,” 2016. [Online]. Available: http://arxiv.org/abs/1608.08445
[31] Y. Jiang and P. J. Emstad, Analysis of Stochastic Service Guarantees in Communication Networks: A Server Model,
H. de Meer and N. Bhatti, Eds.
Berlin, Heidelberg: Springer, 2005.
[32] M. Fidler, “An end-to-end probabilistic network calculus with moment generating functions,” in Proc. 14th IEEE
Inter.Workshop on Quality of Service, June 2006, pp. 261–270.
[33] ——, “A network calculus approach to probabilistic quality of service analysis of fading channels,” in Proc. IEEE Global
Commun. Conf. (Globecom), Nov. 2006.
[34] N. Petreska, H. Al-Zubaidy, and J. Gross, “Power minimization for industrial wireless networks under statistical delay
constraints,” in Proc. 26th International Teletraffic Congress (ITC), Sep. 2014.
29
[35] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables.
New York: Dover, 1964.
[36] A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time Wireless Communications, 1st ed. New York, NY, USA:
Cambridge University Press, 2003.
[37] P. A. Dighe, R. K. Mallik, and S. S. Jamuar, “Analysis of transmit-receive diversity in Rayleigh fading,” IEEE Trans.
Commun., vol. 51, no. 4, pp. 694–703, Apr. 2003.
[38] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, “Multiple-antenna channel hardening and its implications for rate feedback
and scheduling,” IEEE Trans. Inf. Theory, vol. 50, no. 9, pp. 1893–1909, Sep. 2004.
[39] M. R. McKay, P. J. Smith, H. A. Suraweera, and I. B. Collings, “On the mutual information distribution of OFDM-based
spatial multiplexing: Exact variance and outage approximation,” IEEE Trans. Inf. Theory, vol. 54, no. 7, pp. 3260–3278,
July 2008.
[40] H. A. David and H. N. Nagaraja, Order Statistics.
Hoboken, NJ, USA: John Wiley & Sons, Inc., 2005.
| 7cs.IT
|
Actors without Borders: Amnesty for Imprisoned State
Elias Castegren
Tobias Wrigstad
Uppsala University, Sweden
In concurrent systems, some form of synchronisation is typically needed to achieve data-race
freedom, which is important for correctness and safety. In actor-based systems, messages
are exchanged concurrently but executed sequentially by the receiving actor. By relying on
isolation and non-sharing, an actor can access its own state without fear of data-races, and
the internal behavior of an actor can be reasoned about sequentially.
However, actor isolation is sometimes too strong to express useful patterns. For example,
letting the iterator of a data-collection alias the internal structure of the collection allows
a more efficient implementation than if each access requires going through the interface of
the collection. With full isolation, in order to maintain sequential reasoning the iterator must
be made part of the collection, which bloats the interface of the collection and means that a
client must have access to the whole data-collection in order to use the iterator.
In this paper, we propose a programming language construct that enables a relaxation of
isolation but without sacrificing sequential reasoning. We formalise the mechanism in a simple
lambda calculus with actors and passive objects, and show how an actor may leak parts of its
internal state while ensuring that any interaction with this data is still synchronised.
1
Introduction
Synchronisation is a key aspect of concurrent programs and different concurrency models handle
synchronisation differently. Pessimistic models, like locks or the actor model [1] serialise computation within certain encapsulated units, allowing sequential reasoning about internal behavior.
In the case of the actor model, for brevity including also active objects (which carry state,
which actor’s traditionally do not), if a reference to an actor A’s internal state is accessible
outside of A, operations inside of A are subject to data-races and sequential reasoning is lost.
The same holds true for operations on an aggregate object behind a lock, if a subobject is leaked
and becomes accessible where the appropriate lock is not held.
In previous work, we designed Kappa [4], a type system in which the boundary of a unit of
encapsulation can be statically identified. An entire encapsulated unit can be wrapped inside some
synchronisation mechanism, e.g., a lock or an asynchronous actor interface, and consequently
all operations inside the boundary are guaranteed to be data-race free. An important goal of
this work is facilitating object-oriented reuse in concurrent programming: internal objects are
oblivious to how their data-race freedom is guaranteed, and the building blocks can be reused
without change regardless of their external synchronisation.
This extended abstract explores two extensions to this system, which we explain in the
context of the actor model (although they are equally applicable to a system using locks). Rather
than rejecting programs where actors leak internal objects, we allow an actor to bestow its
synchronisation mechanism upon the exposed objects. This allows multiple objects to effectively
construct an actor’s interface. Exposing internal operations externally makes concurrency more
fine-grained. To allow external control of the possible interleaving of these operations, we introduce
an atomic block that groups them together. The following section motivates these extensions.
V.T. Vasconcelos and P. Haller (Eds.): Workshop on Programming Language
Approaches to Concurrency- and Communication-cEntric Software (PLACES’17)
EPTCS 246, 2017, pp. 10–20, doi:10.4204/EPTCS.246.4
E. Castegren & T. Wrigstad
class Node[t]
var next : Node[t]
var elem : t
// getters and setters omitted
actor List[t]
var first : Node[t]
def getFirst() : Node[t]
return this.first
def get(i : int) : t
var current = this.first
while i > 0 do
current = current.next
i = i - 1
return current.elem
(a)
11
class Iterator[t]
var current : Node[t]
def init(first : Node[t]) : void
this.current = first
def getNext() : t
val elem = this.current.elem
this.current = this.current.next
return elem
def hasNext() : bool
return this.current != null
actor List[t]
def getIterator() : Iterator[t]
val iter = new Iterator[t]
iter.init(this.first)
return iter
(b)
Figure 1: (a) A list implemented as an actor. (b) An iterator for that list.
2
Breaking Isolation: Motivating Example
We motivate breaking isolation in the context of an object-oriented actor language, with actors
serving as the units of encapsulation, encapsulating zero or more passive objects. Figure 1a shows
a Kappa program with a linked list in the style of an actor with an asynchronous external interface.
For simplicity we allow asynchronous calls to return values and omit the details of how this is
accomplished (e.g., by using futures, promises, or by passing continuations).
Clients can interact with the list for example by sending the message get with a specified
index. With this implementation, each time get is called, the corresponding element is calculated
from the head of the list, giving linear time complexity for each access. Iterating over all the
elements of the list has quadratic time complexity.
To allow more efficient element access, the list can provide an iterator which holds a pointer
to the current node (Figure 1b). This allows constant-time access to the current element, and
linear iteration, but also breaks encapsulation by providing direct access to nodes and elements
without going through the list interface. List operations are now subject to data-races.
A middle ground providing linear time iteration without data-races can be implemented
by moving the iterator logic into the list actor, so that the calls to getNext and hasNext are
synchronised in the message queue of the actor. This requires a more advanced scheme to map
different clients to different concurrent iterators, clutters the list interface, creates unnecessary
coupling between List and Iterator, and complicates support of e.g., several kinds of iterators.
Another issue with concurrent programs is that interleaving interaction with an actor makes
it hard to reason about operations that are built up from several smaller operations. For example,
a client might want to access two adjacent nodes in the list and combine their elements somehow.
When sending two get messages, there is nothing that prevents other messages from being
processed by the list actor after the first one, possibly removing or changing one of the values.
12
Actors without Borders: Amnesty for Imprisoned State
actor List[t]
...
def getIterator() : B(Iterator[t])
val iter = new Iterator[t]
iter.init(this.first)
return bestow iter
val iter = list!getIterator()
while iter!hasNext() do
val elem = iter!getNext()
...
Figure 2: A list actor returning a bestowed iterator, and the code for a client using it
Again, unless the list actor explicitly provides an operation for getting adjacent values, there is
no way for a client to safely express this operation.
3
Bestowing and Grouping Activity
Encapsulating state behind a synchronisation mechanism allows reasoning sequentially about
operations on that state. However, since Kappa lets us identify the encapsulation boundary of
the data structure [4], it is possible to bestow objects that are leaked across this boundary with a
synchronisation wrapper. Statically, this means changing the type of the returned reference to
reflect that operations on it may block. Dynamically it means identifying with what and how the
leaked object shall synchronise.
For clarity, we explicate this pattern with a bestow operation. In the case of actors, an actor
a that performs bestow on some reference r creates a wrapper around r that makes it appear like
an actor with the same interface as r, but asynchronous. Operations on the bestowed reference
will be relayed to a so that the actor a is the one actually performing the operation. If r was
leaked from an enclosure protected by a lock l, r’s wrapper would instead acquire and release l
around each operation.
Figure 2 shows the minimal changes needed to the code in Figure 1b, as well as the code for
a client using the iterator. The only change to the list is that getIterator() returns a bestowed
iterator (denoted by wrapping the return type in B(...)1 ), rather than a passive one. In the
client code, synchronous calls to hasNext() and getNext() become asynchronous message sends.
These messages are handled by the list actor, even though they are not part of its interface. This
means that any concurrent usages of iterators are still free from data-races.
It is interesting to ponder the difference between creating an iterator inside the list and
bestowing it, or creating an iterator outside the list, and bestowing each individual list node it
traverses. In the former case, getNext() is performed without interleaved activities in the same
actor. In the latter case, it is possible that the internal operations are interleaved with other
operations on list. The smaller the object returned, the more fine-grained is the concurrency.
Sometimes it is desirable that multiple operations on an object are carried out in a noninterleaved fashion. For this purpose, we use an atomic block construct that operates on a an
actor or a bestowed object, cf. Figure 3. In the case of operations on an actor, message sends
inside an atomic block are batched and sent as a single message to the receiver. In the case of
operations on an object guarded by a lock, we replace each individual lock–release by a single
lock–release wrapping the block. It is possible to synchronise across multiple locked objects in a
single block.
1
If desired, this type change can be implicit through view-point adaptation [9].
E. Castegren & T. Wrigstad
class Iterator[t]
var current : B(Node[t])
def getNext() : t
val elem = this.current ! elem()
// Possible interleaving of other messages
this.current = this.current ! next()
return elem
13
class Iterator[t]
var current : B(Node[t])
def getNext() : t
atomic c <- this.current
val elem = c ! elem()
this.current = c ! next()
return elem
Figure 3: Fine-grained (left) and coarse-grained (right) concurrency control.
An atomic block allows a client to express new operations by composing smaller ones. The
situation sketched in § 2, where a client wants to access two adjacent nodes in the list actor
without interleaving operations from other clients is easily resolved by wrapping the two calls to
get (or getNext, if the iterator is used) inside an atomic block. This will batch the messages and
ensure that they are processed back to back:
atomic it <- list ! getIterator()
val e1 <- it.getNext()
val e2 <- it.getNext()
4
=⇒ (e1, e2) =
list ! λ
{val it
val e1
val e2
return
this .
= this.getIterator();
= it.getNext();
= it.getNext();
(e1, e2)}
Formalism
To explain bestow and atomic we use a simple lambda calculus with actors and passive objects. We
abstract away most details that are unimportant when describing the behavior of bestowed objects.
For example, we leave out classes and actor interfaces and simply allow arbitrary operations on
values. By disallowing sharing of (non-bestowed) passive objects, we show that our language is
free from data-races (cf. § 4.4).
The syntax of our calculus is shown in Figure 4. An expression e is a variable x, a function
application e e0 or a message send e!v. Messages are sent as anonymous functions, which are
executed by the receiving actor. We abstract updates to passive objects as e.mutate(), which has
no actual effect in the formalism, but is reasoned about in § 4.4. A new object or actor is created
with new τ and a passive object can be bestowed by the current actor with bestow e. We don’t
need a special atomic construct in the formalism as this can be modeled by composing operations
in a single message as sketched in the end of the previous section.
Statically, values are anonymous functions or the unit value (). Dynamically, id is the identifier
of an actor, ι is the memory location of a passive object, and ιid is a passive object ι bestowed
by the actor id. A type is an active type α, a passive type p, a function type τ → τ , or the Unit
type. An active type is either an actor type c or a bestowed type B(p). Note that for simplicity,
e
v
::=
::=
x | e e | e!v | e.mutate() | new τ | bestow e | v
λx : τ.e | () | id | ι | ιid
τ
α
::=
::=
α | p | τ → τ | Unit
c | B(p)
Figure 4: The syntax of a simple lambda calculus with actors, bestow and atomic.
14
Actors without Borders: Amnesty for Imprisoned State
p and c are not meta-syntactic variables; every passive object has type p, every actor has type c,
and every bestowed object has type B(p).
Γ`e:τ
(Expressions)
e - a p p ly
e - va r
Γ(x) = τ
Γ`x :τ
Γ ` e : τ0 → τ
Γ ` e0 : τ 0
Γ ` e e0 : τ
e - n e w - pa s s i v e
e - n e w - ac t o r
Γ ` new p : p
Γ ` new c : c
e-send
e - m u tat e
Γ`e:p
Γ ` e. mutate() : Unit
e-fn
Γ, x : τ ` e : τ 0
Γ ` (λx : τ.e) : τ → τ 0
Γα , x : p ` e 0 : τ 0
6 ∃ ι . ι ∈ e0
Γ ` e!λx : p.e 0 : Unit
Γ`e:α
e - b e s t ow
Γ`e:p
Γ ` bestow e : B(p)
e-unit
e-loc
e-id
e - b e s t ow e d
Γ ` () : Unit
Γ`ι:p
Γ ` id : c
Γ ` ιid : B(p)
Figure 5: Static semantics. Γ maps variables to types. Γα contains only the active types α of Γ.
4.1
Static Semantics
The typing rules for our formal language can be found in Figure 5. The typing context Γ maps
variables to types. The “normal” lambda calculus rules E - VA R and E - A P P LY are straightforward.
The new keyword can create new passive objects or actors ( E - N E W - * ) . Passive objects may be
mutated ( E - M U TAT E ) , and may be bestowed activity ( E - B E S T OW ) .
Message sends are modeled by sending anonymous functions which are run by the receiver
( E - S E N D ) . The receiver must be of active type (i.e., be an actor or a bestowed object), and the
argument of the anonymous function must be of passive type p (this can be thought of as the
this of the receiver). Finally, all free variables in the body of the message must have active type
to make sure that passive objects are not leaked from their owning actors. This is captured by
Γα which contains only the active mappings _ : α of Γ. Dynamically, the body may not contain
passive objects ι. Typing values is straightforward.
4.2
Dynamic Semantics
Figure 6 shows the small-step operational semantics for our language. A running program is a
heap H, which maps actor identifiers id to actors (ι, L, Q, e), where ι is the this of the actor, L is
the local heap of the actor (a set containing the passive objects created by the actor), Q is the
message queue (a list of lambdas to be run), and e is the current expression being evaluated.
An actor whose current expression is a value may pop a message from its message queue and
apply it to its this ( E VA L - AC T O R - M S G ) . Any actor in H may step its current expression, possibly
also causing some effect on the heap ( E VA L - AC T O R - RU N ) . The relation id ` hH, ei ,→ hH 0 , e0 i
denotes actor id evaluating heap H and expression e one step.
E. Castegren & T. Wrigstad
15
H ,→ H 0
(Evaluation)
e va l - ac t o r - ru n
e va l - ac t o r - m s g
H (id) = (ι, L, Q, e)
id ` hH , ei ,→ hH 0 , e 0 i
0
H (id) = (ι, L0 , Q 0 , e)
H 00 = H 0 [id 7→ (ι, L0 , Q 0 , e 0 )]
H ,→ H 00
0
H (id) = (ι, L, Q v , v)
H = H [id 7→ (ι, L, Q, v 0 ι)]
H ,→ H 0
0
id ` hH, ei ,→ hH 0 , e0 i
(Evaluation of expressions)
e va l - s e n d - ac t o r
H (id 0 ) = (ι, L, Q, e)
H 0 = H [id 0 7→ (ι, L, v Q, e)]
id ` hH , id 0 !vi ,→ hH 0 , ()i
e va l - a p p ly
e 0 = e[x 7→ v]
id ` hH , (λx : τ.e) vi ,→ hH , e 0 i
e va l - n e w - pa s s i v e
H (id) = (ι, L, Q, e)
ι0 fresh
0
0
H = H [id 7→ (ι, L ∪ {ι }, Q, e)]
id ` hH , new pi ,→ hH 0 , ι0 i
e va l - s e n d - b e s t ow e d
H (id 0 ) = (ι0 , L, Q, e)
H 0 = H [id 0 7→ (ι0 , L, (λx : p.v ι) Q, e)]
id ` hH , ιid 0 !vi ,→ hH 0 , ()i
e va l - m u tat e
e va l - b e s t ow
id ` hH , ι. mutate()i ,→ hH , ()i
id ` hH , bestow ιi ,→ hH , ιid i
e va l - n e w - ac t o r
id 0 fresh
ι0 fresh
0
0
H = H [id 7→ (ι , {ι0 }, , ())]
id ` hH , new αi ,→ hH 0 , id 0 i
0
e va l - c o n t e x t
id ` hH , ei ,→ hH 0 , e 0 i
id ` hH , E[e]i ,→ hH 0 , E[e 0 ]i
E[•] ::= • e | v • | • !v | • .mutate() | bestow •
Figure 6: Dynamic semantics.
Sending a lambda to an actor prepends this lambda to the receiver’s message queue and
results in the unit value ( E VA L - S E N D - AC T O R ) . Sending a lambda v to a bestowed value instead
prepends a new lambda to the queue of the actor that bestowed it, which simply applies v to the
underlying passive object ( E VA L - S E N D - B E S T OW E D ) .
Function application replaces all occurrences of the parameter x in its body by the argument
v ( E VA L - A P P LY ) . Mutation is a no-op in practice ( E VA L - M U TAT E ) . Bestowing a passive value ι
in actor id creates the bestowed value ιid ( E VA L - B E S T OW ) .
Creating a new object in actor id adds a fresh location ι0 to the set of the actors passive
objects L and results in this value ( E VA L - N E W - PA S S I V E ) . Creating a new actor adds a new actor
with a fresh identifier to the heap. Its local heap contains only the fresh this, its queue is empty,
and its current expression is the unit value ( E VA L - N E W - AC T O R ) .
We handle evaluation order by using an evaluation context E ( E VA L - C O N T E X T ) .
4.3
Well-formedness
A heap H is well-formed if all its actors are well-formed with respect to H, and the local heaps
Li and Lj of any two different actors are disjoint ( W F - H E A P ) . We use LH(H(id)) to denote the
local heap of actor id. An actor is well-formed if its this is in its local heap L and its message
16
`H
Actors without Borders: Amnesty for Imprisoned State
H ` (ι, L, Q, e)
H `Q
(Well-formedness)
w f - ac t o r
ι∈ L
wf-heap
∀id 1 6= id 2 . LH(H (id 1 )) ∩ LH(H (id 2 )) = ∅
∀id ∈ dom (H ) . H ` H (id)
`H
H;L ` Q
`e:τ
∀ι ∈ e . ι ∈ L
∀id ∈ e . id ∈ dom (H )
∀ιid ∈ e . ι ∈ LH(H (id))
H ` (ι, L, Q, e)
w f - q u e u e - m e s s ag e
H;L ` Q
x :p`e:τ
∀ι ∈ e . ι ∈ L
∀id ∈ e . id ∈ dom (H )
∀ιid ∈ e . ι ∈ LH(H (id))
H ; L ` (λx : p.e) Q
wf-queue-empty
H;L `
Figure 7: Well-formedness rules. LH gets the local heap from an actor: LH((ι, L, Q, e)) = L
queue Q is well-formed. The current expression e must be typable in the empty environment,
and all passive objects ι that are subexpressions of e must be in the local heap L. Similarly, all
actor identifiers in e must be actors in the system, and all bestowed objects must belong to the
local heap of the actor that bestowed it ( W F - AC T O R ) .
A message queue is well-formed if all its messages are well-formed ( W F - Q U E U E - * ) . A message
is well-formed if it is a well-formed anonymous function taking a passive argument, and has a
body e with the same restrictions on values as the current expression in an actor.
4.4
Meta Theory
We prove soundness of our language by proving progress and preservation in the standard fashion:
Progress: A well-formed heap H can either be evaluated one step, or only has actors
with empty message queues and fully reduced expressions:
` H =⇒ (∃H 0 . H ,→ H 0 ) ∨ (∀id ∈ dom(H) . H(id) = (ι, L, , v))
Preservation: Evaluation preserves well-formedness of heaps: ` H ∧ H ,→ H 0 =⇒ ` H 0
Both properties can be proven to hold with straightforward induction.
The main property that we are interested in for our language is data-race freedom. As we
don’t have any actual effects on passive objects, we show this by proving that if an actor is about
to execute ι.mutate(), no other actor will be about to execute mutate on the same object:
Data-race freedom: Two actors will never mutate the same active object
id1 6= id2
0
∧ H(id1 ) = (ι1 , L1 , Q1 , ι.mutate()) =⇒ ι 6= ι
∧ H(id2 ) = (ι2 , L2 , Q2 , ι0 .mutate())
E. Castegren & T. Wrigstad
17
This property is simple to prove using two observations on what makes a well-formed heap:
1. An actor will only ever access passive objects that are in its local heap
2. The local heaps of all actors are disjoint ( W F - H E A P ) .
( W F - AC T O R ) .
The key to showing preservation of the first property is in the premise of rule E - S E N D which states
that all free variables and values must be active objects (Γα , x : p ` e0 : τ 0 and 6 ∃ι . ι ∈ e0 ). This
prevents sending passive objects between actors without bestowing them first. Sending a message
to a bestowed object will always relay it to the actor that owns the underlying passive object
(by the premise of W F - AC T O R : ∀ιid ∈ e . ι ∈ LH(H(id))). Preservation of the second property
is simple to show since local heaps grow monotonically, and are only ever extended with fresh
locations ( E VA L - N E W - PA S S I V E ) .
Having made these observations, it is trivial to see that an actor in a well-formed heap H
that is about to execute ι.mutate() must have ι in its own local heap. If another actor is about to
execute ι0 .mutate(), ι0 must be in the local heap of this actor. As the local heaps are disjoint, ι
and ι0 must be different. Since well-formedness of heaps are preserved by evaluation, all programs
are free from data-races.
5
Related Work
An important property of many actor-based systems is that a single actor can be reasoned about
sequentially; messages are exchanged concurrently but executed sequentially by the receiving
actor. For this property to hold, actors often rely on actor isolation [10], i.e., that the state of
one actor cannot be accessed by another. If this was the not the case, concurrent updates to
shared state could lead to data-races, breaking sequential reasoning.
Existing techniques for achieving actor isolation are often based on restricting aliasing, for
example copying all data passed between actors [2], or relying on linear types to transfer ownership
of data [3, 5, 6, 10]. Bestowed objects offer an alternative technique which relaxes actor isolation
and allows sharing of data without sacrificing sequential reasoning. Combining bestowed objects
with linear types is straightforwand and allows for both ownership transfer and bestowed sharing
between actors in the same system.
Miller et al.propose a programming model based on function passing, where rather than passing
data between concurrent actors, functions are sent to collections of stationary and immutable
data called silos [7]. Bestowed objects are related in the sense that sharing them doesn’t actually
move data between actors. In the function passing model, they could be used to provide an
interface to some internal part of a silo, but implicitly relay all functions passed to it to its owning
silo. While the formalism in § 4 also works by passing functions around, this is to abstract away
from unimportant details, and not a proposed programming model.
References to bestowed objects are close in spirit to remote references in distributed programming or eventual references in E [8]. In the latter case, the unit of encapsulation, e.g., an actor
or an aggregate object protected by a lock, acts similar to a Vat in E, but with an identifiable
boundary and an identity with an associated interface. By bestowing and exposing sub-objects, a
unit of encapsulation can safely delegate parts of its interface to its inner objects, which in turn
need not be internally aware of the kind of concurrency control offered by their bestower.
18
6
Actors without Borders: Amnesty for Imprisoned State
Discussion
Although our formal description and all our examples focus on actors, bestow also works with
threads and locks. An object protected by a lock can share one of its internal objects while
requiring that any interaction with this object also goes via this lock. We believe there is also
a straightforward extension to software transactional memory. In the future, we would like to
study combinations of these.
Bestowed objects lets an actor expose internal details about its implementation. Breaking
encapsulation should always be done with care as leaking abstractions leads to increased coupling
between modules and can lead to clients observing internal data in an inconsistent state. The
latter is not a problem for bestowed objects however; interactions with bestowed objects will be
synchronised in the owning actor’s message queue, so as long as data is always consistent between
messages, we can never access data in an inconsistent state (if your data is inconsistent between
messages, you have a problem with or without bestowed objects).
Sharing bestowed objects may increase contention on the owner’s message queue as messages
to a bestowed object are sent to its owner. Similarly, since a bestowed object is protected by the
same lock as its owner, sharing bestowed objects may lead to this lock being polled more often.
As always when using locks there is a risk of introducing deadlocks, but we do not believe that
bestowed objects exacerbate this problem. Deadlocks caused by passing a bestowed object back
to its owner can be easily avoided by using reentrant locks (as accessing them both would require
taking the same lock twice).
When using locks, atomic blocks are very similar to Java’s synchronized-blocks. With actors,
an atomic block groups messages into a single message. For fairness, it may make sense to only
allow atomic blocks that send a limited number of messages.
It is possible to synchronise on several locked objects by simply grabbing several locks.
Synchronising on several actors is more involved, as it requires actors to wait for each other and
communicate their progress so that no actor starts or finishes before the others. The canonical
example of this is atomically withdrawing and depositing the same amount from the accounts
of two different actors. Interestingly, if the accounts are bestowed objects from the same actor
(e.g., some bank actor), this atomic transaction can be implemented with the message batching
approach suggested in this paper. We leave this for future work.
6.1
Implementation
We are currently working on implementing bestowed objects and atomic blocks in the context of
Encore [3], which uses active objects for concurrency. In Encore, each object (passive or active)
has an interface defined by its class, and only the methods defined therein may be invoked. Thus
it does not follow the formal model from § 4, where message passing is implemented by sending
anonymous functions. It does however use the same approach for the implementation of bestowed
objects and atomic blocks.
We extend each active class with an implicit method perform which takes a function, applies
it to the this of the receiver, and returns the result wrapped in a future. A bestowed object is
logically implemented as an object with two fields owner and object. A message send x ! foo() to a
bestowed object is translated into the message send x.owner ! perform((λ _ . x.object.foo())).
The atomic block can be implemented as sketched in the end of § 3, where messages are batched
and sent as a single message:
E. Castegren & T. Wrigstad
atomic x <- e
x ! foo(42)
x ! bar(-42)
19
=⇒ e ! perform(λ this . {this.foo(42); this.bar(-42)})
This implementation works for the use-cases discussed here, but is somewhat limiting as it
doesn’t allow the caller to react to intermediate values. We are therefore exploring an alternative
approach where we temporarily switch the message queue of an active object to one that only
the caller can submit messages to. Other messages passed to the active object will end up in the
original message queue, and will be processed first when the atomic block finishes.
Each active object would implicitly be extended with two methods override, which switches
the current message queue to a new one, and resume, which discards the temporary queue and
resumes execution with the original queue. Logically, the translation could look like this:
val q = new MessageQueue()
atomic x <- e
e ! override(q) // 1
val v1 = x ! foo(42)
val v1 = q.enqueue(("foo", [42]))
=⇒
val v2 = this.bar(v1)
val v2 = this.bar(v1)
x ! baz(v2)
q.enqueue(("baz", [v2]))
q.enqueue(("resume", [])) // 2
When the message at 1 is processed by receiver, it stops reading from its regular message
queue and instead starts using the queue provided by the caller. Rather than sending messages
normally, the caller interacts with x through this queue (waiting for responses if necessary). When
the message at 2 has been processed by the receiver, it goes back to reading messages normally.
6.2
Abstracting Over Synchronisation Methods
Finally, we note the connection to the safe type qualifier introduced by the Kappa type system [4],
which ranges over both actors and locks (and immutables etc.). A value with a safe type can be
accessed concurrently without risk of data-races, but how this is achieved depends on the type of the
value at runtime. Let x have the type safe τ . Now, z = x.foo() is equivalent to z = x!foo().get()
when x is an actor returning a future value, and get() is a blocking read on the future. When
x is protected by a lock l, the same access is equivalent to lock(l); z = x.foo(); unlock(l);.
When x is immutable, no special synchronisation is needed.
Consequently, the safe qualifier can be used to express operations on objects with concurrency
control abstracted out, without losing safety. An atomic block can be used to atomically compose
operations on a safe object, and the choice of concurrency control mechanism can be relegated to
the runtime. Similarly, bestowed objects internally has no knowledge about their own concurrency
control. Thus, when a bestowed object is used as a safe object, neither the object itself nor its
client needs knows how the interaction is made safe.
7
Conclusion
Actor isolation is important to maintain sequential reasoning about actors’ behavior. By bestowing
activity on its internal objects, an actor can share its representation without losing sequential
reasoning and without bloating its own interface. With atomic blocks, a client can create new
behavior by composing smaller operations. The bestowed objects themselves do not need to know
why access to them is safe. They can just trust the safety of living in a world where actors have
no borders.
20
Actors without Borders: Amnesty for Imprisoned State
References
[1] G. Agha (1986): Actors: a Model of Concurrent Computation in Distributed Systems, Series
in Artificial Intelligence. MIT Press 11.
[2] J. Armstrong (2007): A History of Erlang. In: HOPL III, doi:10.1145/1238844.1238850.
[3] S Brandauer et al. (2015): Parallel Objects for Multicores: A Glimpse at the Parallel Language
Encore. In: Formal Methods for Multicore Programming, doi:10.1007/978-3-319-18941-3_1.
[4] E. Castegren & T. Wrigstad (2016): Reference Capabilities for Concurrency Control. In:
ECOOP, doi:10.4230/LIPIcs.ECOOP.2016.5.
[5] S. Clebsch, S. Drossopoulou, S. Blessing & A. McNeil (2015): Deny Capabilities for Safe,
Fast Actors. In: AGERE, doi:10.1145/b2824815.2824816.
[6] P. Haller & M. Odersky (2010): Capabilities for Uniqueness and Borrowing. In: ECOOP,
doi:10.1007/978-3-642-14107-2_17.
[7] Heather Miller, Philipp Haller, Normen Müller & Jocelyn Boullier (2016): Function
Passing: A Model for Typed, Distributed Functional Programming.
In: Onward!,
doi:10.1145/2986012.2986014.
[8] M. Miller (2006): Robust Composition: Towards a Unified Approach to Access Control and
Concurrency Control. Ph.D. thesis, Johns Hopkins University, USA.
[9] P. Müller (2002): Modular Specification and Verification of Object-oriented Programs.
Springer-Verlag, Berlin, Heidelberg, doi:10.1007/3-540-45651-1.
[10] S. Srinivasan & A. Mycroft (2008): Kilim: Isolation-Typed Actors for Java. In: ECOOP,
doi:10.1007/978-3-540-70592-5_6.
| 6cs.PL
|
arXiv:1804.02482v1 [math.ST] 6 Apr 2018
High-dimensional Adaptive Minimax
Sparse Estimation with Interactions
Chenglong Ye, Yuhong Yang
School of Statistics
University of Minnesota
313 Ford Hall, 224 Church St SE
Minneapolis, MN 55455, USA
e-mail: [email protected]; [email protected]
Abstract: High-dimensional linear regression with interaction effects is
broadly applied in research fields such as bioinformatics and social science.
In this paper, we first investigate the minimax rate of convergence for regression estimation in high-dimensional sparse linear models with two-way
interactions. We derive matching upper and lower bounds under three types
of heredity conditions: strong heredity, weak heredity and no heredity. From
the results: (i) A stronger heredity condition may or may not drastically
improve the minimax rate of convergence. In fact, in some situations, the
minimax rates of convergence are the same under all three heredity conditions; (ii) The minimax rate of convergence is determined by the maximum
of the total price of estimating the main effects and that of estimating the
interaction effects, which goes beyond purely comparing the order of the
number of non-zero main effects r1 and non-zero interaction effects r2 ; (iii)
Under any of the three heredity conditions, the estimation of the interaction
terms may be the dominant part in determining the rate of convergence for
two different reasons: 1) there exist more interaction terms than main effect
terms or 2) a large ambient dimension makes it more challenging to estimate
even a small number of interaction terms. Second, we construct an adaptive
estimator that achieves the minimax rate of convergence regardless of the
true heredity condition and the sparsity indices r1 , r2 .
MSC 2010 subject classifications: Primary 62C20; secondary 62J05.
Keywords and phrases: Minimax rate of convergence, sparsity, highdimensional regression, quadratic model, interaction selection, heredity condition, hierarchical structure, adaptive estimation.
1. Introduction
High-dimensional data are increasingly prevalent in various areas such as bioinformatics, astronomy, climate science and social science. When the number of
variables p is larger than the sample size n in the linear regression setting, statistical estimation of the regression function often requires some crucial conditions.
One common condition is the sparsity of the data generating model, under which
only a small portion of the variables are important to affect the response variable. Under this condition, both sparse estimation of high-dimensional linear
regression functions and variable selection have been well studied with fruitful theoretical understandings in the recent decade. Minimax estimation of
the regression function with main effects only are well investigated under lq 1
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
2
sparsity constraints with 0 ≤ q ≤ 1 (e.g., van de Geer 2007, Candes & Tao 2007,
Bunea et al. 2007, Zhang & Huang 2008, van de Geer 2008, van de Geer & Bhlmann
2009, Bickel et al. 2009, Zhang 2010a, Knight & Fu 2000, Raskutti et al. 2011,
Rigollet & Tsybakov 2011, Wang et al. 2014); model selection consistency results are also obtained for various model selection procedures (e.g., Fan & Li
2001, Zhao & Yu 2006, Zhang & Huang 2008, Zou & Yuan 2008, Lv & Fan 2009).
However, models with only main effects are often not adequate to fully capture the nature of the data. Interaction terms may be necessary to not only
improve the prediction performance but also enhance the understanding of the
relationships among the variables, especially in areas such as genetics, medicine
and behavioristics, where interaction effects between the covariates are of enormous interest. Hierarchical constraints are often imposed to describe the underlying structure of models with interaction effects, such as the marginality
principle (Nelder 1977), the effect heredity principle (Hamada & Wu 1992) and
the “well-formulated models” (Peixoto 1987). We follow a popular naming convention of heredity conditions as adopted in Chipman (1996): strong heredity
and weak heredity. Strong heredity assumes that if an interaction term is in
the model, then both of its corresponding main effects should also be included,
while weak heredity only requires that at least one of its main effects should
be included. In practice, it is possible that, compared to the interaction terms,
some main effects are so small that including them in modeling may not be beneficial from the perspective of estimation variability. Thus, in this work we take
into consideration the additional case where no heredity condition is imposed at
all, also for the purpose of theoretical comparison with the other two heredity
conditions.
Many approaches are proposed for interaction selection, most of which can
be categorized into two types: joint selection and stage-wise selection. The joint
selection approach selects the main and interaction terms simultaneously by
searching over all possible models with interactions. A typical way of joint selection is to use regularization methods with specially designed penalty terms. For
example, Yuan et al. (2009) introduced a family of shrinkage estimators, which
incorporate the hierarchical structures through linear equality constraints on the
coefficients and possess both selection consistency and root-n estimation consistency under fixed p. Choi et al. (2010) re-parameterized the regression model
with interactions and applied an adaptive L1 -norm penalty. The estimators have
the oracle property (Fan & Li 2001) when p = o(n1/10 ). Hao et al. (2017) proposed a computationally efficient regularization algorithm under marginality
principle (RAMP) that simultaneously selects the main effects, interaction effects and quadratic effects for high-dimensional data p ≫ n. They also verified
the interaction selection consistency property of the two-stage LASSO under
some sensible conditions. The stage-wise selection procedure first performs a
main effect selection (by excluding the interaction terms) to reduce the dimension of variables and then carries out a joint selection on the reduced dimension of variables, which is computationally feasible and effective. For example,
viewing the sliced inverse regression (Li 1991) from a likelihood perspective,
Jiang & Liu (2014) suggested a stage-wise variable selection algorithm (SIRI)
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
3
via inverse regression, which is able to detect higher order interactions without
imposing any hierarchical structures. Hao & Zhang (2014) proposed two stagewise interaction selection procedures, IFORT and IFORM, both of which enjoy
sure screening property in the first stage. Fan et al. (2016) proposed a method,
named the interaction pursuit, that incorporates both screening and variable selection in ultra-high dimensions. The method possesses both the sure screening
property and the oracle property in the two stages respectively. For some other
works on interaction selection, see Zhao et al. 2009, Li et al. 2012, Bien et al.
2013, Hall & Xue 2014. While having the aforementioned good properties, both
types of interaction selection approaches have their own disadvantages as well.
The joint selection is usually computational infeasible (insufficient storage) when
p is large; the stage-wise selection, as pointed out in Hao & Zhang (2014), may
be very difficult to be theoretically justified under general conditions.
Although there have been many novel developments on selection of interaction terms as described above, little work has been done on the estimation
of the regression function when interactions exist. In this paper, we present
some theoretical results on the minimax rate of convergence for estimating the
high-dimensional regression function with interaction terms under three different hierarchical structures. Regardless of the heredity condition, our results show
that the minimax rate is determined by the maximum of the total estimation
price of the main effects and that of the interaction effects. Heredity conditions
enter the minimax rate of convergence in terms of the estimation price of the
interaction effects, namely r2 (1 + log(K/r2 ))/n, where r2 is the number of nonzero interaction effects and K is the number of eligible candidate interaction
terms under the different heredity conditions. Consequently, a stronger heredity
condition leads to possibly faster minimax rate of convergence. For example,
when the underlying model has no more than r1 non-zero main effects, at most
K = r21 interaction terms are allowed to enter the model under strong heredity, compared to K = r1 (pn − (r1 + 1)/2) under weak heredity. As will be seen,
only in certain situations is the minimax rate improved by imposing the strong
heredity, although strong heredity allows fewer eligible interaction terms than
the other two heredity conditions. Also, from the perspective of estimation,
there may be no difference in rate of convergence between weak heredity and
no heredity in many situations. Our results provide a complete characterization
and comparison of the minimax rates of convergence under the three heredity
conditions.
In real applications, since one does not know the true heredity condition behind the data (or practically the best heredity condition to describe the data at
the given sample size), it is desirable to construct an estimator that performs
optimally no matter which of the three heredity conditions holds. Such an estimator that adapts to the true heredity condition as well as the unknown number
of main and interaction effects will be obtained in this paper.
The remainder of the paper is organized as follows. In Section 2, we introduce
the model setup, the loss function and the heredity conditions for the problem. In
Section 3, after stating the required assumption, we present our main results of
the minimax rate of convergence under strong heredity. The theoretical results
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
4
under weak heredity and no heredity are presented in Section 4. Section 5.1
provides detailed rates of convergence under different heredity conditions in
relation to the sparsity indices, the ambient dimension and the sample size,
followed by Section 5.2 where we present some interesting implications of the
detailed results. In Section 6, we extend our results to quadratic models in which
both quadratic and interaction effects are considered. In Section 7, we construct
an adaptive estimator that achieves the minimax rate of convergence without
knowledge of the type of the heredity condition or the sparsity indices (r1 and
r2 ). The proofs of our results and some technical tools are presented in the
Appendix.
2. Preliminaries
Model Setup Suppose the dataset is composed of (X, Y), where X =
(x1 , ..., xp ) is a n × p matrix with n observations on p covariates and Y =
(y1 , ..., yn )T is the response vector. We start by considering a linear regression
model with both main effects and two-way interaction effects:
Y
=
Zβ + ǫ,
(2.1)
where β = ((β (1) )T , (β (2) )T )T is the overall coefficient vector, Z = (X, [XX]) ∈
p2 +p
Rn×( 2 ) is the full design matrix, and the random noise vector ǫ ∼ N (0, σ 2 In )
p
with known σ. More specifically, β (1) ∈ Rp and β (2) ∈ R(2) are the coefficients of
the main effects and the two-way interaction effects respectively.
Here we define
[XX] = (x1 ◦ x2 , ..., x1 ◦ xp , ..., xp−1 ◦ xp )T as the n × p2 matrix that contains
all the two-way interaction terms, where ◦ denotes the point-wise product of
two vectors.
In this paper, our focus is on the fixed design, i.e., the covariates are considered given. Our goal is to estimate the mean regression function by a linear
combination of the covariates and interaction terms.
2
Loss Function Denote h(·) : R(p +p)/2 → R as the mean regression function,
2
i.e., h(z) = zT β for z ∈ R(p +p)/2 . Denote ĥ(z) = zT β̂ as an estimated function
of h(z). In our fixed design setting, we focus on the prediction loss (or the
Averaged Squared Error) L(h, ĥ) := n1 ||Zβ − Zβ̂||22 , where k·k2 is the Euclidean
norm. Set the index sets for the main effects and the interaction effects as
Imain = {1, ..., p} and Iint = {(i, j) : 1 ≤ i < j ≤ p} respectively.
Let I = (I1 , I2 ) ⊂ Imain Iint ( is the Cartesian product) be the index set
of a model with |I1 | non-zero main effects and |I2 | non-zero interaction effects.
In this paper, we consider the data generating model (2.1) with at least two
main effects and one interaction effect purely for convenience, which does not
affect the conclusions. Let ZI be the n × |I| submatrix of Z that corresponds
to the model index I. Its corresponding least squares estimator PI Y is used to
estimate Zβ, where PI is the projection matrix onto the column space of ZI .
The loss function of using model I is denoted as L(I) := n1 ||PI Y − Zβ||22 .
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
Heredity Conditions Denote the space of all the p+
with a hierarchical notation of the subscripts as
5
-dimensional vectors
p
2
p
R̈p = {β ∈ Rp+(2) |β = (β1 , ..., βp , β1,2 , ..., βp−1,p )}.
We refer to β (1) = (β1 , ..., βp ) as the subvector consisting of the first p elements
in β, and β (2) = (β1,2 , ..., βp−1,p ) as the subvector containing the rest of the
elements. We introduce the following two vector spaces:
o
n
R̈pweak = β ∈ R̈p |1βi,j 6=0 ≤ 1βi 6=0 ∨ 1βj 6=0 , 1 ≤ i < j ≤ p
and
o
n
R̈pstrong = β ∈ R̈p |1βi,j 6=0 ≤ 1βi 6=0 · 1βj 6=0 , 1 ≤ i < j ≤ p .
The space R̈pstrong captures the strong heredity condition that if the interaction
term is in the model, then both of its corresponding main effects should also
be included. The space R̈pweak characterizes the weak heredity condition that if
the interaction is in the model, then at least one of its main effects should be
included. As pointed out in Hao & Zhang (2016), the sign of the main effect
coefficients are not invariant of linear transformation of the covariates individually due to the existence of the interaction terms. Heredity conditions are consequently meaningless without the specification of the model parametrization.
In our paper, we stick to the parameterization Z and include the no heredity
condition by considering the vector space R̈p . Define the l0 -normP
of a vector
p
a = (a1 , ..., ap ) as the number
of
its
non-zero
elements,
i.e.,
kak
=
0
i=1 1ai 6=0 .
n
o
For a vector space S ∈ R̈pstrong , R̈pweak , R̈p , define the corresponding l0 -ball
and l0 -hull of S as
o
n
(2.2)
B0 (r1 , r2 ; S) = β = (β (1) , β (2) ) ∈ S, β (1) ≤ r1 , β (2) ≤ r2
0
and
0
F0 (r1 , r2 ; S) = h : h(z) = zT β, β ∈ B0 (r1 , r2 ; S)
respectively. Note that B0 (r1 , r2 ; S) represents the collection of coefficients β
with at most r1 non-zero main effects and r2 non-zero interaction effects under a certain hierarchical constraint S. And F0 (r1 , r2 ; S) denotes the collection of linear combinations of the covariates with coefficients β ∈ B0 (r1 , r2 ; S).
Throughout this paper, we assume that r1 + r2 ≤ n (otherwise the minimax risk
may not converge or the rate may not be optimal), r1 ≥ 2 and r2 ≥ 1.
Minimax Risk It is helpful to consider the uniform performance of a modeling procedure when we have plentiful choices of modeling procedures during the
analysis of a statistical problem. The minimax framework seeks an estimator
that minimizes the worst performance (in statistical risk) assuming that the
truth belongs to a function class W. The minimax risk we consider is
min max EL(ĥ, h),
ĥ
h∈W
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
6
where ĥ is over all estimators, and min and max may refer to inf and sup,
more formally speaking. In our work, we assume that the true mean regression
function has a hierarchical
structure by imposing W = F0 (r1 , r2 ; S), with S ∈
n
o
R̈pstrong , R̈pweak , R̈p .
In this paper, we will use the notation bn an or an bn to represent
an = O(bn ). If both bn an and an bn hold, we denote an ≍ bn to indicate
that an and bn are of the same order. If an bn holds without an ≍ bn , we use
the notation an ≻ bn or bn ≺ an .
3. Minimax Rate of Convergence under Strong Heredity
3.1. Assumption
We start by stating an assumption required for our result of the minimax rate
of convergence under strong heredity. In this paper, we use pn to indicate that
the number of main effects p can go to infinity as n increases. We also allow r1
and r2 to increase with the sample size n as well.
Sparse Reisz Condition (SRC) For some l1 , l2 > 0, there exist constants
b1 , b2 > 0 (not depending on n) such that for any β = (β (1) , β (2) ) with β (1) 0 ≤
min(2l1 , pn ) and β (2) 0 ≤ min 2l2 , p2n , we have
1
b1 kβk2 ≤ √ kZβk2 ≤ b2 kβk2 .
n
(3.1)
The SRC assumption requires that the eigenvalues of ZTI ZI for any sparse
submatrix ZI of Z are bounded above and away from 0. It was first proposed by Zhang & Huang (2008). It is similar to the sparse eigenvalue conditions in Zhang (2010b), Raskutti et al. (2011), quasi-isometry condition
in Rigollet & Tsybakov (2011); it is also related to the more stringent restricted isometry property (which requires the constants b1 , b2 are close
to 1) in Candes & Tao (2007). Such assumptions are standard in the l1 regularization analysis like LASSO and the Dantzig selector. See Bickel et al.
(2009), Meinshausen & Yu (2009), van de Geer (2007), Koltchinskii (2009) for
more references.
3.2. Minimax Rate
Now we present our main result of the minimax rate of convergence under
strong heredity. A simple estimator
an effective minimax upper
Pn is enoughI for
2
(Y
−
Ŷ
)
be
the model that minimizes
bound. Let Î = arg minI∈Istrong
i
i
i=1
r1 ,r2
the residual sum of squares over all the models that have exactly r1 non-zero
main effects and r2 non-zero interaction effects under strong heredity, denoted
I
as Istrong
r1 ,r2 , where Ŷ = PI Y is the projection of Y onto the column space of
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
7
the design matrix ZI . For lower bounding the minimax risk, the informationtheoretical tool of using Fano’s inequality with metric entropy understanding
(Yang & Barron 1999) plays an important role in the proof.
Theorem
3.1. Under the Sparse Reisz Condition with l1 = r1 ≤ pnn∧ n, l2 =
r2 ≤ r21 ∧ n and the strong heredity condition W = F0 (r1 , r2 ; R̈pstrong
), the
minimax risk is upper bounded by
!!
r1
cσ 2
pn
2
min max EL(ĥ, h) ≤ sup E(L(Î)) ≤
r1 1 + log
+ r2 1 + log
,
n
r1
r2
ĥ h∈W
h∈W
(3.2)
where c is a pure constant; the minimax risk is lower bounded by
!!
r1
pn
σ2
2
r1 1 + log
∨ r2 1 + log
(3.3)
min max EL(ĥ, h) ≥ c1
n
r1
r2
ĥ h∈W
for some positive constant c1 that only depends on the constants b1 and b2 in
the SRC assumption.
From the theorem, under the SRC and the strong heredity condition, the min2
imax rate of convergence scales as: minĥ maxh∈W EL(ĥ, h) ≍ σn (r1 (1 + log prn1 )∨
r2 (1 + log( r21 /r2 ))).
Remark 1. The term r1 (1 + log(pn /r1 ))/n = rn1 + rn1 log(pn /r1 ) reflects two
aspects in the estimation of the main effects: the price of searching among prn1
possible models, which is of order r1 log(pn /r1 )/n, and the price of estimating
the r1 main effect coefficients after the search. Thus r1 (1 + log(pn /r1 ))/n
is the
total price of estimating the main effects. Similarly, r2 1 + log r21 /r2 /n is
the total price of estimating the interaction effects.
Remark 2. Our result of the upper bound is general and does not require the
sparsity condition of r1 ≺ pn , although it may be needed for fast rate of convergence.
4. Minimax Rate of Convergence under Weak Heredity and No
Heredity
Similar results are obtained under weak heredity and no heredity. The minimax
rate of convergence is still determined by the maximum of the total price of
estimating the main effects and that of the interaction effects. When the heredity
condition changes, the total price of estimating the interaction effects may differ,
possibly substantially.
Theorem 4.1. Under the Sparse Reisz Condition with l1 = r1 ≤ pn ∧ n, l2 =
n
), the
r2 ≤ (r1 pn ) ∧ n and the weak heredity condition W = F0 (r1 , r2 ; R̈pweak
minimax risk is of order
σ2
pn
r1 · pn
min max EL(ĥ, h) ≍
r1 1 + log
∨ r2 1 + log
.
(4.1)
n
r1
r2
ĥ h∈W
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
8
Theorem
4.2. Under the Sparse Reisz Condition with l1 = r1 ≤ pn ∧ n, l2 =
r2 ≤ p2n ∧ n and the no heredity condition W = F0 (r1 , r2 ; R̈pn ), the minimax
risk is of order
!!
pn
pn
σ2
2
r1 1 + log
∨ r2 1 + log
.
(4.2)
min max EL(ĥ, h) ≍
n
r1
r2
ĥ h∈W
5. Comparisons and Insights
In this section, we summarize the consequences of our main results in three scenarios for an integrated understanding. For brevity, we introduce the following
notation. For a, b ∈ N+ and a ≥ b, define the quantity ξba := b(1 + log(a/b)).
The total price of estimating the main effects and the interaction effects are
then denoted as σ 2 ξrp1n /n and σ 2 ξrK2 /n respectively, where K depends on pn , r1
and the heredity condition. We also use the notation KS (7.5) to indicate that
K depends on the heredity condition S. Let
M(S) := min max EL(ĥ, h)
ĥ
h∈F0 (r1 ,r2 ;S)
denote the minimax risk under the heredity condition S.
5.1. Detailed Rates of Convergence
Since the minimax rate of convergence depends on the maximum of ξrp1n and ξrK2 ,
we discuss the cases where one of the two quantities is greater than the other.
Scenario 1: r2 r1
When there are more main effects than interaction effects in the sense that r2
r1 , the minimax rate of convergence is not affected by the heredity conditions.
When log(pn /r1 ) log r1 , we always have ξrp1n ξrK2 regardless of the heredity
conditions. When log(pn /r1 ) ≺ log r1 , it depends on the order of r2 to further
decide which estimation price is larger. When log(pn /r1 ) ≺ log r1 , let r∗ be such
r2
that ξrp1n ≍ ξr∗1 . If r∗ r2 , we have ξrp1n ξrK2 ; otherwise ξrp1n ≺ ξrK2 .
In summary, given that r2 r1 , the minimax risk is of order
( 2 2
σ r1
ξ , if r∗ r2 r1 and log prn1 ≺ log r1 ,
M(S) ≍ σn2 rp2
n
n ξr1 , otherwise,
n
o
for S ∈ R̈pstrong , R̈pweak , R̈p .
Remark 3. This scenario also includes the special case when pn = O(1), where
we must have r1 = O(1) and r2 = O(1). The minimax rate of convergence is of
the standard parametric order 1/n regardless of the heredity conditions.
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
9
Scenario 2: r1 r2 and log pn r1 When there exist more interaction
terms, i.e., r1 r2 , under weak or no heredity, the quantity ξrK2 is always no less
than (in order) ξrp1n .
For strong heredity, we discuss case by case. When log(pn /r1 ) ≺ log r1 , we
r2
always have ξrp1n ξr21 . When log(pn /r1 ) log r1 , it depends on the order of r2
to decide which estimation price is larger in terms of order. When log(pn /r1 )
r2
r2
log r1 , let r∗′ be such that ξrp1n ≍ ξr∗′1 . If r2 r∗′ , we have ξrp1n ξr21 ; otherwise
r2
ξrp1n ≻ ξr21 . In summary, given that r1 r2 and log pn r1 , the minimax risk is
of order
( σ2
pn
if r1 r2 r∗′ and log prn1 log r1 ,
n ξr1 ,
pn
2
M(R̈strong ) ≍ σ2 r1
n ξ(r2 ∧r 2 ) , otherwise,
1
n
)≍
M(R̈pweak
σ 2 r1 pn
,
ξ
n (r2 ∧r1 pn )
M(R̈pn ) ≍
σ 2 p2n
ξ .
n (r2 )
K
Remark 4. The term ξ(r
deals with the case where r2 is inactive in the
2 ∧K)
sense that
r
exceeds
K
under
the
specific heredity condition. For example, with
2
r2 ≥ r21 , the upper bound r2 in (2.2) does not provide any new information of
the number of non-zero interaction effects for strong heredity. Thus
the l0 -ball
B0 (r1 , r2 ; R̈pstrong ) is automatically reduced to a subset B0 (r1 , r21 ; R̈pstrong ).
Scenario 3: r1 r2 and log pn r1 When the number of the main effects
pn is at least exponentially as many as the non-zero main effects in the sense
that log pn r1 , ξrp1n is always no less than ξrK2 in terms of order. In fact, in this
scenario, the results of the minimax rates under weak or no heredity are exactly
the same as those in Scenario 2. For completeness, we still present the results.
Specifically, the minimax risk is of order
n
M(R̈pstrong
)≍
n
)≍
M(R̈pweak
σ 2 pn
ξ ,
n r1
σ 2 r1 pn
,
ξ
n (r2 ∧r1 pn )
M(R̈pn ) ≍
σ 2 p2n
ξr .
n 2
5.2. Interesting Implications
1. Comparing the results for weak heredity and no heredity, we may or may
not have distinct rates of convergence. When there exists a small constant
c > 0 such that log r2 ≤ (1 − c) · log(r1 pn ) for large enough n, there is
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
2.
3.
4.
5.
6.
10
no difference between weak heredity and no heredity from the perspective
of rate of convergence in estimation. It still remains an open question
how they are different for the problem of model identification. Without
the above relationship between r1 and r2 , there is no guarantee that the
rates of convergence are the same under weak heredity and no heredity.
For example, when r2 = r1 pn / log r1 , if in addition we have r1 = pn ≤
n1/2 , the minimax rates are the same under weak and no heredity, at
n
) ≍ M(R̈pn ) ≍ r1 pn log log r1 /(n log r1 ). In contrast, if instead
M(R̈pweak
√
n
)≍
we have r1 = pn , then the minimax rates are different, with M(R̈pweak
pn
r1 pn log log r1 /(n log r1 ) and M(R̈ ) ≍ r1 pn /n.
Heredity conditions do not affect the rates of convergence in some situations. For example, when there exist more main effects than interaction
effects (Scenario 1), the minimax rates of convergence are the same under
all three heredity conditions.
From the detailed rates of convergence, under any of the three heredity
conditions, the estimation of the interaction terms ξrK2 /n may become the
dominating part. There are two different reasons why the price of estimating the interaction terms becomes higher than that for the main effect
terms. One is that the number of interaction terms is more than that of
the main effect terms. The other reason is that although the main effect terms outnumber the interaction terms, the ambient dimension is so
large that even estimating a small number of the interaction terms is more
challenging than estimating the main effects.
How much can the rate of convergence be improved by imposing strong
heredity? We quantify this improvement by taking the ratio of two
minimax rates of convergence given the ambient dimension pn , i.e.,
n
n
n
M(R̈pstrong
) and M(R̈pstrong
)/M(R̈pweak
)/M(R̈pn ). In Scenario 2 (r1 r2
pn
n
and log pn r1 ), we have M(R̈strong )/M(R̈pweak
) log pn /pn , where
the maximal improvement happens when r1 ≍ log pn and r2 ≍ r1 pn .
That is, the minimax rate of convergence under strong heredity is up
to log pn /pn times faster than that under weak heredity. Similarly we
n
)/M(R̈pn ) log2 pn /p2n , where the maximal improvement
have M(R̈pstrong
log2 pn /p2n happens at r1 ≍ log pn and r2 ≍ p2n .
In Scenario 3 (r1 r2 and log pn r1 ), the improvement
n
n
) log pn /pn , where the maximal improvement
)/M(R̈pweak
M(R̈pstrong
happens when r2 r1 pn . In this scenario, the maximal improvement of
the minimax rate from weak heredity to strong heredity depends on the
ambient dimension pn . In other words, the larger the ambient dimension is,
the more improvement of minimax rate of convergence we have from weak
n
)/M(R̈pn )
heredity to strong heredity. Similarly we have M(R̈pstrong
2
2
log pn /pn , where the equality holds if r1 = O(1) and r2 ≍ pn .
If r2 is active for all three heredity conditions, i.e., r2 ≤ r21 , the maximal improvement of minimax rate from weak/no heredity to strong
n
n
heredity turns out to be consistent. That is, M(R̈pstrong
) ≍
)/M(R̈pweak
pn
pn
M(R̈strong )/M(R̈ ) 1/ log pn , where the maximal improvement hap-
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
11
pens at r1 ≍ log pn and r2 ≍ r12 .
6. Extension to Quadratic Models
Our aforementioned results do not consider quadratic effects. When both
quadratic and two-way interaction effects are included in a model (called a
quadratic model), it is easy to see the rates of convergence in the theorems still
apply under both strong heredity and weak heredity. However, in the case of no
heredity, the number of quadratic terms enters into the minimax rate. Assume
one model has at most r3 extra non-zero quadratic terms. We need the following
assumption.
Sparse Reisz Condition 2 (SRC2) For some l1 , l2 , l3 > 0, there exist constants b1 , b2 > 0 (not depending on n) such that for any β =
(β (1) , β (2) , β (3) ) with β (1) 0 ≤ min(2l1 , pn ), β (2) 0 ≤ min(2l2 , p2n ) and
β (3) 0 ≤ min(2l3 , pn ), we have
1
b1 kβk2 ≤ √ kZ∗ βk2 ≤ b2 kβk2 ,
n
where Z∗ = (X, [XX], X2 ) is the new design matrix, with X2 representing the
n × p matrix that contains all the quadratic terms.
Next we state the minimax results for quadratic models. Strong heredity and
weak heredity are exactly the same condition since a quadratic term has only one
corresponding main effect term. That is, both strong and weak heredity require
that if a quadratic term X12 has a non-zero coefficient, then X1 must also have
a non-zero coefficient. Similarly, under SRC2 with l1 = r1 , l2 = r2 , l3 = r3 ,
the minimax rate of convergence under strong/weak heredity for the quadratic
model stays the order
!
r1
pn
σ2
2
r1 (1 + log ) ∨ r2 (1 + log
) ;
(6.1)
n
r1
r2
under no heredity, its order becomes
σ2
n
!
pn
pn
2
r̄(1 + log ) ∨ r2 (1 + log
) ,
r̄
r2
(6.2)
where r̄ = r1 ∨ r3 .
7. Adaptation to Heredity Conditions and Sparsity Indices
In the previous sections, we have determined the minimax rates of convergence
for estimating the linear regression function with interactions under different
sizes of sparsity indices r1 , r2 and heredity conditions S. These results assume
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
12
that r1 , r2 and S are known. However, in practice, we usually have no prior information about the underlying heredity condition nor the sparsity constraints.
Thus it is necessary and appealing to build an estimator that adaptively achieves
the minimax rate of convergence without the knowledge of S, r1 and r2 . We construct such an adaptive estimator as below.
To achieve our goal, we consider one specific model and three types of models
together as the candidate models:
no
weak
F̄ = {Ipn ,(p2n −pn )/2 } ∪ {Istrong
k1 ,k2 } ∪ {Ik1 ,k2 } ∪ {Ik1 ,k2 },
where Ipn ,(p2n −pn )/2 denotes the full model with pn main effects and all the p2n
interaction effects. It is included so that the risk of our estimator will not be
worse than order RZ /n, in which RZ is the rank of the full design matrix. With
weak
no
a slight abuse of the notation, we use Istrong
k1 ,k2 , Ik1 ,k2 and Ik1 ,k2 to represent
a model with k1 main effects and k2 interaction effects under strong heredity,
weak heredity and no heredity respectively. Note that some models appear more
than once in F̄ , which does not cause any problem for the goal of estimating
the regression function. The details of the range of k1 and k2 for each model
class are shown in (7.2), (7.3) and (7.4).
To choose a model from the candidate set, we apply the ABC criterion in
Yang (1999). For a model I in F̄ , the criterion value is
ABC(I) =
n
X
i=1
(Yi − ŶiI )2 + 2rI σ 2 + λσ 2 CI ,
(7.1)
where ŶI = PI Y is the projection of Y onto the column space of the design matrix ZI with rank rI , CI is the descriptive complexity of model I and
λ > 0 is a constant. The model descriptive complexity satisfies CI > 0 and
P
I∈F̄ exp(−CI ) ≤ 1.
The model descriptive complexity is crucial in building the adaptive model.
Let π0 , π1 , π2 , π3 ∈ (0, 1) be four constants such that π0 + π1 + π2 + π3 = 1. Set
CIpn ,(p2 −pn )/2 = − log π0 for the full model,
n
k1
pn
k1
2
+log
∧ n +log
(7.2)
CIstrong = − log π1 +log(pn ∧n)+log
k1 ,k2
k1
k2
2
for 1 ≤ k1 ≤ pn ∧ n and 0 ≤ k2 ≤ k21 ∧ n,
= − log π2 + log(pn ∧ n) + log (K ∧ n) + log
CIweak
k ,k
1
2
1
2
(7.3)
− k1 for 1 ≤ k1 ≤ pn ∧ n and 0 ≤ k2 ≤ K ∧ n, and
pn
pn
pn
2
+log
∧ n +log
, (7.4)
= − log π3 +log(pn ∧n)+log
k1
k2
2
with K = k1 pn −
CIno
k ,k
k1
2
K
pn
+ log
k1
k2
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
13
for 1 ≤ k1 ≤ pn ∧n and 0 ≤ k2 ≤ p2n ∧n. This complexity assignment recognizes
that there are three types of models under the different heredity conditions.
Let Î = arg minI∈F̄ ABC(I) denote the model that minimizes the ABC criterion over the candidate model set F̄ and ŶÎ := PÎ Y denote the least squares
estimate of Y using the model Î. Then we have the following oracle inequality.
Theorem 7.1. When λ ≥ 5.1/ log 2, the worst risk of the ABC estimator ŶÎ
is upper bounded by
pn
KS
cσ 2
RZ ∧ r1 1 + log
+ r2 1 + log
,
sup E(L(Î)) ≤
n
r1
r2
h∈F0 (r1 ,r2 ;S)
with
,
KS = r1 pn ,
pn
2 ,
r1
2
if S = R̈pstrong ,
if S = R̈pweak ,
if S = R̈p ,
(7.5)
where RZ is the rank of the full design matrix Z and the constant c only depends
on the constant λ.
From the theorem, without any prior knowledge of the sparsity indices, the
constructed ABC estimator adaptively achieves the minimax upper bound regardless of the heredity conditions. The result also indicates a major difference
between estimation and model identification. For estimation, from the result,
we are able to achieve adaptation with respect to the heredity condition without any additional assumption. For model identification, although we are not
aware of any work that addresses the task of adaptation over the unknown
heredity nature, it seems certain that much stronger assumptions than those for
consistency under an individual heredity condition will be necessary to achieve
adaptive selection consistency. Achieving adaptive model selection consistency
under different types of conditions remains an important open problem on model
selection theory and methodology.
Remark 5. We do not require any assumptions on the relationship among the
variables for the upper bound in the theorem. In particular, the variables may
be arbitrary correlated.
Remark 6. The order RZ /n is achievable when we use the projection estimator
from the full model. Thus the minimax rate of convergence is no slower than
the order RZ /n. As is known, the rank of the design matrix plays an important
role in determining the minimax rate of convergence under fixed design (Yang
1999, Rigollet & Tsybakov 2011, Wang et al. 2014). For our result, when pn , r1
and r2 together make the total estimation price of the true model small enough,
the upper
bound will be improved from RZ /n to (r1 (1 + log(pn /r1 )) ∨ r2 (1 +
log( r21 /r2 )))/n.
Remark 7. The ABC estimator may not be practical when pn is large. In such
case, stochastic search instead of all subset selection can be used for implementation.
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
14
Remark 8. The term “RZ ∧ ” automatically applies to the lower bound under
whichever heredity condition, since under the
SRC assumption, it intrinsically
requires that r1 (1+log(pn /r1 ))∨r2 (1+log( r21 /r2 )) is no larger than RZ in terms
of order. Otherwise, the lower bound (r1 (1+log(pn /r1 ))∨r2 (1+log( r21 /r2 )))/n
by our proof will exceed the upper bound RZ /n, which leads to a contradiction.
We give a specific example in Appendix F to illustrate this requirement.
Appendix A: Proof of the upper bound in Theorem 3.1
Proof of (3.2)
Recall that h(z) = zT β and ĥ(z) = zT β̂. Set hI := PI h as the estimator by
model I, where we use the bold-face h = (h(zT1 ), ..., h(zTn ))T to denote the mean
regression function vector and zi is the i-th row of the full design matrix Z. We
first prove that Î is equivalently an ABC estimator over the candidate set we
consider. The SRC assumption with l1 = r1 , l2 = r2 assures that r1 + r2 ≤ n.
It follows that, for any model I = (I1 , I2 ) with |I1 |0 = r1 , |I2 |0 = r2 , the
corresponding submatrix ZI is full rank, i.e., rI = r1 + r2 . Thus,
Î =
arg min
I∈F
n
X
i=1
n
X
(Yi − ŶiI )2
(Yi − ŶiI )2 + 2rI σ 2 + λσ 2 CI
=
arg min
=
arg min ABC(I),
I∈F
i=1
I∈F
where F is the collection of models that have r1 non-zero main effects and r2
non-zero interaction effects with 0 ≤ r1 ≤ pn , 1 ≤ r2 ≤ r21 , and all the models
in F share the same model descriptive complexity
r1
pn
2
+
log
=
log
.
CIstrong
r1 ,r2
r1
r2
The ABC criterion and the model descriptive complexity are introduced near
(7.1). Therefore, Î is an ABC estimator over the candidate set F .
Next we prove the upper bound. Since Î is an ABC estimator over the candidate set F , by Theorem 1 in Yang (1999), we have:
σ 2 rI
λσ 2 CI
1
2
,
(A.1)
khI − hk2 +
+
E(L(Î)) ≤ c inf
I∈F
n
n
n
where c is a positive constant that depends on the constant λ only. When
n
), there exists a specific model in F such that the
h ∈ W = F0 (r1 , r2 ; R̈pstrong
projection estimator of this model is equal to h. We consider the RHS of (A.1)
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
15
evaluated at such a model, where we still denote it as Ir1 ,r2 for convenience.
Thus,
!
λσ 2 CIr1 ,r2
σ 2 rIr1 ,r2
2
+
E(L(Î)) ≤ c
hIr1 ,r2 − h +
n
n
c 2
σ rIr1 ,r2 + λσ 2 CIr1 ,r2 .
=
|n
{z
}
(i)
The term (i) is bounded as follows:
r1
pn
c1 λ 2 1
2
(i) ≤
σ
(r1 + r2 ) + log
+ log
r1
r2
n
λ
!!
r1
c1 λ 2 1
pn
≤
+ r2 1 + log 2
σ
(r1 + r2 ) + r1 1 + log
n
λ
r1
r2
!!
r1
pn
c2 2
+ r2 1 + log 2
.
σ r1 1 + log
≤
n
r1
r2
Therefore,
c2 · σ 2
E(L(Î)) ≤
n
pn
r1 1 + log
+ r2
r1
1 + log
r1
2
!!
r2
.
Thus we have
c2 · σ 2
min max EL(ĥ, h) ≤ max E(L(Î)) ≤
h∈W
n
ĥ h∈W
pn
r1 1 + log
+ r2
r1
1 + log
r1
2
!!
r2
where the above c1 , c2 are universal constants.
Appendix B: Proof of the lower bound in Theorem 3.1
Before stating the proof of (3.3), we introduce the local metric entropy, two
important sets that aid the understanding of the metric entropy of the regression
function space, together with the lemmas in relation to these two sets.
B.1. Metric Entropy
Metric entropy plays a central role in minimax theory, through the concepts
of packing and covering. It provides a way to understand the “cardinality” of
a set with infinitely many elements. In deriving the lower bound, information
theoretic techniques play a key role, such as the local metric entropy, Fano’s
inequality, Shannon’s mutual information and Kullback–Leibler divergence. We
begin by introducing the definition of the local metric entropy.
,
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
16
Definition 1 (Local Metric Entropy). Given a metric space (X , ρ), let B(x, ǫ) =
{x′ ∈ X |ρ(x, x′ ) ≤ ǫ} be a ǫ-ball around x. For 0 < a < 1, the a-local ǫ-entropy
at x, denoted as log Mxa (ǫ; X , ρ), is defined as the aǫ-packing entropy of B(x, ǫ).
a
The a-local ǫ-entropy, denoted as log Mlocal
(ǫ; X , ρ), is then defined as the
maximum (or supremum if maximum does not exist) of log Mxa (ǫ; X , ρ) over all
a
x in X , i.e., log Mlocal
(ǫ; X , ρ) = max log Mxa (ǫ; X , ρ).
x∈X
B.2. Important Subsets
Set the Hamming distance between any two vectors v, v ′ ∈ Rd as ρH (v, v ′ ) =
Pd
′
i=1 1vi 6=vi . Consider the set
n
pn
n
H = β ∈ R̈pstrong
: β ∈ {−1, 0, 1}pn+( 2 ) , β (1)
0
≤ r1 , β (2)
0
≤ r2
o
and let H1 denote a subset of H where the the first r1 coordinates are fixed,
i.e.,
H1 = β ∈ H : β (1) = (1, ..., 1, 0, ..., 0), β (2) = r2 .
| {z } | {z }
0
r1
pn −r1
Let H2 denote another subset of H where no interaction effect exists, i.e.,
n
o
H2 = β ∈ H : β (1) = r1 , β (2) = 0 ,
0
0
The following two lemmas of the metric entropy of the subsets H1 and H2
are needed in the proof of (3.3).
exists a subset of H1 with its cardinality
Lemma 1. If r2≤ 32 r21 , then there
r1
−r
/2
)
(
2
such that the pairwise Hamming distance of
no less than exp r22 log 2 r2
the points in this subset is greater than r2 /2.
Proof. The proof is presented in Appendix B.4.
Lemma 2. If r1 ≤ 2pn /3, then there exists a subset of H2 with its cardinality
1 /2
such that the pairwise Hamming distance of
no less than exp r21 log pn −r
r1
the points in this subset is greater than r1 /2.
Proof. The proof is similar to that of Lemma 1.
B.3. Proof of (3.3)
r1
r1
2
/4.
Since
r
(1
+
log(
It suffices
to
prove
under
r
≤
r
−
r
/r
))
≍
for
2
2
1
2
1
2
2
r1
1 r1
risk in the function class
2 2 ≤ r2 ≤ 2 , the monotonicity of the minimax
reduces the proof to the case r2 ≤ r12 − r1 /4. Similarly it suffices to prove
under r1 ≤ pn /2.
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
n
Recall that B0 (r1 , r2 ; R̈pstrong
)=
n
n
β ∈ R̈pstrong
: β (1)
0
≤ r1 , β (2)
17
0
≤ r2
o
n
is the coefficient space of interest and F0 (r1 , r2 ; R̈pstrong
)
=
o
n
pn
T
is the mean regression function
h : h(z) = z β, β ∈ B0 (r1 , r2 ; R̈strong )
space. For convenience, let hθ , hϑ denote the regression functions with
coefficents θ, ϑ respectively, i.e., hθ (z) = zT θ, hϑ (z) = zT ϑ. Let
o
n
n
n
), kβk2 ≤ ǫ
)(ǫ) = β : β ∈ B0 (r1 , r2 ; R̈pstrong
B0 (r1 , r2 ; R̈pstrong
n
be an l2 -ball of radius ǫ around 0 in B0 (r1 , r2 ; R̈pstrong
) and
n
o
n
n
), d(h′ , h) ≤ ǫ0
)(h, ǫ0 ) = h′ : h′ (z) = zT β, β ∈ B0 (r1 , r2 ; R̈pstrong
F0 (r1 , r2 ; R̈pstrong
be the ball of radius ǫ0 around the underlying regression function h. Without
loss of generality,
we assume h = 0. The square root of the empirical l2 -norm loss
q P
n
1
d(hθ , hϑ ) := n i=1 (hθ (zi ) − hϑ (zi ))2 = √1n kZ(θ − ϑ)k2 is used to measure
the distance between any two functions hθ , hϑ . We prove the following two cases
separately.
Case 1: r21 log((pn − r1 /2)/r1 ) ≤ r22 log(( r21 − r2 /2)/r2 ). We consider the
n
)(ǫ), where ◦ is the
subset H1′ = {ǫ ◦ β : β ∈ H1 } of the l2 -ball B0 (r1 , r2 ; R̈pstrong
point-wise product of two vectors,
√
√
√
√
ǫ
ǫ = √ (1/ r1 , ..., 1/ r1 , 1/ r2 , ..., 1/ r2 )
{z
} |
{z
}
2 |
pn
and
H1 =
(p2n −pn )/2
β ∈ H : β (1) = (1, ..., 1, 0, ..., 0), β (2)
| {z } | {z }
r1
pn −r1
0
= r2
.
From Lemma 1, there exists a subset Hsub of H1 such that |Hsub | ≥
(r1 )−r2 /2
exp( r22 log 2 r2
) and the pairwise Hamming distance of the elements within
′
Hsub is greater than r2 /2. Set Hsub
:= {ǫ ◦ β : β ∈ Hsub }. For any θ′ , ϑ′ ∈
′
Hsubp
, there exist θ, ϑ ∈p Hsub such that kθ′ − ϑ′ k2 = kǫ ◦ θ − ǫ ◦ ϑk2 ≥
ǫ
′
√ǫ
| = |Hsub | since
ρH (θ, ϑ) ≥ √2r
r2 /2 = 2ǫ . We also have |Hsub
2r2
2
′
′
it is a one-to-one mapping from Hsub to Hsub . Thus, we have Hsub
⊆
pn
′
is
B0 (r1 , r2 ; R̈strong )(ǫ) and the pairwise l2 -distance of the elements in Hsub
greater than ǫ/2.
n
′
)(ǫ), let hθ′ , hϑ′ be such that
For any θ′ , ϑ′ ∈ Hsub
⊆ B0 (r1 , r2 ; R̈pstrong
T
′
T
′
hθ′ (z) = z θ , hϑ′ (z) = z ϑ . By SRC assumption with l1 = r1 , l2 = r2 , we
have
ǫ
b1 ≤ b1 k(θ′ − ϑ′ )k2 ≤ d(hθ′ , hϑ′ )
2
d(h, hϑ′ ) ≤ b2 k(0 − ϑ′ )k2 ≤ b2 ǫ.
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
18
n
)(h, ǫ0 ) has a subset
Let ǫ0 = b2 ǫ, it follows that F0 (r1 , r2 ; R̈pstrong
′
Fsub := h′ : h′ (z) = zT β, β ∈ Hsub
, d(h′ , h) ≤ ǫ0 ,
in which the pairwise distance (in terms of d) of the functions are no
b1
b1
ǫ0 . This implies that the 2b
-local ǫ0 -packing entropy of
less than 2b
2
2
pn
′
F0 (r1 , r2 ; R̈strong )(h, ǫ0 ) is lower bounded by log |Fsub | = log |Hsub
| ≥
r1
−r
/2
(2) 2
b1 /(2b2 )
r2
n
) is no less than
. So log Mlocal
(ǫ0 ) of F0 (r1 , r2 ; R̈pstrong
2 log
r2
r2
2
Yang
&
Barron
(1999),
the minimax
log((r
−
r
−
r
)/2r
).
Then
by
(7)
in
1
2
2
1
2
risk is lower bounded by
c1
σ 2 r22 log(
r12 −r1 −r2
)
2r2
n
σ2
= c1
n
r1
pn − r1 /2 r2
∨
log
log
2
r1
2
r1
2
− r2 /2
r2
!
,
where c1 > 0 is a constant that depends on b1 and
b2 only.
Case 2: r21 log((pn − r1 /2)/r1 ) ≥ r22 log(( r21 − r2 /2)/r2 ). We consider the
√
n
)(ǫ), where ǫ′1 = ǫ/ r1 and
subset H2′ = ǫ′1 H2 of B0 (r1 , r2 ; R̈pstrong
n
H2 := β ∈ H : β (1)
0
= r1 , β (2)
0
o
=0 .
Following the same arguments above, we conclude that the minimax is lower
bounded by
!
r1
σ 2 r1
pn − r1 /2
pn − r1 /2 r2
σ 2 r1
2 − r2 /2
,
= c2
∨
log
log
log
c2
n 2
r1
n
2
r1
2
r2
where c2 > 0 is a constant that depends on b1 and b2 only.
1
(1 + log(p
Notice that when pn /r1 ≥ 2, we have log(pn /r1 − 12 ) ≥10
n /r1 )).
r1
r1
1
1
Similarly, we have log 2 /r2 − 2 ≥ 10 1 + log 2 /r2 when r21 /r2 ≥ 2.
Together with the fact that the lower bounds for the two cases are the same,
the minimax risk is lower bounded by
!
r1
pn
σ2
2
r1 (1 + log( )) ∨ r2 (1 + log
) .
c
n
r1
r2
Thus the desired lower bound holds.
Remark 9. One way to interpret the imposition of the SRC assumption
2
is that kZθ − Zβk2 is indeed up to a constant of the Kullback-Leibler divergence between two joint densities (the joint distribution of the response
variable y under fixed design) parameterized with θ and β respectively. To
see this, let Q
zi be the i-th row of Z and we have the joint density Pθ =
n
(2π)−n/2 σ −n i=1 exp(− 12 (yi − zi θ)2 /σ 2 ) with parameter θ. The K-L distance
Pn
2
is then D(Pθ ||Pβ ) = 2σ1 2 i=1 (zi β − zi θ)2 = 2σ1 2 kZθ − Zβk2 .
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
19
B.4. Proof of Lemma 1
r
2
1 )/2
2 2 since the main effects are fixed. Fix z ∈ H1 ,
First we have |H1 | = (r1 −r
r2
let A denote the collection of all the points in H1 that are within r22 Hamming distances to z, i.e., A = {z ′ ∈ H1 : ρH (z, z ′ ) ≤ r2 /2}. It follows that the
cardinality of A is bounded above:
r1
2
3r2 /2 .
|A| ≤
r2 /2
For this upper bound, since the main effects are fixed for any point in H1 , we
only need to pick r2 /2 positions of the interaction effects where z ′ is different
from z. In the remaining interaction effect positions, z ′ is the same as z. It
r1
gives us at most r( 2/2) possible choices of the r2 /2 positions out of the r21
2
coordinates. For these r2 /2 positions, z ′ can take any values in {−1, 1, 0}, thus
the desired upper bound follows.
r1
r1
Let B be a subset of H such that |B| ≤ m := ( 2 ) / ( 2 ) . Consider the
1
r2
r2 /2
collection of the points in H1 that are within r2 /2 Hamming distance to some
element in B, i.e., {z ∈ H1 : ρH (z, z ′ ) ≤ r22 f or some z ′ ∈ B}. We have
o
n
r2
f or some z ′ ∈ B
z ∈ H1 : ρH (z, z ′ ) ≤
2
≤ |B| |A|
(r21 ) r1
r2
2
3r2 /2
·
≤
(r21 )
r2 /2
r2 /2
r1
2
<
2 r2
r2
= |H1 | .
The strictly less inequality implies that for any set B ⊂ H1 with |B| ≤ m,
∃z ∈ H1 such that ρH (z, z ′ ) > 21 r2 for all z ′ ∈ B. By induction, we can create a
set B ⊂ H1 with |B| > m such that Hamming distance between any two elements
in B exceeds 12 r2 . Next, we introduce one useful inequality. When 0 ≤ B ≤ 23 A
for A, B ∈ N, we have
A
B
A
B
2
=
B/2
B/2
Y A−B +
Y A−B +j
(A − B2 )!( B2 )!
≥
=
B
B
B
(A − B)!(B)!
2 +j
2 + 2
j=1
j=1
B
2
When r2 ≤ (r12 − r1 )/3, we have
m
=
(r21 )
r2
≥
(r1 )
2
r2 /2
r1
2
− r2 /2
r2
!r2 /2
.
=(
A−
B
B
2 B/2
)
.
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
20
Thus,
log m
≥
r2
log
2
r1
2
−
r2
r2
2
.
The desired result follows.
Appendix C: Proof of Theorem 4.1
Proof. The proofs are similar to the arguments for strong heredity with slight
differences.
To prove the upper bound
heredity, we instead consider the
Pn under weak
I 2
(Y
−
Ŷ
)
that
minimizes the residual sum of
model Î = arg minI∈Iweak
i
i
i=1
r1 ,r2
squares over all the models that have r1 non-zero main effects and r2 non-zero
interaction effects under weak heredity. The model descriptive complexity
is
thus
K
pn
with
+
log
different from the strong heredity. In this case, CIweak
=
log
r1
r1 ,r2
r2
r1
K = r1 (pn − (r1 + 1)/2) for 1 ≤ r1 ≤ pn ∧ n and 0 ≤ r2 ≤ (r1 pn − 2 − r1 ) ∧ n.
The ABC criteria for the models are defined as in (7.1). The same arguments
in the proof of (3.2) can then be used.
To prove the lower bound under weak heredity, we consider the set
o
n
pn
n
: β ∈ {−1, 0, 1}pn+( 2 ) , β (1) ≤ r1 , β (2) ≤ r2 .
Hweak = β ∈ R̈pweak
0
0
Then the two important subsets are instead
H1 = β ∈ Hweak : β (1) = (1, ..., 1, 0, ..., 0), β (2)
| {z } | {z }
r1
and
n
H2 = β ∈ Hweak : β (1)
0
0
= r2
pn −r1
= r1 , β (2)
0
o
=0 .
Similar metric entropy results of the above two subsets can be derived in the
same fashion as in Lemmas 1 and 2. Other arguments are the same as in the
proof of (3.3).
Appendix D: Proof of Theorem 4.2
Proof. For theP
upper bound under no heredity, we consider the model Î =
n
I 2
=
arg minI∈Ino
i=1 (Yi − Ŷi ) with the model descriptive complexity CIno
r1 ,r2
r1 ,r2
pn
p
pn
log r1 + log ( r22 ) for 1 ≤ r1 ≤ pn ∧ n and 0 ≤ r2 ≤ 2n ∧ n. The ABC criteria
for the models are defined as in (7.1).
For the lower bound under no heredity, we consider the set
o
n
pn
Hno = β ∈ R̈pn : β ∈ {−1, 0, 1}pn+( 2 ) , β (1) ≤ r1 , β (2) ≤ r2 .
0
0
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
Then the two important subsets are instead
H1 = β ∈ Hno : β (1) = (1, ..., 1, 0, ..., 0), β (2)
| {z } | {z }
r1
and
n
H2 = β ∈ Hno : β (1)
0
0
= r2
pn −r1
= r1 , β (2)
0
o
=0 .
21
Similar metric entropy results of the above two subsets can be derived in the
same fashion as Lemmas 1 and 2.
Other arguments are the same as in the proofs of (3.2) and (3.3).
Appendix E: Proof of Theorem 7.1
The model descriptive complexity term λσ 2 CI plays a fundamental role in
model selection theory (Barron & Cover 1991, Barron et al. 1999, Yang 1999,
Wang et al. 2014). Since we are considering models with interaction terms, the
model descriptive complexity CI reflects our comprehension of the model complexity other than the total number of parameters only. The detailed designation
of the descriptive complexity usually depends on the class of models of interest.
Instead of interpreting CI as the code length (or description length) of describing
the model index, one can also treat exp(−CI ) as the prior probability assigned
to the model from a Bayesian viewpoint.
Proof. The candidate set can be represented as the union of the candidate sets
under three heredity conditions, i.e., F̄ = Fstrong ∪ Fweak ∪ Fno , with
Fstrong := {Ipn ,(p2n −pn )/2 } ∪ {Istrong
k1 ,k2 },
Fweak := {Ipn ,(p2n −pn )/2 } ∪ {Iweak
k1 ,k2 },
Fno := {Ipn ,(p2n −pn )/2 } ∪ {Ino
k1 ,k2 }.
n
When h ∈ F0 (r1 , r2 ; R̈pstrong
), there exists a specific model in Fstrong such that
the projection estimator of this model is equal to h. Also, the projection of h
onto the full design matrix is still h. We denote the two models as Ir1 ,r2 and
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
22
Ipn ,(pn )(pn −1)/2 respectively. It follows that
E(L(Ŷ )) ≤
F̄
≤
≤
=
σ 2 rI
λσ 2 CI
1
2
khI − hk2 +
+
c inf
n
n
n
I∈F̄
2
1
σ rI
λσ 2 CI
2
c inf
(E.1)
khI − hk2 +
+
I∈Fstrong
n
n
n
!
σ 2 rIr1 ,r2
λσ 2 CIr1 ,r2
2
c
hIr1 ,r2 − h +
+
n
n
σ 2 RZ
−λσ 2 log π0
2
∧ c hIpn ,pn (pn −1)/2 − h +
+
n
n
c
c 2
σ rIr1 ,r2 + λσ 2 CIr1 ,r2 ∧
σ 2 RZ − λσ 2 log π0 , (E.2)
n
n
|
{z
} |
{z
}
(i)
(ii)
where RZ is the rank of the full design matrix, the first inequality follows from
(A.1), the second inequality follows from Fstrong ⊆ F̄ and the third inequality
results from the evaluation of (E.1) at Ir1 ,r2 and Ipn ,pn (pn −1)/2 . The two terms
(i) and (ii) are bounded as follows:
r1
pn
c1 λ 2 r1 + r2
r1
2
+ log
+ log
σ
− log π1 + log pn + log
(i) ≤
r1
n
λ
r2
2
c1 λ 2 r1 + r2
pn
≤
σ
− log π1 + r1 (1 + log )
n
λ
r1
!
r1
p
n
+ log r12 + r1 (1 + log ) + r2 (1 + log 2 )
r1
r2
!
r1
c2 2
pn
2
≤
) ,
σ r1 (1 + log ) + r2 (1 + log
n
r1
r2
and
c 2
σ RZ − λσ 2 log π0
n
c3 2
σ RZ .
n
(ii) ≤
≤
Therefore, we have
max(c2 , c3 ) · σ 2
E(L(Ŷ F̄ )) ≤
n
"
#
!
r1
pn
2
r1 (1 + log ) + r2 (1 + log
) ∧ RZ ,
r1
r2
where c1 , c2 , c3 are some constants that depend only on the constant λ. Thus
the desired minimax upper bounded follows.
n
) or h ∈ F0 (r1 , r2 ; R̈pn ), with I ∈ Fweak or I ∈ Fno
When h ∈ F0 (r1 , r2 ; R̈pweak
replacing I ∈ Fstrong in (E.1), the quantity (i) in (E.2) will instead be no greater
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
23
than
r1 + r2
K
pn
+ log
− log π2 + log pn + log K + log
λ
r1
r2
n
with K = r1 pn − r21 − r1 under weak heredity h ∈ F0 (r1 , r2 ; R̈pweak
), or
pn
pn
pn
c1 λ 2 r1 + r2
2
+ log
+ log
σ
− log π3 + log pn + log
(i) ≤
r1
r2
n
λ
2
c1 λ 2
σ
n
under no heredity h ∈ F0 (r1 , r2 ; R̈pn ). The different constants π2 , π3 does not
affect the conclusion in terms of order. Following the same arguments in the
proof of strong heredity, the desired results follow when the underlying heredity
condition is weak heredity or no heredity.
Appendix F: An Example When SRC is not Satisfied
For simplicity, let us consider an example where the regression mean function
includes only one main effect term, i.e., r1 = 1, r2 = 0. The corresponding SRC
assumption with l1 = r1 = 1, l2 = r2 = 0 will be that there exist constants
b1 , b2 > 0 (not depend on n) such that for any β ∈ Rpn with kβk0 ≤ 2, we have
1
b1 kβk2 ≤ √ kZβk2 ≤ b2 kβk2 ,
n
(F.1)
where the design matrix Z = X is the matrix that contains the main effects.
Assume the first RZ columns of Z are linearly independent and denote Z =
(Z1 , Z2 ), where Z1 = (Z1 , ..., ZRZ ) is the n×RZ submatrix with rank(Z1 ) = RZ .
Suppose the submatrix Z1 satisfies the SRC assumption. Assume that
√ kZi k2 =
f (n) for 1 ≤ i ≤ pn . For the purpose of illustration, we set f (n) = n.
Let A be the collection of all columns in Z2 : A = {z|z = Z1 α, α ∈
RZ
R , kzk2 = f (n)}. Then A should satisfy that ∀z, z ′ ∈ A, we have b1 ≤
√1 ka1 z + a2 z ′ k ≤ b2 for all a1 , a2 ∈ R and a2 + a2 = 1. We know
1
2
2
n
1
1
√ ka1 z + a2 z ′ k2 = √
n
n
q
2
2
a21 kzk2 + a22 kz ′ k2 + 2a1 a2 kzk2 kz ′ k2 cos θ,
where θ is the angle between two n-dimensional vectors z and z ′ .
Thus we have
q
1
2
2
√
a21 kzk2 + a22 kz ′ k2 + 2a1 a2 kzk2 kz ′ k2 cos θ
n
q
f (n)
a21 + a22 + 2a1 a2 cos θ
= √
n
p
1 + 2a1 a2 cos θ.
=
√
Then 1 + 2a1 a2 cos θ ≥ b1 for all a21 + a22 = 1 (otherwise √1n ka1 z + a2 Zi k2
is less than b1 , which violates the SRC assumption). Since −1 ≤ 2a1 a2 ≤ 1
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
24
p
for a21 + a22 = 1, we have b1 ≤ 1 − |cos θ|, which implies |cos θ| ≤ 1 − b21 .
That means the
p between any two elements in A should be
√ pairwise l2 distance
greater than 2b1 and less than 4 − 2b21 . It is well known that the ǫ-covering
entropy
of the RZ -dimensional
unit ball
√ B is of order RZ log(1/ǫ). We denote
√
√
nB as a ball of radius
n.
Let
ǫ
=
2b1 /2,
constant
√
√
√ there exists a positive
c1 such that log N (ǫ; nB, l2 ) ≤ c1 RZ log( n/ǫ) √
= c1 RZ log( 2n/b1 ). Since
A is a 2ǫ-packing set
√ of a ball of radius f (n) = n, its cardinality satisfies
log |A| ≤ log M (2ǫ; nB, l2 ). The covering number and the packing number
are closely related as in the well-known inequality
( 2ǫ ; X , ρ) ≤
√ M (ǫ; X , ρ) ≤ N √
ǫ
M ( 2 ; X , ρ). Thus we have log |A| ≤ log M (2ǫ; nB, l2 ) ≤ log N (ǫ; nB, l2 ) ≤
√
√
c 1 RZ
c1 RZ log( 2n/b1 ), which implies A has at most
elements under
1)
√ ( 2n/b
c 1 RZ
, the SRC assumption
the SRC assumption. Thus, as long as pn > ( 2n/b1 )
will not be satisfied because the SRC assumption requires that (F.1) must hold
for any pair of columns in Z. In this case, the lower bound r1 (1 + log(pn /r1 ))/n
in our theorems does not apply.
References
Barron, A., Birgé, L. & Massart, P. (1999), ‘Risk bounds for model selection via
penalization’, Probability theory and related fields 113(3), 301–413.
Barron, A. R. & Cover, T. M. (1991), ‘Minimum complexity density estimation’,
IEEE transactions on information theory 37(4), 1034–1054.
Bickel, P. J., Ritov, Y. & Tsybakov, A. B. (2009), ‘Simultaneous analysis of
lasso and dantzig selector’, The Annals of Statistics pp. 1705–1732.
Bien, J., Taylor, J. & Tibshirani, R. (2013), ‘A lasso for hierarchical interactions’,
Annals of statistics 41(3), 1111.
Bunea, F., Tsybakov, A. B., Wegkamp, M. H. et al. (2007), ‘Aggregation for
gaussian regression’, The Annals of Statistics 35(4), 1674–1697.
Candes, E. & Tao, T. (2007), ‘The dantzig selector: Statistical estimation when
p is much larger than n’, The Annals of Statistics pp. 2313–2351.
Chipman, H. (1996), ‘Bayesian variable selection with related predictors’,
The Canadian Journal of Statistics / La Revue Canadienne de Statistique
24(1), 17–36.
Choi, N. H., Li, W. & Zhu, J. (2010), ‘Variable selection with the strong heredity constraint and its oracle property’, Journal of the American Statistical
Association 105(489), 354–364.
Fan, J. & Li, R. (2001), ‘Variable selection via nonconcave penalized likelihood and its oracle properties’, Journal of the American statistical Association 96(456), 1348–1360.
Fan, Y., Kong, Y., Li, D. & Lv, J. (2016), ‘Interaction pursuit with feature
screening and selection’, arXiv preprint arXiv:1605.08933 .
Hall, P. & Xue, J.-H. (2014), ‘On selecting interacting features from highdimensional data’, Computational Statistics & Data Analysis 71, 694–708.
Hamada, M. & Wu, C. J. (1992), ‘Analysis of designed experiments with complex
aliasing’, Journal of Quality Technology;(United States) 24(3).
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
25
Hao, N., Feng, Y. & Zhang, H. H. (2017), ‘Model Selection for High Dimensional
Quadratic Regression via Regularization’, Journal of the American Statistical
Association 24(4), 0–0.
Hao, N. & Zhang, H. H. (2014), ‘Interaction Screening for UltrahighDimensional Data’, Journal of the American Statistical Association
109(507), 1285–1301.
Hao, N. & Zhang, H. H. (2016), ‘A note on high dimensional linear regression
with interactions’, The American Statistician (just-accepted).
Jiang, B. & Liu, J. S. (2014), ‘Variable selection for general index models via
sliced inverse regression’, The Annals of Statistics 42(5), 1751–1786.
Knight, K. & Fu, W. (2000), ‘Asymptotics for lasso-type estimators’, Annals of
statistics pp. 1356–1378.
Koltchinskii, V. (2009), ‘The dantzig selector and sparsity oracle inequalities’,
Bernoulli 15(3), 799–828.
Li, K.-C. (1991), ‘Sliced inverse regression for dimension reduction’, Journal of
the American Statistical Association 86(414), 316–327.
Li, R., Zhong, W. & Zhu, L. (2012), ‘Feature screening via distance correlation
learning’, Journal of the American Statistical Association 107(499), 1129–
1139.
Lv, J. & Fan, Y. (2009), ‘A unified approach to model selection and sparse
recovery using regularized least squares’, The Annals of Statistics pp. 3498–
3528.
Meinshausen, N. & Yu, B. (2009), ‘Lasso-type recovery of sparse representations
for high-dimensional data’, The Annals of Statistics pp. 246–270.
Nelder, J. A. (1977), ‘A Reformulation of Linear Models’, Journal of the Royal
Statistical Society. Series A (General) 140(1), 48–31.
Peixoto, J. L. (1987), ‘Hierarchical variable selection in polynomial regression
models’, The American Statistician 41(4), 311–313.
Raskutti, G., Wainwright, M. J. & Yu, B. (2011), ‘Minimax rates of estimation
for high-dimensional linear regression over ℓq -balls’, IEEE transactions on
information theory 57(10), 6976–6994.
Rigollet, P. & Tsybakov, A. (2011), ‘Exponential screening and optimal rates of
sparse estimation’, Annals of Statistics 39(2), 731–771.
van de Geer, S. A. (2007), The deterministic lasso.
van de Geer, S. A. (2008), ‘High-dimensional generalized linear models and the
lasso’, Ann. Statist. 36(2), 614–645.
van de Geer, S. A. & Bhlmann, P. (2009), ‘On the conditions used to prove
oracle results for the lasso’, Electron. J. Statist. 3, 1360–1392.
Wang, Z., Paterlini, S., Gao, F. & Yang, Y. (2014), ‘Adaptive minimax regression estimation over sparse lq-hulls.’, Journal of Machine Learning Research
15(1), 1675–1711.
Yang, Y. (1999), ‘Model selection for nonparametric regression’, Statistica Sinica
pp. 475–499.
Yang, Y. & Barron, A. (1999), ‘Information-theoretic determination of minimax
rates of convergence’, The Annals of Statistics 27(5), 1564–1599.
Yuan, M., Joseph, V. R. & Zou, H. (2009), ‘Structured variable selection and
C. Ye and Y. Yang/Adaptive Minimax Estimation with Interactions
26
estimation’, The Annals of Applied Statistics pp. 1738–1757.
Zhang, C.-H. (2010a), ‘Nearly unbiased variable selection under minimax concave penalty’, Ann. Statist. 38(2), 894–942.
Zhang, C.-H. & Huang, J. (2008), ‘The sparsity and bias of the lasso selection in
high-dimensional linear regression’, The Annals of Statistics pp. 1567–1594.
Zhang, T. (2010b), ‘Analysis of multi-stage convex relaxation for sparse regularization’, Journal of Machine Learning Research 11(Mar), 1081–1107.
Zhao, P., Rocha, G. & Yu, B. (2009), ‘The composite absolute penalties family for grouped and hierarchical variable selection’, The Annals of Statistics
pp. 3468–3497.
Zhao, P. & Yu, B. (2006), ‘On model selection consistency of lasso’, Journal of
Machine learning research 7(Nov), 2541–2563.
Zou, H. & Yuan, M. (2008), ‘Composite quantile regression and the oracle model
selection theory’, Ann. Statist. 36(3), 1108–1126.
| 10math.ST
|
k-NN Estimation of Directed Information
Yonathan Murin
Department of Electrical Engineering, Stanford University, USA
arXiv:1711.08516v1 [cs.IT] 22 Nov 2017
[email protected]
Abstract
This report studies data-driven estimation of the directed information (DI) measure between
twoem discrete-time and continuous-amplitude random process, based on the k-nearest-neighbors
(k-NN) estimation framework. Detailed derivations of two k-NN estimators are provided. The two
estimators differ in the metric based on which the nearest-neighbors are found. To facilitate the
estimation of the DI measure, it is assumed that the observed sequences are (jointly) Markovian
of order m. As m is generally not known, a data-driven method (that is also based on the k-NN
principle) for estimating m from the observed sequences is presented. An exhaustive numerical
study shows that the discussed k-NN estimators perform well even for relatively small number
of samples (few thousands). Moreover, it is shown that the discussed estimators are capable of
accurately detecting linear as well as non-linear causal interactions.
1
Introduction
Detection and estimation of causality relationships between two random processes is a fundamental
problem in many natural and social sciences [1]. This task is in particular challenging as in many
real-life scenarios one does not have a good underlying statistical model for the considered process,
e.g., in the fields of neuroscience, financial markets, meteorology, etc. For such scenarios it is
desirable to use a non-parametric estimator for the causal influence between two observed timeseries. The common approach for quantifying the causal influence between two time-series dates
back to the seminal work of Granger [2] where {Xn } is said to have a Granger-causal influence on
{Yn } if:
Given the past of Yn , the past of Xn helps in predicting future samples of Yn , e.g., Yn+1 .
While this approach is indeed general, the common formulation of Granger causality (GC) assumes
that the time-series obey a linear structure, and therefore it does not follow the non-parametric
approach mentioned above. A possible alternative to GC is the information theoretic measure of
directed information (DI) [3, 4], which is closely related to the transfer entropy (TE) functional
[5].
In this report we discuss the estimation of DI between two discrete-time and continuous-amplitude
sequences (time-series). The estimated DI can then be used as a measure of the causal influence
between the random processes underlying the observed sequences.1 Note that DI is a deterministic
function of the joint density of the underlying random processes. Therefore, the DI can be estimated by first estimating the (local) joint densities, and then using these densities to estimate the
DI. This approach was taken in [7] where it was proposed to use a kernel density estimator (KDE)
for estimating the local densities, and in [8] that suggested to estimate the local the densities using
1 Note that the work [6] presented scenarios where DI (or TE) fail to detect or quantify the causal influence between a
pair of time-series. The conclusions of [6] also hold for the GC measure.
1
correlation integrals. On the other hand, the estimation method we discuss in the current report is
based on the k-nearest-neighbors (k-NN) principle. Before delving into the technical details of estimating the DI functional, we emphasize that when estimating statistical functionals it is common
to assume that the underlying process are stationary, ergodic, and smooth. In the rest of this report
we build upon these assumptions without verifying their validity.
The rest of this report is organized as follows: In Section 2 we discuss k-NN estimation of
differential entropy. This estimation approach is extended to estimating the mutual information
(MI) functional in Section 3. Estimating the DI is discussed in Section 4, and a numerical study is
presented in Section 5.
Notation: We denote random variables (RVs) by upper case letters, X, and their realizations
with the corresponding lower case letters. We use the short-hand notation X1n to denote the sequence {X1 , X2 , . . . , Xn }. We denote random processes using boldface letters, e.g., X. We denote
sets by calligraphic letters, e.g., S, where R denotes the set of real numbers. fX (x) denotes the
probability density function (PDF) of a continuous RV X on R, and log(·) denotes the natural basis
logarithm. Finally, we use h(·) and I(·; ·) to denote differential entropy and mutual information as
defined in [9, Ch. 8].
2 k-NN Estimation of Differential Entropy
We introduce the concept of k-NN estimation of information theoretic measures by first discussing
the estimation of differential entropy. A detailed discussion regarding methods (including k-NN) for
estimating differential entropy, MI, and the Kullback–Leibler (KL) divergence is provided in [10].
A popular approach for estimating information theoretic functionals is the re-substitution method
where first the density is locally (around each of the data points) estimated, and then the functional
is estimated via empirical averaging [10, Sec. 2.2.1].
Specifically, let Xi ∈ Rd , i = 1, 2, . . . , N , be independent and identically distributed (i.i.d.)
samples of X with PDF fX (x), and consider estimating the differential entropy of X from X1N . Let
fˆX (Xi ), i ∈ {1, 2, . . . , N }, be a local estimation of the density around the ith sample. Recalling that
the differential entropy is defined as h(X) , −E{log fX (x)}, see [9, Ch. 8], the re-substitution
estimator ĥ(X) of h(X) is given by:
ĥ(X) = −
N
1 X
log fˆX (Xi ).
N i=1
(1)
While there are several popular techniques for estimating the local density, e.g., kernel density
estimation [11] and correlation integrals [12], in this report we focus on estimation algorithms
based on the k-NN principle. Before presenting the estimation strategy, we provide several definitions. Let p ≥ 1. For x ∈ Rd , the `p -norm of x is defined as:
||x||p ,
d
X
! p1
|xi |p
,
i=1
where the `p -distance between xi ∈ Rd and xj ∈ Rd is defined as dp (xi , xj ) , ||xi − xj ||p . Let
n
sort(x),
x ∈ R , denotethe sorted version of the vector x, in ascending order, and define ρi,p ,
sort {dp (xi , xj )}N
j=1,j6=i to be the (sorted) vector of distances, in `p -norm, of all the samples from
xi . In particular, ρk,i,p denotes the distance from xi to its k-NN.2
2 Unless otherwise stated, in this report we assume that k is a fixed and relatively small number (independent of N ) at
the range 4, 5, . . . , 10.
2
Assuming a uniform local density in a small environment around each of the samples (this
assumption implies that the density fX (x) is smooth), ρk,i,p can be used to estimate fX (Xi ). Let
Γ(a) denote the Euler’s gamma function [13, eq. (5.2.1)], and let cd,p denote the volume of the
unit lp -ball in d dimensions, given by [14]:
cd,p = 2d
(Γ(1 + p1 ))d
Γ(1 + p1 )
.
Since in the ball of radius ρk,i,p there are k samples out of N in total, we can approximate the local
density via:
fˆX (Xi ) ≈
k
.
N cd,p (ρk,i,p )d
(2)
Substituting (2) into (1) we obtain the following entropy estimator:
N
1 X
N cd,p (ρk,i,p )d
ĥ(X) =
.
log
N i=1
k
(3)
The work [15], by Kozachenko and Leonenko (KL), showed that the simple estimator (3) is biased,
and suggested the following bias-corrected estimator:
N
N cd,p (ρk,i,p )d
1 X
ĥKL (X) =
log
+ log(k) − ψ(k)
N i=1
k
N
d X
log(ρk,i,p ),
= log(N ) − ψ(k) + log(cd,p ) +
N i=1
(4)
where ψ(k) is the digamma function [13, Ch. 5.4].
Remark 1 (The case of k dependent of N ). If k is chosen as a function of N , then ψ(k) converges to
log(k) and no bias-correction is required to obtain a consistent estimator. On the other hand, if k is
fixed, then the correction term log(k) − ψ(k) is crucial for consistency.
Next, we discuss k-NN estimation of mutual information (MI).
3
From Entropy to Mutual Information
Consider two RVs X ∈ Rdx and Y ∈ Rdy . Observing N i.i.d pairs (Xi , Yi ) from the joint density
fX,Y (x, y), we are interested in estimating the MI I(X; Y ) = h(X) + h(Y ) − h(X, Y ). Fixing k and
estimating each of the entropy terms via ĥKL (·), one obtains the following consistent estimator (the
consistency follows from the consistency of the ĥKL (·) estimator), denoted by Iˆ3KL (X; Y ):
Iˆ3KL (X; Y ) = ĥKL (X) + ĥKL (Y ) − ĥKL (X, Y ).
(5)
Note that in the estimator (5) each entropy term is estimated separately, with a different k-NN
distance. Thus, even though the estimator (5) is consistent, since the estimators are not coupled,
for a finite number of samples the bias can be non-negligible. This motivated the work of Kraskov,
Stögbauer and Grassberger (KSG) [16], that presented a modification of the Iˆ3KL (X; Y ) estimator,
and empirically showed that this modification improves performance (higher accuracy) when the
number of samples is finite. In the following we refer to this estimator as IˆKSG (X; Y ).
3
The main idea behind IˆKSG (X; Y ) is to modify the estimators of the individual entropy terms,
h(X) and h(Y ), such that their correlation with the estimator of the joint entropy h(X, Y ) term
will be higher, leading to a smaller bias. This interpretation was recently presented in [17]. Let
ρk,i,p (X, Y ) denote the distance from the pair (Xi , Yi ) to its k-NN. Further define I(·) to be the
PN
indicator function and let nx,i,p , j=1,j6=i I(dp (xi , xj ) ≤ ρk,i,p (X, Y )). ny,i,p is defined similarly.
Note that nx,i,p is the number of samples which are within a distance ρk,i,p (X, Y ) from Xi , where
the distance is measured in the X-plane. Thus, as this places no limitation on the distance in the
Y -plane, it follows that nx,i,p ≥ k. Using these definitions, and setting p = ∞, the KSG estimator is
given by:
N
1 X
ˆ
(ψ(nx,i,∞ + 1) + ψ(ny,i,∞ + 1)).
IKSG (X; Y ) = ψ(k) + log(N ) −
N i=1
(6)
Note that when p = ∞, then dX,Y = dX + dY , cdX,Y ,∞ = cdX ,∞ · cdY ,∞ , and ρk,i,∞ (X, Y ) =
max{ρk,i,∞ (X), ρk,i,∞ (Y )}. Hence, the KSG estimator in (6) estimates the individual entropy h(X)
via:
ĥKSG (X) = log(N ) + log(cdX ,∞ ) +
N
1 X
(dX log(ρk,i,∞ (X, Y )) − ψ(nx,i,∞ + 1)) ,
N i=1
(7)
which is sample dependent. On the other hand, the estimation of the joint entropy term is identical
to the one in (4). The work [17] showed that IˆKSG (X; Y ) is consistent and derived the order of its
bias.
Remark 2 (The inherent bias in (7)). The estimator (6) uses a hyper-cube of radius ρk,i,p (X, Y )
around each sample point. Since ρk,i,∞ (X, Y ) = max{ρk,i,∞ (X), ρk,i,∞ (Y )}, it follows that for one
of the individual terms, e.g., X, ρk,i,∞ (X, Y ) is exactly the distance to the (nx,i,∞ + 1)th -NN, while
for the other term this is not the case, see [16, Fig. 1] for illustration of this observation. Thus,
letting (Xi , Yi )k be the k−NN of (Xi , Yi ), and assuming that (Xi , Yi )k lies on the X-boundary of the
1
hyper-cube around (Xi , Yi ), the bias of the estimator (7) is of the order of ny,i,∞
. This inherent bias
is partially addressed in the second estimator introduced in [16, eqn. (9)]. This estimator uses a
hyper-rectangle instead of a hyper-cube, yet, using hyper-rectangles requires other approximations.
Thus, none of the two estimators presented in [16] is uniformly better than the other. We refer the
reader to [16, Fig. 3] for a detailed discussion regarding the approximations used as apart of the
second estimator of [16].
Motivated by the idea of using the sample dependent nx,i,∞ and ny,i,∞ , the work [17] proposed
a different method to tackle the inherent bias discussed in Remark 2. The idea is to use an `2 ball
instead of the `∞ ball used in (6). With this choice ρk,i,2 (X, Y ) is not the X-boundary distance nor
the Y -boundary distance, see [17, Fig. 5]. However, when the Euclidean norm is used, namely
p = 2, a new relationship should be derived between nx,i,2 (number of points in the X-space) and
fX (x). Such a relationship is formulated in [17, Thm. 9], which essentially states the following
approximation:
E{nx,i,2 } ≈ N fX (x)cdX ,2 (ρk,i,2 (X, Y ))dX .
(8)
This approximation motivates estimating log fX (x) via:
log\
fX (x) = log(nx,i,2 ) − log N − log cdX ,2 − dX log ρk,i,2 (X, Y ).
4
(9)
Plugging this estimation to the re-substitution (1) results in the Gao-Oh-Viswanath (GOV) entropy
estimator:
ĥGOV (X) = log(N ) + log(cdX ,2 ) +
N
1 X
(dX log(ρk,i,2 (X, Y )) − log(nx,i,2 )) .
N i=1
(10)
Finally, to obtain an estimator for the MI, we recall that dX,Y = dX + dY and cdX,Y ,2 = cdX +dY ,2 .
Using the estimator (4) to estimate the joint entropy and the estimators (10) for the individual
entropy terms one obtains:
IˆGOV (X; Y ) = log(N ) + ψ(k) + log
Comparing (11) and (6), log
cdX ,2 ·cdY ,2
cdX +dY ,2
cdX ,2 · cdY ,2
cdX +dY ,2
−
N
1 X
(log(nx,i,2 ) + log(ny,i,2 )).
N i=1
(11)
can be viewed as a correction term for using the Euclidean norm. Finally, we note that the consistency of IˆGOV (X; Y ) is stated in [17, Thm. 10].
4
4.1
Directed Information
Definitions and Background
Let X and Y be arbitrary discrete-time continuous-amplitude random processes, and let X N ∈ RN
and Y N ∈ RN , be N -length sequences. The directed information from X N to Y N is defined as [3]:
I(X N → Y N ) ,
n
X
I(X1i ; Yi |Y1i−1 )
i=1
=
N
X
h(Yi |Y1i−1 ) − h(Yi |Y1i−1 , X1i )
i=1
(a)
= h(Y N ) − h(Y N ||X N ),
(12)
PN
where (a) follows by defining h(Y N ||X N ) , i=1 h(Yi |Y1i−1 , X1i ), and h(·) is the differential entropy. This definition implies that I(X N → Y N ) = 0 when Yi is independent of X1i , given Y1i−1 . DI
can be viewed as quantifying the causal influence of the sequence X N on the sequence Y N . Therefore, it is not surprising that in contrast to MI, DI is not symmetric. The directed information rate
[4] between the processes X and Y is defined as:
1
I(X N → Y N ),
N →∞ N
I(X → Y) , lim
(13)
provided that this limit exists. We now make the following assumptions regarding the processes X
and Y.
A1) The random processes X and Y are assumed to be stationary, ergodic, and Markovian of
order m in the observed sequences. The stationarity assumption implies that the statistics of
the considered random processes is constant throughout the observed sequences. Note that
formally speaking, the random processes X and Y should be stationary in order to ensure
the existence of the DI rate. From practical perspective, it is required that the sequences
X N and Y N will be stationary to ensure that the causal influence does not change over
5
the observed sequences. Ergodicity is assumed to ensure that the observed sequences truly
represent the underlying processes. Finally, the Markovity assumption is common in modeling
real-life systems which have finite memory. For example, [18, 19, 7] used Markov models in
analysis of neural recordings, [20] and [21] used Markov models in financial modeling, and
[22]–[23] in social networks dynamics. We formulate the assumption of Markovity of order
m in the observed sequences via:
i−1
i−1
i−1
∀i > m, f (yi |Y1i−1 ) = f (yi |Yi−m
) and f (yi |Y1i−1 , X1i ) = f (yi |Yi−m
, Xi−m
).
(14)
In (14) we use the simplifying assumption that the dependence of yi on past samples of Y1i−1
and X1i is of the same order. These definitions and the estimation methods defined in the
sequel can be easily extended to two different orders. Moreover, in (14) we implicitly assume
i−1
i−1
that given (Yi−m
, Xi−m
), yi is independent of Xi which reflects a setting where Xi and Yi are
simultaneously measured.
A2) The entropy of the first sample y1 exists, i.e., |h(Y1 )| < ∞.
A3) The following conditional entropy exists: |h(Ym+1 |Y1m , X1m )| < ∞.
Assumptions A2) and A3) are required to mathematically insure that the DI rate exists and is
equal to a simple expression that depends on the finite memory length m. Moreover, Assumptions
A2) and A3) prevent the degenerate case of deterministic Y1 or deterministic relationship between
Ym+1 and Y1m , X1m (this is one of the scenarios discussed in [6]). Under these assumptions, [7,
Lemmas 3.1 and 3.2] imply that I(X → Y) exists and is equal to:
i−1
i−1
I(X → Y) = I(Xi−m
; Yi |Yi−m
),
i > m,
(15)
In view of (15), I(X → Y) can be given the following interpretation:
Given the past of the sequence Y , namely Y1i−1 , how much the past of the sequence X, namely
helps in predicting the next sample of Y , Yi ?
It can be observed that (15) is a function of m, the Markov order. When one is interested in a
data-driven estimator of the DI, m is unknown and must be estimated from the observed sequences.
X1i−1 ,
4.2
Estimating the Markov Order m
While it is a reasonable assumption that the observed time-series have finite memory and thus obey
a Markov model, the order m should be estimated from the data.
A possible approach for estimating m is to choose the value that facilitates the best prediction of
future samples (see [24] and references therein). Specifically, let M be a (finite) set of candidate
Markov orders. m̂ is estimated to be the value in this set that minimizes a pre-defined loss function
in predicting the next sample of Y1N from m̂ previous samples. While in [5] it was proposed to use
the prediction method of [25], i.e., use k-NN prediction of the next sample in Y1N based on the
past samples of Y1N , this approach ignores the dependency between Y1N and X1N . To account for this
dependency, we propose to predict the next sample of Y1N based on the past m̂ samples of both Y1N
and X1N .
i−1
i−1
Let ϕ : R2M → R be a prediction function that predicts Yi+1 from (Yi−m
, Xi−m
). To measure
the quality of prediction we use the `2 -loss (or distance), e.g., mean-square-error. Thus, the model
order is estimated via:
i−1
i−1
m̂ = argmin E d2 (Yi+1 , ϕ((Yi−m
, Xi−m
)) ,
(16)
m∈M
6
Figure 1: 3-NN prediction. Illustration of 3-NN prediction of y4 , assuming m = 3.
where the expectation averages over everything that is random, and can be approximated via averaging. Note that any prediction method can be used in (16). In fact, in [24] it was proposed to use
an ensemble of predictors in order to increase the prediction power. Yet, this comes at the cost of
much higher computational complexity. Moreover, the numerical study in [24] indicates that when
the Markov order m is small, k-NN is a very efficient predictor that out-performs significantly more
complicated models such as support-vector-regression and regression based on non-linear terms.
On the other hand, when m is large (and N is fixed), k-NN suffers from the curse of dimensionality
[26, Section 6.3] and performs poorly. Note that, as stated in [27, Sec. V.E], the number of samples
required for accurate estimation of DI grows exponentially with the Markov order m (this follows as
increasing m can be viewed as increasing the state space). Therefore, as in many practical settings
the number of samples is limited [7, 28, 29], we focus on settings where m is relatively small.3 .
Fig. 1 illustrates the procedure for estimating y4 , for k = 3 and m = 3. The estimator first
finds the k = 3 tuples that are nearest to the tuple (x1 , x2 , x3 , y1 , y2 , y3 ). Each of these tuples has a
response variable; y12 , y13 and y20 in Fig. 1. These responses are used to predict y4 , resulting in the
loss δ4 . Note that instead of averaging one can use weighted averaging based on the distances from
(x1 , x2 , x3 , y1 , y2 , y3 ). This procedure is repeated for every tuple (note that we do not search NN
among the tuples that overlap with the current tuple). The calculated loss values δi are averaged
resulting in ∆m , m ∈ M. This procedure is repeated for all m ∈ M, and the value that minimize
the average loss is declared as m̂.
Next, we discuss the estimation of the DI measure, while assuming that the Markov order m
was correctly estimated.
4.3
Estimating Directed Information
i−1
i−1
To simplify the notation we let Xi− , Xi−m
denote the past of Xi , and Yi− , Yi−m
denote the past
of Yi . Using this notation, we write (15) as:
i−1
i−1
I(X → Y) = I(Xi−m
; Yi |Yi−m
)
= h(Y, Y − ) − h(Y − ) − h(Y, Y − , X − ) + h(Y − , X − ).
(17)
3 When this does not hold one can either down-sample the observed sequences or decimate (Y i−1 , X i−1 ) when estii−m
i−m
mating the DI, see the discussion in [7]
7
4.3.1
The KSG Estimator
We begin with an estimator that extends the approach used to estimate MI in (6). This estimator
was presented in [5] for estimating the TE measure. Let ρk,i,p (X − , Y − , Y ) denote the distance from
the tuple (Xi− , Yi− , Yi ) to its k-NN. In the following we refer to this distance as ρk,i,p . Recalling that
the dimensions of Xi− and Yi− are m, we note that ρk,i,p is calculated in a space with dimension
2m + 1. Similarly to (7), the individual entropies are estimated using ρk,i,p , the distance calculated
in the largest space (Xi− , Yi− , Yi ). The resulting estimators of the individual entropy terms are then
given by:
ĥKSG (Y − ) =
ĥKSG (Y − , Y ) =
N
X
1
log N + log cm,∞ + m log ρk,i,∞ − ψ(nY − ,i,∞ + 1)
N − m i=m+1
(18a)
N
X
1
log N + log(cm,∞ · c1,∞ )
N − m i=m+1
+ (m + 1) log ρk,i,∞ − ψ(n(Y − ,Y ),i,∞ + 1)
ĥKSG (Y − , X − ) =
N
X
1
log N + log(cm,∞ · cm,∞ )
N − m i=m+1
+ 2m log ρk,i,∞ − ψ(n(Y − ,X − ),i,∞ + 1)
ĥKSG (Y − , X − , Y ) =
(18b)
(18c)
N
X
1
log N + log(cm,∞ · cm,∞ · c1,∞ )
N − m i=m+1
+ (2m + 1) log ρk,i,∞ − ψ(k) .
(18d)
Combining the entropy estimators in (18) we obtain:
IˆKSG (X → Y) = ψ(k)+
N
X
1
ψ(nY − ,i,∞ +1)−ψ(n(Y − ,Y ),i,∞ +1)−ψ(n(Y − ,X − ),i,∞ +1) . (19)
N −m i=m+1
Similarly to the KSG estimator for MI in (6), the joint entropy h(Y − , X − , Y ) is estimated using
the KL estimator, see (18d), while the other entropy terms are estimated using sample dependent
expressions, see (18a)–(18c).
4.3.2
The GOV Estimator
The second DI estimator extends the approach used to estimate MI in (11). Following the steps
leading to (4) and (10), we obtain the following estimators:
N
X
1
ĥGOV (Y ) =
log N +log cm,2 +m log ρk,i,2 −log(nY − ,i,2 )
N −m i=m+1
−
ĥGOV (Y − , Y ) =
N
X
1
log N +log(cm+1,2 )+(m+1) log ρk,i,2 −log(n(Y − ,Y ),i,2 )
N −m i=m+1
8
(20a)
(20b)
N
X
1
ĥGOV (Y , X ) =
log N +log(c2m,2 )+2m log ρk,i,2 −log(n(Y − ,X − ),i,2 )
N −m i=m+1
−
−
ĥGOV (Y − , X − , Y ) =
N
X
1
log N +log(c2m+1,2 )+(2m+1) log ρk,i,2 −ψ(k) .
N −m i=m+1
(20c)
(20d)
Combining the entropy estimators in (20) we obtain:
cm+1,2 · c2m,2
IˆGOV (X → Y) = ψ(k)+log
c2m+1,2 · c1,2
+
4.3.3
N
X
1
log(nY − ,i,2 )−log(n(Y − ,Y ),i,2 )−log(n(Y − ,X − ),i,2 ) .
N −m i=m+1
(21)
On the Statistical Significance of the Estimated DI
Since I(X → Y) is estimated from a finite number of samples, one may want to assess the statistical
significance of this estimation. A lack of statistical significance may imply that there is no causal
influence between the underlying random processes, and the estimated values are due to either
noise or estimation error. In such a case one may choose to set the estimated DI to zero. While for
estimating GC there are known methods for quantifying the statistical significance via the respective
null-distribution [30, Sec. 2.5], the null-distribution is not known for non-parametric estimation
of DI from continuous alphabet sequences. An alternative method for evaluating the statistical
significance is via a non-parametric bootstrapping procedure in the spirit of [31]. Specifically, the
idea is to randomly shuffle and re-sample X1N such that the causal interactions from X to Y are
destroyed (Y is not changed). We repeat this shuffle and re-sample procedure L times and for
each l ∈ {1, 2, . . . , L} we estimate the DI I(X(l) → Y), where X(l) denotes the shuffled sequence.
Since this construction destroys the causal influence, the L new estimated DI values are assumed
to be taken from the null-distribution of no causal influence. The statistical significance is then
determined by the resulting P-value with parameter εp . Finally, we note that the main drawback
of such a bootstrapping procedure is the significant increase in computational complexity, since
applying such a procedure amounts to multiplying the computational complexity by a factor of at
least 20 (for the common value of εp = 0.05).
4.3.4
The JVHW Estimator
The above KSG and GOV estimators allow for a continuous input alphabet. A different approach is
to first quantize the observed sequences into only a few bins in order to estimate an empirical probability mass functions (PMFs), and then estimate the DI from these PMFs. Such an estimator was
developed by Jiao, Venkat, Han, and Weissman (JVHW) [32], based on min-max optimal estimation
of the entropy of discrete distributions [27]. Specifically, the estimator uses the relationship (17) to
independently estimate each of the entropy terms. Its main advantages are the proven optimality
in estimating the entropy terms (of the discrete distributions) as well as its linear time complexity
(in contrast to the universal estimators detailed in [33]). On the other hand, when the number
of available samples is small, quantizing the input sequences can lead to significant performance
degradation (see the numerical study in Section 5). To quantize the observed sequences one may
use the optimal Lloyd-Max scalar quantizer. Our numerical study showed that when the cardinality
of the discrete alphabet is large enough, Lloyd-Max quantization yields performance very similar to
a naive quantization based on equal-size bins.4
4 Here
we assumed that the probability of observing samples with very large magnitude is negligible.
9
To evaluate the statistical significance of the estimated values we use the same bootstrapping
procedure discussed in the previous section. Note that while for discrete-alphabet samples the
work [34] derived the asymptotic (N → ∞) null-distribution, this result may not be used when the
number of samples is small.
5
Numerical Study
In this section we examine the performance of the discussed estimators in several simulated scenarios. We also consider an estimation of the GC via the toolbox [30]. This is motivated by the fact
that GC is widely used to quantify causal influence, and in particular when X and Y are generated
through a multivariate auto-regressive process, then GC is equal to twice the DI [35]. We consider
linear as well as non-linear scenarios, where some of the scenarios were taken from [7], and focus
on the regime where N is at the order of few thousands (in contrast to [7] that focused on the
case of N = 105 ). Relatively small values of N are motivated by practical applications where the
number of samples is limited, e.g., [28, 36]. Our simulation results indicate that the discussed k-NN
estimators (KSG and GOV) achieve roughly the same accuracy as the estimator presented in [7],
but with an order of magnitude less samples.
5.1
Linear Interaction
Let xi , zi , i = 1, 2, . . . , N , be i.i.d. zero-mean and unit-variance Gaussian RVs, where xi is independent of zi . We generate the sequence {yi }N
i=1 via:
yi = β1 xi−1 + β2 xi−2 + zi ,
0 ≤ β ≤ 1.
(22)
Observe that in (22) X causally influence Y while Y has no causal influence on X. As was shown
in [7, Appendix B], when β1 , β2 6= 0 then I(X → Y) is given by:
2
β1 + β22 + 1
1
−1
.
(23)
I(X → Y) = log2 (e) log (|β1 β2 |) + cosh
2
2|β1 β2 |
When β1 = 0, β2 6= 0, then I(X → Y) = 21 log2 1 + β22 , and when β1 6= 0, β2 = 0, then I(X →
Y) = 21 log2 1 + β12 . Here, the DI is measured in bits.
Fig. 2 depicts the mean of the estimated DIs versus the value of β1 , for β2 = β1 and for
β2 = 1 − β1 , where β1 ∈ {0, 0.1, . . . , 1}. The DI values were estimated from N = 3000 samples, and
the mean was calculated over 48 independent (with different random generator seeds) trials. For
both plots in Fig. 2, the average standard deviation for the GC estimator is about 0.01 for I(X → Y)
and 0.0006 for I(Y → X); the average standard deviation for the JVHW estimator is about 0.05 for
I(X → Y) and 0.03 for I(Y → X); the average standard deviation for the KSG estimator is about
0.02 for I(X → Y) and 0.006 for I(Y → X); and, the average standard deviation for the GOV
estimator is about 0.02 for I(X → Y) and 0.002 for I(Y → X). The statistical significance of the
estimated values was evaluated using the approach discussed in Section 4.3.3 (for the GC estimates
we used the null-distribution as stated in [30, Sec. 2.5]). For β2 = β1 and I(X → Y), when
β1 ≥ 0.5, all the estimators yield significant estimations in all trials. While the GC estimator yield
significant estimations in all trials when β1 ≥ 0.1, the KSG and GOV estimators yield significant
estimations in all trials when β1 ≥ 0.3. For β1 = 0.1 and β1 = 0.2 the KSG and GOV estimators
yield significant estimations in about 20% and 90% of the trials, respectively. The significance of
the JVHW estimator gradually increases to 100% in the range 0 ≤ β1 ≤ 0.5. For β2 = β1 and
I(X → Y), as the true DI values are relatively large, all the estimators yield significant estimations
in all the trials.
10
0.6
0.8
Theory X → Y
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
Average DI
0.6
0.5
0.5
0.4
Average DI
0.7
Theory X → Y
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
0.2
0.4
β1
0.6
0.8
0
1
(a)
0
0.2
0.4
β1
0.6
0.8
1
(b)
Figure 2: Average DI estimates versus β1 for the interaction (23), N = 3000. Four quantization levels are used in the
JVHW estimator. In the KSG and GOV estimators K = 8. In sub-figure (a) β2 = β1 while in sub-figure (b) β2 = 1 − β1 .
Comparing the estimation results in Fig. 2 to the estimation results reported in [7, Fig. 2], one
can observe that the curves are very similar (KSG and GOV versus the KDE estimator of [7]). Yet,
there is a fundamental difference between the estimators: in [7, Fig. 2] the number of samples is
N = 105 , while in the above Fig. 2 the number of samples is N = 3000. Thus, in the considered
scenario, to achieve the same accuracy the KSG and GOV estimators require an order of magnitude
less samples compared to the KDE estimator.
To explore the performance of the discussed estimators when N is larger, we repeated the experiment specified in (22) when N = 20000, and β2 = β1 . The results are depicted in Fig. 3.
Observing Fig. 3a we note that the accuracy of the KSG and GOV estimator is significantly higher
(compared to Fig. 2) when β1 ≥ 0.4. Moreover, the standard deviations of the estimated values are
smaller by about an order of magnitude. On the other hand, for 0.1 ≤ β1 ≤ 0.3, one can observe
an apparent bias. This bias is due to an inaccurate estimation of the Markov order. When β1 ≤ 0.3
the noise level is significantly higher than the signal level and the estimation method discussed in
Section 4.2 under estimates the true Markov order. To verify this observation, in Fig. 3b we present
the average estimated DI, for the KSG estimator, and for several alternatives for the Markov order
estimation:
• The approach discussed in Section 4.2. The corresponding curve is denoted by “Baseline”.
• The approach proposed in [25] where a k-NN estimator is used to estimate the Markov order
based only on the past samples of Y, ignoring the past samples of X. The corresponding curve
is denoted by “Ragwitz”.
• Fixing the Markov order to the correct order according to (22), i.e., setting m = 2. The
corresponding curve is denoted by “m = 2”.
• Setting the Markov order to m = 4. The corresponding curve is denoted by “m = 4”.
It can be observed that the method by [25] performs much worse than the method discussed in
Section 4.2. In fact, the curve corresponding to the “Ragwitz” method is almost identical to a curve
generated by setting m = 1. Setting m = 2, the true order, leads to almost no errors comparing to
the theoretical values. Finally, setting m = 4 works better for small values of β1 (when the problem
11
0.7
0.5
0.4
0.5
0.3
0.4
0.3
0.2
0.2
0.1
0.1
0
0
0.2
Theory X → Y
m = 2, X → Y
m = 2, Y → X
Baseline X → Y
Baseline Y → X
Ragwitz X → Y
Ragwitz Y → X
m = 4, X → Y
m = 4, Y → X
0.6
Average DI
0.6
Average DI
0.7
Theory X → Y
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
0.4
β1
0.6
0.8
0
1
0
0.2
(a)
0.4
β1
0.6
0.8
1
(b)
Figure 3: Average DI estimates versus β1 for the interaction (22), N = 20000. Six quantization levels are used in the
JVHW estimator. In the KSG and GOV estimators K = 8. (a) Average estimated DI of the JVHW, GC, KSG, and GOV
estimators, when β2 = β1 . (b) Average estimated DI, of the KSG estimator with different Markov order estimation methods,
when β2 = β1
of estimating the Markov order is very challenging), yet a significant performance degradation can
be observed for larger values of β1 .
5.2
Quadratic Interaction
Next, we consider a quadratic dependency between X and Y. Again, we let xi , zi , i = 1, 2, . . . , N
be i.i.d. zero-mean and unit-variance Gaussian RVs, where xi is independent of zi . We generate the
sequence {yi }N
i=1 via:
yi = β1 x2i−1 + β2 x2i−2 + zi .
(24)
A similar model was studied in [7, Sec. V.B]. Similarly to Section 5.1, we consider two scenarios:
β2 = β1 and β2 = 1 − β1 . When β2 = β1 we expect I(X → Y) to increase monotonically with β1 ,
and when β2 = 1 − β1 we expect to observe a “U” shape as in Fig. 2b. Similarly to Fig. 2 there
should be no causal influence from Y to X. We note that calculating the theoretical DI values of
this model seems intractable.
Fig. 4 depicts the mean of the estimated DIs versus the value of β1 , for β2 = β1 and for
β2 = 1 − β1 . N = 3000 samples were used to estimated the DI values. The mean values reported
in Fig. 4 were calculated over 48 independent trials. Similarly to the linear scenario, for both plots
the average standard deviation for the GC estimator is about 0.001 for I(X → Y) and 0.0008 for
I(Y → X); the average standard deviation for the JVHW estimator is about 0.04 for I(X → Y)
and 0.03 for I(Y → X); the average standard deviation for the KSG estimator is about 0.02 for
I(X → Y) and 0.006 for I(Y → X); and, the average standard deviation for the GOV estimator
is about 0.02 for I(X → Y) and 0.002 for I(Y → X). Except the GC estimator, the statistical
significance is also similar to the linear scenario. It can be observed that as expected when β2 = β1 ,
the estimated DI monotonically increases with β1 . Moreover, the results are similar to [7, Fig.
3], again with the exception that significantly smaller number of samples was used for estimation
(N = 3000 versus N = 105 ). Finally, Fig. 4 indicates that the JVHW estimator is able to capture
the causal influence from X to Y, although the estimated values are smaller than those estimated
12
0.6
0.6
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
Average DI
0.4
0.5
0.4
Average DI
0.5
0.3
0.3
0.2
0.2
0.1
0.1
0
0
0.2
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
0.4
β1
0.6
0.8
0
1
0
0.2
0.4
0.6
0.8
1
β1
(a)
(b)
Figure 4: Average DI estimates versus β1 for the interaction (24), N = 3000. Four quantization levels are used in the
JVHW estimator. In the KSG and GOV estimators K = 8. In sub-figure (a) β2 = β1 while in sub-figure (b) β2 = 1 − β1 .
by the k-NN estimators. On the other hand, as expected and as indicated in [7], GC is not able to
capture the causal influence from X to Y. This follows as GC is based on a linear model while (24)
obeys a quadratic one.
Next, we discuss two non-linear coupling maps that were discussed in [37].
5.3
Noisy Hénon Map
We first study a model that consists of unidirectional coupling via the Hénon map [38]. Let x̃j , ỹj ∼
U(0, 1), j = 1, 2, and let Ñ = 105 + 3000. We generate the sequence {ỹi , x̃i }Ñ
i=3 via:
x̃i+2 = 1.4 − x̃2i+1 + 0.3x̃i ,
ỹi+2 = 1.4 − (β x̃i+1 + (1 − β)ỹi+1 )ỹi+1 + 0.3ỹi ,
(25)
for 0 ≤ β ≤ 1. Next, let zx,i , zy,i , i = 1, 2, . . . , Ñ be i.i.d. zero-mean and unit-variance Gaussian RVs,
where zx,i is independent of zy,i . We generate the sequence {yi , xi }Ñ
i=1 via:
yi = ỹi + γ · zy,i ,
xi = x̃i + γ · zx,i ,
(26)
where γ is a gain parameter that sets the signal-to-noise-ratio. The DI is estimated from the last
N = 3000 samples of the sequences {yi , xi }.
We first note that in (25) the parameter β controls the strength of coupling between the two sequences. Moreover, (25) indicates that X causally influences Y, and this influence should increase
with β. On the other hand, based on (25) we do not expect any causal influence from Y to X.
As discussed in [37], for the noiseless scenario, namely γ = 0, when β > 0.7, then X and Y are
fully synchronized and the two sequences are indistinguishable. Thus, in this case the causal flow
of information is zero. Fig. 5a, that considers the almost noiseless setting,5 indicates that the KSG
estimator indeed estimates an increasing causal influence from X to Y for β < 0.7. Moreover, this
causal influence drops to almost zero when β is increased beyond 0.7, indicating on the (almost)
5 Here we do not set γ to exactly zero in order not to violate Assumption A3), and to prevent numerical instabilities when
estimating GC.
13
1.4
1
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
0.5
0.4
Average DI
1.2
Average DI
0.6
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
0.8
0.6
0.3
0.2
0.4
0.1
0.2
0
0
0.2
0.4
β
0.6
0.8
0
1
0
0.2
(a)
0.4
β
0.6
0.8
1
(b)
Figure 5: Average DI estimates versus β for the interaction (25)–(26), N = 3000. Four quantization levels are used in
the JVHW estimator. In the KSG and GOV estimators K = 8. (a) γ = 0.001. (b) γ = 0.3.
deterministic relationship between X and Y. This result is in full correspondence to [37, Figs. 2a
and 2b]. Fig. 5a also indicates that the KSG estimator correctly infers a negligible causal influence
from Y to X for all values of β. It can further be observed from Fig. 5a that the curve corresponding
to the GOV estimator is biased, and that the JVHW estimator misses the drop in causal influence
for β > 0.7.
In Fig. 5a we consider the case of γ = 0.3. It can be observed that for the KSG estimator the
results are similar to the case of γ = 0.001, with the exception that the DI, in both directions, is
significantly larger than zero for β > 0. This exactly matches the desired behavior as indicated in
[37, Figs. 4a and 4b]. For this setting, the GOV estimator seems to be less biased, while the JVHW
still misses the decrease in causal influence for β > 0.7.
Finally, we note that for both γ = 0.001 and γ = 0.3, GC fails to infer any significant causal
influence from X to Y or from Y to X.
5.4
Sigmoid Coupling
In this sub-section we discuss a scenario where X drives Y through a sigmoid function [37, Sec.
E], defined as:
Sigmoid(θ) =
1
.
1 + exp{−θ}
Let x1 , y1 ∼ U(0, 1). We generate the sequence {yi , xi }N
i=2 via:
25xi
+ 2 cos(1.2 · i) + zx,i ,
4(x2i + 1)
= 0.1yi2 − β Sigmoid(xi )2 − 0.3 + zy,i ,
xi+1 = 0.125xi +
yi+1
(27)
where 0 ≤ β ≤ 1, and zx,i , zy,i , i = 2, 3, . . . , N , are i.i.d. zero-mean and unit-variance Gaussian RVs
(zx,i is independent of zy,i ). (27) indicates that X causally influences Y, and this influence should
increase with β. On the other hand, based on (27) we do not expect any causal influence from Y
to X.
14
0.1
JVHW X → Y
JVHW Y → X
2GC X → Y
2GC Y → X
KSG X → Y
KSG Y → X
GOV X → Y
GOV Y → X
0.09
0.08
AverageDI
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0
0.2
0.4
β
0.6
0.8
1
Figure 6: Average DI estimates versus β for the interaction (27), N = 3000. Four quantization levels are used in the
JVHW estimator. In the KSG and GOV estimators K = 8.
Fig. 6 depicts the average estimated DI versus β. It can be observed that for all four estimators
ˆ to Y) increases with β. We note here that the curve
(KSG, GOV, GC, and JVHW) the average I(X
corresponding to the GC estimator is somewhat surprising as the interaction (27) is clearly nonˆ
linear. When examining the average I(Y
to X) it can be observed that both GOV and JVHW find
a non-negligible causal influence (which does not exist according to (27) and [37, Sec. E]). On
the other hand, the KSG estimator estimates a negligible amount of causal influence from Y to X.
Thus, the KSG estimator is the only one to follow the expected results as indicated in [37, Figs. 11a
and 11c].
References
[1] K. Hlavackova-Schindler, M. Palus, M. Vejmelka, and J. Bhattacharya, “Causality detection
based on information-theoretic approaches in time series analysis,” Physics Reports, vol. 441,
pp. 1–46, 2007.
[2] C. W. J. Granger, “Investigating causal relations by econometric models and cross-spectral
methods,” Econometrica, vol. 37, no. 3, pp. 424–438, 1969.
[3] J. L. Massey, “Causality, feedback, and directed information,” in Proc. Int. Symp. Inf. Theory
Appl., Waikiki, HI, USA, 1990, pp. 303–305.
[4] G. Kramer, “Directed information for channels with feedback,” Ph.D. dissertation, Swiss Federal Institute of Technology, 1998.
[5] M. Wibral, R. Vicente, and M. Lindner, “Transfer entropy in neuroscience,” Directed Information Measures in Neuroscience, Understanding Complex Systems, pp. 3–36, 2014.
[6] D. Janzing, D. Balduzzi, M. Grosse-Wentrup, and B. Scholkopf, “Quantifying causal influences,” The Annals of Statistics, vol. 41, no. 5, pp. 2324–2358, 2013.
15
[7] R. Malladi, G. Kalamangalam, N. Tandon, and B. Aazhang, “Identifying seizure onset zone
from the causal connectivity inferred using directed information,” IEEE Jour. of Sel. Topics in
Sig. Proc., vol. 10, no. 7, pp. 1267–1283, 2016.
[8] S. Sabesan, L. B. Good, K. S. Tsakalis, A. Spanias, D. M. Treiman, and L. D. Iasemidis, “Information flow and application to epileptogenic focus localization from intracranial eeg,” IEEE
Trans. Neural Syst. Rehabil. Eng., vol. 17, no. 3, pp. 244–253, 2009.
[9] T. M. Cover and J. A. Thomas, Elements of Information Theory 2nd Edition, 2nd ed.
Interscience, 2006.
Wiley-
[10] Q. Wang, S. R. Kulkarni, and S. Verduú, “Universal estimation of information measures for
analog sources,” Foundations and Trends in Communications and Information Theory, vol. 5,
no. 3, pp. 265–353, 2008.
[11] Y. Moon, B. Rajagopalan, and U. Lall, “Estimation of mutual information using kernel density
estimators,” Physical Review E, vol. 52, no. 3, pp. 2318–2321, 1995.
[12] P. Grassberger and I. Procaccia, “Measuring of strangeness of strange attractors,” Physica D,
vol. 9, no. 1-2, pp. 189–208, 1983.
[13] F. W. J. Olver, D. W. Lozier, R. F. Boisvert, and C. W. Clark, Eds., NIST Handbook of Mathematical Functions, 1st ed. Cambridge University Press, 2010.
[14] X. Wang, “Volumes of generalized unit balls,” Mathematics Magazine, vol. 78, no. 5, pp. 390–
395, 2005.
[15] L. F. Kozachenko and N. N. Leonenko, “Sample estimate of the entropy of a random vector,”
Problemy Peredachi Informatsii, vol. 23, no. 2, pp. 9–16, 1987.
[16] A. Kraskov, H. Stögbauer, and P. Grassberger, “Estimating mutual information,” Physical Review
E, vol. 69, no. 6, 2004.
[17] W. Gao, S. Oh, and P. Viswanath, “Demystifying fixed k-nearest neighbor information estimators,” 2016, available at https://arxiv.org/abs/1604.03006.
[18] D. Lederman and J. Tabrikian, “Classification of multichannel eeg patterns using parallel hidden markov models,” Medical & Biological Engineering & Computing, vol. 50, no. 4, pp. 319–
328, Apr. 2012.
[19] T. Wissel, T. Pfeiffer, R. Frysch, R. T. Knight, E. F. Chang, H. Hinrichs, J. W. Rieger, , and
G. Rose, “Hidden markov model and support vector machine based decoding of finger movements using electrocorticography,” Journal of Neural Engineering, vol. 10, no. 5, Oct. 2013.
[20] C. Erlwein, “Applications of hidden markov models in financial modelling,” Ph.D. dissertation,
Brunel University, 2008.
[21] J. G. Dias, J. K. Vermunt, and S. Ramos, “Mixture hidden markov models in finance research,”
Advances in Data Analysis, Data Handling and Business Intelligence, pp. 451–459, Jul. 2009.
[22] T. A. B. Snijders, “The statistical evaluation of social network dynamics,” Sociological Methodology, vol. 31, no. 1, pp. 361–395, 2001.
[23] C. Heaukulani and Z. Ghahramani, “Dynamic probabilistic models for latent feature propagation in social networks,” in Proc. 30th Int. Conf. Machine Learning, Atlanta, GA, USA, 2013,
pp. 275–283.
16
[24] Y. Murin, A. Goldsmith, and B. Aazhang, “Estimating the memory order of electrocorticography recordings,” To be submitted to IEEE Transactions on Biomedical Engineering, 2017.
[25] M. Ragwtiz and H. Kantz, “Markov models from data by simple nonlinear time series predictors in delay embedding spaces,” Physical Review E, vol. 65, no. 5, pp. 1–12, 2002.
[26] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning, 2nd ed.
Springer, 2009.
[27] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Minimax estimation of functionals of discrete
distributions,” IEEE Transactions on Information Theory, vol. 61, no. 5, pp. 2835–2885, May
2015.
[28] Y. Murin, J. Kim, J. Parvizi, and A. Goldsmith, “Sozrank: A new approach for localizing the
epileptic seizure onset zone,” Submitted to PLoS Comput. Biol., 2017.
[29] Y. Murin, J. Kim, and A. Goldsmith, “Tracking epileptic seizure activity via information theoretic graphs,” in Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA,
USA, 2016.
[30] L. Barnett and A. Sheth, “The MVGC multivariate Granger causality toolbox: A new approach
to Granger-causal inference,” Journal of Neuroscience Methods, vol. 223, pp. 50–68, 2014.
[31] C. Diks and J. DeGoede, A general nonparametric bootstrap test for Granger causality. Institute
of Physics Publishing, 2001, in Global Analysis of Dynamical Systems.
[32] “The Jiao–Venkat–Han–Weissman (JVHW) shannon entropy, renyi entropy, and mutual information estimator,” http://web.stanford.edu/∼tsachy/index jvhw.html, accessed: 2017-0913.
[33] J. Jiao, H. H. Permuter, L. Zhao, Y.-H. Kim, and T. Weissman, “Universal estimation of directed
information,” IEEE Transactions on Information Theory, vol. 59, no. 10, pp. 6220–6242, Oct.
2013.
[34] I. Kontoyiannis and M. Skoularidou, “Estimating the directed information and testing for
causality,” IEEE Transactions on Information Theory, vol. 62, no. 11, pp. 6053–6067, Nov.
2016.
[35] N. Soltani, Inferring signaling structures in the brain via directed information. Doctoral Thesis,
Stanford University, 2015.
[36] T. Diamandis, Y. Murin, and A. Goldsmith, “Ranking causal influence of financial markets via
directed information graphs,” in 52nd Annual Conference on Information Sciences and Systems,
Princeton, NJ, USA, 2018, to be submitted to.
[37] K. Ishiguro, N. Otsu, M. Lungarella, and Y. Kuniyoshi1, “Comparison of nonlinear granger
causality extensions for low-dimensional systems,” Physical Review E, vol. 77, no. 3, 2008.
[38] M. Hénon, “A two-dimensional mapping with a strange attractor,” Communications in Mathematical Physics, vol. 50, no. 1, pp. 69–77, 1976.
17
| 7cs.IT
|
Message Passing and Combinatorial Optimization
by
Siamak Ravanbakhsh
A thesis submitted in partial fulfillment of the requirements for the degree of
Doctor of Philosophy
Department of Computing Science
University of Alberta
c Siamak Ravanbakhsh, 2015
Abstract
Graphical models use the intuitive and well-studied methods of graph theory to implicitly represent
dependencies between variables in large systems. They can model the global behaviour of a complex
system by specifying only local factors.This thesis studies inference in discrete graphical models from
an “algebraic perspective” and the ways inference can be used to express and approximate NP-hard
combinatorial problems.
We investigate the complexity and reducibility of various inference problems, in part by organizing
them in an inference hierarchy. We then investigate tractable approximations for a subset of these
problems using distributive law in the form of message passing. The quality of the resulting message
passing procedure, called Belief Propagation (BP), depends on the influence of loops in the graphical
model. We contribute to three classes of approximations that improve BP for loopy graphs (I) loop
correction techniques; (II) survey propagation, another message passing technique that surpasses BP
in some settings; and (III) hybrid methods that interpolate between deterministic message passing and
Markov Chain Monte Carlo inference.
We then review the existing message passing solutions and provide novel graphical models and inference techniques for combinatorial problems under three broad classes: (I) constraint satisfaction problems (CSPs) such as satisfiability, coloring, packing, set / clique-cover and dominating / independent set
and their optimization counterparts; (II) clustering problems such as hierarchical clustering, K-median,
K-clustering, K-center and modularity optimization; (III) problems over permutations including (bottleneck) assignment, graph “morphisms” and alignment, finding symmetries and (bottleneck) traveling
salesman problem. In many cases we show that message passing is able to find solutions that are either
near optimal or favourably compare with today’s state-of-the-art approaches.
ii
To the memory of my grandparents.
iii
Acknowledgements
Many have helped me survive and grow during my graduate studies. First, I would like to thank my
supervisor Russell Greiner, for all I have learned from him, most of all for teaching me the value of
common sense in academia and also for granting me a rare freedom in research. During these years, I
had the chance to learn from many great minds in Alberta and elsewhere. It has been a pleasure working
with David Wishart, Brendan Frey, Jack Tuszynski and Barnabás Póczos. I am also grateful to my
committee members, Dale Schuurmans, Csaba Szepesvári, Mohammad Salavatipour and my external
examiner Cristopher Moore for their valuable feedback.
I have enjoyed many friendly conversations with colleagues: Babak Alipanahi, Nasimeh Asgarian,
Trent Bjorndahl, Kit Chen, Andrew Delong, Jason Grant, Bret Hoehn, Sheehan Khan, Philip Liu,
Alireza Makhzani, Rupasri Mandal, James Neufeld, Christopher Srinivasa, Mike Wilson, Chun-nam Yu
and others.
I am thankful to my friends: Amir, Amir-Mashoud, Amir-Ali, Amin, Arash, Arezoo, Azad, Babak(s),
Fariba, Farzaneh, Hootan, Hossein, Kent, Kiana, Maria, Mariana, Meys, Meysam, Mohsen, Mohammad,
Neda, Niousha, Pirooz, Sadaf, Saeed, Saman, Shaham, Sharron, Stuart, Yasin, Yavar and other for all
the good time in Canada. Finally I like to thank my family and specially Reihaneh and Tare for their
kindness and constant support.
Here, I acknowledge the help and mental stimulation from the playful minds on the Internet, from
Reddit to stackexchange and the funding and computational/technical support from computing science
help desk, Alberta Innovates Technology Futures, Alberta Innovates Center for Machine Learning and
Compute Canada.
iv
Contents
Abstract
ii
Acknowledgements
iv
Contents
v
List of tables
x
List of figures
xi
General notation
xii
Introduction
I
1
Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
Inference by message passing
6
1 Representation, complexity and reducibility
8
1.1
The problem of inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
1.2
The inference hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
1.2.1
Single marginalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
1.2.2
Complexity of general inference classes . . . . . . . . . . . . . . . . . . . . . . . .
18
1.2.3
Complexity of the hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
Polynomial-time inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
1.3
v
1.4
Reductions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
1.4.1
Marginalization and integration . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
Marginalization reduces to integration . . . . . . . . . . . . . . . . . . . . . . . .
26
Integration reduces to marginalization . . . . . . . . . . . . . . . . . . . . . . . .
27
Min-max reductions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
Min-max reduces to min-sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
Min-max reduces to sum-product . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
1.4.2
2 Approximate inference
2.1
2.2
Belief Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
2.1.1
Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
2.1.2
The limits of message passing . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
Tractable factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
2.2.1
Sparse factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
K of N factors for sum-product . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
K of N factors for min-sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
Large number of constraint factors . . . . . . . . . . . . . . . . . . . . . . . . . .
42
Inference as optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
2.3.1
Sum-product BP and friends . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
2.3.2
Min-sum message passing and LP relaxation . . . . . . . . . . . . . . . . . . . .
46
2.3.3
Min-max and other families . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
Loop corrections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
2.4.1
Short loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
2.4.2
Long loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
2.4.3
Both message dependencies and short loops . . . . . . . . . . . . . . . . . . . . .
52
Message region-graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
Grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
2.2.2
2.3
2.4
34
2.4.4
vi
2.5
2.6
Regular Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
Survey Propagation: semirings on semirings . . . . . . . . . . . . . . . . . . . . . . . . .
56
2.5.1
Decomposition of the integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
2.5.2
The new factor-graph and semiring . . . . . . . . . . . . . . . . . . . . . . . . . .
62
2.5.3
The new integral and marginals . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
2.5.4
Counting survey propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
Messages and particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
2.6.1
Markov Chain Monte Carlo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
Metropolis-Hasting Algorithm and Gibbs sampling . . . . . . . . . . . . . . . . .
67
2.6.2
Hybrid methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
68
2.6.3
Perturbed Belief Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
Combining the operators
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
Perturbed survey propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
2.6.4
II
Combinatorial problems
73
3 Constraint satisfaction
3.1
75
Phase transitions in random CSPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
3.1.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
Revisiting survey propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
3.2.1
Flavours of SP-guided decimation
. . . . . . . . . . . . . . . . . . . . . . . . . .
82
3.2.2
Computational Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
83
3.2.3
Perturbed survey propagation for CSP . . . . . . . . . . . . . . . . . . . . . . . .
83
3.3
Satisfiability and coloring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
84
3.4
Clique-cover problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
3.5
Dominating set and set cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
88
3.5.1
89
3.2
Pitfalls of decimation
Factor-graph and complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vii
3.6
3.7
Clique problem, independent set and sphere packing . . . . . . . . . . . . . . . . . . . .
90
3.6.1
Binary variable factor-graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
91
3.6.2
Categorical variable factor-graph . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
3.6.3
Efficient Sphere packing with Hamming distance . . . . . . . . . . . . . . . . . .
95
Optimization variations of CSPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
98
4 Clustering
4.1
K-medians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.1.1
4.2
100
Facility location problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Hierarchical clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.2.1
Spanning and Steiner trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3
K-clustering problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.4
K-center problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.5
Modularity maximization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.5.1
The Potts model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.5.2
Clique model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Factor-graph representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Simplified message update and augmentation . . . . . . . . . . . . . . . . . . . . 109
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5 Permutations
5.1
5.2
112
Matching and permanent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.1.1
Factor-graph and complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.1.2
Arbitrary graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Traveling salesman problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.2.1
Augmentative approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
Simplified messages
Augmentation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
viii
5.2.2
5.3
Using pairwise factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Graph matching problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.3.1
Sub-graph isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.3.2
Subgraph monomorphism and supermorphism
5.3.3
Graph homomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
. . . . . . . . . . . . . . . . . . . 126
Reduction to other CSPs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.4
Finding symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.5
Graph alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Conclusion
136
Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
References
139
ix
List of tables
2.1
The correspondence between BP and SP . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
3.1
Comparison of different methods for K-satisfiability and K-coloring. . . . . . . . . . . .
87
4.1
Comparison of different modularity optimization methods. . . . . . . . . . . . . . . . . . 110
5.1
Summary of message-passing solutions to combinatorial problems. . . . . . . . . . . . . 137
x
List of figures
1.1
A factor-graph with variables as circles and factors as squares. . . . . . . . . . . . . . .
11
2.1
Belief Propagation on a loop-free factor-graph . . . . . . . . . . . . . . . . . . . . . . . .
35
2.2
Illustration of distributions used for loop correction. . . . . . . . . . . . . . . . . . . . .
51
2.3
An example for generalized loop correction method. . . . . . . . . . . . . . . . . . . . .
53
2.4
Run-time and accuracy at different levels of difficult for various loop correction methods.
57
2.5
Run-time versus accuracy over different problem sizes for various loop correction methods. 58
2.6
Survey propagation factor-graph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
2.7
Comparison of marginalization accuracy for BP, Gibbs sampling and Perturbed BP. . .
71
3.1
Factor-graph and set of solutions for a simple 3-SAT problem. . . . . . . . . . . . . . . .
75
3.2
Comparison of perturbed BP and BP-dec on real-world benchmark CSPs. . . . . . . . .
79
3.3
A schematic view of phase transitions in rCSPs . . . . . . . . . . . . . . . . . . . . . . .
80
3.4
Condensation phase and perturbed BP’s behaviour.
. . . . . . . . . . . . . . . . . . . .
81
3.5
comparison of different methods on K-satisfiability and K-coloring . . . . . . . . . . . .
85
3.6
An example of induced set cover and its factor-graph. . . . . . . . . . . . . . . . . . . .
88
3.7
Example of message passing for K-packing. . . . . . . . . . . . . . . . . . . . . . . . . .
93
3.8
The factor-graph for sphere-packing with Hamming distances, where the distance factors
are white squares and z-factors are in black. . . . . . . . . . . . . . . . . . . . . . . . . .
96
4.1
experimental results for K-clustering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2
The factor-graph for K-center clustering. . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
xi
4.3
Experimental results for K-center clustering and min-max facility location.
4.4
Example of maximum Modularity clustering by message passing. . . . . . . . . . . . . . 110
5.1
The factor-graph for max-weighted matching . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2
An example of application of augmentative message passing to solve TSP. . . . . . . . . 118
5.3
Experimental results for TSP using message passing on various benchmark instances. . . 119
5.4
Experimental results for bottleneck TSP. . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.5
Quality of BP marginals and integral in graph endomorphism. . . . . . . . . . . . . . . . 128
5.6
Counter-example for relaxed conditions of conjecture 5.3.4. . . . . . . . . . . . . . . . . 130
5.7
An example of finding approximate symmetry using endomorphism marginals. . . . . . 131
5.8
An example of sparse graph matching by message passing. . . . . . . . . . . . . . . . . . 135
xii
. . . . . . . 106
Introduction
Many complicated systems can be modeled as a graphical structure with interacting local functions. Many fields have (almost independently) discovered this: graphical models have been used
in bioinformatics (protein folding, medical imaging and spectroscopy, pedagogy trees, regulatory networks [27, 178, 224, 255, 323]), neuroscience (formation of associative memory and neuroplasticity [10, 65]), communication theory (low density parity check codes [106, 290]), statistical physics
(physics of dense matter and spin-glass theory [211]), image processing (inpainting, stereo/texture reconstruction, denoising and super-resolution [98, 101]), compressed sensing [84], robotics [294] (particle
filters), sensor networks [76, 143], social networks [203, 307], natural language processing [200], speech
recognition [73, 131], artificial intelligence (artificial neural networks, Bayesian networks [244, 316]) and
combinatorial optimization. This thesis is concerned with the application of graphical models in
solving combinatorial optimization problems [122, 222, 287], which broadly put, seeks an “optimal”
assignment to a discrete set of variables, where a brute force approach is infeasible.
To see how the decomposition offered by a graphical model can model a complex system, consider
a joint distribution over 200 binary variables. A naive way to represent this would require a table with
2200 entries. However if variables are conditionally independent such that their dependence structure
forms a tree, we can exactly represent the joint distribution using only 200 × 22 values. Operations such
as marginalization, which require computation time linear in the original size, are now reduced to local
computation in the form of message passing on this structure (i.e., tree), which in this case, reduces
the cost to linear in the new exponentially smaller size. It turns out even if the dependency structure
has loops, we can use message passing to perform “approximate” inference.
Moreover, we approach the problem of inference from an algebraic point of view [7]. This is in
contrast to the variational perspective on local computation [303]. These two perspectives are to some
extent “residuals” from the different origins of research in AI and statistical physics.
In the statistical study of physical systems, the Boltzmann distribution relates the probability of each
state of a physical system to its energy, which is often decomposed due to local interactions [208, 211].
These studies have been often interested in modeling systems at the thermodynamic limit of infinite
variables and the average behaviour through the study of random ensembles. Inference techniques with
this origin (e.g., mean-field and cavity methods) are often asymptotically exact under these assumptions.
2
Most importantly these studies have reduced inference to optimization through the notion of free energy
–a.k.a. variational approach.
In contrast, graphical models in the AI community have emerged in the study of knowledge representation and reasoning under uncertainty [245]. These advances are characterized by their attention
to the theory of computation and logic [17], where interest in computational (as opposed to analytical)
solutions has motivated the study of approximability, computational complexity [74, 267] and invention
of inference techniques such as belief propagation that are efficient and exact on tree structures. Also,
these studies have lead to algebraic abstractions in modeling systems that allow local computation
[185, 279].
The common foundation underlying these two approaches is information theory, where derivation of
probabilistic principles from logical axioms [146] leads to notions such as entropy and divergences that
are closely linked to their physical counter-parts i.e., entropy and free energies in physical systems. At
a less abstract level, it was shown that inference techniques in AI and communication are attempting
to minimize (approximations to) free energy [6, 326].
Another exchange of ideas between the two fields was in the study of critical phenomenon in random
constraint satisfaction problems by both computer scientists and physicists [103, 215, 216]; satisfiability
is at the heart of theory of computation and an important topic to investigate reasoning in AI. On
the other hand, the study of critical phenomena and phase transitions is central in statistical physics
of disordered systems. This was culminated when a variational analysis lead to discovery of survey
propagation [213] for constraint satisfaction, which significantly advanced the state-of-the-art in solving
random satisfiability problems.
Despite this convergence, variational and algebraic perspectives are to some extent complementary –
e.g., the variational approach does not extend beyond (log) probabilities, while the algebraic approach
cannot justify application of message passing to graphs with loops. Although we briefly review the
variational perspective, this thesis is mostly concerned with the algebraic perspective. In particular,
rather than the study of phase transitions and the behaviour of the set of solutions for combinatorial
problems, we are concerned with finding solutions to individual instances.
Part I starts by expressing the general form of inference, proposes a novel inference hierarchy and
studies its complexity in chapter 1.
Here, we also show how some of these problems are reducible to
others and introduce the algebraic structures that make efficient inference possible. The general form
of notation and the reductions that are proposed in this chapter are used in later chapters.
Chapter 2 studies some forms of approximate inference, by first introducing belief propagation. It
then considers the problems with intractably large number of factors and factors with large cardinality,
then proposes/reviews solutions to both problems. We then study different modes of inference as
optimization and review alternatives such as convergent procedures and convex and linear programming
relaxations for some inference classes in the inference hierarchy. Standard message passing using belief
propagation is only guaranteed to be exact if the graphical structure has no loops. This optimization
3
perspective (a.k.a. variational perspective) has also led to design of approximate inference techniques
that account for short loops in the graph. A different family of loop correction techniques can account
for long loops by taking message dependencies into account. This chapter reviews these methods and
introduces a novel loop correction scheme that can account for both short and long loops, resulting in
more accurate inference over difficult instances.
Message passing over loopy graphs can be seen as a fixed point iteration procedure, and the existence
of loops means there may be more than one fixed point. Therefore an alternative to loop correction is to
in some way incorporate all fixed points. This can be performed also by a message passing procedure,
known as survey propagation. The next section of this chapter introduces survey propagation from
a novel algebraic perspective that enables performing inference on the set of fixed points. Another
major approach to inference is offered by Markov Chain Monte Carlo (MCMC) techniques. After a
minimal review of MCMC, the final section of this chapter introduces a hybrid inference procedure,
called perturbed belief propagation that interpolates between belief propagation and Gibbs sampling.
We show that this technique can outperform both belief propagation and Gibbs sampling in particular
settings.
Part II of this thesis uses the inference techniques derived in the first part to solve a wide range
of combinatorial problems. We review the existing message passing solutions and provide novel formulations for three broad classes of problems: 1) constraint satisfaction problems (CSPs), 2) clustering
problems and 3) combinatorial problems over permutations.
In particular, in chapter 3 we use perturbed belief propagation and perturbed survey propagation to
obtain state-of-the-art performance in random satisfiability and coloring problems. We also introduce
novel message passing solutions and review the existing methods for sphere packing, set-cover, cliquecover, dominating-set and independent-set and several of their optimization counterparts. By applying
perturbed belief propagation to graphical representation of packing problem, we are able to compute
long “optimal” nonlinear binary codes with large number of digits.
Chapter 4 proposes message passing solutions to several clustering problems such as K-clustering,
K-center and Modularity optimization and shows that message passing is able to find near-optimal
solutions on moderate instances of these problems. Here, we also review the previous approaches to
K-median and hierarchical clustering and also the related graphical models for minimum spanning tree
and prize-collecting Steiner tree.
Chapter 5 deals with combinatorial problems over permutations, by first reviewing the existing
graphical models for matching, approximation of permanent, and graph alignment and introducing
two novel message passing solutions for min-sum and min-max versions of traveling salesman problem
(a.k.a. bottleneck TSP). We then study graph matching problems, including (sub-)graph isomorphism,
monomorphism, homomorphism, graph alignment and “approximate” symmetries. In particular, in
the study of graph homomorphism we show that its graphical model generalizes that of of several
other problems, including Hamiltonian cycle, clique problem and coloring. We further show how graph
4
homomorphism can be used as a surrogate for isomorphism to find symmetries.
Contributions and acknowledgment
The results in this thesis are a joint work with my supervisor Dr. Greiner and other researchers. In
detail, the algebraic approach to inference is presented in [256]; the loop correction ideas are published
in [254]; perturbation schemes for CSP are presented in [253]; performing min-max inference was first
suggested by Dr. Brendan Frey and Christopher Srinivasa, and many of the related ideas including
min-max reductions are presented in our joint paper [258]. Finally, the augmentation scheme for TSP
and Modularity maximization is discussed in [257].
The contribution of this thesis, including all the published work is as follows:
– Generalization of inference problems in graphical models including:
• The inference hierarchy.
• The limit of distributive law on tree structures.
• All the theorems, propositions and claims on complexity of inference, including
◦ NP-hardness of inference in general commutative semirings.
– A unified treatment of different modes of inference over factor-graphs and identification of their
key properties (e.g., significance of inverse operator) in several settings including:
• Loop correction schemes.
• Survey propagation equations.
– Reduction of min-max inference to min-sum and sum-product inference.
– Simplified form of loop correction in Markov networks and their generalization to incorporate short
loops over regions.
– A novel algebraic perspective on survey propagation.
– Perturbed BP and perturbed SP and their application to constraint satisfaction problems.
– Factor-graph augmentation for inference over intractably large number of constraints.
– Factor-graph formulation for several combinatorial problems including
• Clique-cover.
• Independent-set, set-cover and vertex cover.
• Dominating-set and packing (the binary-variable model)
• Packing with hamming distances
• K-center problem, K-clustering and clique model for modularity optimization.
• TSP and bottleneck TSP.
– The general framework for study of graph matching, including
• Subgraph isomorphism.1
1
Although some previous work [44] claim to address the same problem, we note that their formulation is for sub-graph
monomorphism rather than isomorphism.
5
• Study of message passing for Homomorphism and finding approximate symmetries.
• Graph alignment with a diverse set of penalties.
Part I
Inference by message passing
7
This part of the thesis first studies the representation formalism, hierarchy of inference problems,
reducibilities and the underlying algebraic structure that allows efficient inference in the form of message passing in graphical models, in chapter 1. By viewing inference under different lights, we then
review/introduce procedures that allow better approximations in chapter 2.
Chapter 1
Representation, complexity and
reducibility
In this chapter, we use a simple algebraic structure – i.e., commutative semigroup – to express a general
form for inference in graphical models. To this end, we first introduce the factor-graph representation
and formalize inference in section 1.1. Section 1.2 focuses on four operations defined by summation,
multiplication, minimization and maximization, to construct a hierarchy of inference problems within
PSPACE, such that the problems in the same class of the hierarchy belong to the same complexity
class. Here, we encounter some new inference problems and establish the completeness of problems at
lower levels of hierarchy w.r.t. their complexity classes. In section 1.3 we augment our simple structures
with two properties to obtain message passing on commutative semirings. Here, we also observe that
replacing a semigroup with an Abelian group, gives us normalized marginalization as a form of inference
inquiry. Here, we show that inference in any commutative semiring is NP-hard and postpone further
investigation of message passing to the next chapter. Section 1.4 shows how some of the inference
problems introduced so far are reducible to others.
1.1
The problem of inference
We use commutative semigroups to both define what a graphical model represents and also to define
inference over this graphical model. The idea of using structures such as semigroups, monoids and
semirings in expressing inference has a long history[33, 185, 275]. Our approach, based on factorgraphs [181] and commutative semigroups, generalizes a variety of previous frameworks, including
Markov networks [68], Bayesian networks [243], Forney graphs [100], hybrid models [83], influence
diagrams [137] and valuation networks [278].
In particular, the combination of factor-graphs and semigroups that we consider here generalizes
the plausibility, feasibility and utility framework of Pralet et al. [250], which is explicitly reduced to the
1.1. THE PROBLEM OF INFERENCE
9
graphical models mentioned above and many more. The main difference in our approach is in keeping
the framework free of semantics (e.g., decision and chance variables, utilities, constraints), that are
often associated with variables, factors and operations, without changing the expressive power. These
notions can later be associated with individual inference problems to help with interpretation.
Definition 1.1.1. A commutative semigroup is a pair G = (Y ∗ , ⊗), where Y ∗ is a set and ⊗ :
Y ∗ × Y ∗ → Y ∗ is a binary operation that is (I) associative: a ⊗ (b ⊗ c) = (a ⊗ b) ⊗ c and (II) commutative:
a ⊗ b = b ⊗ a for all a, b, c ∈ Y ∗ . A commutative monoid is a commutative semigroup plus an identity
⊗
⊗
element 1 such that a ⊗ 1 = a. If every element a ∈ Y ∗ has an inverse a−1 (often written a1 ), such that
⊗
⊗
a ⊗ a−1 = 1, and a ⊗ 1 = a, the commutative monoid is an Abelian group.
Here, the associativity and commutativity properties of a commutative semigroup make the operations invariant to the order of elements. In general, these properties are not “vital” and one may define
inference starting from a magma.1
Example 1.1.1. Some examples of semigroups are:
– The set of strings with the concatenation operation forms a semigroup with the empty string as
the identity element. However this semigroup is not commutative.
– The set of natural numbers N with summation defines a commutative semigroup.
– Integers modulo n with addition defines an Abelian group.
– The power-set 2S of any set S, with intersection operation defines a commutative semigroup with
S as its identity element.
– The set of natural numbers with greatest common divisor defines a commutative monoid with 0 as
its identity. In fact any semilattice is a commutative semigroup [79].
– Given two commutative semigroups on two sets Y ∗ and Z ∗ , their Cartesian product is also a
commutative semigroup.
Let x = (x1 , . . . , xN ) be a tuple of N discrete variables xi ∈ Xi , where Xi is the domain of xi
and x ∈ X = X1 × . . . × XN . Let I ⊆ N = {1, 2, . . . , N } denote a subset of variable indices and
xI = {xi | i ∈ I} ∈ XI be the tuple of variables in x indexed by the subset I. A factor fI : XI → YI is a
function over a subset of variables and YI = {fI (xI ) | xI ∈ XI } is the range of this factor.
Definition 1.1.2. A factor-graph is a pair (F, G ) such that
• F = {fI } is a collection of factors with collective range Y =
S
I YI .
• |F| = Poly(N ).
1
A magma [247] generalizes a semigroup, as it does not require associativity property nor an identity element. Inference
in graphical models can be also extended to use magma (in definition 1.1.2). For this, the elements of Y ∗ and/or X should
be ordered and/or parenthesized so as to avoid ambiguity in the order of pairwise operations over the set. Here, to avoid
unnecessary complications, we confine our treatment to commutative semigroups.
10
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
• fI has a polynomial representation in N and it is possible to evaluate fI (xI ) ∀I, xI in polynomial
time.
• G = (Y ∗ , ⊗) is a commutative semigroup, where Y ∗ is the closure of Y w.r.t. ⊗.
The factor-graph compactly represents the expanded (joint) form
q(x) =
O
fI (xI )
(1.1)
I
Note that the connection between the set of factors F and the commutative semigroup is through
the “range” of factors. The conditions of this definition are necessary and sufficient to 1) compactly
represent a factor-graph and 2) evaluate the expanded form, q(x), in polynomial time. A stronger
condition to ensure that a factor has a compact representation is |XI | = Poly(N ), which means fI (xI )
can be explicitly expressed for each xI ∈ XI as an |I|-dimensional array.
F can be conveniently represented as a bipartite graph that includes two sets of nodes: variable
nodes xi , and factor nodes I. A variable node i (note that we will often identify a variable xi with its
index “i”) is connected to a factor node I if and only if i ∈ I –i.e., I is a set that is also an index. We will
use ∂ to denote the neighbours of a variable or factor node in the factor graph – that is ∂I = {i | i ∈ I}
(which is the set I) and ∂i = {I | i ∈ I}. Also, we use ∆i to denote the Markov blanket of node xi –
i.e., ∆i = {j ∈ ∂I | I ∈ ∂i, j 6= i}.
Example 1.1.2. Figure 1.1 shows a factor-graph with 12 variables and 12 factors. Here x = (xi , xj , xk , xe , xm , xo , xr
I = ∂I = {i, j, k}, xK = x{k,w,v} and ∂j = {I, V, W}. Assuming Ge = (R, min), the expanded form represents
q(x) = min{fI (xI ), fJ (xJ ), . . . , fZ (xZ )}.
Now, assume that all variables are binary – i.e., X = {0, 1}12 and q(x) is 12-dimensional hypercube, with one assignment at each corner. Also assume all the factors count the number of nonzero variables – e.g., for z W = (1, 0, 1) ∈ XW we have fW (z W ) = 2. Then, for the complete assignment z = (0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1) ∈ X , it is easy to check that the expanded form is q(z) =
min{2, 0, 1, . . . , 1} = 0.
A marginalization operation shrinks the expanded form q(x) using another commutative semigroup
with binary operation ⊕. Inference is a combination of an expansion and one or more marginalization
operations, which can be computationally intractable due to the exponential size of the expanded form.
Definition 1.1.3. Given a function q : XJ → Y, and a commutative semigroup G = (Y ∗ , ⊕), where Y ∗
is the closure of Y w.r.t. ⊕, the marginal of q for I ⊂ J is
q(xJ\I )
def
=
M
xI
q(xJ )
(1.2)
1.1. THE PROBLEM OF INFERENCE
11
Figure 1.1: A factor-graph with variables as circles and factors as squares.
where
L
xI
q(xJ ) is short for
L
xI ∈XI
q(xJ\I , xI ), and it means to compute q(xJ\I ) for each xJ\I , one
should perform the operation ⊕ over the set of all the assignments to the tuple xI ∈ XI .
We can think of q(xJ ) as a |J|-dimensional tensor and marginalization as performing ⊕ operation
over the axes in the set I. The result is another |J \ I|-dimensional tensor (or function) that we call the
marginal. Here if the marginalization is over all the dimensions in J, we denote the marginal by q(∅)
instead of q(x∅ ) and call it the integral of q.
Now we define an inference problem as a sequence of marginalizations over the expanded form of a
factor-graph.
Definition 1.1.4. An inference problem seeks
q(xJ0 ) =
M
−1
M
M
M
xJ
M
xJ
...
M −1
1
M
O
xJ
1
fI (xI )
I
where
1
M
• Y ∗ is the closure of Y (the collective range of factors), w.r.t. ⊕, . . . , ⊕ and ⊗.
m
• Gm = (Y ∗ , ⊕) ∀1 ≤ m ≤ M and Ge = (Y ∗ , ⊗) are all commutative semigroups.
• J0 , . . . , JL partition the set of variable indices N = {1, . . . , N }.
• q(xJ0 ) has a polynomial representation in N – i.e., |XJ0 | = Poly(N )
(1.3)
12
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
1
M
Note that ⊕, . . . , ⊕ refer to potentially different operations as each belongs to a different semigroup.
When J0 = ∅, we call the inference problem integration (denoting the inquiry by q(∅)) and otherwise
we call it marginalization. Here, having a constant sized J0 is not always enough to ensure that q(xJ0 )
has a polynomial representation in N . This is because the size of q(xJ0 ) for any individual xJ0 ∈ XJ0
may grow exponentially with N (e.g., see Claim 1.2.1). In the following we call Ge = (Y ∗ , ⊗) the
m
expansion semigroup and Gm = (Y ∗ , ⊕) ∀1 ≤ m ≤ M the marginalization semigroup.
Example 1.1.3. Going back to example 1.1.3, the shaded region in figure 1.1 shows a partitioning of
the variables that we use to define the following inference problem:
q(xJ0 ) = max
xJ
3
X
min min fI (xI )
xJ
xJ
1
I
2
We can associate this problem with the following semantics: we may think of each factor as an agent,
where fI (xI ) is the payoff for agent I, which only depends on a subset of variables xI . We have adversarial
variables (xJ1 ), environmental or chance variables (xJ2 ), controlled variables (xJ3 ) and query variables
(xJ0 ). The inference problem above for each query xJ0 seeks to maximize the expected minimum payoff
of all agents, without observing the adversarial or chance variables, and assuming the adversary makes
its decision after observing control and chance variables.
Example 1.1.4. A “probabilistic” graphical model is defined using a expansion semigroup Ge =
(R≥0 , ×) and often a marginalization semigroup Gm = (R≥0 , +). The expanded form represents the unQ
normalized joint probability q(x) = I fI (xI ), whose marginal probabilities are simply called marginals.
Replacing the summation with marginalization semigroup Gm = (R≥0 , max), seeks the maximum probQ
ability state and the resulting integration problem q(∅) = maxx I fI (xI ) is known as maximum a
posteriori (MAP) inference.
Alternatively by adding a second marginalization operation to the
summation, we get the marginal MAP inference
q(xJ0 )
=
max
xJ
2
XY
xJ
1
where here
N
=
fI (xI ).
(1.4)
I
1
2
Q L
P
L
,
=
and
= max.
If the object of interest is the negative log-probability (a.k.a. energy), the product expansion semigroup is replaced by Ge = (R, +). Instead of sum marginalization semigroup, we can use the log-sumdef
exp semigroup, Gm = (R, +) where a ⊕ b = log(e−a + e−b ). The integral in this case is the log-partition
function. If we change the marginalization semigroup to Gm = (R, min), the integral is the minimum
energy (corresponding to MAP).
A well-known example of a probabilistic graphical model is the Ising model of ferromagnetism.
This model is an extensively studied in physics, mainly to model the phase transition in magnets. The
1.2. THE INFERENCE HIERARCHY
13
model consists of binary variables (xi ∈ {−1, 1}) – denoting magnet spins – arranged on the nodes of a
graph G = (V, E) (usually a grid or Cayley tree).
The energy function (i.e., Hamiltonian) associated
with a configuration x is the joint form
e(x)
=
q(x)
=
−
X
xi Jij xj −
X
hi
(1.5)
i∈V
(i,j)∈E
Variable interactions are denoted by J and h is called the local field. Here each Ji,j defines a factor
over xi , xj : f{i,j} (x{i,j} ) = −xi Jij xj and local fields define local factors f{i} (xi ) = −hi xi .
Depending on the type of interactions, we call the resulting Ising model:
• ferromagnetic, if all Jij > 0. In this setting, neighbouring variables are likely to take similar
values.
• anti-ferromagnetic, if all Jij < 0.
• non-ferromagnetic, if both kind of interactions are allowed. In particular, if the ferromagnetic and
anti-ferromagnetic interactions have comparable frequency, the model is called spin-glass. This class
of problem shows most interesting behaviours, which is not completely understood [212]. As we will see,
the studied phenomena in these materials have important connections to difficult inference problems
including combinatorial optimization problems. Two well studied models of spin glass are EdwardAnderson (EA [93]) and Sherrington-Kirkpatrick (SK [280]) models. While the EA model is defined on
a grid (i.e., spin-glass interactions over a grid), the SK model is a complete graph.
1.2
The inference hierarchy
Often, the complexity class is concerned with the decision version of the inference problem in definition 1.1.4. The decision version of an inference problem asks a yes/no question about the integral:
?
q(∅) ≥ q for a given q.
Here, we produce a hierarchy of inference problems in analogy to polynomial [288], the counting [302]
and arithmetic [264] hierarchies.
To define the hierarchy, we assume the following in definition 1.1.4:
l
l+1
– Any two consecutive marginalization operations are distinct (⊕ =
6 ⊕ ∀1 ≤ l < M ).
– The marginalization index sets Jl ∀1 ≤ l ≤ M are non-empty. Moreover if |Jl | = O(log(N )) we
call this marginalization operation a polynomial marginalization as here |XJl | = Poly(N ).
– In defining the factor-graph, we required each factor to be polynomially computable. In building
the hierarchy, we require the operations over each semigroup to be polynomially computable as well.
To this end we consider the set of rational numbers Y ∗ ⊆ Q≥0 ∪{±∞}. Note that this automatically
eliminates semigroups that involve operations such as exponentiation and logarithm (because Q
is not closed under these operations) and only consider summation, product, minimization and
maximization.
14
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
We can always re-express any inference problem to enforce the first two conditions and therefore
they do not impose any restriction. In the following we will use the a language to identify inference
problems for an arbitrary set of factors F = {fI }. For example, sum-product refers to the inference
?
P Q
problem x I fI (xI ) ≥ q. In this sense the rightmost “token” in the language (here product) identifies
Q
the expansion semigroup Ge = (Q, ) and the rest of tokens identify the marginalization semigroups
over Q in the given order. Therefore, this minimal language exactly identifies the inference problem. The
only information that affects the computational complexity of an inference problem but is not specified
in this language is whether each of the marginalization operations are polynomial or exponential.
We define five inference families: Σ, Π, Φ, Ψ, ∆. The families are associated with that “outermost”
M
marginalization operation – i.e., ⊕ in definition 1.1.4). Σ is the family of inference problems where
M
⊕ = sum. Similarly, Π is associated with product, Φ with minimization and Ψ with maximization. ∆
is the family of inference problems where the last marginalization is polynomial (i.e., |JM | = O(log(N ))
M
regardless of ⊕).
Now we define inference classes in each family, such that all the problems in the same class have
the same computational complexity. Here, the hierarchy is exhaustive – i.e., it includes all inference
problems with four operations sum, min, max and product whenever the integral q(∅) has a polynomial
representation (see Claim 1.2.1). Moreover the inference classes are disjoint. For this, each family is
parameterized by a subscript M and two sets S and D (e.g., ΦM (S, D) is an inference “class” in family
Φ). As before, M is the number of marginalization operations, S is the set of indices of the (exponential)
sum-marginalization and D is the set of indices of polynomial marginalizations.
Example 1.2.1. Sum-min-sum-product identifies the decision problem
X
xJ
3
min
xJ
2
XY
xJ
1
?
fI (xI ) ≥ q
I
where J1 , J2 and J3 partition N . Assume J1 = {2, . . . , N2 }, J2 = { N2 + 1, . . . , N } and J3 = {1}.
Since we have three marginalization operations M = 3. Here the first and second marginalizations are
exponential and the third one is polynomial (since |J3 | is constant). Therefore D = {3}. Since the only
1
L
P
exponential summation is x = x , S = {1}. In our inference hierarchy, this problem belongs to
the class ∆3 ({1}, {3}).
J1
J1
Alternatively, if we use different values for J1 , J2 and J3 that all linearly grow with N , the corresponding inference problem becomes a member of Σ3 ({1, 3}, ∅).
Remark 1. Note that arbitrary assignments to M , S and D do not necessarily define a valid inference
class. For example we require that S ∩ D = ∅ and no index in D and S are is larger than M . Moreover,
the values in S and D should be compatible with the inference class. For example, for inference class
ΣM (S, D), M is a member of S. For notational convenience, if an inference class notation is invalid we
1.2. THE INFERENCE HIERARCHY
15
equate it with an empty set – e.g., Ψ1 ({1}, ∅) = ∅, because S = {1} and M = 1 means the inference
class is Σ rather than Ψ.
In the definition below, we ignore the inference problems in which product appears in any of the
marginalization semigroups (e.g., product-sum). The following claim, explains this choice.
Claim 1.2.1. For ⊕M = prod, the inference query q(xJ0 ) can have an exponential representation in
N.
Proof. The claim states that when the product appears in the marginalization operations, the marginal
(and integral) can become very large, such that we can no longer represent them in polynomial space in
N . We show this for an integration problem. The same idea can show the exponential representation
of a marginal query.
To see why this integral has an exponential representation in N , consider its simplified form
q(∅) =
Y
q(xI )
xI
M
where q(x) here is the result of inference up to the last marginalization step ⊕, which is product, where
XI grows exponentially with N . Recall that the hierarchy is defined for operations on Q≥0 . Since q(xI )
for each xI ∈ XI has a constant size, say c, the size of representation of q(∅) using a binary scheme is
dlog2 (q(∅))e =
log2
Y
xI
X
q(xI )
c
=
= dc|XI |e
xI
which is exponential in N .
Define the base members of families as
def
def
def
def
Σ0 (∅, ∅) = {sum} Φ0 (∅, ∅) = {min}
(1.6)
Ψ0 (∅, ∅) = {max} Π0 (∅, ∅) = {prod}
def
∆0 (∅, ∅) = ∅ ∆1 (∅, {1}) = {sum − sum, min − min, max − max}
where the initial members of each family only identify the expansion semigroup – e.g., sum in Σ0 (∅, ∅)
P
identifies q(x) = I fI (xI ). Here, the exception is ∆1 (∅, {1}), which contains three inference problems.2
Let ΞM (S, D) denote the union of corresponding classes within all families:
ΞM (S, D) = ΣM (S, D) ∪ ΠM (S, D) ∪ ΦM (S, D) ∪ ΨM (S, D) ∪ ∆M (S, D)
2
We treat M = 1 for ∆ specially as in this case the marginalization operation can not be polynomial. This is because
if |J1 | = O(log(N )), then |J0 | = Ω(N ) which violates the conditions in the definition of the inference problem.
16
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
Now define the inference family members recursively, by adding a marginalization operation to all
the problems in each inference class. If this marginalization is polynomial then the new class belongs
to the ∆ family and the set D is updated accordingly. Alternatively, if this outermost marginalization
is exponential, depending on the new marginal operation (i.e., min, max, sum) the new class is defined
to be a member of Φ, Ψ or Σ. For the case that the last marginalization is summation set S is updated.
• Adding an exponential marginalization ∀ |XJM | = Poly(N ), M > 0
def
ΣM +1 (S ∪ {M + 1}, D) = sum − ξ | ξ ∈ ΞM (S, D) \ ΣM (S, D)}
def
ΦM +1 (S, D) = min −ξ | ξ ∈ ΞM (S, D) \ ΦM (S, D)
def
ΨM +1 (S, D) = max −ξ | ξ ∈ ΞM (S, D) \ ΨM (S, D)
(1.7)
def
ΠM +1 (S, D) = ∅
• Adding a polynomial marginalization
def
∆M +1 (S, D ∪ {M + 1}) =
1.2.1
∀ |XJM | = Poly(N ), M > 1
⊕ −ξ | ξ ∈ ΞM (S, D) , ⊕ ∈ {min, max, sum}
(1.8)
Single marginalization
The inference classes in the hierarchy with one marginalization are
∆1 (∅, {1}) = {min − min, max − max, sum − sum}
(1.9)
Ψ1 (∅, ∅) = {max − min, max −sum, max −prod}
(1.10)
Φ1 (∅, ∅) = {min − max, min −sum, min −prod}
(1.11)
Σ1 ({1}, ∅) = {sum − prod, sum − min, sum − max}
(1.12)
Now we review all the problems above and prove that ∆1 , Ψ1 , Φ1 and Σ1 are complete w.r.t. P, NP,
coNP and PP respectively. Starting from ∆1 :
Proposition 1.2.2. sum-sum, min-min and max-max inference are in P.
Proof. To show that these inference problems are in P, we provide polynomial-time algorithms for them:
• sum − sum is short for
q(∅) =
XX
x
fI (xI )
I
which asks for the sum over all assignments of x ∈ X , of the sum of all the factors. It is easy to see
that each factor value fI (xI ) ∀I, XI is counted |X\I | times in the summation above. Therefore we can
1.2. THE INFERENCE HIERARCHY
17
rewrite the integral above as
q(∅) =
X
X
|X\I |
fI (xI )
xI
I
where the new form involves polynomial number of terms and therefore is easy to calculate.
• min − min (similar for max − max) is short for
q(∅) = min min fI (xI )
x
I
where the query seeks the minimum achievable value of any factor. We can easily obtain this by seeking
the range of all factors and reporting the minimum value in polynomial time.
Max-sum and max-prod are widely studied and it is known that their decision version are NPcomplete [281]. By reduction from satisfiability we can show that max-min inference [258] is also
NP-hard.
?
Proposition 1.2.3. The decision version of max-min inference that asks maxx minI fI (xI ) ≥ q is NPcomplete.
Proof. Given x it is easy to verify the decision problem, so max-min decision belongs to NP. To show
NP-completeness, we reduce the 3-SAT to a max-min inference problem, such that 3-SAT is satisfiable
iff the max-min value is q(∅) ≥ 1 and unsatisfiable otherwise.
Simply define one factor per clause of 3-SAT, such that fI (xI ) = 1 if xI satisfies the clause and
any number less than one otherwise. With this construction, the max-min value maxx minI∈F fI (xI ) is
one iff the original SAT problem was satisfiable, otherwise it is less than one. This reduces 3-SAT to
Max-Min-decision.
This means all the problems in Ψ1 (∅, ∅) are in NP (and in fact are complete w.r.t. this complexity
class). In contrast, problems in Φ1 (∅, ∅) are in coNP, which is the class of decision problems in which
the “NO instances” result has a polynomial time verifiable witness or proof. Note that by changing the
?
?
decision problem from q(∅) ≥ q to q(∅) ≤ q, the complexity classes of problems in Φ and Ψ family are
reversed (i.e., problems in Φ1 (∅, ∅) become NP-complete and the problems in Ψ1 (∅, ∅) become coNPcomplete).
Among the members of Σ1 ({1}, ∅), sum-product is known to be PP-complete [194, 267]. It is easy
to show the same result for sum-min (sum-max) inference.
Proposition 1.2.4. The sum-min decision problem
P
x minI fI (xI )
?
≥ q is PP-complete for Y = {0, 1}.
PP is the class of problems that are polynomially solvable using a non-deterministic Turing machine,
where the acceptance condition is that the majority of computation paths accept.
18
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
Proof. To see that
P
?
x minI fI (xI )
≥ q is in PP, enumerate all x ∈ X non-deterministically and for each
assignment calculate minI fI (xI ) in polynomial time (where each path accepts iff minI fI (xI ) = 1) and
accept iff at least q of the paths accept.
Given a matrix A ∈ {0, 1}N ×N the problem of calculating its permanent
perm(A) =
N
X Y
Ai,zi
z∈SN i=1
where SN is the set of permutations of 1, . . . , N is #P-complete and the corresponding decision problem
is PP-complete [297]. To show completeness w.r.t. PP it is enough to reduce the problem of computing
the matrix permanent to sum-min inference in a graphical model.
The problem of computing the
permanent has been reduced to sum-product inference in graphical models [139]. However, when fI (xI ) ∈
{0, 1} ∀I, sum-product is isomorphic to sum-min. This is because y1 × y2 = min(y1 , y2 )∀yi ∈ {0, 1}.
Therefore, the problem of computing the permanent for such matrices reduces to sum-min inference in
the factor-graph of [139].
1.2.2
Complexity of general inference classes
Let f(.) denote the complexity class of an inference class in the hierarchy. In obtaining the complexity
class of problems with M > 1, we use the following fact, which is also used in the polynomial hierarchy:
A
A
PNP = PcoNP [14]. In fact PNP = PcoNP , for any oracle A. This means that by adding a polynomial marginalization to the problems in ΦM (S, D) and ΨM (S, D), we get the same complexity class
f(∆M +1 (S, D ∪ {M + 1})). The following gives a recursive definition of complexity class for problems
in the inference hierarchy.3 Note that the definition of the complexity for each class is very similar to
the recursive definition of members of each class in equations (1.7) and (1.8)
Theorem 1.2.5. The complexity of inference classes in the hierarchy is given by the recursion
f(ΦM +1 (S, D)) = coNPf(ΞM (S,D)\ΦM (S,D))
(1.13)
f(ΨM +1 (S, D)) = NPf(ΞM (S,D)\ΨM (S,D))
(1.14)
f(ΣM +1 (S ∪ {M + 1}, D)) = PPf(ΞM (S,D)\ΣM (S,D))
(1.15)
f(∆M +1 (S, D ∪ {M + 1})) = Pf(ΞM (S,D))
(1.16)
where the base members are defined in equation (1.6) and belong to P.
Proof. Recall that our definition of factor graph ensures that q(x) can be evaluated in polynomial time
and therefore the base members are in P (for complexity of base members of ∆ see proposition 1.2.2).
3
We do not prove the completeness w.r.t. complexity classes beyond the first level of the hierarchy and only assert the
membership.
1.2. THE INFERENCE HIERARCHY
We use these classes as the base of our induction and assuming the complexity classes above are correct
for M we show that are correct for M + 1. We consider all the above statements one by one:
• Complexity for members of ΦM +1 (S, D):
Adding an exponential-sized min-marginalization to an inference problem with known complexity A,
requires a Turing machine to non-deterministically enumerate z JM ∈ XJM possibilities, then call the A
oracle with the “reduced factor-graph” – in which xJM is clamped to z JM – and reject iff any of the
calls to oracle rejects. This means f(ΦM +1 (S, D)) = coNPA .
Here, equation (1.13) is also making another assumption expressed in the following claim.
Claim 1.2.6. All inference classes in ΞM (S, D) \ ΦM (S, D) have the same complexity A.
• M = 0: the fact that q(x) can be evaluated in polynomial time means that A = P.
• M > 0: ΞM (S, D) \ ΦM (S, D) only contains one inference class – that is exactly only one of the
following cases is correct:
– M ∈ S ⇒ ΞM (S, D) \ ΦM (S, D) = ΣM (S, D)
– M ∈ D ⇒ ΞM (S, D) \ ΦM (S, D) = ∆M (S, D)
– M∈
/ S ∪ D ⇒ ΞM (S, D) \ ΦM (S, D) = ΨM (S, D).
(in constructing the hierarchy we assume two consecutive marginalizations are distinct and
the current marginalization is a minimization.)
But if ΞM (S, D) \ ΦM (S, D) contains a single class, the inductive hypothesis ensures that all
problems in ΞM (S, D) \ ΦM (S, D) have the same complexity class A.
This completes the proof of our claim.
• Complexity for members of ΨM +1 (S, D):
Adding an exponential-sized max -marginalization to an inference problem with known complexity A,
requires a Turing machine to non-deterministically enumerate z JM ∈ XJM possibilities, then call the
A oracle with the reduced factor-graph and accept iff any of the calls to oracle accepts. This means
f(ΨM +1 (S, D)) = NPA . Here, an argument similar to that of Claim 1.2.6 ensures that ΞM (S, D) \
ΨM (S, D) in equation (1.14) contains a single inference class.
• Complexity for members of ΣM +1 (S ∪ {M + 1}, D):
Adding an exponential-sized sum-marginalization to an inference problem with known complexity A,
requires a Turing machine to non-deterministically enumerate z JM ∈ XJM possibilities, then call the A
oracle with the reduced factor-graph and accept iff majority of the calls to oracle accepts. This means
f(ΨM +1 (S, D)) = PPA .
• M = 0: the fact that q(x) can be evaluated in polynomial time means that A = P.
• M > 0:
19
20
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
– M ∈ D ⇒ ΞM (S, D) \ ΣM (S, D) = ∆M (S, D).
– M ∈
/ D∪S
⇒ ΞM (S, D) \ ΣM (S, D) = ΨM (S, D) ∪ ΦM (S, D): despite the fact that
A = f(ΨM (S, D)) is different from A0 = f(ΦM (S, D)), since PP is closed under complement,
which means PPA = PPA and the recursive definition of complexity equation (1.15) remains
correct.
• Complexity for members of ∆M +1 (S, D ∪ {M + 1}):
Adding a polynomial-sized marginalization to an inference problem with known complexity A, requires
a Turing machine to deterministically enumerate z JM ∈ XJM possibilities in polynomial time, and each
time call the A oracle with the reduced factor-graph and accept after some polynomial-time calculation.
This means f(ΨM +1 (S, D)) = PA . Here, there are three possibilities:
• M = 0: here again A = P.
• M ∈ S ⇒ ΞM (S, D) = ΣM (S, D).
• M ∈ D ⇒ ΞM (S, D) = ∆M (S, D).
B
B
• M ∈
/ D ∪ S ⇒ ΞM (S, D) = ΨM (S, D) ∪ ΦM (S, D), in which case since PPNP = PPcoNP , the
recursive definition of complexity in equation (1.16) remains correct.
Example 1.2.2. Consider the marginal-MAP inference of equation (1.4). The decision version of this
?
problem, q(∅) ≥ q, is a member of Ψ2 ({1}, ∅) which also includes max −sum−min and max −sum−max.
The complexity of this class according to equation (1.14) is f(Ψ2 ({1}, ∅)) = NPPP . However, marginalMAP is also known to be “complete” w.r.t. NPPP [241]. Now suppose that the max-marginalization over
xJ2 is polynomial (e.g., |J2 | is constant). Then marginal-MAP belongs to ∆2 ({1}, {2}) with complexity
PPP . This is because a Turing machine can enumerate all z J2 ∈ XJ2 in polynomial time and call its PP
oracle to see if
?
where
q(xJ0 | z J2 ) ≥ q
XY
q(xJ0 | z J2 ) =
fI (xI\J2 , z I∩J2 )
xJ
2
I
and accept if any of its calls to oracle accepts, and rejects otherwise. Here, fI (xI\J2 , z I∩J2 ) is the reduced
factor, in which all the variables in xJ2 are fixed to z J2 ∩I .
The example above also hints at the rationale behind the recursive definition of complexity class for
each inference class in the hierarchy. Consider the inference family Φ:
Here, Toda’s theorem [296] has an interesting implication w.r.t. the hierarchy. This theorem states
that PP is as hard as the polynomial hierarchy, which means min − max − min − . . . − max inference for
1.2. THE INFERENCE HIERARCHY
21
an arbitrary, but constant, number of min and max operations appears below the sum-product inference
in the inference hierarchy.
1.2.3
Complexity of the hierarchy
By restricting the domain Y ∗ to {0, 1}, min and max become isomorphic to logical AND (∧) and OR (∨)
respectively, where 1 ∼
= true, 0 ∼
= false. By considering the restriction of the inference hierarchy to
these two operations we can express quantified satisfiability (QSAT) as inference in a graphical model,
where ∧ ∼
= ∀ and ∨ ∼
= ∃. Let each factor fI (xI ) be a disjunction –e.g., f(xi,j,k ) = xi ∨ ¬xj ∨ ¬xk . Then
we have
∀xJ ∃xJ
M
M −1
. . . ∃xJ ∀xJ
2
^
1
fI (xI ) ∼
= min max . . . max min min fI (xI )
xJ
I
M
xJ
xJ
M −1
2
xJ
1
I
By adding the summation operation, we can express the stochastic satisfiability [194] and by generalizing
the constraints from disjunctions we can represent any quantified constraint problem (QCP) [36]. QSAT,
stochastic SAT and QCPs are all PSPACE-complete, where PSPACE is the class of problems that can
be solved by a (non-deterministic) Turing machine in polynomial space. Therefore if we can show that
inference in the inference hierarchy is in PSPACE, it follows that inference hierarchy is in PSPACEcomplete as well.
Theorem 1.2.7. The inference hierarchy is PSPACE-complete.
Proof. (theorem 1.2.7) To prove that a problem is PSPACE-complete, we have to show that 1) it is
in PSPACE and 2) a PSPACE-complete problem reduces to it. We already saw that QSAT, which is
PSPACE-complete, reduces to the inference hierarchy. But it is not difficult to show that inference
hierarchy is contained in PSPACE. Let
q(xJ0 ) =
M
−1
M
M
M
xJ
M
xJ
...
M −1
1
M
O
xJ
1
fI (xI )
I
be any inference problem in the hierarchy. We can simply iterate over all values of z ∈ X in nested
loops or using a recursion. Let (i) : {1, . . . , N } → {1, . . . , M } be the index of the marginalization
that involves xi – that is i ∈ J(i) . Moreover let i1 , . . . , iN be an ordering of variable indices such that
(ik ) ≤ (ik+1 ). Algorithm 1 uses this notation to demonstrate this procedure using nested loops. Note
that here we loop over individual domains Xik rather than XJm and track only temporary tuples qik , so
that the space complexity remains polynomial in N .
22
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
input :
M
−1
L
M
L
xJ
xJ
M
...
1
L
xJ
M −1
N
I fI (xI )
1
output: q(xJ0 )
for each z J0 ∈ XJ0 do // loop over the query domain
for each ziN ∈ XiN do // loop over XiN
.
.
.
for each zi1 ∈ Xi1 do // loop over Xi1
N
q1 (zi1 ) := I fI (z I );
end
(i
L1 )
qi2 (zi2 ) :=
xi q1 (xi1 )
1
.
.
.
qN (ziN ) :=
(iL
N −1 )
xiN −1 qN −1 (xiN −1 )
end
q(z J0 ) :=
(i
N)
L
xiN qN (xiN )
end
Algorithm 1: inference in PSPACE
1.3
Polynomial-time inference
Our definition of inference was based on an expansion operation ⊗ and one or more marginalization
1
M
operations ⊕, . . . , ⊕. If we assume only a single marginalization operation, polynomial time inference is
still not generally possible. However, if we further assume that the expansion operation is distributive
over marginalization and the factor-graph has no loops, exact polynomial time inference is possible.
Definition 1.3.1. A commutative semiring S = (Y ∗ , ⊕, ⊗) is the combination of two commutative
semigroups Ge = (Y ∗ , ⊗) and Gm = (Y ∗ , ⊕) with two additional properties
⊕
⊗
⊕
⊗
⊕
• identity elements 1 and 1 such that 1 ⊕ a = a and 1 ⊗ a = a. Moreover 1 is an annihilator for
⊕
⊕
Ge = (⊗, Y ∗ ): a ⊗ 1 = 1
∀a ∈ Y ∗ .4
• distributive property:
a ⊗ (b ⊕ c) = (a ⊗ b) ⊕ (a ⊗ b) ∀a, b, c ∈ Y ∗
4
⊕
That is when dealing with reals, this is 1 = 0; this means a × 0 = 0.
1.3. POLYNOMIAL-TIME INFERENCE
23
The mechanism of efficient inference using distributive law can be seen in a simple example: instead
of calculating min(a + b, a + c), using the fact that summation distributes over minimization, we may
instead obtain the same result using a + min(b, c), which requires fewer operations.
Example 1.3.1. The following are some examples of commutative semirings:
– Sum-product (R≥0 , +, ×).
– Max-product (R≥0 ∪ {−∞}, max, ×) and ({0, 1}, max, ×).
– Min-max (S, min, max) on any ordered set S.
– Min-sum (R ∪ {∞}, min, +) and ({0, 1}, min, +).
– Or-and ({true, false}, ∨, ∧).
– Union-intersection (2S , ∪, ∩) for any power-set 2S .
– The semiring of natural numbers with greatest common divisor and least common multiple (N , lcm, gcd).
– Symmetric difference-intersection semiring for any power-set (2S , ∇, ∩).
Many of the semirings above are isomorphic –e.g., y 0 ∼
= − log(y) defines an isomorphism between
min-sum and max-product. It is also easy to show that the or-and semiring is isomorphic to minsum/max-product semiring on Y ∗ = {0, 1}.
The inference problems in the example above have different properties indirectly inherited from their
commutative semirings: for example, the operation min (also max) is a choice function, which means
∈ A. The implication is that if sum of the semiring is min (or max), we can replace it with
N
max and (if required) recover q(∅) using q(∅) = I fI (x∗ ) in polynomial time.
mina∈A a
argxJ
M
As another example, since both operations have inverses, sum-product is a field [247]. The availability of inverse for ⊗ operation – i.e., when Ge is an Abelian group – has an important implication
for inference: the expanded form of equation (1.1) can be normalized, and we may inquire about normalized marginals
p(xJ )
M
=
p(x)
(1.17)
x\J
where
p(x)
O
1
⊗
fI (xI )
q(∅)
def
=
⊕
if q(∅) 6= 1
(1.18)
I
p(x)
def
=
⊕
⊕
1
if q(∅) = 1
(1.19)
where p(x) is the normalized joint form. We deal with the case where the integral evaluates to the
annihilator as a special case because division by annihilator may not be well-defined. This also means,
⊗
L
when working with normalized expanded form and normalized marginals, we always have x p(xJ ) = 1
J
Example 1.3.2. Since Ge = (R>0 , ×) and Ge = (R, +) are both Abelian groups, min-sum and sumsum
product inference have normalized marginals. For min-sum inference this means minxJ p(xJ ) = 1 = 0.
However, for min-max inference, since (S, max) is not Abelian, normalized marginals are not defined.
24
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
We can apply the identity and annihilator of a commutative semiring to define constraints.
⊗ ⊕
Definition 1.3.2. A constraint is a factor fI : XI → { 1, 1} whose range is limited to identity and
annihilator of the expansion monoid.5
⊕
⊗
Here, fI (x) = 1 iff x is forbidden and fI (x) = 1 iff it is permissible. A constraint satisfaction
problem (CSP) is any inference problem on a semiring in which all factors are constraints. Note that
this allows definition of the “same” CSP on any commutative semiring. The idea of using different
semirings to define CSPs has been studied in the past [33], however its implication about inference on
commutative semirings has been ignored.
Theorem 1.3.1. Inference in any commutative semiring is NP-hard under randomized polynomial-time
reduction.
⊕ ⊗
Proof. To prove that inference in any semiring S = (Y ∗ , 1, 1) is NP-hard under randomized polynomial
reduction, we deterministically reduce unique satisfiability (USAT) to an inference problems on any
semiring. USAT is a so-called “promise problem”, that asks whether a satisfiability problem that is
promised to have either zero or one satisfying assignment is satisfiable. Valiant and Vazirani [298] prove
that a polynomial time randomized algorithm (RP) for USAT implies a RP=NP.
For this reduction consider a set of binary variables x ∈ {0, 1}N , one per each variable in the given
⊗
instance of USAT. For each clause, define a constraint factor fI such that fI (xI ) = 1 if xI satisfies
⊕
that clause and fI (xI ) = 1 otherwise. This means, x is a satisfying assignment for USAT iff q(x) =
⊗
⊕
L ⊕ ⊕
N
x 1 = 1 (by definition of 1). If the
I fI (xI ) = 1. If the instance is unsatisfiable, the integral q(∅) =
⊗
instance is satisfiable there is only a single instance x∗ for which q(x∗ ) = 1, and therefore the integral
⊗
evaluates to 1. Therefore we can decide the satisfiability of USAT by performing inference on any
semiring, by only relying on the properties of identities. The satisfying assignment can be recovered
using a decimation procedure, assuming access to an oracle for inference on the semiring.
Example 1.3.3. Inference on xor-and semiring ({true, false}, xor, ∧), where each factor has a disjunction form, is called parity-SAT, which asks whether the number of SAT solutions is even or odd. A
corollary to theorem 1.3.1 is that parity-SAT is NP-hard under randomized reduction, which is indeed
the case [298].
5
Recall that a monoid is a semigroup with an identity. The existence of identity here is a property of the semiring.
1.4. REDUCTIONS
25
We find it useful to use the same notation for the identity function 1(condition):
def
1(cond.) =
cond. = true
cond. = false
(+, ×) (min, +) (min, max)
1
0
−∞
0
+∞
+∞
(1.20)
where the intended semiring for 1(.) function will be clear from the context.
1.4
Reductions
Several of the inference problems over commutative semirings are reducible to each other. Section 1.4.1
reviews the well-known reduction of marginalization to integration for general commutative semirings.
We use this reduction to obtain approximate message dependencies in performing loop corrections in
section 2.4.
In equation (1.22), we introduce a procedure to reduce integration to that of finding normalized
marginals. The same procedure, called decimation, reduces sampling to marginalization. The problem
of sampling from a distribution is known to be almost as difficult as sum-product integration [151]. As
we will see in chapter 3, constraint satisfaction can be reduced to sampling and therefore marginalization.
In section 2.6.3 we introduce a perturbed message passing scheme to perform approximate sampling and
use it to solve CSPs. Some recent work perform approximate sampling by finding the MAP solution
in the perturbed factor-graph, in which a particular type of noise is added to the factors [125, 239].
Approximate sum-product integration has also been recently reduced to MAP inference [96, 97]. In
section 2.3, we see that min-max and min-sum inference can be obtained as limiting cases of min-sum
and sum-product inference respectively.
Section 1.4.2 reduces the min-max inference to min-sum also to a sequence of CSPs (and therefore
sum-product inference) over factor-graphs. This reduction gives us a powerful procedure to solve minmax problems, which we use in part II to solve bottleneck combinatorial problems.
In contrast to this type of reduction between various modes of inference, many have studied reductions of different types of factor-graphs [90]. Some examples of these special forms are factor-graphs with:
binary variables, pairwise interactions, constant degree nodes, and planar form. For example Sanghavi
et al. [274] show that min-sum integration is reducible to maximum independent-set problem. However
since a pairwise binary factor-graph can represent a maximum independent-set problem (see section 3.7),
this means that min-sum integration in any factor-graph can be reduced to the same problem on a pairwise binary model.
These reductions are in part motivated by the fact that under some further restrictions the restricted
factor-graph allows more efficient inference. For example, (I) it is possible to calculate the sum-product
integral of the planar spin-glass Ising model (see example 1.1.4) in polynomial time, in the absence of
26
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
local fields [99]; (II) the complexity of the loop correction method that we study in section 2.4.2 grows
exponentially with the degree of each node and therefore it may be beneficial to consider reduced factorgraph where |∂i| = 3; and (III) if the factors in a factor-graphs with pairwise factors satisfy certain
metric property, polynomial algorithms can obtain the exact min-sum integral using graph-cuts [41].
1.4.1
Marginalization and integration
This section shows how for arbitrary commutative semirings there is a reduction from marginalization
to integration and vice versa.
Marginalization reduces to integration
For any fixed assignment to a subset of variables xA = z A (a.k.a. evidence), we can reduce all the
factors fI (xI ) that have non-empty intersection with A (i.e., I ∩ A 6= ∅) accordingly:
fI\A (xI\A | z A )
def
M
=
fI (xI ) ⊗ 1(xI∩A = z I∩A )
∀I s.t. A ∩ I 6= ∅
(1.21)
xI∩A
where the identity function 1(.) is defined by equation (1.20). The new factor graph produced by
clamping all factors in this manner, has effectively accounted for the evidence. Marginalization or
integration, can be performed on this reduced factor-graph. We use similar notation for the integral and
marginal in the new factor graph – i.e., q(∅ | xA ) and q(xB | xA ). Recall that the problem of integration
is that of calculating q(∅). We can obtain the marginals q(z A ) by integration on reduced factor-graphs
for all z A ∈ XA reductions.
Claim 1.4.1.
q(z A )
=
q(∅ | z A )
(1.22)
Proof.
q(z A ) =
MO
x\A
fI (xI\A , z A∩I )
I
M
O
=
1(xA = z A ) ⊗
fI (xI\A , z A∩I )
x
I
MO
=
fA (xA ) ⊗ 1(xI∩A = z I∩A )
x
=
I
MO
x
I
fI\A (xI\A | z A )
=
q(∅ | z A )
1.4. REDUCTIONS
27
where we can then normalize q(xI ) values to get p(xI ) (as defined in equation (1.17)).
Integration reduces to marginalization
Assume we have access to an oracle that can produce the normalized marginals of equation (1.17).
We show how to calculate q(∅) by making N calls to the oracle. Note that if the marginals are not
L
normalized, the integral is trivially given by q(∅) = x q(xJ )
J
Start with t = 1, B(t = 0) = ∅ and given the normalized marginal over a variable p(xi(t) ), fix the
xi(t) to an arbitrary value zi(t) ∈ Xi(t) . Then reduce all factors according to equation (1.21). Repeat this
process of marginalization and clamping N times until all the variables are fixed. At each point, B(t)
denotes the subset of variables fixed up to step t (including i(t)) and p(xi(t) | z B(t−1) ) =
q(xi(t) |z B(t−1) )
q(∅|z B(t−1) )
refers to the new marginal. Note that we require i(t) ∈
/ B(t − 1) – that is at each step we fix a different
variable.
⊕
⊕
We call an assignment to xi(t) = zi(t) invalid, if p(zi(t) | z B(t) ) = 1. This is because 1 is the annihilator
of the semiring and we want to avoid division by the annihilator. Using equations (1.17) to (1.19), it is
⊕
L
easy to show that if q(∅) 6= 1, a valid assignment always exists (this is because xi(t) p(xi(t) | z B(t−1) ) =
⊗
⊕
1). Therefore if we are unable to find a valid assignment, it means q(∅) = 1.
Let z = z B(N +1) denote the final joint assignment produced using the procedure above.
Proposition 1.4.2. The integral in the original factor-graph is given by
!
q(∅)
=
O
I
fI (z I )
−1
O
⊗
p(zi(t) | z B(t−1) )
(1.23)
1≤t≤N
where the inverse is defined according to ⊗-operation.
Proof. First, we derive the an equation for “conditional normalized marginals” for semirings where ⊗
defines an inverse.
Claim 1.4.3. For any semiring with normalized joint form we have
p(x) = p(xi ) ⊗ p(x\i | xi )
where p(x\i | xi ) =
q(x\i |xi )
q(∅|xi )
To arrive at this equality first note that since x = x\i , xi , q(x\i | xi ) = q(x). Then multiply both
28
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
sides by q(xi ) = q(∅ | xi ) (see Claim 1.4.1) to get
q(x) ⊗ q(∅ | xi ) = q(xi ) ⊗ q(x\i | xi )
⇒
q(x)
q(xi ) q(x\i | xi )
=
⊗
q(∅)
q(∅)
q(∅ | xi )
⇒
p(x) = p(xi ) ⊗ p(x\i | xi )
where we divided both sides by q(∅) and moved a term from left to right in the second step.
Now we can apply this repeatedly to get a chain rule for the semiring:
p(x)
=
p(xi1 ) ⊗ p(xi2 | xi1 ) ⊗ p(xi3 | x{i1 ,i2 } ) ⊗ . . . ⊗ p(xiN | x{i1 ,...,iN −1 } )
which is equivalent to
p(x)
=
p(xi(1) ) ⊗ p(xi(2) | xB(1) ) ⊗ . . . ⊗ p(xi(N ) | xB(N −1) )
=
O
p(xi(t) | xB(t−1) )
1≤t≤N
Simply substituting this into definition of p(x) (equation (1.18)) and re-arranging we get
O
1≤t≤N
q(∅)
1 O
fI (xI ) ⇒
q(∅)
I∈F
−1
!
O
O
fI (z I ) ⊗
p(zi(t) | z B(t−1) )
p(xi(t) | xB(t−1) )
=
I
=
1≤t≤N
The procedure of incremental clamping is known as decimation, and its variations are typically
used for two objectives: (I) recovering the MAP assignment from (max) marginals (assuming a maxproduct semiring). Here instead of an arbitrary z J ∈ XJ , one picks z J = argxJ max p(xJ ). (II) producing
an unbiased sample from a distribution p(.) (i.e., assuming sum-product semiring). For this we sample
from p(xI ): z J ∼ p(xJ ).
1.4.2
Min-max reductions
The min-max objective appears in various fields, particularly in building robust models under uncertain
and adversarial settings. In the context of probabilistic graphical models, several min-max objectives
different from inference in min-max semiring have been previously studied [140, 162] (also see section 2.1.2). In combinatorial optimization, min-max may refer to the relation between maximization
and minimization in dual combinatorial objectives and their corresponding linear programs [276], or it
may refer to min-max settings due to uncertainty in the problem specification [5, 16].
1.4. REDUCTIONS
29
In part II we will see that several problems that are studied under the class of bottleneck problems can be formulated using the min-max semiring. Instances of these problems include bottleneck
traveling salesman problem [242], K-clustering [119], K-center problem [87, 164] and bottleneck assignment problem [121].
Edmonds and Fulkerson [92] introduce a bottleneck framework with a duality theorem that relates
the min-max objective in one problem instance to a max-min objective in a dual problem. An intuitive
example is the duality between the min-max cut separating nodes a and b – the cut with the minimum
of the maximum weight – and min-max path between a and b, which is the path with the minimum
of the maximum weight [104]. Hochbaum and Shmoys [136] leverages the triangle inequality in metric
spaces to find constant factor approximations to several NP-hard min-max problems under a unified
framework.
The common theme in a majority of heuristics for min-max or bottleneck problems is the relation
of the min-max objective to a CSP [136, 237]. We establish a similar relation within the context of
factor-graphs, by reducing the min-max inference problem on the original factor-graph to inference over
a CSP factor-graph (see section 1.3) on the reduced factor-graph in equation (1.25). In particular, since
we use sum-product inference to solve the resulting CSP, we call this reduction, sum-product reduction
of min-max inference.
Min-max reduces to min-sum
Here, we show that min-max inference reduces to min-sum, although in contrast to the sum-product
reduction of the next subsection, this is not a polynomial time reduction. First, we make a simple
S
observation about min-max inference. Let Y = I∈F YI denotes the union over the range of all factors.
The min-max value belongs to this set maxI∈F fI (x∗I ) ∈ Y. In fact for any assignment x, maxI∈F fI (xI ) ∈
Y.
Now we show how to manipulate the factors in the original factor-graph to produce new factors over
the same domain such that the min-max inference on the former corresponds to the min-sum inference
on the later.
Lemma 1.4.4. Any two sets of factors, {fI } and {gI }, over the identical domains {XI } have identical
min-max solutions
argx min max fI (xI ) = argx min max gI (xI )
I
I
if ∀I, J ∈ F, xI ∈ XI , xJ ∈ XJ
fI (xI ) < fJ (xJ )
⇔
gI (xI ) < gJ (xJ )
30
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
Proof. Assume they have different min-max assignments6 –i.e., x∗ = argx min maxI fI (xI ), x0∗ = argx min maxI gI (xI
and x∗ 6= x0∗ . Let y ∗ and y 0∗ denote the corresponding min-max values.
Claim 1.4.5.
y ∗ > max fI (x0∗
I )
I
∗
y <
max fI (x0∗
I )
I
⇔
⇔
y 0∗ < max gI (x∗I )
I
0∗
y > max fI0 (x∗I )
I
This simply follows from the condition of the Lemma. But in each case above, one of the assignments
y∗
or y 0∗ is not an optimal min-max assignment as there is an alternative assignment that has a lower
maximum over all factors.
This lemma simply states that what matters in the min-max solution is the relative ordering in the
factor-values.
Let y[1] ≤ . . . ≤ y[|Y|] be an ordering of elements in Y, and let r(fI (xI )) denote the rank in
{1, . . . , |Y|} of yI = fI (xI ) in this ordering. Define the min-sum reduction of {fI }I∈F as
gI (xI ) = 2r(fI (xI ))
∀I ∈ F
Theorem 1.4.6.
argx min
X
gI (xI )
=
I
argx min max fI (xI )
I
where {gI }I is the min-sum reduction of {fI }I .
Proof. First note that since g(z) = 2z is a monotonically increasing function, the rank of elements in
the range of {gI }I is the same as their rank in the range of {fI }I . Using Lemma 1.4.4, this means
argx min max gI (xI )
I
Since 2z >
Pz−1
l=0
argx min max fI (xI ).
I
(1.24)
2l , by definition of {gI } we have
max gI (xI ) >
I∈F
6
=
X
gI (xI )
where I∗ = argI max gI (xI )
I∈F \I∗
For simplicity, we are assuming each instance has a single min-max assignment. In case of multiple assignments there is
a one-to-one correspondence between them. Here the proof instead starts with the assumption that there is an assignment
x∗ for the first factor-graph that is different from all min-max assignments in the second factor-graph.
1.4. REDUCTIONS
31
It follows that for x1 , x2 ∈ X ,
max gI (x1I ) < max gI (x2I )
I
X
⇔
I
gI (x1I ) <
X
I
gI (x2I )
I
Therefore
argx min max gI (xI )
argx min
=
I
X
gI (xI ).
I
This equality, combined with equation (1.24), prove the statement of the theorem.
An alternative approach is to use an inverse temperature parameter β and re-state the min-max
objective as the min-sum objective at the low temperature limit
lim argx min
X
β→+∞
fIβ (xI )
=
argx min max fI (xI )
I
I
(1.25)
Min-max reduces to sum-product
S
Recall that Y = I∈F YI denote the union over the range of all factors. For any y ∈ Y, we reduce the
original min-max problem to a CSP using the following reduction.
Definition 1.4.1. For any y ∈ Y, py -reduction of the min-max problem:
x∗
=
argx min max fI (xI )
I∈F
(1.26)
is given by
py (x)
def
=
1 Y
1(fI (xI ) ≤ y)
qy (∅)
(1.27)
I∈F
where qy (∅) is the normalizing constant.7
This distribution defines a CSP over X , where py (x) > 0 iff x is a satisfying assignment. Moreover,
qy (∅) gives the number of satisfying assignments. The following theorem is the basis of our reduction.
Theorem 1.4.7. Let x∗ denote the min-max solution and y ∗ be its corresponding value –i.e., y ∗ =
maxI fI (x∗I ). Then py (x) is satisfiable for all y ≥ y ∗ (in particular py (x∗ ) > 0) and unsatisfiable for all
y < y∗.
7
To always have a well-defined probability, we define
0 def
=
0
0.
32
CHAPTER 1. REPRESENTATION, COMPLEXITY AND REDUCIBILITY
Proof. (A) py for y ≥ y ∗ is satisfiable: It is enough to show that for any y ≥ y ∗ , py (x∗ ) > 0. But since
py (x∗ )
1 Y
1(fI (x∗I ) ≤ y)
qy (∅)
=
I
and fI (x∗I ) ≤ y ∗ ≤ y, all the indicator functions on the rhs evaluate to 1, showing that py (x∗ ) > 0.
(B) py for y < y ∗ is not satisfiable: Towards a contradiction assume that for some y < y ∗ , py is
satisfiable. Let x denote a satisfying assignment –i.e., py (x) > 0. Using the definition of py -reduction,
this implies that 1(fI (xI ) ≤ y) > 0 for all I ∈ F. However this means that maxI fI (xI ) ≤ y < y ∗ , which
means y ∗ is not the min-max value.
This theorem enables us to find a min-max assignment by solving a sequence of CSPs. Let y[1] ≤
. . . ≤ y[|Y|] be an ordering of y ∈ Y. Starting from y = y[dN/2e], if py is satisfiable then y ∗ ≤ y. On
the other hand, if py is not satisfiable, y ∗ > y. Using binary search, we need to solve log(|Y|) CSPs
to find the min-max solution. Moreover at any time-step during the search, we have both upper and
lower bounds on the optimal solution. That is y < y ∗ ≤ y, where py is the latest unsatisfiable and py
is the latest satisfiable reduction.
However, finding an assignment x∗ such that py (x∗ ) > 0 or otherwise showing that no such assignment exists, is in general, NP-hard. Instead, we can use an incomplete solver [160], which may find a
solution if the CSP is satisfiable, but its failure to find a solution does not guarantee unsatisfiability.
By using an incomplete solver, we lose the lower bound y on the optimal min-max solution.8 However
the following theorem states that, as we increase y from the min-max value y ∗ , the number of satisfying
assignments to py -reduction increases, making it potentially easier to solve.
Proposition 1.4.8.
y1 < y2
⇒
qy1 (∅) ≤ qy2 (∅)
∀y1 , y2 ∈ Y
where qy (∅) ( i.e., partition function) is the number of solutions of py -reduction.
Proof. Recall the definition qy (∅) =
fI (xI ) ≤ y1
→
P Q
x
I 1(fI (xI )
≤ y). For y1 < y2 we have:
fI (xI ) ≤ y2
1(fI (xI ) ≤ y1 ) ≤ 1(fI (xI ) ≤ y2 )
XY
XY
1(fI (xI ) ≤ y1 ) ≤
1(fI (xI ) ≤ y2 )
x
I
x
I
qy1 (∅) ≤ qy2 (∅)
8
To maintain the lower bound one should be able to correctly assert unsatisfiability.
⇒
⇒
⇒
1.4. REDUCTIONS
This means that the sub-optimality of our solution is related to our ability to solve CSP-reductions
– that is, as the gap y − y ∗ increases, the py -reduction potentially becomes easier to solve.
33
Chapter 2
Approximate inference
2.1
Belief Propagation
A naive approach to inference over commutative semirings
q(xJ )
=
MO
x\J
fI (xI )
(2.1)
I
or its normalized version (equation (1.17)), is to construct a complete N -dimensional array of q(x) using
N
the tensor product q(x) =
I fI (xI ) and then perform ⊕-marginalization. However, the number of
elements in q(x) is |X |, which is exponential in N , the number of variables.
If the factor-graph is loop free, we can use distributive law to make inference tractable. Assuming
q(xK ) (or q(xk )) is the marginal of interest, form a tree with K (or k) as its root. Then starting from
the leaves, using the distributive law, we can move the ⊕ inside the ⊗ and define “messages” from leaves
towards the root as follows:
qi→I (xi )
=
O
qJ→i (xi )
(2.2)
J∈∂i\I
qI→i (xi )
=
M
x\i
fI (xI )
O
qj→I (xj )
(2.3)
j∈∂I\i
where equation (2.2) defines the message from a variable to a factor, closer to the root and similarly
equation (2.3) defines the message from factor I to a variable i closer to the root. Here, the distributive
law allows moving the ⊕ over the domain XI\i from outside to inside of equation (2.3) – the same way
⊕ moves its place in (a ⊗ b) ⊕ (a ⊗ c) to give a ⊗ (b ⊕ c), where a is analogous to a message.
By starting from the leaves, and calculating the messages towards the root, we obtain the marginal
2.1. BELIEF PROPAGATION
35
Figure 2.1: The figure shows a loop-free factor-graph and the direction of messages sent between variable and
factor nodes in order to calculate the marginal over the grey region.
over the root node as the product of incoming messages
q(xk )
=
O
qI→k (xk )
(2.4)
I∈∂k
In fact, we can assume any subset of variables xA (and factors within those variables) to be the root.
Then, the set of all incoming messages to A, produces the marginal
q(xA )
O
=
fI (xI )
I⊆A
O
qJ→i (xi )
i∈A,J∈∂i,J6⊆A
Example 2.1.1. Consider the joint form represented by the factor-graph of figure 2.1
q(x)
O
=
fA (xA )
A∈{I,J,K,L,O,T,U,V,W,X,Y,Z}
and the problem of calculating the marginal over x{i,j,k} (i.e., the shaded region).
q(x{i,j,k} )
=
M
O
fA (xA )
x\{i,j,k} A∈{I,J,K,L,O,T,U,V,W,X,Y,Z}
We can move the ⊕ inside the ⊗ to obtain
q(x{i,j,k} )
=
fI (xI ) ⊗ qL→i (xi ) ⊗ qK→i (xi ) ⊗ qV→j (xj ) ⊗ qW→j (xj ) ⊗ qK→k (xk )
(2.5)
36
CHAPTER 2. APPROXIMATE INFERENCE
where each term qA→i factors the summation on the corresponding sub-tree. For example
qL→i
=
M
fL (xL )
xw
Here the message qW→j is itself a computational challenge
qW→j
=
M
O
fA (xA )
x\j A∈{W,U,Y,X,O,T,Z}
However we can also decompose this message over sub-trees
qW→j
=
M
fA (xA ) ⊗ qe→W (xe ) ⊗ qr→W (xr )
x\j
where again using distributive law qe→W and qr→W further simplify based on the incoming messages to
the variable nodes xr and xe .
This procedure is known as Belief Propagation (BP), which is sometimes prefixed with the corresponding semiring e.g., sum-product BP. Even though BP is only guaranteed to produce correct answers
when the factor-graph is a tree (and few other cases [8, 22, 310, 313]), it performs surprisingly well when
applied as a fixed point iteration to graphs with loops [106, 225]. In the case of loopy graphs the message
updates are repeatedly applied in the hope of convergence. This is in contrast with BP on trees, where
the messages – from leaves to the root – are calculated only once. The message update can be applied
to update the messages either synchronously or asynchronously and the update schedule can play an
important role in convergence (e.g., [94, 173]). Here, for numerical stability, when the ⊗ operator has
an inverse, the messages are normalized. We use ∝ to indicate this normalization according to the mode
of inference
b
pI→i (xi )
∝
M
fI (xI )
x\i
b
pi→I (xi )
∝
O
b
pj→I (xj )
∝ PI→i (b
p∂I\i→I )(xi )
(2.6)
∝ Pi→I (b
p∂i\I→i )(xi )
(2.7)
j∈∂I\i
O
b
pJ→i (xi )
J∈∂i\I
b
p(xI )
∝
fI (xI )
∝
O
O
b
piI(xi )
(2.8)
i∈∂I
b
p(xi )
b
pI→i (xi )
(2.9)
I∈∂i
Here, for general graphs, b
p(xi ) and b
p(xI ) are approximations to p(xi ) and p(xI ) of equation (1.17). The
functionals Pi→I (b
p∂i\I→i )(.) and PI→i (b
p∂I\i→I )(.) cast the BP message updates as an operator on a
subset of incoming messages – i.e., b
p∂i\I→i = {b
pJ→i | J ∈ ∂i \ I}. We use these functional notation in
2.1. BELIEF PROPAGATION
37
presenting the algebraic form of survey propagation in section 2.5.
Another heuristic that is often employed with sum-product and min-sum BP is the Damping of
messages. This often improves the convergence when BP is applied to loopy graphs. Here a damping
parameter λ ∈ (0, 1] is used to partially update the new message based on the old one – e.g., for
sum-produt BP we have
b
pI→i (xi )
∝
λb
pI→i (xi ) + (1 − λ)
X
fI (xI )
x\i
b
pi→I (xi )
∝
λb
pi→I (xi ) + (1 − λ)
Y
Y
b
pj→I (xj )
(2.10)
j∈∂I\i
b
pJ→i (xi )
(2.11)
J∈∂i\I
(2.12)
where as an alternative one may use the more expensive form of geometric damping (where λ appears
in the power) or apply damping to either variable-to-factor or factor-to-variable messages but not both.
Currently – similar to several other ideas that we explore in this thesis – damping is a “heuristic”, which
has proved its utility in applications but lacks theoretical justification.
2.1.1
Computational Complexity
The time complexity of a single variable-to-factor message update (equation (2.2)) is O(|∂i| |Xi |). To
save on computation, when a variable has a large number of neighbouring factors, and if none of the
⊕
message values is equal to the annihilator 1 (e.g., zero for the sum-product), and the inverse of ⊗ is
defined, we can derive the marginals once, and produce variable-to-factor messages as
b
pi→I (xi )
=
b
p(xi ) ⊗ (b
pI→i (xi ))−1
∀I ∈ ∂i
(2.13)
This reduces the cost of calculating all variable-to-factor messages leaving a variable from O(|Xi | |∂i|2 )
to O(|Xi | |∂i|). We call this type of BP update, variable-synchronized (v-sync) update. Note that
since max is not Abelian on any non-trivial ordered set, min-max BP does not allow this type of variablesynchronous update. This further motivates using the sum-product reduction of min-max inference. The
time complexity of a single factor-to-variable message update (equation (2.3)) is O(|XI |). However as
we see in section 2.2, sparse factors allow much faster updates. Moreover in some cases, we can reduce
the time-complexity by calculating all the factor-to-variable messages that leave a particular factor at
the same time (e.g., section 5.2). We call this type of synchronized update, factor-synchronized
(f-sync) update.
38
CHAPTER 2. APPROXIMATE INFERENCE
2.1.2
The limits of message passing
By observing the application of distributive law in semirings, a natural question to ask is: can we use
distributive law for polynomial time inference on loop-free graphical models over any of the inference
problems at higher levels of inference hierarchy or in general any inference problem with more than one
marginalization operation? The answer to this question is further motivated by the fact that, when loops
exists, the same scheme may become a powerful approximation technique. When we have more than
one marginalization operations, a natural assumption in using distributive law is that the expansion
operation distributes over all the marginalization operations – e.g., as in min-max-sum (where sum
distributes over both min and max), min-max-min, xor-or-and. Consider the simplest case with three
1
2
1
2
1
M
O
operators ⊕, ⊕ and ⊗, where ⊗ distributes over both ⊕ and ⊕. Here the integration problem is
q(∅)
2
M
=
xJ
xJ
2
1
fI (xI )
I
where J1 and J2 partition {1, . . . , N }.
1
2
In order to apply distributive law for each pair (⊕, ⊗) and (⊕, ⊗), we need to be able to commute
1
2
⊕ and ⊕ operations. That is, we require
1
M
2
M
xA
xB
g(xA∪B ) =
2
1
M
M
xB
xA
g(xA∪B ).
for the specified A ⊆ J1 and B ⊆ J2 .
Now, consider a simple case involving two binary variables xi and xj , where g(x{i,j} ) is
xj
xi
0
1
0
a
b
1
c
d
Applying equation (2.14) to this simple case (i.e., A = {i}, B = {j}), we require
1
2
1
(a⊕b)⊕(c⊕d)
=
2
1
2
(a⊕b)⊕(c⊕d).
The following theorem leads immediately to a negative result:
Theorem 2.1.1. [91]:
1
2
1
2
1
2
(a⊕b)⊕(c⊕d) = (a⊕b)⊕(c⊕d)
⇔
1
2
⊕=⊕
∀a, b, c
(2.14)
2.2. TRACTABLE FACTORS
39
which implies that direct application of distributive law to tractably and exactly solve any inference
problem with more than one marginalization operation is unfeasible, even for tree structures. This
limitation was previously known for marginal MAP inference [241].
Min and max operations have an interesting property in this regard. Similar to any other operations
for min and max we have
min max g(xI∪J ) 6= max min g(xI∪J )
xJ
However,
xI
xI
xJ
if we slightly change the inference problem (from pure assignments xJl ∈ XJl to a
distribution over assignments; a.k.a. mixed strategies), as a result of the celebrated minimax theorem
[300], the min and max operations commute – i.e.,
min max
s(xJ ) s(xI )
X
s(xJ )g(xI∪J )s(xI )
xI∪J
=
max min
s(xI ) s(xJ )
X
s(xI )g(xI∪J )s(xJ1 )
xI∪J
where s(xJ1 ) and s(xJ2 ) are mixed strategies. This property has enabled addressing problems with
min and max marginalization operations using message-passing-like procedures. For example, Ibrahimi
et al. [140] solve this (mixed-strategy) variation of min-max-product inference. Message passing procedures that operate on graphical models for game theory (a.k.a. “graphical games”) also rely on this
property [161, 232].
2.2
Tractable factors
The applicability of graphical models to discrete optimization problems is limited by the size and
number of factors in the factor-graph. In section 2.2.1 we review some of the large order factors that
allow efficient message passing, focusing on the sparse factors used in part II to solve combinatorial
problems. In section 2.2.2 we introduce an augmentation procedure similar to cutting plane method to
deal with large number of “constraint” factors.
2.2.1
Sparse factors
The factor-graph formulation of many interesting combinatorial problems involves sparse (high-order)
factors. Here, either the factor involves a large number of variables, or the variable domains, Xi , have
large cardinality. In all such factors, we are able to significantly reduce the O(|XI |) time complexity of
calculating factor-to-variable messages. Efficient message passing over such factors is studied by several
works in the context of sum-product and min-sum inference classes [123, 249, 269, 291, 292]. Here we
confine our discussion to some of the factors used in part II.
The application of such sparse factors are common in vision. Many image labelling solutions to
problems such as image segmentation and stereo reconstruction, operate using priors that enforce sim-
40
CHAPTER 2. APPROXIMATE INFERENCE
ilarity of neighbouring pixels. The image processing task is then usually reduced to finding the MAP
solution. However pairwise potentials are insufficient for capturing the statistics of natural images and
therefore higher-order-factors have been employed [168–170, 174, 183, 234, 268].
The simplest form of sparse factor in combinatorial applications is the Potts factor, f{i,j} (xi , xj ) =
1(xi = xj ). This factor assumes the same domain for all the variables (Xi = Xj ∀i, j) and its tabular
form is non-zero only across the diagonal. It is easy to see that this allows the marginalization of
equation (2.3) to be performed in O(|Xi |) rather than O(|Xi | |Xj |). Another factor of similar form is the
inverse Potts factor, f{i,j} (xi , xj ) = 1(xi 6= xj ), which ensures xi 6= xj . In fact any pair-wise factor that
is a constant plus a band-limited matrix allows O(|Xi |) inference (e.g., factors used for bottleneck
TSP in section 5.2.2).
Another class of sparse factors is the class of cardinality factors, where Xi = {0, 1} and the factor
P
is defined based on only the number of non-zero values –i.e., fI (xI ) = g( i∈∂I xi ). Gail et al. [105]
P
proposes a simple O(|∂I| K) method for f(xI ) = 1(( i∈∂I xi ) = K). We refer to this factor as K-of-N
P
factor and use similar algorithms for at-least-K-of-N fI (xI ) = 1(( i∈∂I xi ) ≥ K) and at-most-K-of-N
P
fI (xI ) = 1(( i∈∂I xi ) ≤ K) factors.
An alternative is the linear clique potentials of Potetz and Lee [249].
O(|∂I| |Xi
The authors propose a
|2 )
(assuming all variables have the same domain Xi ) marginalization scheme for a general
P
family of factors, called linear clique potentials, where fI (xI ) = g( i∈∂I xi wi ) for a nonlinear gI (.). For
sparse factors with larger non-zero values (i.e., larger k), more efficient methods evaluate the sum of
pairs of variables using auxiliary variables forming a binary tree and use the Fast Fourier Transform to
reduce the complexity of K-of-N factors to O(|∂I| log(|∂I|)2 ) (see [292] and references in there).
Here for completeness we provide a brief description of efficient message passing through at-leastK-of-N factors for sum-product and min-sum inference.
K of N factors for sum-product
Since variables are binary, it is convenient to assume all variable-to-factor messages are normalized such
that b
pj→I (0) = 1. Now we calculate b
pI→i (0) and b
pI→i (1) for at-least-K-of-N factors, and then normalize
them such that b
pI→i (0) = 1.
In deriving b
pI→i (0), we should assume that at least K other variables that are adjacent to the factor
fI are nonzero and extensively use the assumption that b
pj→I (0) = 1. The factor-to-variable message of
equation (2.6) becomes
b
pI→i (0)
=
Y
X X
b
1 (
xj ) ≥ K
pj→I (xj )
x\i
j∈∂I\i
=
X
Y
A⊆∂I\i, |A|≥K
j∈A
j∈∂I\i
b
pj→I (1)
(2.15)
2.2. TRACTABLE FACTORS
41
where the summation is over all subsets A of ∂I \ i that have at least K members.
pI→i
Then, to calculate b
(1) we follow the same procedure, except that here the factor is replaced by
P
1 ( j∈∂I\i xj ) ≥ K − 1 . This is because here we assume xi = 1 and therefore it is sufficient for K − 1
other variables to be nonzero.
Note that in equation (2.15), the sum iterates over “all” A ⊆ ∂I \ i of size at least K. For high-order
factors fI (where |I| is large), this summation contains an exponential number of terms. Fortunately,
we can use dynamic programming to perform this update in O(|∂I| K). The basis for the recursion of
dynamic programming is that, starting from B = I \ i, a variable xk ∈ xK can be either zero or one
X
Y
b
pj→I (1) =
A∈{K⊆B,|K|≥k} j∈A
X
Y
b
pj→I (1) + b
pk→I (1)
A∈{K⊆B\k, |K|≥K} j∈A
X
Y
b
pj→I (1)
A∈{K⊆B\k, |A|≥K−1} j∈A
where each summation on the r.h.s. can be further decomposed using similar recursion. Here, dynamic
program reuses these terms so that each is calculated only once.
K of N factors for min-sum
Here again, it is more convenient to work with normalized variable-to-factor messages such that b
pj→I (0) =
⊗
1 = 0. Moreover in computing the factor-to-variable message b
pI→i (xi ) we also normalize it such that
b
pI→i (0) = 0. Recall the objective is to calculate
b
pI→i (xi )
=
X
X
b
pj→I (xj )
xj ) = K
min 1 (
x\i
j∈∂I
j∈∂I\i
for xi = 0 and xi = 1.
We can assume the constraint factor is satisfied, since if it is violated, the identity function evaluates
to +∞ (see equation (1.20)). For the first case, where xi = 0, K out of |∂I \ i| neighbouring variables
P
to factor I should be non-zero (because 1(( j∈∂I xj ) = K) and xi = 0). The minimum is obtained if
we assume the neighbouring variables with smallest b
pj→I (1) are non-zero and the rest are zero. For
xi = 1, only K − 1 of the remaining neighbouring variables need to be non-zero and therefore we need
to find K − 1 smallest of incoming messages (b
pj→I (1) ∀j ∈ ∂I \ i) as the rest of messages are zero due
to normalization.
By setting the b
pI→i (0) = 0, and letting A(K) ⊂ ∂I \ i identify the set of K smallest incoming
42
CHAPTER 2. APPROXIMATE INFERENCE
messages to factor I, the b
pI→i (1) is given by
b
pI→i (1)
=
X
b
pj→I (1) −
j∈A(K)
X
b
pj→I (1)
=
b
pj K →I (1)
j∈A(K−1)
where j K is the index of K th smallest incoming message to I, excluding b
pi→I (1). A similar procedure
can give us the at-least-K-of-N and at-most-K-of-N factor-to-variable updates.
If K is small (i.e., a constant) we can obtain the K th smallest incoming message in O(K |∂I|) time,
and if K is in the order of |∂I| this requires O(|∂I| log(|∂I|)) computations. For both min-sum and
sum-product, we incur negligible additional cost by calculating “all” the outgoing messages from factor
I simultaneously (i.e., f-sync update).
2.2.2
Large number of constraint factors
We consider a scenario where an (exponentially) large number of factors represent hard constraints
(see definition 1.3.2) and ask whether it is possible to find a feasible solution by considering only
a small fraction of these constraints. The idea is to start from a graphical model corresponding to a
computationally tractable subset of constraints, and after obtaining a solution for a sub-set of constraints
(e.g., using min-sum BP), augment the model with the set of constraints that are violated in the current
solution. This process is repeated in the hope that we might arrive at a solution that does not violate
any of the constraints, before augmenting the model with “all” the constraints. Although this is not
theoretically guaranteed to work, experimental results suggest this can be very efficient in practice.
This general idea has been extensively studied under the term cutting plane methods in different
settings. Dantzig et al. [77] first investigated this idea in the context of TSP and Gomory et al.
[118] provided a elegant method to identify violated constraints in the context of finding integral
solutions to linear programs (LP). It has since been used to also solve a variety of nonlinear optimization
problems. In the context of graphical models, Sontag and Jaakkola [284] (also [286]) use cutting plane
method to iteratively tighten the marginal polytope – that enforces the local consistency of marginals;
see section 2.3 – in order to improve the variational approximation. Here, we are interested in the
augmentation process that changes the factor-graph (i.e., the inference problem) rather than improving
the approximation of inference.
The requirements of the cutting plane method are availability of an optimal solver – often an LP
solver – and a procedure to identify the violated constraints. Moreover, they operate in real domain
Rd ; hence the term “plane”. However, message passing can be much faster than LP in finding approximate MAP assignments for structured optimization problems [325]. This further motivates using
augmentation in the context of message passing.
In sections 4.5 and 5.2, we use this procedure to approximately solve TSP and graph-partitioning
respectively. Despite losing the guarantees that make cutting plane method very powerful, augmentative
2.3. INFERENCE AS OPTIMIZATION
43
message passing has several advantages: First, message passing is highly parallelizable. Moreover by
directly obtaining integral solutions, it is much easier to find violated constraints. (Note the cutting
plane method for combinatorial problems operates on fractional solutions, whose rounding may eliminate
its guarantees.) However, due to non-integral assignments, cutting plane methods require sophisticated
tricks to find violations. For example, see [12] for application of cutting plane to TSP.
2.3
Inference as optimization
The variational approach is concerned with probabilities, and therefore this section is limited to operations on real domain. In the variational approach, sum-product inference is expressed as
b
p
=
argbp min
D(b
p | pβ )
(2.16)
where D(b
p | pβ ) is the KL-divergence between our approximation b
p and the true distribution p at
inverse temperature β (see example 1.1.4). Here b
p is formulated in terms of desired marginals.
Expanding the definition of KL-divergence and substituting p from equation (1.18), equation (2.16)
becomes
b
p
=
argbp min
X
x
x
argbp min
X
X
b
p(x) log(b
p(x)) − β
X
b
p(x) log(p(x)) ≡
(2.17)
!
b
p(x) log(b
p(x)) − β
x
X
b
p(x)
x
log(fI (xI ))
(2.18)
I
P Q
β from equation (2.17)
where we have removed the log partition function log(q(∅, β)) = log
f
(x
)
I
I
x
I
because it does not depend on b
p. This means that the minimum of equation (2.18) is log(q(∅, β)), which
appears when D(b
p | pβ ) = 0 – i.e., b
p = pβ .
The quantity being minimized in equation (2.18), known as variational free energy, has two
terms: the (expected) energy term
!
def
U(b
p, p) = −
X
b
p(x)
x
X
log(fI (xI ))
I
and the entropy term
H(b
p) = −
X
b
p(x) log(b
p(x)) .
x
Different families of representations for b
p(.) (in terms of its marginals) produces different inference
procedures such as BP, Generalized BP and Mean-field method [303].
Max-product (or min-sum) inference is retrieved as zero-temperature limit of sum-product infer-
44
CHAPTER 2. APPROXIMATE INFERENCE
ence:
argbp min βU(b
p, p) − H(b
p)
X
X
b
≡ argbp min −
p(x)
log(fI (xI ))
b
p(.) =
lim
β→+∞
x
(2.19)
I
where the energy term is linear in b
p and therefore the optima will be at a corner of probability simplex,
reproducing the MAP solution.
Here by defining f 0 (x) ←
1
f(x) ,
we get the min-sum form
X
b
p(.) = argbp min
b
p(xI )
X
x
log(fI0 (xI )
I
We observe that using a second parameter α, min-max inference is also retrievable1
b
p(.) =
lim
α→+∞
argbp min
≡ argbp min
X
X
b
p(x)
x
X
log(fI (xI ))α
I
b
p(xI ) max log(fI (xI ))
(2.20)
I
x
where again due to the linearity of the objective in b
p, the optima are at the extreme points of the
probability simplex.
To retrieve sum-product BP update equations from divergence minimization of equation (2.16), we
p(xi ). Here, we present this reparametrization
will reparameterize b
p using its marginals b
p(xI ) and b
in a more general form, as it holds for a any commutative semiring where (Y ∗ , ⊗) is an abelian group.
Proposition 2.3.1. If the ⊗ operator of the semiring has an inverse and the factor-graph is loop-free,
we can write p(x) as
N
b
p(x) = N
i
p(xI )
Ib
(2.21)
b
p(xi ) (|∂i| − 1)
def
where the inverse is w.r.t ⊗ and the exponentiation operator is defined as a b = |a ⊗ .{z
. . ⊗ a}.
b times
Proof. For this proof we use the exactness of BP on trees and substitute BP marginals equation (2.9)
1
Here we assume there are no ties at the min-max solution i.e., fI (x∗I ) > fJ (x∗J ) ∀J 6= I.
2.3. INFERENCE AS OPTIMIZATION
45
into equation (2.21):
N
N
p(xI )
Ib
=
p(xi ) (|∂i| − 1)
i b
N
N
pi→I (xi )
I fI (xI )
i∈∂I b
N N
pi→I (xi )
i
I∈∂i b
N
I fI (xI )
pi→I (xi )
i∈∂I b
=
pI→i (xi ) (|∂i| − 1)
I∈∂i b
N N
i
=
N
O
I
fI (xI )
=
p(x)
where we substituted the variable-to-factor messages in the denominator with factor-to-variable mes⊗
sages according to equation (2.2) and used the definition of inverse (i.e., a ⊗ a−1 = 1) to cancel out the
denominator.
Intuitively, the denominator is simply cancelling the double counts – that is since b
p(xi ) is counted
once for any I ∈ ∂i in the nominator, the denominator removes all but one of them.
2.3.1
Sum-product BP and friends
Q
Rewriting equation (2.21) for sum-product ring b
p(x) =
Q
energy minimization, we get
b
p
argbp min
=
β
X
p(xI )
Ib
p(xi )|∂i|−1
ib
and replacing b
p in the variational
b
p(xI )fI (xI )
(2.22)
I
−
XX
I
such that
X
b
p(xI )) −
p(xI ) log(b
xI
b
p(xI ) = b
p(xi )
!
X
i
∀i, I ∈ ∂i
(1− | ∂i |)
X
b
p(xi ) log(b
p(xi ))
(2.23)
xi
(2.24)
x\i
X
b
p(xi ) = 1
(2.25)
xi
where the energy term equation (2.22) is exact and the quantity that is minimized is known as Bethe
free energy [30, 328]. The constraints equations (2.24) and (2.25) ensure that marginals are consistent
and sum to one. Following the lead of Yedidia et al. [328], Heskes [132] showed that stable fixed points
of sum-product BP are local optima of Bethe free energy.
The optimization above approximates the KL-divergence minimization of equation (2.18) in two
ways: (I) While the marginal constraint ensure local consistency, for general factor-graphs there is no
guarantee that even a joint probability b
p with such marginals exists (i.e., local consistency conditions
outer-bound marginal polytope; the polytope of marginals realizable by a join probability p(x)).
(II) Bethe entropy is not exact for loopy factor-graphs. Using the method of Lagrange multipliers to
enforce the local consistency constraints and setting the derivatives of equation (2.23) w.r.t. b
p to zero,
recovers sum-product BP updates [327, 328]. This optimization view of inference has inspired many
46
CHAPTER 2. APPROXIMATE INFERENCE
sum-product inference techniques with convex entropy approximations and convergence guarantees [113,
126, 134, 206, 293, 304, 319, 329].
2.3.2
Min-sum message passing and LP relaxation
LP relaxation of min-sum problem seeks marginals b
p(xI )∀I
b
p
=
X
argbp min
b
p(xI )fI (xI )
(2.26)
∀i, I ∈ ∂i
(2.27)
I
such that
X
b
p(xI ) = b
p(xi )
x\i
X
b
p(xi ) = 1
xi
If integral (i.e., b
p(xi ) = 1(xi = x∗i ) for some x∗ = {x∗1 , . . . , x∗N }), this LP solution is guaranteed to be
optimal (i.e., identical to equation (2.19)). Taking the zero temperature limit (lim β → ∞) of the Bethe
free energy of equations (2.22) and (2.23), for any convex entropy approximation [126, 133, 304, 305],
ensures that sum-product message passing solution recovers the Linear Programming (LP) solution [311].
Moreover, replacing the summation with maximization (which again corresponds to temperature limit)
in the resulting convex message passing, produces the convex min-sum message passing, which agrees
with LP relaxations, under some conditions (e.g., when there are no ties in beliefs). The general interest
in recovering LP solutions by message passing is to retain its optimality guarantees while benefiting from
the speed and scalability of message passing that stems from exploitation of graphical structure [324].
One may also interpret some of these convex variations as replicating variables and factors while keeping
the corresponding messages identical over the replicates [271]. After obtaining message updates, the
number of replicates are allowed to take rational values (Parisi introduced a similar trick for estimation
of the partition function using replica trick [55, 211]).
Another notable variation for approximate MAP inference is max-product-linear-program , which
performs block coordinate descend in the space of duals for equation (2.26). MPLP is guaranteed to
converge and is often able to recover LP solution [114, 285]. Finally dual (and primal) decomposition methods minimize factors separately and combine their estimates in a way that agrees with
sub-gradient in each iteration [42, 155, 175, 176].
2.3.3
Min-max and other families
By rephrasing the variational min-max inference of equation (2.20)
b
p = argbp min
X
x
b
p(xI ) max log(fI (xI ))
I
2.4. LOOP CORRECTIONS
47
in terms of marginals b
p(xI ) and enforcing marginal consistency constraints, we obtain the following LP
relaxation
b
p
=
such that
argbp min y
X
b
p(xI )fI (xI ) ≤ y
(2.28)
∀I
xI
X
∀i, I ∈ ∂i
b
p(xi )
p(xI ) = b
x\i
X
b
p(xi ) = 1
xi
which surprisingly resembles our sum-product reduction of min-max inference in equation (1.25). Here
P
p(xI )fI (xI ) ≤ y is a relaxation of our sum-product factor 1(fI (xI ) ≤ y) in equation (1.27).
x b
I
Claim 2.3.2. y in equation (2.28) lower bounds the min-max objective y ∗ . Moreover, if b
p is integral,
then y = y ∗ and x∗ is the optimal min-max assignment.
Proof. The integral solution b
p, corresponds to the following optimization problem
x∗
such that
=
argx min
y
(2.29)
fI (xI ) ≤ y
≡
argx min max fI (xI )
I
which is the exact min-max inference objective. Therefore, for integral b
p, we obtain optimal min-max
solution. On the other hand by relaxing the integrality constraint, because of the optimality guarantee
of LP, the LP solution y can not be worse than the integral solution and its corresponding value y ∗ .
This lower bound complements the upper bound that we obtain using a combination of sum-product
reduction and an incomplete solver (such as perturbed BP of section 2.6.3) and can be used to assess
the optimality of a min-max solution.
The only other extensively studied inference problem in the inference hierarchy of section 1.2 is maxsum-product (a.k.a. marginal MAP) inference [80, 152, 202]. In particular variational formulation
of max-sum-product inference [195], substitutes the entropy term in equation (2.17) with conditional
entropy.
2.4
Loop corrections
In this section we first review the region-based methods that account for short loops in section 2.4.1
and then show how to perform loop correction by taking into account the message dependencies in
section 2.4.2. In section 2.4.3 we introduce a loop correction method that can benefit from both types
48
CHAPTER 2. APPROXIMATE INFERENCE
of loop corrections, producing more accurate marginals. While the region-based techniques can be used
to directly estimate the integral, the approximation techniques that take message dependencies into
account are only applied for estimation of marginals.
2.4.1
Short loops
We consider a general class of methods that improve inference in a loopy graphical model by performing
exact inference over regions that contain small loops.
The earliest of such methods is junction-tree [148, 186], which performs exact inference with
computation cost that grows exponentially in the size of largest region –i.e., tree width [57]. Here,
regions form a tree and the messages are passed over regional intersections. While this algorithm is still
popular in applications that involve certain class of graphs [35] or when the exact result is required,
most graphs do not have a low tree width [157, 167].
An extension to junction tree is the junction graph method [6] that removes the requirement for
the regions to form a tree. For this, the proxy between two regions is a subset of their intersection
(rather than the whole intersection) and one still requires the regions that contain a particular variable
to form a tree. Similar ideas are discussed under the name of cluster graphs in [171].
Inspired by the connection between Bethe free energy and belief propagation (see section 2.3),
Yedidia et al. [328] proposed Generalized BP that minimizes Kikuchi approximation to free energy
(a.k.a. Cluster Variational Method [165, 246]). Here the entropy approximation is obtained from a
region-graph.
A region ρ is a collection of connected variables V(ρ) and a set of factors F(ρ) such that each
participating factor depends only on the variables included in the region. To build the CVM regiongraph2 , one starts with predefined top (or outer) regions such that each factor is included in at least
one region. Then, we add the intersection of two regions (including variables and factors) recursively
until no more sub(inner)-region can be added. Each region is then connected to its immediate parent.
A region-graph, reparameterizes b
p(.) in terms of its marginals over the regions
b
p(x)
=
Y
b
p(xV(ρ) )c(ρ)
(2.30)
ρ
where c(ρ) is the counting number for region ρ and ensures that each variable and factor is counted
only once. This number is recursively defined by Möbius formula for inner regions:
c(ρ)
=
1 −
X
c(ρ)
ρ0 ⊃ρ
where ρ0 is an ancestors of ρ in the region graph.3
2
3
Here we are making a distinction between a general region graph and a CVM region-graph.
More accurately V(ρ) ⊆ V(ρ0 ).
(2.31)
2.4. LOOP CORRECTIONS
49
Similar to BP, by substituting the reparametrization of equation (2.30) into the variational free
energy minimization of equation (2.18) we get
b
p = argbp min
s.t.
X
X
c(ρ)
ρ
X
b
p(xV(ρ) )
xV(ρ)
b
p(xV(ρ0 ) ) = b
p(xV(ρ) )
X
fI (xI ) − b
p(xV(ρ) ) log(b
p(xV(ρ) ))
I∈F (ρ)
∀ρ ⊂ ρ0
(2.32)
x\V(ρ)
which is known as Kikuchi approximation to free energy [165]. The constraints of equation (2.32)
ensure that marginals are consistent across overlapping regions. Solving this constraint optimization
using the method of Lagrange multipliers, yields a set of recursive equations that are known as Generalized BP equations [328]. Again a region-based approximation is exact only if the region-graph has
no loops.
A region-graph without restrictions on the choice of regions generalizes junction-graph method as
well. The general construction of the region graph only requires that the counting numbers of all the
regions to which a variable (or a factor) belong, sum to 1 [327]. For different criteria on the choice of
regions see also [235, 314, 317].
2.4.2
Long loops
In graphical models with long-range correlation between variables, region-based methods are insufficient.
This is because their complexity grows exponentially with the number of variables in each region and
therefore they are necessarily inefficient account for long loops in the graph.
A class of methods for reducing long-range correlations are methods based on cut-set conditioning
[245], where a subset of variables are clamped, so as to remove the long-range correlations that are formed
through the paths that include the cut-set. For example, consider a Markov network in the form of
a cycle. By fixing any single variable, the reduced factor graph becomes a tree and therefore allows
exact inference. Several works investigate more sophisticated ideas in performing better inference by
clamping a subset and the resulting theoretical guarantees [78, 89, 312]. A closely related idea is RaoBlackwellization (a.k.a. collapsed MCMC; see section 2.6.1), a hybrid approach to inference [108]
where particles represent a partial assignment of variables and inference over the rest of variables is
performed using a deterministic method. The deterministic inference method such as BP is used to
calculate the value of the partition function, for each possible joint assignment of the variables that are
not collapsed. Then collapsed particles are sampled accordingly. This process in its general form is very
expensive, but one could reduce the cost, depending on the structure of the network [32].
The loop calculus of Chertkov [60][61] expands the free energy around the Bethe approximation,
with one term per each so-called generalized loop in the graph. Since the number of loops grows
exponentially in the number of variables, this expansion does not provide a practical solution. Some
50
CHAPTER 2. APPROXIMATE INFERENCE
attempts have been made to to make this method more practical by truncating the loop series [117].
While the original loop series was proposed for binary valued and pairwise factors, it has been generalized
to arbitrary factor-graphs [315, 320] and even region-graphs [334].
Another class of approximate inference methods perform loop correction by estimating the message
dependencies in a graphical model [217, 220, 262]. These methods are particularly interesting as they
directly compensate for the violated assumption of BP – i.e., corresponding to independent set of
incoming messages.
For the benefit of clarity, we confine the loop correction equations in this section and its generalization
in the next section to Markov networks (i.e., |∂I| = 2); see [254] for our factor-graph versions. Although
the previous works on loop corrections have been only concerned with sum-product inference, here we
present loop corrections for a general commutative semiring (Y ∗ , ⊕, ⊗) in which the operation ⊗ has
an inverse (i.e., (Y ∗ , ⊗) is a group). In particular this means these loop corrections may be used for
min-sum class of inference.
Here we rewrite BP update equations (2.2) and (2.3) for Markov networks4
b
pi→j (xi )
∝
M O
f{k,i} (x{k,i} ) ⊗ b
pk→i (xk )
(2.33)
\xi k∈∆i\j
b
p(xi )
∝
MO
f{k,i} (xk,i ) b
pk→i (xk )
(2.34)
\xi k∈∆i
Figure 2.2(left) shows the BP messages on a part of Markov network. Here if the Markov network
is a tree, BP’s assumption that b
ps→i , b
po→i and b
pn→i are independent is valid, because these messages
summarize the effect of separate sub-trees on the node i. However if the graph has loops, then we use
b
h(x∆i\j ) to denote message dependencies. If we had access to this function, we could easily change
the BP message update of equation (2.33) to
b
pi→j (xi )
∝
b
h(x∆i\j ) ⊗
M O
f{k,i} (x{k,i} ) ⊗ b
pk→i (xk )
\xi k∈∆i\j
Since it is not clear how to estimate h(x∆i\j ), we follow a different path, and instead estimate the
so-called cavity distribution, which we denote by b
h(x∆i ), which is simply the joint marginal over the
Markov blanket after making a cavity – i.e., removing a variable i and its neighbouring factors fI ∀I ∈ ∂i.
However, since the missing information is the “dependence” between the messages, b
h(x∆(i) ) has a degree
N
of freedom w.r.t. individual marginals – i.e., it can be inaccurate by a factor of
j∈∆(i) gj (xj ) for
some gj ∀j, without affecting the loop-corrected message passing procedure. This degree of freedom is
essentially equivalent to the freedom in initialization of BP messages. In the following, we show the
resulting loop-corrected message passing. But first we write BP updates in a different form.
4
Note that b
pi→j (xi ) is over Xi rather the conventional way of defining it on Xj . This formulation is the same as original
BP equation for Markov network if the graph does not have a loop of size two.
2.4. LOOP CORRECTIONS
51
Figure 2.2: (left) BP messages on a Markov network and the ideal way dependencies should be taken into
account. (right) BP marginal over extended Markov blanket ∇j for node j and the message dependencies over
the Markov blanket ∆i for node i.
Define the extended Markov blanket ∇i = ∆i ∪ {i} be the Markov blanket of i plus i itself, see
figure 2.2 (right). We can write BP marginals over ∇i
b
p(x∇i ) ∝
O
f{i,k} (x{i,k} ) ⊗ b
pk→i (xk )
(2.35)
M
b
p(x∇i )/f{i,j} (x{i,j} )
(2.36)
b
p(x∇i )
(2.37)
k∈∆i
Using this, equation (2.33) simplifies to:
b
pi→j (xi )
∝
\xi
b
p(xk )
∝
M
\xi
Now assume we are given the dependency between the messages b
pk→i (xk ) ∀k ∈ ∆i in the form of
b
h(x∆i ). This means we can re-express equation (2.35) as
b
p(x∇i )
∝
b
h(x∆i )
O
f{i,k} (x{i,k} ) ⊗ b
pk→i (xk )
(2.38)
k∈∆i
By enforcing the marginal consistency of b
p(x∇i ) and b
p(x∇j ) over xi (and xj )
M
x\i,j
b
p(x∇i )/f{i,j} (xi , xj )
=
M
b
p(x∇j )/f{i,j} (xi , xj )
x\i,j
we retrieve a message update similar to that of BP (in equation (2.33)) that incorporates the dependency
52
CHAPTER 2. APPROXIMATE INFERENCE
between BP messages
(t+1)
b
pi→j (xi )
L
∝L
\xi
b
p(x∇i )/f{i,j} (x{i,j} )
p(x∇j )/f{i,j} (x{i,j} )
\xi b
(t)
⊗b
pi→j (xi )
(2.39)
It is easy to verify that this update reduces to BP updates (equation (2.33)) when b
h(x∂i ) is uniform
– that is we do not have any dependency between messages. The loop-correction method of Mooij
et al. [220] is similar, however this interpretation does not apply to their updates for factor graphs.
We extend the same idea to perform loop-correction for overlapping regions of connected variables in
section 2.4.3 where we pass the messages from one region to the outer boundary of another region.
The main computational cost in these loop correction methods is estimating b
h(x ), the message
∆i
dependencies. We use clamping to perform this task. For this we remove xi and all the immediately
depending factors from the graph. Then we approximate the marginal b
h(x∆i ) by reduction to integrab
tion; see section 1.4.1. Note that the h(x ) obtained this way contains not only dependencies but also
∆i
the individual marginals in the absence of node i (b
h(xj ) ∀j ∈ ∆i). However since the messages updates
for b
pj→i (xj ) ∀j ∈ ∆i, perform “corrections” to this joint probability, we do not need to divide b
h(x∆i )
by the individual marginals.
2.4.3
Both message dependencies and short loops
Section 2.4.1 presented loop correction methods that improve loopy BP by considering interactions
within small clusters of variables, thus taking small loops within these clusters into account. The
previous section showed how to account for dependency between BP messages – thus taking long-range
correlations into account. In this section we introduce a generalization that performs both types of loop
correction.
The basic idea is to form regions, and perform exact inference over regions, to take short loops
into account. However in performing message passing between these regions, we introduce a method to
perform loop correction over these messages.
We start by defining a region ρ = {i, . . . , l} as a set of connected variables. Note that this definition
is different from definition of region for region-based methods as it only specifies the set of variables,
and not factors. Let ∆ρ = {i ∈ ∆j, i 6∈ ρ | j ∈ ρ} be the Markov blanket of region ρ, and as before let
∇ρ = ρ ∪ ∆ρ.
def
Region ρ1 is a neighbour of ρ2 with neighbourhood ρ1:2 iff ρ1:2 = (∆ρ1 )∩ρ2 6= ∅ – i.e., the Markov
blanket of ρ1 intersects with ρ2 (note that ρ1:2 is different from ρ2:1 ). The messages are exchanged on
these neighbourhoods and b
p1→2 (xρ1:2 ) is a message from region ρ1 to ρ2 .
Example 2.4.1. Figure 2.3 shows a set of neighbouring regions (indexed by 1, 2, 3, 4, 5, 6 and 7).
Here the region ρ1 receives “overlapping” messages from four other regions. For example the message
b
p6→1 (xd,i ) overlaps with the message b
p5→1 (xd,e ) as well as b
p7→1 (xh,i ). Therefore simply writing p(x∇ρ )
2.4. LOOP CORRECTIONS
Figure 2.3: (top) seven regions ρ1 , . . . , ρ7 and the domain of messages sent from each region to ρ1 . Here ρ5:1 is
the domain of message from region 5 to region 1. (bottom) the message region-graph shows how these overlapping
messages are combined to prevent double-counts.
in terms of the factors inside ∇ρ and the incoming messages (as in equation (2.5)) will double-count
some variables.
Message region-graphs
Here, similar to section 2.4.1, we construct a region-graph to track the double-counts. However, here we
have one message-region-graph per region ρ. The construction is similar to that of cluster variational
methods; we start with the original message-domains (e.g., ρ2:1 ) and recursively add the intersections,
until no more message-region γ can be added. Each message-region γ is connected to its immediate
parent. figure 2.3 shows the two-layered message-region-graph for region ρ1 . Here, for discussions
around a particular message-region-graph we will drop the region-index 1.
def
Let m(xγ ) = b
pπ→1 (xρπ:1 ) ∀π be the top regions in the region-graph, consisting of all the incoming
messages to region of interest ρ1 . The Möbius formula (equation (2.31)) gives counting number for
message-region (here again the top regions’ counting number is one). A downward pass, starting
from top regions, calculates the belief m(xγ ) over each message-region, as the average of beliefs over its
parents.
53
54
CHAPTER 2. APPROXIMATE INFERENCE
Example 2.4.2. In figure 2.3 the belief m(xe ) is the average of beliefs over m(x{d,e} ) = b
p5→1 (xρ5:1 )
and m(x{e,h} ) = b
p4→1 (xρ4:1 ) when marginalized over xe . Here the counting number of xe is c({e}) =
1 − (1 + 1) = −1.
We require the ⊗ operator of the semiring to have an inverse. Recall the power operator
def
y k =
y ⊗ . . . ⊗ y . For sum-product and min-sum semirings, this operator corresponds to exponentiation and
| {z }
k times
product respectively and it is well-defined also for rational numbers k. Now define the average as
def
avg({y1 , . . . , yk }) =
O 1
yi
k
(2.40)
i
Using Pa(γ) to denote the parents of region γ, in the downward pass
m(xγ ) ∝
O
M
avg(
m(γ 0 ))
γ 0 ∈Pa(γ)
(2.41)
x\γ
where m(xγ ) for top regions are just the incoming messages.
N
Let fA (xA ) = I⊆A fI (xI ) be the semiring-product of all the factors defined over a subset of A. For
example in figure 2.3, f∇ρa (x∇ρa ) is the product of 9 pairwise factors (i.e., all the edges in the figure).
After a downward pass, the belief over ∇ρ (analogous to equation (2.35)):
∝
b
p(x∇ρ )
f∇ρ (x∇ρ ) ⊗
O
m(xγ ) c(γ)
(2.42)
ρ
that is b
p(x∇ρ ) is the semiring product of all the factors inside this region and all the beliefs over messageregions inside its message-region-graph, where double counts are taken into account. For our example,
assuming sum-product ring,
−1
−1
b
p(x∇ρa ) = f∇ρa (x∇ρa )m(x4,5 ) . . . m(x9,4 ) m(x4 ) . . . m(x9 )
where the semiring-product and inverse are product and division on real domain.
At this point we can also introduce an estimate for message dependencies b
h(x∂ρ ) into the equation
above, and generalize the update of equation (2.39) to
L
(t+1)
b
pb→a (xρb:a ) ∝ L
\xρ
b:a
\xsρb:a
b
p(x∇ρb )/f∇ρa ∩∇ρb (x∇ρa ∩∇ρb )
b
p(x∇ρa )/f∇ρa ∩∇ρb (x∇ρa ∩∇ρb )
(t)
⊗ p̃b→a (xρb:a )
(2.43)
(t)
One last issue to resolve is to define the “effective” message p̃b→a (xρb:a ), which is different from
(t)
(t)
b
pb→a (xρb:a ). Since we did not directly use b
pb→a (xρb:a ) in the previous iteration, we should not include it
directly in this update. Instead we use the message region-graph for region a to calculate the effective
2.4. LOOP CORRECTIONS
55
message:
p̃b→a (xρb:a ) =
O
m(xγ ) c(γ)
(2.44)
γ⊆ρb:a
The effective message, as defined above (equation (2.44)), can be efficiently calculated in an upward
pass in the message region-graph. Starting from the parents of the lowest regions, update the belief
m(xγ ) obtained in downward pass equation (2.41) using the new beliefs over its children:
m̃(xγ )
=
m(xγ )
O
L
γ 0 ∈Ch(γ)
m̃(xγ 0 )
x 0 m(xγ )
(2.45)
\γ
where Ch(γ) is the set of children of γ in the message region-graph. After the upward pass, the new
beliefs over top regions gives us the effective messages
p̃a→b (xρa:b ) = m̃(xρa:b )
(2.46)
Example 2.4.3. In our example of figure 2.3(bottom), assuming a sum-product ring, since the messageregion-graph only has two layers, we can write the effective message p̃h→a (x{4,9} ) as
m(x4 )m(x9 )
P
p̃h→a (x{4,9} ) = b
ph→a (x{4,9} ) P
b
ph→a (x{4,9} ))
( x4 ph→a (x{4,9} ))( x9 b
(2.47)
This form of loop correction, which we call Generalized Loop Correction (GLC), generalizes both
correction schemes of sections 2.4.1 and 2.4.2. The following theorem makes this relation to generalized
BP more explicit.
Theorem 2.4.1.
5
For Markov networks, if the regions {ρ} partition the variables, then any Generalized
BP fixed point of a particular region-graph construction is also a fixed point for GLC, when using uniform
message dependencies ( i.e., b
h(x ) ∝ 1 ∀i, x ).
∆i
2.4.4
∆i
Experiments
This section compares different variations of our generalized loop correction method (GLC) for sumproduct ring against BP as well as Cluster Variational Method (CVM; section 2.4.1), loop correction
of [220] (LCBP), which does not exactly account for short loops and the Tree Expectation Propagation
(TreeEP) [214] method, which also performs some kind of loop correction. For CVM, we use the doubleloop algorithm [133], which is slower than Generalized BP but has better convergence properties.6 We
5
See [254] for the proof.
All methods are applied without any damping. We stop each method after a maximum of 1E4 iterations or if the
change in the probability distribution (or messages) is less than 1E-9.
6
56
CHAPTER 2. APPROXIMATE INFERENCE
report the time in seconds and the error for each method as the average of absolute error in single
P P
variable marginals – i.e., N1 i \xi | b
p(xi ) − p(xi ) |. For each setting, we report the average results
over 10 random instances of the problem. We experimented with grids and 3-regular random graphs.7
Both LCBP and GLC can be used without any information on message dependencies. with an
initial cavity distribution estimated via clamping cavity variables. In the experiments, full means
message dependencies b
h was estimated while uniform means b
h was set to uniform distribution (i.e., loopcorrection was over the regions only). We use GLC to denote the case where the regions were selected
such that they have no overlap (i.e., ρa ∩ ρb = ∅ ∀a, b) and GLC+ when overlapping clusters of some
form are used. For example, GLC+(Loop4, full) refers to a setting with message dependencies that
contains all overlapping loop clusters of length up to 4. If a factor does not appear in any loops, it
forms its own cluster. The same form of clusters are used for CVM.
Grids
We experimented with periodic spin-glass Ising grids of example 1.1.4. In general, smaller local fields
and larger variable interactions result in more difficult problems. We sampled local fields independently
from N (0, 1) and interactions from N (0, β 2 ). Figure 2.4a summarizes the results for 6x6 grids for
different values of β.
We also experimented with periodic grids of different sizes, generated by sampling all factor entries
independently from N (0, 1). Figure 2.5a compares the computation time and error of different methods
for grids of sizes that range from 4x4 to 10x10.
Regular Graphs
We generated two sets of experiments with random 3-regular graphs (all nodes have degree 3) over 40
variables. Here we used Ising model when both local fields and couplings are independently sampled
from N (0, β 2 ). Figure 2.4b shows the time and error for different values of β. Figure 2.5b shows time
versus error for graph size between 10 to 100 nodes for β = 1.
Our results suggest that by taking both long and short loops into account we can significantly
improve the accuracy of inference at the cost of more computation time. In fact both plots in figure 2.5
show a log-log trend in time versus accuracy which suggests that taking short and long loops into
account has almost independently improved the quality of inference.
2.5
Survey Propagation: semirings on semirings
Survey propagation (SP) was first introduced as a message passing solution to satisfiability [48] and was
later generalized to general CSP [47] and arbitrary inference problems over factor-graphs [208]. Several
7
The evaluations are based on implementation in libdai inference toolbox [221].
2.5. SURVEY PROPAGATION: SEMIRINGS ON SEMIRINGS
(a) spin-glass Ising grid
(b) spin-glass Ising model on a 3-regular graph
Figure 2.4: Average Run-time and accuracy for 6x6 spinglass Ising grids and 3-regular Ising model for different
values of β. Variable interactions are sampled from N (0, β 2 ) and local fields are sampled from N (0, 1).
57
58
CHAPTER 2. APPROXIMATE INFERENCE
(a) spin-glass Ising grid
(b) spin-glass Ising model on 3-regular graph
Figure 2.5: Time vs error for Ising grids and 3-regular Ising models with local field and interactions sampled
from a standard normal. Each method in the graph has 10 points, each representing an Ising model of different
size (10 to 100 variables).
works offer different interpretations and generalizations of survey propagation [46, 177, 199]. Here,
we propose a generalization based the same notions that extends the application of BP to arbitrary
commutative semirings. Our derivation closely follows and generalizes the variational approach of
Mézard and Montanari [208], in the same way that the algebraic approach to BP (using commutative
semirings) generalizes the variational derivation of sum-product and min-sum BP.
As a fixed point iteration procedure, if BP has more than one fixed points, it may not converge at
all. Alternatively, if the messages are initialized properly BP may converge to one of its fixed points.
SP equations, take “all” BP fixed points into account. In our algebraic perspective, this accounting of
_
_
all fixed points is using a third operation ⊕. In particular, we require that ⊗ also distribute over ⊕,
forming a second commutative semiring. We refer to the this new semiring as SP semiring.
Let b
p·→· be a BP fixed point – that is
b
p·→· = {∀i, I ∈ ∂i b
pi→I = Pi→I (b
p∂i\I→i ), b
pI→i = PI→i (b
p∂I\i→I )}
and denote the set of all such fixed points by P. Each BP fixed point corresponds to an approximation to
the q(∅), which we denote by Q(b
p·→· )(∅) – using this functional form is to emphasize the dependence of
this approximation on BP messages. Recall that in the original problem, X is the domain of assignments,
q(x) is the expanded form and ⊕-marginalization is (approximately) performed by BP. In the case of
survey propagation, P is domain of assignments and the integral Q(b
p·→· )(∅) evaluates a particular
2.5. SURVEY PROPAGATION: SEMIRINGS ON SEMIRINGS
59
Table 2.1: The correspondence between BP and SP
Belief Propagation
Survey Propagation
domain:
x
∀i
b
p·→·
xi
b
pi→I , b
pI→i
X
P
∀i, I ∈ ∂i
expanded form:
q(x)
Q(b
p·→· )(∅)
x q(x)
Q(∅)(∅) =
marginalization:
L
p(xi ) ∝ x\i p(x)
S(b
pI→i ) ∝
integration:
q(∅) =
L
_
L
p·→· )(∅)
b
p·→· Q(b
_
L
p·→· )
\b
pI→i P(b
factors:
∀I fI (xI )
e I (b
e i (b
e i↔I (b
P
p∂I→I )(∅), P
p∂i\I→i )(∅) and P
pi→I , b
pI→i )(∅)−1
∀i, I ∈ ∂i
p·→· )(∅) is the new expanded form.
assignment b
p·→· to all the messages – i.e., Q(b
_
In this algebraic perspective, SP efficiently performs a second integral using ⊕ over all fixed points:
Q(∅)(∅) =
_
M
b
p·→· ∈P
Q(q·→· )(∅)
(2.48)
Table 2.1 summarizes this correspondence.
Our derivation requires (Y ∗ , ⊗) to be an Abelian group (i.e., every element of Y ∗ has an inverse
w.r.t. ⊗). The requirement for invertablity of ⊗ is because we need to work with normalized BP and
SP messages. In section 2.5.4 we introduce another variation of SP that simply counts the BP fixed
points and relaxes this requirement.
2.5.1
Decomposition of the integral
In writing the normalized BP equations in section 2.1, we hid the normalization constant using ∝ sign.
Here we explicitly define the normalization constants or local integrals by defining unnormalized
60
CHAPTER 2. APPROXIMATE INFERENCE
messages, based on their normalized version
e
pI→i (xi )
def
M
=
fI (xI )
x\i
e
pi→I (xi )
def
def
b
pj→I (xj )
=
e I→i (b
P
p∂I\i→I )(xi )
(2.49)
e i→I (b
P
p∂i\I→i )(xi )
(2.50)
j∈∂I\i
O
=
O
def
b
pJ→i (xi )
=
J∈∂i\I
e
pI (xI )
def
=
fI (xI )
def
O
O
def
b
pi→I (xi )
=
e I (b
P
p∂I→I )(xI )
(2.51)
e i (b
P
p∂i→i )(xi )
(2.52)
i∈∂I
e
pi (xi )
=
def
b
pI→i (xi )
=
I∈∂i
where each update also has a functional form on the r.h.s. In each case, the local integrals are simply
L
pI→i (xi ).
the integral of unnormalized messages or marginals – e.g., e
pI→i (∅) = xi e
e
Define the functional Pi↔I (b
pi→I , b
pI→i ) as the product of messages from i to I and vice versa
def
e
pi↔I (xi )
b
pi→I (xi ) ⊗ b
pI→i (xi )
=
def
=
e i↔I (b
P
pi→I , b
pI→i )(xi )
(2.53)
Theorem 2.5.1. If the factor-graph has no loops and (Y ∗ , ⊗) is an Abelian group, the global integral
decomposes to local BP integrals as
−1
q(∅)
=
O
e
pI (∅)
O
e
pi (∅)
i
I
O
e
pi↔I (∅)
(2.54)
i,I∈∂i
or in other words q(∅) = Q(b
p·→· )(∅) where
−1
Q(b
p·→· )(∅) =
O
e I (b
P
p∂I→I )(∅)
O
e i (b
P
p∂i→i )(∅)
i
I
O
e i↔I (b
P
pi→I , b
pI→i )(∅)
(2.55)
i,I∈∂i
Proof. For this proof we build a tree around an root node r that is connected to one factor. (Since the
factor-graph is a tree such a node always exists.) Send BP messages from the leaves, up towards the
root r and back to the leaves. Here, any message qi→I (xi ), can give us the integral for the sub-tree that
L
contains all the nodes and factors up to node i using qi→I (∅) = xi qi→I (xi ). Noting that the root is
connected to exactly one factor, the global integral is
M
xr
q(xr ) =
M O
xr
I∈∂r
qI→r (xr ) = qI→r (∅)
(2.56)
On the other hand, We have the following relation between qi→I and b
pi→I (also corresponding
2.5. SURVEY PROPAGATION: SEMIRINGS ON SEMIRINGS
61
factor-to-variable message)
qi→I (xi ) = b
pi→I (xi ) ⊗ qi→I (∅)
∀i, I ∈ ∂i
(2.57)
qI→i (xi ) = b
pI→i (xi ) ⊗ qI→i (∅)
∀i, I ∈ ∂i
(2.58)
Substituting this into BP equation (2.2) we get
qi→I (xi )
O
=
qJ→i (∅)b
pJ→i (xi )
(2.59)
J∈∂i\I
qI→i (xi )
M
=
O
fI (xI )
x\i
qj→I (∅)b
pj→I (xj )
(2.60)
j∈∂I\i
By summing over both l.h.s and r.h.s in equations above and substituting from equation (2.50) we get
M
xi
qi→I (xi )
=
O
qJ→i (∅) ⊗
M
J∈∂i\I
qi→I (∅)
=
O
e
pi→I (∅)
xi
O
b
pJ→i (xi ) →
J∈∂i\I
qJ→i (∅)
(2.61)
J∈∂i\I
and similarly for equation (2.60) using integration and substitution from equation (2.49) we have
M
xi
qI→i (xi )
=
O
M
b
fI (xI )
pj→I (xj ) →
qj→I (∅) ⊗
O
xI
j∈∂I\i
qI→i (∅)
=
e
pI→i (∅)
O
j∈∂I\i
qj→I (∅)
(2.62)
j∈∂I\i
Equation (2.61) and 2.61 are simply recursive integration on a tree, where the integral up to node
i (i.e., qi→I (∅) in equation (2.61)) is reduced to integral in its sub-trees. By unrolling this recursion we
see that qi→I (∅) is simply the product of all e
pI→i (∅) and e
pI→i (∅) in its sub-tree, where the messages
are towards the root. Equation (2.56) tells us that the global integral is not different. Therefore,
equation (2.61) we can completely expand the recursion for the global integral. For this, let ↑ i restrict
the ∂i to the factor that is higher than variable i in the tree (i.e., closer to the root r). Similarly let ↑ I
be the variable that is closer to the root than I. We can write the global integral as
q(∅) =
O
i,I=↑i
O
e
pi→I (∅)
I,i=↑I
e
pI→i (∅)
(2.63)
Proposition 2.5.2 shows that these local integrals can be written in terms of local integrals of interest
62
CHAPTER 2. APPROXIMATE INFERENCE
– i.e.,
e
pI→i (∅)
=
e
pI (∅)
e
pi↔I (∅)
and e
pi→I (∅)
e
pi (∅)
e
pi↔I (∅)
=
Substituting from the equations above into equation (2.63) we get the equations of theorem 2.5.1.
Proof. (proposition 2.5.2) By definition of e
pI (xI ) and b
pi→I (xi ) in equation (2.49)
e
pI (xi ) = e
pI→i (xi ) ⊗ b
pi→I (xi )
M
e
b
pI (∅) = e
pI→i (∅) ⊗
pI→i (xi ) ⊗ b
pi→I (xi )
xi
→
M
→
e
pI (∅) = e
pI→i (∅) ⊗ e
pi↔I (∅)
xi
e
pI (xi ) =
M
xi
e
pI→i (xi ) ⊗ b
pi→I (xi )
→
where in the last step we used equation (2.53).
Similarly for the second statement of the proposition we have
e
pi (xi ) = e
pi→I (xi ) ⊗ b
pI→i (xi )
M
e
b
pi (∅) = e
pi→I (∅) ⊗
pI→i (xi ) ⊗ b
pi→I (xi )
xi
2.5.2
M
→
M
→
e
pi (∅) = e
pi→I (∅) ⊗ e
pi↔I (∅)
xi
e
pi (xi ) =
xi
e
pi→I (xi ) ⊗ b
pI→i (xi )
→
The new factor-graph and semiring
The decomposition of integral in theorem 2.5.1 means Q(b
p·→· )(∅) has a factored form. Therefore, a
factor-graph with b
p·→· as the set of variables and three different types of factors corresponding to
e I (b
e i (b
e i↔I (b
different terms in the decomposition – i.e., P
p
)(∅), P
p
)(∅) and P
pi→I , b
pI→i )(∅)−1 can
∂I→I
∂i\I→i
represent Q(b
p·→· )(∅).
Figure 2.6 shows a simple factor-graph and the corresponding SP factor-graph. The new factor-graph
has one variable per each message in the original factor-graph and three types of factors as discussed
above. Survey propagation is simply belief propagation applied to the this new factor-graph using the
new semiring. As before BP messages are exchanged between variables and factors. But here, we can
simplify BP messages by substitution and only keep two types of factor-to-factor messages. We use Si→I
and SI→i to denote these two types of SP messages. These messages are exchanged between two types
e I (b
e i (b
e i↔I (b
of factors, namely P
p∂I→I )(∅) and P
p∂i\I→i )(∅). Since the third type of factors P
pi→I , b
pI→i )(∅)−1
are always connected to only two variables, b
pi→I and b
pI→i , we can simplify their role in the SP message
2.5. SURVEY PROPAGATION: SEMIRINGS ON SEMIRINGS
63
Figure 2.6: Part of a factor-graph (left) and the corresponding SP factor-graph on the right. The variables in
e I (.)(∅),
SP factor-graph are the messages in the original graph. The SP factor-graph has three type of factors: (I)P
−1
e i (.)(∅) and (III)P
e i↔I (.)(∅) . As the arrows suggest, SP message updates are simplified so that only two
(II) P
type of messages are exchanged: Si→I and SI→i between factors of type (I) and (II).
update to get
Si→I (b
pi→I , b
pI→i ) ∝
SI→i (b
pi→I , b
pI→i ) ∝
_
M
e i (b
P
p∂i→i )(∅)
\b
pi→I ,b
pI→i
_
M
!
O
J∈∂i\I
e i↔I (b
P
pi→I , b
pI→i )(∅)
e I (b
O
)(∅)
P
p
∂I→I
\b
pi→I ,b
pI→i
e i↔I (b
P
pi→I , b
pI→i )(∅)
j∈∂I\i
SJ→i (b
pi→J , b
pJ→i )
(2.64)
!
Sj→I (b
pj→I , b
pI→j )
(2.65)
where in all cases we are assuming the messages b
p·→· ∈ P are consistent with each other – i.e., satisfy BP
equations on the original factor-graph. Note that, here again we are using the normalized BP message
update and the normalization factor is hidden using ∝ sign. This is possible because we assumed ⊗ has
an inverse. We can further simplify this update using the following proposition.
Proposition 2.5.2. for b
p·→· ∈ P
e i (b
P
p∂i→i )(∅)
e i↔I (b
P
pI→i , b
pi→I )(∅)
e i→I (b
= P
p∂i\I→i )(∅)
(2.66)
and
e I (b
P
p∂I→I )(∅)
e i↔I (b
P
pI→i , b
pi→I )(∅)
e I→i (b
= P
p∂I\i→I )(∅)
(2.67)
64
CHAPTER 2. APPROXIMATE INFERENCE
Proof. By definition of e
pI (xI ) and b
pi→I (xi ) in equation (2.49)
e
pI (xi ) = e
pI→i (xi ) ⊗ b
pi→I (xi )
M
e
b
pI (∅) = e
pI→i (∅) ⊗
pI→i (xi ) ⊗ b
pi→I (xi )
xi
M
→
M
→
e
pI (∅) = e
pI→i (∅) ⊗ e
pi↔I (∅)
xi
e
pI (xi ) =
xi
e
pI→i (xi ) ⊗ b
pi→I (xi )
→
where in the last step we used equation (2.53).
Similarly for the second statement of the proposition we have
e
pi (xi ) = e
pi→I (xi ) ⊗ b
pI→i (xi )
M
e
b
pi (∅) = e
pi→I (∅) ⊗
pI→i (xi ) ⊗ b
pi→I (xi )
xi
→
M
→
e
pi (∅) = e
pi→I (∅) ⊗ e
pi↔I (∅)
xi
e
pi (xi ) =
M
e
pi→I (xi ) ⊗ b
pI→i (xi )
xi
→
The term on the l.h.s. in the proposition above appear in equation (2.64) and the terms on the
r.h.s are local message integrals given by equation (2.49). We can enforce b
p·→· ∈ P, by enforcing BP
pI→i = PI→i (b
p∂I\→I ) “locally”, during the message updates in the
updates b
pi→I = Pi→I (b
p∂i\I→i ) and b
new factor-graph. Combining this constraint with the simplification offered by proposition 2.5.2 gives
us the SP message updates
_
M
Si→I (b
pi→I ) ∝
O
1 b
pi→I = Pi→I (b
p∂i\I→i ) ⊗ Pi→I (b
p∂i\I→i )(∅)
J∈∂i\I
SJ→i (b
pJ→i )
(2.68)
Sj→I (b
pj→I )
(2.69)
b
p
∂i\I→i
_
M
SI→i (b
pI→i ) ∝
pI→i
1 b
O
p∂I\i→I )(∅)
= PI→i (b
p∂I\i→I ) ⊗ PI→i (b
j∈∂I\i
b
p
∂I\i→I
⊗
_
⊕
where 1(.) is the identity function on the SP semiring, where 1(true) = 1 and 1(false) = 1.
Here each SP message is a functional over all possible BP messages between the same variable and
factor. However, in updating the SP messages, the identity functions ensure that only the messages
that locally satisfy BP equations are taken into account. Another difference from the updates of equation (2.64) is that SP messages have a single argument. This is because the new local integrals either
depend on b
pi→I or b
pI→i , and not both.
Example 2.5.1. In variational approach, survey propagation comes in two variations: entropic SP(ξ)
and energetic SP(y) [208]. For the readers familiar with variational derivation of SP, here we express
the relation to the algebraic approach. According to the variational view, the partition function of the
P
entropic SP is bp eξ log(Q(bp·→· )(∅)) , where Q(b
p·→· )(∅) is the partition function for the sum-product
·→·
semiring. The entropic SP has an inverse temperature parameter, a.k.a. Parisi parameter, ξ ∈ R. It
_
is easy to see that ξ = 1 corresponds to ⊕ = +, ⊕ = + and ⊗ = × in our algebraic approach.
The
2.5. SURVEY PROPAGATION: SEMIRINGS ON SEMIRINGS
65
_
limits of ξ → ∞ corresponds to ⊕ = max. On the other hand, the limit of ξ → 0 amounts to ignoring
Q(b
p·→· )(∅) and corresponds to the counting SP; see section 2.5.4.
P
The energetic SP(y) is different only in the sense that Q(b
p·→· )(∅) in bp e−y log(Q(bp·→· )(∅)) is the
·→·
_
P
ground state energy. This corresponds to ⊕ = +, ⊕ = max and ⊗ = , and the limits of the inverse
_
P
temperature parameter y → ∞ is equivalent to ⊕ = min, ⊕ = min and ⊗ = . By taking an algebraic
view we can choose between both operations and domains. For instance, an implication of algebraic
view is that all the variations of SP can be applied to the domain of complex numbers Y ∗ = C.
2.5.3
The new integral and marginals
Once again we can use theorem 2.5.1, this time to approximate the SP integral Q(∅)(∅) =
_
L
p·→· )(∅)
b
p·→· Q(b
using local integral of SP messages.
The SP marginal over each BP message b
pi→I or b
pI→i is the same as the corresponding SP message
– i.e., S(b
pi→I ) = Si→I (b
pi→I ). To see this in the factor-graph of figure 2.6, note that each message
variable is connected to two factors, and both of these factors are already contained in calculating one
SP messages.
Moreover, from the SP marginals over messages we can recover the SP marginals over BP marginals
which we denote by S(b
p)(xi ). For this, we simply need to enumerate all combinations of BP messages
that produce a particular marginal
S(b
p)(xi ) ∝
2.5.4
_
M
b
p∂i→i
O
1(b
p(xi ) = P(b
p∂i→i )(xi ))
I∈∂i
SI→i (b
pI→i )
(2.70)
Counting survey propagation
Previously we required the ⊗ operator to have an inverse, so that we can decompose the BP integral q(∅)
into local integrals. Moreover, for a consistent decomposition of the BP integral, SP and BP semiring
previously shared the ⊗ operation.8
Here, we lift these requirements by discarding the BP integrals altogether. This means SP semiring
could be completely distinct from BP semiring and (Y ∗ , ⊗) does not have to be an Abelian group. This
setting is particularly interesting when the SP semiring is sum-product over real domain
Si→I (b
pi→I ) ∝
X
Y
1 b
pi→I = Pi→I (b
p∂i\I→i )
SJ→i (b
pJ→i )
b
p∂i\I→i
SI→i (b
pI→i ) ∝
X
b
p∂I\i→I
8
_
(2.71)
J∈∂i\I
1 b
pI→i = PI→i (b
p∂I\i→I )
Y
Sj→I (b
pj→I )
(2.72)
j∈∂I\i
This is because if the expansion operation ⊗ was different from the expansion operation of BP, ⊗, the expanded form
Q(b
p·→· ) in the SP factor-graph would not evaluate the integral q(∅) in the BP factor-graph, even in factor-graphs without
any loops.
66
CHAPTER 2. APPROXIMATE INFERENCE
Here, the resulting SP integral Q(b
p·→· ) =
P
b
p·→·
1(b
p·→· ∈ P) simply “counts” the number of BP fixed
points and SP marginals over BP marginals (given by equation (2.70)) approximates the frequency of
a particular marginal. The original survey propagation equations in [48], that are very successful in
solving satisfiability correspond to counting SP applied to the or-and semiring.
Example 2.5.2. Interestingly, in all min-max problems with discrete domains X , min-max BP messages
can only take the values that are in the range of factors – i.e., Y ∗ = Y. This is because any ordered set
is closed under min and max operations. Here, each counting SP message Si→I (b
pi→I ) : Y |Xi | → R is a
discrete distribution over all possible min-max BP messages. This means counting survey propagation
where the BP semring is min-max is computationally “tractable”. In contrast (counting) SP, when
applied to sum-product BP over real domains is not tractable. This is because in this case each SP
message is a distribution over a uncountable set: Si→I (b
pi→I ) : R|Xi | → R.
In practice, (counting) SP is only interesting if it remains tractable. The most well-known case
corresponds to counting SP when applied to the or-and semiring. In this case the factors are constraints
and the domain of SP messages is {true, false}|Xi | . Our algebraic perspective extends this set of
tractable instances. For example, it show that counting SP can be used to count the number of fixed
points of BP when applied to xor-and or min-max semiring.
2.6
Messages and particles
The contrasting properties of stochastic and deterministic approximations make a general hybrid method
desirable. After reviewing the basics of MCMC in section 2.6.1, we discuss some particle-based approaches to message passing and introduce our hybrid inference method that combines message passing
and Gibbs sampling in section 2.6.3. The discussions of this section are limited to sum-product inference.
2.6.1
Markov Chain Monte Carlo
Markov Chain Monte Carlo (MCMC) is a technique to produce samples from a target distribution p,
by exploring a Markov Chain which is constructed such that more probable areas are visited more often
[11, 227, 263].
A Markov Chain is a stochastic process x(0) , . . . , x(t) in which:
p(x(t) | x(t−1) , . . . , x(1) )
=
kt (x(t) , x(t−1) )
that is the current state x(t) is independent of all the history, given only the previous state x(t−1) .
a homogeneous Markov chain, the transition kernel kt
and under some
9
assumptions9 ,
(x(t) , x(t−1) )
(2.73)
For
is the same for all t. In this case
starting from any arbitrary distribution x(0) ∼ p(0) (x) after at least Tmix
The assumptions are: (I) Irreducibility: There is a non-zero probability of reaching all states starting with any
arbitrary state and (II) Aperiodicity: The chain does not trap in cycles.
2.6. MESSAGES AND PARTICLES
67
transitions by the chain, we have x(Tmix ) ∼ p(x). Given a set of particles x[1], . . . , x[L] sampled from
p(x), we can estimate the marginal probabilities (or any other expectation) as
L
b
p(xi )
1X
1(x[n]i = xi )
L
∝
(2.74)
n=1
For a given transition kernel, the following condition, known as detailed balance, identifies the
stationary distribution p:
p(x(t) )k(x(t) , x(t−1) )
=
p(x(t) )
=
p(x(t−1) )k(x(t−1) , x(t) )
X
p(x) k(x(t) ; x)
⇒
(2.75)
x
which means that p(.) is the left eigenvector of k(., .) with eigenvalue 1. All the other eigenvalues are
less than one and the mixing time, Tmix , of the chain depends on the second largest eigenvalue; the
smaller it is, the faster consecutive transition by k(., .) shrinks the corresponding components, retaining
only p.
Metropolis-Hasting Algorithm and Gibbs sampling
Many important MCMC algorithms can be interpreted as a special case of Metropolis-Hasting (MH)
[124, 207]. Similar to importance sampling, MH uses proposal distribution m(x(t) | x(n−1) ), but in this
case, the proposal distribution is to help with the design of transition kernel k(., .). After sampling x(t)
from the proposal x(t) ∼ m(x), it is accepted with probability
(
min 1 ,
m(x(t)
p(x(t) )/p(x(t−1) )
| x(t−1) )/m(x(t−1) | x(t) )
)
(2.76)
where, if the proposed sample is not accepted, x(t) = x(t−1) .
The kernel resulting from this procedure admits the detailed balance condition w.r.t. the stationary
distribution p(x). An important feature of MCMC, which allows for its application in graphical models,
is the possibility of building valid transition kernels as the mixtures and cycles of other transition
kernels. If p, is the stationary distribution for k1 and k2 , then it is also the stationary distribution for
k1 k2 (cycle) and λk1 + (1 − λ)k2 , 0 ≤ λ ≤ 1 (mixture) [295].
Cycling of kernels gives us Gibbs sampling in graphical models [109], when kernels are
(t)
(t−1)
ki (xi , x∆i
)
=
(t)
(t−1)
p(xi | x∆i
)
∀i
(2.77)
where as before ∆i is the Markov blanket of node i. It is also possible to use block MH-kernels
with graphical models. In MH-samplers, when highly correlated variables are blocked together, mixing
68
CHAPTER 2. APPROXIMATE INFERENCE
properties improve. In fact, the Gibbs sampler is such a method, with the proposal distribution that
results in acceptance with probability 1.
Similar to the general MH-methods, Gibbs sampling can fail if the kernel does not mix properly.
This could happen if variables are strongly correlated. In principle one can assemble neighbouring
variables into blocks and update them as one [147]. However in difficult regimes the number of variables
that should be flipped to move from one local optima to another, is in the order of total number of
variables [208], which makes this approach intractable.
Mixture of kernels can be used to combine a global proposal with a local proposal (e.g., [81, 88]). In
fact if we could view a message passing operator as a transition kernel (at least when message passing
is exact), then the mixture of kernels – e.g., with Gibbs sampling – could produce interesting hybrid
methods. In section 2.6.3, by combining BP and Gibbs sampling operator (when rephrased as message
update) we introduce a new hybrid method.
2.6.2
Hybrid methods
Stochastic methods are slow in convergence but they are guaranteed to converge. Even if the kernel
is reducible, samples will cover a subset of the true support – i.e., MCMC still converges to a single
sub-measure when the Gibbs measure is not unique.
On the other hand, deterministic approximations are fast but non-convergent in difficult regimes.
Modifications that result in convergence are either generally intractable (e.g., SP), slow (e.g., loop
corrections and the methods that tighten a bound over the free energy) and/or degrade the quality of
solutions.
Moreover, sampling methods are flexible in representing distributions. This has motivated growing
interest in nonparametric approach to variational inference [110] and in particular variations of Belief
Propagation [141, 142, 144, 172, 231, 282, 283]. However, in the sense that these methods do not rely
on a Markov chain for inference, they are closer to variational inference than MCMC methods.
To better appreciate this distinction, consider two closely related methods: Gibbs Sampling and
hard-spin mean-field [10], that uses the following update equation
b
p(xi )
∝
XY
fI (xI\i , xi )
x\i I∈∂i
Y
b
p(xj )
j∈I\i
Interestingly, the detailed balance condition of equation (2.75) for Gibbs sampler gives us the same
equation:
b
p(xi )
∝
XY
x\i I∈∂i
b
p(xi | xI\i )
Y
b
p(xj )
j∈I\i
However, given enough iterations, Gibbs Sampling can be much more accurate than hard-spin mean
2.6. MESSAGES AND PARTICLES
69
field method. Here the difference is that, with Gibbs sampler, this equation is enforced by the chain
rather than explicit averaging of distributions, which means the correlation information is better taken
into account.
2.6.3
Perturbed Belief Propagation
Consider a single particle x̂ = x[1] in Gibbs Sampling. At any time-step t, x̂i is updated according to
(t)
x̂i
∼ b
p(xi )
Y
∝
(t−1)
fI (xi , x̂∂I\i )
(2.78)
I∈∂i
Here we establish a correspondence between a particle in Gibbs Sampling and a set of variable-topi→I (.)}i,I∈∂i , by defining all the messages leaving variable xi as a
factor messages in BP –i.e., x̂ ⇔ {b
delta-function
b
pi→I (xi )
def
=
def
1(xi = x̂i )
=
Gi→I (b
p∆i→∂i )(xi )
∀I ∈ ∂i
(2.79)
where Gi→I (b
p∆i→∂i ) is the Gibbs sampling operator that defines variable-to-factor message b
pi→I as a
function of all the messages from Markov blanket of i (∆i) to its adjacent factors (∂i). To completely
define this random operator, note that x̂i is a sample from the conditional distribution of Gibbs sampling
x̂i ∼ b
p(xi )
Y
∝
fI (xi , x̂∂I\i )
J∈∂i
∝
YX
fI (xI )
I∈∂i x\i
Y
b
pj→I (xj )
(2.80)
j∈∂I\i
Combining the operators
Now, lets write the BP updates for sum-product semiring once again; by substituting the factor-tovariable messages (equation (2.6)) into variable-to-factor messages and the marginals (equations (2.7)
and (2.9)) we get
b
pi→I (xi )
∝
Y X
fJ (xJ )
J∈∂i\I X∂J\i
b
p(xi )
∝
Y X
I∈∂i X∂I\i
fI (xI )
Y
b
pj→J (xj )
∝
Pi→I (b
p∆i→∂i )(xi )
(2.81)
j∈∂J\i
Y
b
pj→I (xj )
(2.82)
j∈∂I\i
where, similar to equation (2.7), Pi→I (.) denotes the message update operator, with the distinction that
here, the arguments are also variable-to-factor messages (rather than factor-to-variable messages).
By this rewriting of BP updates, the BP marginals equation (2.82) are identical in form to the Gibbs
sampling distribution of equation (2.80). This similar form allows us to combine the operators linearly
70
CHAPTER 2. APPROXIMATE INFERENCE
to get perturbed BP operator:
def
Xi→I (b
p∆i→∂i ) = γ Gi→I (b
p∆i→∂i ) + (1 − γ)Pi→I (b
p∆i→∂i )
∀i, I ∈ ∂i
(2.83)
The Perturbed BP operator Xi→I (.) updates each message by calculating the outgoing message
according to BP and GS operators and linearly combines them to get the final massage. During T
iterations of Perturbed BP, the parameter γ is gradually and linearly changed from 0 towards 1.10
Algorithm 2 summarizes this procedure. Note that the updates of perturbed BP are compatible with
variable synchronous message update (see section 2.1.1).
input : a factor graph, number of iterations T.
output: a sample x̂.
Initialize messages
γ←0
repeat
for each variable xi do
calculate b
p(xi ) using equation (2.80)
calculate BP messages b
pi→I (.) using equation (2.81) ∀I ∈ ∂i
sample x̂i ∼ b
p(xi )
combine BP and Gibbs sampling messages:
b
pi→I (xi ) ← γ b
pi→I (xi ) + (1 − γ) 1(xi = x̂i )
(2.84)
end
1
γ ← γ + T −1
until T iterations
return x̂
Algorithm 2: Perturbed Belief Propagation
Experiments
Perturbed BP is most successful in solving CSPs; see chapter 3. However we can also use it for
marginalization by sampling (equation (2.74)). Here we use the spin-glass Ising model on 8 × 8 grids
and Erdős-Rény (ER) random graphs with N = 50 and 150 edges. We sampled local fields independently
from N (0, 1) and interactions from N (0, θ2 ), where we change θ to control the problem difficulty – higher
values correspond to more difficult inference problems. We then compared the average of the logarithm
10
Perturbed BP updates can also be used with any fixed γ ∈ [0, 1].
2.6. MESSAGES AND PARTICLES
71
(base 10) of mean (over N variables) marginal error
log(
1 X
|p(xi ) − b
p(xi )|)
N
(2.85)
i,xi
(a) Ising grid
(b) Random ER graph
Figure 2.7: Log mean marginal error (x and y axes) comparison between Perturbed BP, BP and Gibbs sampling
for (left) 8x8 periodic Ising grid; (right) random graph with 50 variables, 150 edges and spin-glass interactions.
The size of each circle is proportional to the difficulty of that problem.
Figure 2.7 compares perturbed BP, Gibbs sampling and BP where larger circles correspond to
more difficult instances θ ∈ {0.5, 1, 2, 4, 8}. All methods are given a maximum of 10, 000 iterations.
Perturbed BP and Gibbs sampling use T = 100 iterations to obtain each sample, while BP is ran once
until convergence or until the maximum number of iterations is reached.
The results shows that Perturbed BP as a sampling method is generally better than Gibbs Sampling.
Also, for some cases in which BP’s result is very close to random (i.e., ∼ −.3 in log marginal error),
Perturbed BP produces relatively better results. Note that for difficult instances, increasing T even
by 100 folds does not significantly improve the results for either Gibbs sampling or perturbed BP. For
Gibbs sampling, this can be explained by formation of pure states that result in exponential mixing
time [192, 208].
2.6.4
Perturbed survey propagation
_
_
When the SP operators are sum-product – i.e., ⊕ = sum and ⊗ = prod – we can apply a perturbation
scheme similar to perturbed BP.
Recall that sum-product SP defines a distribution over BP fixed point. Therefore sampling from
this distribution amounts to randomly selecting a single BP fixed point. This corresponds to sampling
a single message b
pi→I [1] ∼ Si→I (b
pi→I )11 and bias the SP message Si→I (.) towards this random choice –
11
Recall that the SP marginal over each message b
pi→I is identical to the corresponding message Si→I (b
pi→I ).
72
CHAPTER 2. APPROXIMATE INFERENCE
i.e., (analogous to equation (2.84) in algorithm 2)
Si→I (b
pi→I )
←
γ Si→I (b
pi→I ) + (1 − γ)1(b
pi→I = b
pi→I [1])
(2.86)
where b
pi→I [1] ∼ Si→I (b
pi→I )
An alternative form of perturbation is to perturb SP messages using implicit SP marginals. Recall
that in using counting SP, the SP marginals over BP marginals (S(b
p)(xi ); see equation (2.70)) are simply
the frequency of observing a particular marginal in BP fixed points. This implicitly defines SP marginal
over the original domains Xi ; ∀i, which we denote by S(xi )
S(xi ) ∝
X
S(b
p)(xi )
(2.87)
b
p
After obtaining a sample x̂i ∼ S(xi ), we bias all the outgoing SP messages accordingly
Si→I (b
pi→I )
where
←
γ Si→I (b
pi→I ) + (1 − γ) 1 b
pi→I (.) = 1(x̂i , .)
∀I ∈ ∂i
(2.88)
x̂i ∼ Si (xi )
where, similar to perturbed BP, γ is gradually increased from 0 to 1 during T iterations of Perturbed
SP. We use this form of perturbation in section 3.3 to obtain a satisfying assignment x̂, to CSPs. We
show that although computationally more expensive than perturbed BP, this method often outperforms
all the other well-known methods in solving random CSPs.
Part II
Combinatorial problems
74
Message Passing algorithms of different semirings are able to solve a variety of combinatorial problems: (I) To solve constraint satisfaction problems (CSPs) the sum-product message passing is
often used, where p(x) defines a uniform distribution over solutions and the objective is to produce
a single assignment x∗ s.t. p(x∗ ) > 0. (II) Here the estimates of the partition function, either using
the approximation given by the Bethe free energy (section 2.3.1) or the decomposition of integral in
section 2.5.1, is used for approximate counting of the number of solutions. This estimate to the
partition function is also used for integration problems such approximating the permanent of a matrix
(see chapter 5). (III) The min-sum semiring is often used for (constrained) optimization and we
formulate (IV) bottleneck problems as min-max inference.
This part of the thesis studies the message passing solutions to combinatorial problems under three
broad categories. (1) Chapter 3 studies the constraint satisfaction problems, where we use perturbed
message passing (section 2.6.3 and 2.6.4) to produce state-of-the-art results in solving random instances
of satisfiability and coloring problems in section 3.3. This chapter then studies several other NP-hard
problems including set-cover, independent set, max-clique, clique-cover and packing for construction of
non-linear codes. (2) Chapter 4 studies variations of clustering problems including k-median, k-center,
k-clustering, hierarchical clustering and modularity optimization. (3) In chapter 5 we study problems
that involve enumeration, constraint satisfaction or constrained optimization over permutations. This
includes (bottleneck) travelling salesman problem, matching, graph alignment, graph isomorphism and
finding symmetries.
Note that this classification of combinatorial problems into three categories is superficial and is made
solely to provide some organization. In several places we violate this categorization in favour of better
flow. For example we study some constraint satisfaction problems such as (sub)-graph isomorphism and
Hamiltonian cycle in chapter 5 rather than chapter 3. We investigate the “optimization” counterpart
of some CSPs in chapter 3 and review message passing solutions to finding trees rather than clusters in
chapter 4. Moreover, many of the graphical models presented here are proposed by other researchers and
they are included here only for completeness. As a final remark, we note that many of the statements
in the following are assuming P 6= NP.
Chapter 3
Constraint satisfaction
We saw in section 1.3 that “any” semiring can formulate Constraint Satisfaction Problems (CSPs). In
particular, as we saw in section 1.3, several semirings are isomorphic to the and-or ({false, true}, ∨, ∧)
semiring and therefore result in equivalent BP procedures.
The BP message update over the and-or
semiring is called warning propagation (WP). WP marginals indicate whether or not a particular
assignment to each variable is allowed, and therefore indicate a cluster of solutions. However, the success
of warning propagation highly depends on initialization of messages. In contrast, if convergent, the fixed
points of BP on the sum-product semiring (R≥0 , +, ×) are less dependent on initialization.
Example 3.0.1. K-coloring:
Given a graph G = (V, E), the K-coloring (K-COL) problem asks
whether it is possible to assign one color (out of K) to each node s.t. no two adjacent nodes have the
same color. Here, xi ∈ Xi = {1, . . . , q} is a K-ary variable for each i ∈ N , and we have M = |E|
constraints; each constraint fi,j (xi , xj ) = 1(xi 6= xj ) depends only on two variables and is satisfied iff
the two variables have different values. Here the identity function 1(xi 6= xj ) depends on the semiring
(see section 1.3).
Figure 3.1: (a) The set of all possible assignments to 3 variables. The solutions to the 3-SAT problem of
equation (3.1) are in white circles. (b) The factor-graph corresponding to the 3-SAT problem of equation (3.1).
Here each factor prohibits a single assignment.
76
CHAPTER 3. CONSTRAINT SATISFACTION
Example 3.0.2. K-Satisfiability (K-SAT): Given a conjunction of disjunctions with K literals,
K-satisfiability seeks an assignment that evaluates to true. Here, all variables are binary (Xi =
{true, false}) and each clause (factor fI ) depends on K =| ∂I | variables. A clause evaluates to
zero only for a single assignment out of 2K possible assignment of variables [107].
Consider the following (or-and semiring formulation of) 3-SAT problem over 3 variables with 5
clauses:
q(x) = (¬xi ∨ ¬xj ∨ xk ) ∧ (¬xi ∨ xj ∨ xk ) ∧ (xi ∨ ¬xj ∨ xk ) ∧
{z
} |
{z
} |
{z
}
|
fI
fJ
(3.1)
fK
(¬xi ∨ xj ∨ ¬xk ) ∧ (xi ∨ ¬xj ∨ ¬xk )
|
{z
} |
{z
}
fL
fH
⊗
The factor fI corresponding to the first clause takes the value 1 (for or-and semiring this corresponds
∧
⊕ ∨
to 1 = true), except for xI = (true, true, false), in which case it is equal to 1 (1 = false). Figure 3.1
shows this factor-graph and its set of solutions:
S = (true, true, true), (false, false, false), (false, false, true) .
When using sum-product semiring, where fI (x) ∈ YI = {0, 1}, p(x) (equation (1.17)) defines a
uniform distribution over the set of solutions and the partition function q(∅) counts the number of
solutions. The challenge is then to sample from this distribution (or estimate q(∅)). The common
approach to sample from p(x) is to use decimation (see equation (1.22)). Here one repeatedly applies
sum-product BP to estimate marginals b
p(xi ). Then one fixes a subset of variables xA according to
their marginals. For this one may sample x∗i ∼ b
p(xi ) or select xi with maximum marginal x∗i =
p(xi ). Sum-product BP is then applied to the reduced factor-graph in which fI (xI ) is replaced
argxi max b
by fI (xI )1(xI∩A = x∗I∩A ). This process, called sum-product BP-guided-decimation (BP-dec), is
repeated to obtain a complete joint assignment x∗ .
However, using BP-guided-decimation is solving a more difficult problem of marginalization. In fact,
in equation (1.22) we showed how using decimation one may estimate the partition function (which is
a #P problem). This suggests that decimation may not be the most efficient approach to solving NPcomplete CSPs. Here, instead we consider using the perturbed belief propagation (section 2.6.3) to
sample from the set of solutions, where the semiring used by perturbed BP is the same as sum-product
BP-dec.
To better understand warning propagation, sum-product BP-dec, and perturbed BP, when applied
to CSPs, consider the following examples.
Example 3.0.3. Here we apply three different message passing methods to solve the simple 3-SAT
example of figure 3.1.
77
(I) Warning Propagation:
We use the max-product semiring ({0, 1}, max, prod) version of warning propagation for this example.
As figure 3.1 suggests, the set of solutions S clusters into two subsets
{{true, true, true}} and {{false, false, false}, {false, false, true} . Here, each of the clusters
is a fixed point for WP – e.g., the cluster with two solutions corresponds to the following fixed point
b
pi→A (true) = b
p(xi = true) = 0
∀A ∈ ∂i
b
pi→A (false) = b
p(xi = false) = 1
b
pj→A (true) = b
p(xj = true) = 0
b
pj→A (false) = b
p(xj = false) = 1
∀A ∈ ∂j
b
pk→A (true) = b
p(xk = true) = 1
b
pk→A (false) = b
p(xk = false) = 1
∀A ∈ ∂k
where the messages indicate the allowed assignments within this particular cluster of solutions. Depending on the initialization, WP messages may converge to any of its fixed points that also include
the trivial cluster, where all (alternatively none) of the assignments are allowed.
(II) BP-dec:
Applying BP to this 3-SAT problem (starting from uniform messages) takes 20 iterations to converge –
i.e., for the maximum change in the marginals to be below = 10−9 . Here the message, b
pI→i (xi ), from
fI to xi is:
b
pI→1 (xi )
∝
X
pj→I (xj ) b
pk→I (xk )
fI (x1,2,3 ) b
xj,k
Similarly, the message in the opposite direction, b
pk→I (xi ) is
b
pi→I (xi )
∝
b
pJ→i (xi ) b
pK→i (xi ) b
pL→i (xi ) b
pH→i (xi )
Here BP gives us the following approximate marginals: b
p(xi = true) = b
p(xj = true) = .319
and b
p(xk = true) = .522. From the set of solutions, we know that the correct marginals are p(xi =
true) = p(xj = true) = 1/3 and p(xk = true) = 2/3. The error of BP is caused by influential loops
in the factor-graph of figure 3.1(b). Here the error is rather small; it can be arbitrarily large in some
instances; sometimes it ca prevent converging at all.
By fixing the value of xi to false, the SAT problem of equation (3.1) collapses to:
SAT (x{j,k} | xi = false) = (¬xj ∨ xk ) ∧ (¬xj ∨ ¬xk )
(3.2)
78
CHAPTER 3. CONSTRAINT SATISFACTION
BP-dec applies BP again to this reduced problem, which give b
p(xj = true) = .14 (note here that
p(xj = true) = 0) and b
p(xk = true) = 1/2. By fixing xj to false, another round of decimation yields
a solution x∗ = {false, false, true}.
(III) Perturbed Belief Propagation:
Perturbed BP can find a solution in T = 4 iterations (see algorithm 2). Our implementation shuffles
the order of updates for variables in each iteration.
In the first iteration, γ = 0, which means updates are the same as that of sum-product BP. In
the second iteration, the order of updates is xj , xk , xi and γ = 1/3. At the end of this iteration
b
p(t=2) (xj = true) = .38. Perturbed BP then samples x̂j = false from this marginal. This sample
influences the outgoing message according to the perturbed BP update equation (2.84), which in turn
influences the beliefs for xi and xk . At the end of this iteration b
p(t=2) (xi = true) = .20 and b
p(t=2) (xk =
true) = .53. At the final iteration γ = 1 and the order of updates is xi , xj and xk . At this point
b
p(t=3) (xi = true) = .07 and the sample x̂i = false. This means the outgoing message is deterministic
(i.e., b
pi→A (false) = 1 and b
pi→A (true) = 0, for all A ∈ ∂i). This choice propagates to select x̂j =
(t=3)
false. Finally b
p
(xk = true) = b
p(t=3) (xk = true) = .5, which correctly shows that both choices
for x̂k produce a solution.
To compare the performance of sum-product BP-dec and Perturbed BP on general CSPs, we considered all CSP instances from XCSP repository [189, 270], that do not include global constraints or
complex domains. All instances with intensive constraints (i.e., functional form) were converted into
extensive format for explicit representation using dense factors. We further removed instances containing constraints with more that 106 enteries in their tabular form. We also discarded instances that
collectively had more than 108 enteries in the dense tabular form of their constraints.1
Figure 3.2(a,b) compares the time and iterations of BP-dec and Perturbed BP for successful attempts
where both methods satisfied an instance.
2
Overall Perturbed BP, with 284 solved instances, is more successful than BP-dec with 253 successful
runs. On the other hand, the average number of iterations for successful instances of BP-dec is 41, 284,
compared to 133 iterations for Perturbed BP. This makes Perturbed BP 300 times more efficient than
BP-dec.
1
3
Since our implementation represents all factors in a dense tabular form, we had to remove many instances because of
their large factor size. We anticipate that Perturbed BP and BP-dec could probably solve many of these instances using
a sparse representation of factors.
2
We used a convergence threshold of = .001 for BP and terminated if the threshold was not reached after T =
10 × 210 = 10, 240 iterations. To perform decimation, we sort the variables according to their bias and fix ρ fraction of
the most biased variables in each iteration of decimation. This fraction, ρ, was initially set to 100%, and it was divided
by 2 each time BP-dec failed on the same instance. BP-dec was repeatedly applied using the reduced ρ, at most 10 times,
unless a solution was reached – i.e., ρ = .1% at final attempt.
For Perturbed BP, T = 10 at the starting attempt, which was increased by a factor of 2 in case of failure. This was
repeated at most 10 times which means Perturbed BP used T = 10, 240 at its final attempt. Note that Perturbed BP at
most uses the same number of iterations as the maximum iterations per single iteration of decimation in BP-dec.
3
We also ran BP-dec on all the benchmarks with maximum number of iterations set to T = 1000 and T = 100
3.1. PHASE TRANSITIONS IN RANDOM CSPS
79
103
102
101
103
Perturbed BP (time)
Perturbed BP (iters)
104
102
101
100
10-1
10-2
100 0
10
101
102
104
103
BP-dec (iters)
105
106
107
10-3
10-4 -4
10
10-3
10-2
10-1
100 101
BP-dec (time)
102
103
104
105
Figure 3.2: Comparison of number of iterations (left) and time (right) used by BP-dec and Perturbed BP in
benchmark instances where both methods found satisfying assignments.
3.1
Phase transitions in random CSPs
Random CSP (rCSP) instances have been extensively used in order to study the properties of combinatorial problems [1, 103, 179, 215] as well as in analysis and design of algorithms [213, 277].
Studies of rCSP, as a critical phenomena, focus on the geometry of the solution space as a function
of the problem’s difficulty, where rigorous [2, 70] and non-rigorous [209, 210] analyses have confirmed
the same geometric picture.
When working with large random instances, a scalar α associated with a problem instance, a.k.a. control parameter – e.g., the clause to variable ratio in SAT– can characterize that instance’s difficulty
(i.e., larger control parameter corresponds to a more difficult instance) and in many situations it characterizes a sharp transition from satisfiability to unsatisfiability [59].
Example 3.1.1. Random K-satisfiability Random K-SAT instance with N variables and M = αN
constraints are generated by selecting K variables at random for each constraint. Each constraint is set
to zero (i.e., unsatisfied) for a single random assignment (out of 2K ). Here α is the control parameter.
Example 3.1.2. Random K-coloring The control parameter for a random K-COL instances with
N variables and M constraints is its average degree α =
2M
N .
We consider Erdős-Rény random graphs
and generate a random instance by sequentially selecting two distinct variables out of N at random to
generate each of M edges. For large N , this is equivalent to selecting each possible factor with a fixed
probability, which means the nodes have Poisson degree distribution Pr(|∂i| = d) ∝ e−α αd .
iterations. This reduced the number of satisfied instances to 249 for T = 1000 and 247 for T = 100, but also reduced the
average number of iterations to 1570 and 562 respectively, which are still several folds more expensive than Perturbed BP.
see Appendix xyz for more details on these results.
80
CHAPTER 3. CONSTRAINT SATISFACTION
While there are tight bounds for some problems [3], finding the exact location of this transition
for different CSPs is still an open problem. Besides transition to unsatisfiability, these analyses have
revealed several other (phase) transitions [179]. Figure 3.3(a)-(c) shows how the geometry of the set of
solutions changes by increasing the control parameter.
Here we enumerate various phases of the problem for increasing values of the control parameter: (a)
In the so-called Replica Symmetric (RS) phase, the symmetries of the set of solutions (a.k.a. ground
states) reflect the trivial symmetries of problem wrt variable domains. For example, for K-COL the set
of solutions is symmetric wrt swapping all red and blue assignment. In this regime, the set of solutions
form a giant cluster (i.e., a set of neighboring solutions), where two solutions are considered neighbors
when their Hamming distance is one [2] (or non-divergent with number of variables [210]. Local search
methods (e.g., [277]) and BP-dec can often efficiently solve random CSPs that belong to this phase.
Figure 3.3: A 2-dimensional schematic view of how the set of solutions of CSP varies as we increase the control
parameter α from (left) replica symmetric phase to (middle) clustering phase to (right) condensation phase. Here
small circles represent solutions and the bigger circles represent clusters of solutions. Note that this view is very
simplistic in many ways – e.g., the total number of solutions and the size of clusters should generally decrease
from left to right.
(b) In clustering or dynamical transition (1dRSB4 ), the set of solutions decomposes into an
exponential number of distant clusters. Here two clusters are distant if the Hamming distance between
their respective members is divergent (e.g., linear) in the number of variables. (c) In the condensation
phase transition (1sRSB5 ), the set of solutions condenses into a few dominant clusters. Dominant
clusters have roughly the same number of solutions and they collectively contain almost all of the
solutions. While SP can be used even within the condensation phase, BP usually fails to converge
in this regime. However each cluster of solutions in the clustering and condensation phase is a valid
fixed-point of BP. (d) A rigidity transition (not included in figure 3.3) identifies a phase in which a
finite portion of variables are fixed within dominant clusters. This transition triggers an exponential
decrease in the total number of solutions, which leads to (e) unsatisfiability transition.6
This rough
4
1st order dynamical RSB. Symmetry Breaking is a general term indicating a phenomenon during which a system is
breaking the symmetry that governs its behaviour by selecting a particular branch. The term Replica Symmetry Breaking
(RSB) originates from the technique –i.e., Replica trick ([211])– that was first used to analyze this setting. According to
RSB, the trivial symmetries of the problem do not characterize the clusters of solution.
5
1st order static RSB.
6
In some problems, the rigidity transition occurs before condensation transition.
3.2. REVISITING SURVEY PROPAGATION
Figure 3.4: This schematic view demonstrates the clustering during condensation phase. Here assume x and y
axes correspond to xi and xj . Considering the whole space of assignments, xi and xj are highly correlated. The
formation of this correlation between distant variables on a factor-graph breaks BP. Now assume that Perturbed
BP messages are focused on the largest shaded ellipse. In this case the correlation is significantly reduced.
picture summarizes first order Replica Symmetry Breaking’s (1RSB) basic assumptions [208].
3.1.1
Pitfalls of decimation
Previously we gave an argument against decimation, based on the complexity of marginalization and
integration. Some recent analyses draw similarly negative conclusions on the effect of decimation [71,
218, 261]. The general picture is that at some point during the decimation process, variables form
long-range correlations such that fixing one variable may imply an assignment for a portion of variables
that form a loop, potentially leading to contradictions. Alternatively the same long-range correlations
result in BP’s lack of convergence and error in marginals that may lead to unsatisfying assignments.
Perturbed BP avoids the pitfalls of BP-dec in two ways: (I) Since many configurations have non-zero
probability until the final iteration, perturbed BP can avoid contradictions by adapting to the most
recent choices. This is in contrast to decimation in which variables are fixed once and are unable to
change afterwards. Some backtracking schemes [240] attempt to fix this problem with decimation. (II)
We speculate that simultaneous bias of all messages towards sub-regions, prevents the formation of
long-range correlations between variables that breaks BP in 1sRSB; see figure 3.4.
3.2
Revisiting survey propagation
SP is studied on random (hyper) graphs representing CSPs at thermodynamic limit (i.e., as N → ∞).
Large random graphs are locally tree-like, which means the length of short loops are typically in the
order of log(N ) [145]. This ensures that, in the absence of long-range correlations, BP is asymptotically
exact, as the set of messages incoming to each node or factor are almost independent. Although
BP messages remain uncorrelated until the condensation transition [179], the BP equations do not
completely characterize the set of solutions after the clustering transition. This inadequacy is indicated
81
82
CHAPTER 3. CONSTRAINT SATISFACTION
by the existence of a set of several valid fixed points (rather than a unique fixed-point) for WP as an
instance of BP. For a better intuition, consider the cartoons of figure 3.3(middle) and (right). During
the clustering phase (middle), xi and xj (corresponding to the x and y axes) are not highly correlated,
but they become correlated during and after condensation (right). This correlation between variables
that are far apart in the factor-graph results in correlation between BP messages. This is because it
implies that even if loops are long, they remain influential. This violates BP’s assumption that messages
are uncorrelated, which results in BP’s failure in this regime.
This is where survey propagation comes into the picture in solving CSPs. Going back to our algebraic
notation for SP, using counting SP with warning propagation semiring ({0, 1}, max, prod) as the initial
semiring and sum-product (R≥0 , sum, prod) as the SP semiring, is computationally tractable. This is
because Y ∗ = {0, 1} in the initial semiring is finite, and therefore each message can have finite number
of 2|Xi | values
b
pi→I (xi ), b
pI→i (.) ∈ (0, . . . , 0), (0, . . . , 0, 1), (0, . . . , 1, 0), . . . , (1, . . . , 0), . . . , (1, . . . , 1)
∀i, I ∈ ∂i
|Xi |
This means each SP message is a distribution over these possibilities Si→I (b
pi→I ) ∈ R2
. However
since (0, . . . , 0) indicates an unfeasible case, where no assignment is allowed, we explicitly ignore it in
SP message updates. This gives us the following update equations and marginals for SP when applied
to CSPs
Si→I (b
pi→I ) ∝
X
b
p∂i\I→i
SI→i (b
pI→i ) ∝
Y
1 b
pi→I (.) =
X
J∈∂i\I
1 b
pI→i (.) =
X
S(b
pi ) =
b
p∂i→i
_
1 (b
pi (.) =
Y
fI (xI )
SI→i ((0, . . . , 0)) = 0
Y
SJ→i (b
pJ→i )
b
pj→I (.)
j∈∂I\i
Y
b
pI→i (.))
I∈∂i
Y
∀i, I ∈ ∂i
(3.3)
J∈∂i\I
x\i
b
p∂I\i→I
X
b
pJ→i (.)
Y
Sj→I (b
pj→I )
(3.4)
j∈∂I\i
SI→i (b
pI→i )
(3.5)
I∈∂i
and Si→I ((0, . . . , 0)) = 0
Example 3.2.1. Consider the SP message Si→I (b
pi→I ) in factor-graph of figure 3.1. Here the summation
in equation (3.3) is over all possible combinations of max-product BP messages b
pJ→ib
pK→i ,b
pL→i ,b
pH→i .
Since each of these messages can assume one of the three valid values – e.g., b
pJ→i (xi ) ∈ {(0, 1), (1, 0), (1, 1)}
4
– for each particular assignment of b
pi→I , a total of 3 possible combinations are enumerated in the summations of equation (3.3). However only the combinations that form a valid max-product message
update have non-zero contribution in calculating Si→I .
3.2.1
Flavours of SP-guided decimation
The SP-marginal over WP marginals (equation (3.5)) also implies a distribution S(xi ) over the original
domain (see equation (2.87)). Similar to BP-dec we can use either the implicit marginals of equa-
3.2. REVISITING SURVEY PROPAGATION
83
tion (2.87) or the SP marginals of equation (3.5) to perform decimation. In the former case, which we
call SP-dec(S) we select x∗i = argxi max S(xi ) during decimation, and in the later case, which we call
SP-dec(C), we clamp b
p∗i = argbpi max S(b
pi ). This means all the outgoing messages from this variable
node in the factor-graph are clamped in the same way – i.e., Si→I (b
pi→I ) = b
p∗i ∀I ∈ ∂i.
In the first case, SP-dec(S), we expect a single assignment x∗ , while for SP-dec(C) at the end of
decimation we should obtain a cluster of solutions, where a subset of assignments is allowed for each
xi . However, during the decimation process (in both SP-dec(S) and SP-dec(C) ), usually after fixing a
subset of variables, SP marginals, S(xi ), become close to uniform, indicating that clusters of solution
have no preference over particular assignment of the remaining variables. The same happens when
we apply SP to random instances in RS phase (figure 3.3(left)). At this point (a.k.a. paramagnetic
phase) solutions form a giant cluster and a local search method or BP-dec can often efficiently find an
assignment to the variables that are not yet fixed by decimation.
The original decimation procedure for K-SAT [48] corresponds to SP-dec(S). SP-dec(C) for CSP
with Boolean variables is only slightly different, as SP-dec(C) can choose to fix a cluster to b
pi = (1, 1) in
addition to the options of b
pi = (1, 0) and b
pi = (0, 1) (corresponding to xi = 0 and xi = 1 respectively),
available to SP-dec(S). However, for larger domains (e.g., K-COL), SP-dec(C) has a clear advantage.
For example, in 3-COL, SP-dec(C) may choose to fix a variable to b
pi→I = (0, 1, 1) (i.e., the first color is
not allowed) while SP-dec(S) can only choose between b
pi ∈ {(0, 0, 1), (0, 1, 0), (1, 0, 0)}. This significant
difference is also reflected in their comparative success-rate on K-COL.
3.2.2
7
(See section 3.3)
Computational Complexity
The computational complexity of each SP update of equation (3.4) is O(2|Xi | −1)|∂I| as for each particular
value b
pi→I , SP needs to consider every combination of incoming messages, each of which can take 2|Xi |
values (minus the empty set). Similarly, using a naive approach the cost of update of equation (3.3)
is O(2|Xi | − 1)|∂i| . However by considering incoming messages one at a time, we can perform the same
exact update in O(|∂i| 22|Xi | ). In comparison to the cost of BP updates (i.e., O(|∂i| |Xi |) and O(|XI |) for
two types of message update; see section 2.1), we see that SP updates are substantially more expensive
for large domains |Xi | and higher order factors with large |∂I|.
3.2.3
Perturbed survey propagation for CSP
Similar to SP, we use perturbed SP with ({0, 1}, max, prod) as the first semiring and (R, sum, prod) as
the second semiring. Since perturbed SP seeks a single assignment, rather than a cluster of solutions,
it can find satisfying solutions to paramagnetic instances. This is in contrast to SP-dec, which in
paramagnetic cases returns a trivial WP fixed point in which all assignment are allowed. This means as
opposed to SP-dec, which is mostly applied to random CSPs in the clustering and condensation phase,
7
Previous applications of SP-dec to K-COL by [49] used a heuristic for decimation that is similar SP-dec (C).
84
CHAPTER 3. CONSTRAINT SATISFACTION
perturbed SP can be used to solve non-random and also random instances in RS phase.
To demonstrate this, we applied perturbed SP to benchmark CSP instances of figure 3.2, in which
the maximum number of elements in the factor was less than 10.8 Here perturbed SP solved 80 instances
out of 202 cases, in comparison to perturbed BP that solved 78 instances, making perturbed SP slightly
better, also in solving real-world problems.
3.3
Satisfiability and coloring
In examples 3.1.1 and 3.1.2, we introduced the random procedures that are often used to produce
instances of K-satisfiability and K-coloring problems.
Here we report the results on K-SAT for K ∈ {3, 4} and K-COL for K ∈ {3, 4, 9}. We used the
procedures to produce 100 random instances with N = 5, 000 variables for each control parameter α and
here report the probability of finding a satisfying assignment for different methods – i.e., the portion of
100 instances that were satisfied by each method.9
Figure 3.5(first row) visualizes the success rate of different methods on 3-SAT (right) and 3-COL
(left). figure 3.5(second row) reports the number of variables that are fixed by SP-dec(C) and (S)
before calling BP-dec as local search. The third row shows the average amount of time that is used
to find a satisfying solution. This does not include the failed attempts. For SP-dec variations, this
time includes the time used by local search. The final row of figure 3.5 shows the number of iterations
used by each method at each level of difficulty over the successful instances. Note that this does not
include the iterations of local search for SP-dec variations. Here the area of each disk is proportional
to the frequency of satisfied instances with particular number of iterations for each control parameter
and inference method10 .
Here we make the following observations:
(I) Perturbed BP is much more effective than BP-dec, while remaining ten to hundreds of time
more efficient. (II) As the control parameter grows larger, the chance of requiring more iterations to
satisfy the instance increases for all methods. (III) Although computationally very inefficient, BP-dec
is able to find solutions for instances with larger control parameter than suggested by previous results
(e.g., [208]). (IV) For many instances where SP-dec(C) and (S) use few iterations, the variables are
8
The number of iterations and other settings for perturbed SP were identical to the ones used to compare BP-dec and
perturbed BP.
9
For coloring instances, to help decimation, we break the initial symmetry of the problem by fixing a single variable
to an arbitrary value. For BP-dec and SP-dec, we use a convergence threshold of = .001 and fix ρ = 1% of variables
per iteration of decimation. Perturbed BP and Perturbed SP use T = 1000 iterations. Decimation-based methods use a
maximum of T = 1000 iterations per iteration of decimation. If any of the methods failed to find a solution in the first
attempt, T was increased by a factor of 4 at most 3 times – i.e., in the final attempt T = 64, 000. To avoid blow-up in
run-time, for BP-dec and SP-dec, only the maximum iteration, T , during the first iteration of decimation, was increased
(this is similar to the setting of [48] for SP-dec). For both variations of SP-dec (see section 3.2.1) after each decimation
step, if maxi,xi b
p(xi ) − |X1i | < .01 (i.e., marginals are close to uniform) we consider the instance para-magnetic, and run
BP-dec (with T = 1000, = .001 and ρ = 1%) on the simplified instance.
10
The number of iterations are rounded to the closest power of two.
3.3. SATISFIABILITY AND COLORING
85
100
100
3-SAT
3-COL
80
0
BP-dec
SP-dec (C)
SP-dec (S)
Perturbed BP
Perturbed SP(S)
4000
2000
0
5000
2000
0
20000
3-SAT
3-COL
5000
15000
3000
2000
satisfiability
rigidity
satisfiability
4000
10000
5000
1000
0
0
5
105
104
satisfiability
satisfiability
iters when successful
3000
1000
1000
10
40
4000
rigidity
satisfiability
3000
60
20
3-SAT
3-COL
0
time when successful (sec)
BP-dec
SP-dec (C)
SP-dec (S)
Perturbed BP
Perturbed SP(S)
satisfiability
40
satisfiability
60
20
avg. nontrivial fixes
rigidity
satisfiability
success rate
80
103
102
104
103
BP-dec
SP-dec (C)
Perturbed BP
Perturbed SP(S)
4.1
4.0
4.2
4.4
4.3
avg. node degree
4.5
4.6
4.7
4.10
4.15
4.20
clause to variable ratio
BP-dec
SP-dec (S)
Perturbed BP
Perturbed SP(S)
102
4.25
4.30
Figure 3.5: (first row) Success-rate of different methods for 3-COL and 3-SAT for various control parameters.
(second row) The average number of variables (out of N = 5000) that are fixed using SP-dec (C) and (S) before
calling local search, averaged over 100 instances. (third row) The average amount of time (in seconds) used
by the successful setting of each method to find a satisfying solution. For SP-dec(C) and (S) this includes the
time used by local search. (forth row) The number of iterations used by different methods at different control
parameters, when the method was successful at finding a solution. The number of iterations for each of 100
random instances is rounded to the closest power of 2. This does not include the iterations used by local search
after SP-dec.
86
CHAPTER 3. CONSTRAINT SATISFACTION
fixed to a trivial cluster b
pi = (1, 1, . . . , 1), in which all assignments are allowed. This is particularly
pronounced for 3-COL. For instances in which non-trivial fixes are zero, the success rate is solely due to
local search (i.e., BP-dec). (V) While SP-dec(C) and SP-dec(S) have a similar performance for 3-SAT,
SP-dec(C) significantly outperforms SP-dec(S) for 3-COL.
Table 3.1 reports the success-rate as well as the average of total iterations in the successful attempts
of each method, where the number of iterations for SP-dec(C) and (S) is the sum of iterations used
by the method plus the iterations of the following BP-dec. Here we observe that perturbed BP can
solve most of easier instances using only T = 1000 iterations (e.g., see perturb BP’s result for 3-SAT at
α = 4., 3-COL at α = 4.2 and 9-COL at α = 33.4). The results also show that most difficult instances
(that require more time/iterations) for each method approximately correspond to the control parameter
for which half of the instances are satisfied. Larger control parameters usually result in early failure in
satisfiability.
Table 3.1 suggests that, as we speculated in section 3.2, SP-dec(C) is in general preferable to SPdec(S), in particular when applied to the coloring problem. The most important advantage of Perturbed
BP over SP-dec and Perturbed SP is that it can be applied to instances with large factor cardinality
(e.g., 10-SAT) and variable domains (e.g., 9-COL). For example for 9-COL, the cardinality of each SP
message is 29 = 512, which makes SP-dec and Perturbed SP impractical. Here BP-dec is not even able
to solve a single instance around the dynamical transition (as low as α = 33.4) while perturbed BP
satisfies all instances up to α = 34.1.11
3.4
Clique-cover problem
The K-clique-cover C = {C1 , . . . , CK } for a graph G = (V, E) is a partitioning of V to at most K
cliques – i.e., ∀i, j, k
i, j ∈ Ck ⇒ (i, j) ∈ E.
NP-completeness of K-clique-cover can be proved by reduction from K-coloring [159]: A K-cliquecover for G 0 , the complement of G (i.e., G 0 = (V, E 0 = {(i, j) | (i, j) ∈
/ E})), is a K-coloring for G, where
all the nodes in the same clique of G 0 are allowed to have the same color in G.
The relation between K-clique-cover and K-coloring extends to their factor-graphs. While in Kcoloring, factors f{i,j} (xi , xj ) = 1(xi 6= xj ) ∀(i, j) ∈ E ensure that the connected nodes have different
colors, for k-clique-cover factors f{i,j} (xi , xj ) = 1(xi 6= xj ) ∀(i, j) ∈
/ E ensure that nodes that are not
connected can not belong to the same clique. Here xi ∈ {1, . . . , K} represents the clique of node i.
The factors, f{i,j} (xi , xj ) = 1(xi 6= xj ), in both K-clique-cover and K-coloirng are inverse Potts
factors that allow efficient O(K) calculation (see section 2.2.1). Using E(i, ·) = {(i, j) ∈ E} to denote
the set of edges adjacent to node i, the following claim states the complexity of BP updates.
11
Note that for 9-COL condensation transition happens after rigidity transition. So if we were able to find solutions
after rigidity, it would have implied that condensation transition marks the onset of difficulty. However, this did not occur
and similar to all other cases, Perturbed BP failed before rigidity transition.
3.4. CLIQUE-COVER PROBLEM
87
Table 3.1: Comparison of different methods on {3, 4}-SAT and {3, 4, 9}-COL. For each method the success-rate
and the average number of iterations (including local search) on successful attempts are reported. The approximate
location of phase transitions are from [219, 330] .
3.86
3-SAT
4-SAT
success rate
avg. iters.
success rate
avg. iters.
success rate
avg. iters.
success rate
avg. iters.
avg. iters.
Perturbed SP
dynamical and condensation transition
85405
99%
102800
100%
96475
100%
1301
100%
1211
100%
104147
83%
118852
100%
111754
96%
5643
95%
1121
100%
4.2
93904
28%
118288
65%
113910
64%
19227
53%
3415
87%
4.22
100609
12%
112910
33%
114303
36%
22430
28%
8413
69%
4.23
123318
5%
109659
36%
107783
36%
18438
16%
9173
58%
4.24
165710
1%
126794
23%
118284
19%
29715
7%
10147
41%
4.25
N/A
0%
123703
9%
110584
8%
64001
1%
14501
18%
4.26
37396
1%
83231
6%
106363
5%
32001
3%
22274
11%
4.268
satisfiability transition
9.38
dynamical transition
9.547
condensation transition
9.73
134368
8%
119483
32%
120353
35%
25001
43%
11142
86%
9.75
168633
5%
115506
15%
96391
21%
36668
27%
9783
68%
N/A
0%
83720
9%
139412
7%
34001
12%
11876
37%
9.88
rigidity transition
9.931
satisfiability transition
dynamical and condensation transition
4.2
24148
93%
25066
94%
24634
94%
1511
100%
1151
100%
4.4
51590
95%
52684
89%
54587
93%
1691
100%
1421
100%
4.52
61109
20%
68189
63%
54736
1%
7705
98%
2134
98%
4.56
N/A
0%
63980
32%
13317
1%
28047
65%
3607
99%
4.6
N/A
0%
74550
2%
N/A
0%
16001
1%
18075
81%
4.63
N/A
0%
N/A
0%
N/A
0%
48001
3%
29270
26%
N/A
0%
N/A
0%
N/A
0%
40001
2%
88%
71214
93%
1931
100%
1331
4.66
4.66
rigidity transition
N/A
0%
4.687
satisfiability transition
8.353
dynamical transition
8.4
9-COL
Perturbed BP
4.15
4
4-COL
SP-dec(S)
4.1
9.78
3-COL
SP-dec(C)
success rate
ctrl param α
Problem
BP-dec
8.46
64207
92%
72359
100%
dynamical transition
8.55
77618
13%
60802
13%
62876
9%
3041
100%
5577
100%
8.7
N/A
0%
N/A
0%
N/A
0%
50287
14%
N/A
0%
8.83
rigidity transition
8.901
satisfiability transition
33.45
dynamical transition
33.4
N/A
0%
N/A
N/A
N/A
N/A
1061
100%
N/A
N/A
33.9
N/A
0%
N/A
N/A
N/A
N/A
3701
100%
N/A
N/A
34.1
N/A
0%
N/A
N/A
N/A
N/A
12243
100%
N/A
N/A
34.5
N/A
0%
N/A
N/A
N/A
N/A
48001
6%
N/A
N/A
35.0
N/A
0%
N/A
N/A
N/A
N/A
N/A
0%
N/A
N/A
39.87
rigidity transition
43.08
condensation transition
43.37
satisfiability transition
88
CHAPTER 3. CONSTRAINT SATISFACTION
Claim 3.4.1. Each iteration of BP with variable-synchronous message update for K-clique-cover factorP
graph is O(K(N 2 − |E|)), while asynchronous message update is O(K i∈V (N − |E(i, ·)|)2 ).
Proof. Here the complexity of calculating factor-to-variable message (b
p{i,j}→i ) is O(K). Since there
are N 2 − E factors (one for each edge in G 0 ) the total cost of factor-to-variable messages becomes
O(K(N 2 − |E|)).
The time complexity of each variable-to-factor message (b
pi→{i,j} ) is O(K|∆i|), where ∆i, Markov
blanket of i in the factor-graph, is the set of nodes in V that are not adjacent to i in G – i.e., |∆i| =
N − E(i, ·). Using variable-synchronous update the total cost of variable-to-factor messages becomes
P
O(K i∈V N − |E(i, ·)|) = O(K(N 2 − 2|E|)). This means the cost of all messages in BP update is in
the order of O(K(N 2 − |E|)).
However, using asynchronous update, at each node i, we have to calculate N − |E(i, ·)| messages.
Since each of them is O(N − |E(i, ·)|), the total cost of variable-to-factor dominates the cost of BP
P
update which is O(K i∈V (N − |E(i, ·)|)2 ).
Our experimental results for K-clique-cover are within the context of a binary-search scheme, as the
sum-product reduction of the min-max formulation of K-clustering.
3.5
Dominating set and set cover
The K-dominating set of graph G = (V, E) is a subset of nodes D ⊆ V of size |D| = K such that any
node in V \ D is adjacent to at least one member of D – i.e., ∀i ∈ V \ D
∃j ∈ D s.t. (i, j) ∈ E. The
dominating set problem is NP-complete [107] and has simple reductions to and from set cover problem
[158]. As we see, the factor-graph formulations of these problems are also closely related.
Figure 3.6: (left) an induced 2-set-cover problem and the solution D = {i, k}. (right) The factor-graph
representation of the same problem, where leader factors are grey squares, consistency factors are in black and
the K-of-N factor is in white.
3.5. DOMINATING SET AND SET COVER
89
Given universe set V and a set of its subsets S = {V1 , . . . , VM } s.t. Vm ⊆ V, we say C ⊆ S covers
S
V iff each member of V is present in at least one member of C –i.e., Vm ∈C Vm = V. Now we consider
a natural set-cover problem induced by any directed graph.
for each node i ∈ V, define a subset Vi = {j ∈ V
Given a directed-graph G = (V, E),
| (j, i) ∈ E} as the set of all nodes that are
connected to i. Let S = {V1 , . . . , VN } denote all such subsets. An induced K-set-cover of G is a set
C ⊆ S of size K that covers V. Equivalently, The induced K-set-cover of G = (V, E) is a subset of
vertices D ⊆ V, with |D| = K, such that every node not in D is connected to at least one node in D.
For example in figure 3.6(left), S = {{n, m, i}, {i, j}, {j, k, l}, {l, m, n}, {n}} and its induced solution
C = {{n, m, i}, {j, k, l}} is indicated by grey nodes D = {i, k}.
If we consider an undirected graph G as a directed graph with edges in both directions, then Kdominating set of G is equivalent to an induced K-set-cover problem on G. Moreover given any K-setcover problem instance S = {V1 , . . . , Vm }, we can construct a directed graph G such that the “induced”
S
K-set-cover on G is equivalent to the given K-set-cover problem. For this, let V = ( Vm ∈S Vm ) ∪
{u1 , . . . , uM } be the collection of nodes in S plus one node um per each subset Vm ∈ S. Now define the
directed edges in E to connect every i ∈ Vm to its representative um . Moreover connect all representatives
to each other in both directions – i.e., E = {(i, um ) | ∀m, i ∈ Vm } ∪ {(um , um0 ) | ∀m 6= m0 }. It is easy
to show that the induced K-set-cover on this directed graph defines a set-cover for S.
3.5.1
Factor-graph and complexity
For both problems we have one variable per edge xi:j ∈ {0, 1} ∀(i, j) ∈ E. Note that the G for induced
K-set-cover problem is a directed graph, while the G for the K-dominating-set is undirected. This is
the only difference that affects the factor-graph representation of these two problems. Here, xi:j = 1
indicates that node j ∈ D and node i is associated with node j. Three types of constraint factors ensure
the assignments to xi:j define a valid solution to K-dominating-set and induced K-set-cover:
• Leader factors ensure that each node i is associated with at least one node j (where j = i is
admissible). Let E + (i, ·) = {(i, j) ∈ E} ∪ {(i, i)} be the set of edges leaving node i plus (i, i). Then
fE + (i,·) (xE + (i,·) ) = 1((
X
xi:j ) ≥ 1) ∀i ∈ V
(3.6)
(i,j)∈E + (i,·)
is the leader factor associated with node i.
• Consistency factors ensure that if node j is selected as the leader by node i, node j also selects
itself as leader:
f{i:j,j:j} (xi:j , xj:j ) = 1(xi:j = 0 ∨ xj:j = 0) ∀(i, j) ∈ E
(3.7)
An alternative form of this factor is a high-order factor that allows efficient O(|E(·, i)|) factor-to-
90
CHAPTER 3. CONSTRAINT SATISFACTION
variable update
fE(·,i) (xE(·,i) ) = 1(xi:i = 1 ∨
X
xj:i = 0) ∀i ∈ V
(3.8)
(i,j)∈E(·,i)
• At most K-of-N factor ensures that at most K nodes are selected as leaders (|D| ≤ K):
f{i:i,j:j,...,l:l} (x{i:i,j:j,...,l:l} ) = 1(
X
xi:i ≤ K)
(3.9)
i∈V
Figure 3.6 shows an example of induced K-set-cover problem and its corresponding factor-graph.
In section 2.2.1 we saw that it is possible to calculate sum-product factor-to-variable BP messages for
leader factors in O(|E(i, ·)|) and each at-most-K-of-N factor in O(KN ). This cost for consistency factors
is O(1) for the pairwise and O(|E(·, i)|) for the alternative formulation.
Claim 3.5.1. The time-complexity of message passing for the factor-graph above depending on the
update schedule is O(|E| + KN ) for factor-synchronous (f-sync; see section 2.1.1) update and O(KN 2 +
P
2
2
i∈V |E(i, ·)| + |E(·, i)| ) for asynchronous update.
Proof. We assume the consistency factors are in the higher order form of equation (3.8). Here, each
variable xi:j is connected to at most three factors and therefore the cost of variable-to-factor messages is
P
O(1). If we calculate factor-to-variable messages simultaneously, the cost is O( i∈V |E(i, ·)|) for leader
P
and O( i∈V |E(·, i)|), giving a total of O(|E|). Adding this to the cost of K-of-N factor the total cost
per iteration of BP is O(|E| + KN ).
On the other hand, if we update each factor-to-variable separately, the previous costs are multiplied
P
by |∂I|, which gives O(KN 2 + i∈V |E(i, ·)|2 + |E(·, i)|2 ).
3.6
Clique problem, independent set and sphere packing
Given graph G, the K-clique problem asks whether G contains a clique of size at least K. The Kclique problem is closely related to K-independent-set. Given graph G = (V, E), the K-independent
set problem asks whether V contains a subset of size at least K, s.t. there is no connection between
nodes in D – i.e., ∀i, j ∈ D (i, j) ∈
/ E. The relation between K-clique problem and K-independent-set
is analogous to the connection between K-coloring and K-clique-cover problems: the K-clique problem
on G is equivalent to K-independent-set problem on its complement G 0 .
K-independent-set is in turn equivalent to (N-K)-vertex cover problem. A vertex cover D0 ⊆ V is
a subset of nodes such that each edge is adjacent to at least one vertex in D0 . It is easy to see that D
is an independent set iff D0 = V \ D is a vertex cover. Therefore our solution here for independent set
directly extends to vertex cover.
3.6. CLIQUE PROBLEM, INDEPENDENT SET AND SPHERE PACKING
91
K-packing is a special case of sub-graph isomorphism, which asks whether a graph G1 is a sub-graph
of G2 . For packing, G2 is the main graph (G) and G1 is the complete graph of size K (see section 5.3) Kindependent-set and K-clique problems are also closely related to sphere packing and finding nonlinear
codes. To better motivate these problems, here we start with the problem of K-packing formulated as
min-max problem in (several) factor-graphs. We then show that the sum-product reduction of K-packing
is K-independent-set and K-clique problems. By doing this we simultaneously introduce message passing
solutions to all three problems.
Given a symmetric distance matrix A ∈ RN ×N between N data-points and a number of codewords K, the K-packing problem is to choose a subset of K points such that the minimum distance
Ai,j between any two code-words is maximized. Here we introduce two different factor-graphs such that
min-max inference obtains the K-packing solution.
In order to establish the relation between the min-max problem and the CSPs above, we need the
notion of y-neighbourhood graph
Definition 3.6.1. The y-neighborhood graph for (distance) matrix A ∈ RN ×N is defined as the
graph G(A, y) = (V, E), where V = {1, . . . , N } and E = {(i, j) | Ai,j ≤ y}.
3.6.1
Binary variable factor-graph
Let binary variables x = {x1 , . . . , xN } ∈ {0, 1}N indicate a subset of variables of size K that are selected
as code-words. We define the factors such that the min-max assignment
x∗ = argx min max fI (xI )
I∈F
is the K-packing solution. Define the factor-graph with the following two types of factors:
• K-of-N factor (section 2.2.1)
fN (x) = 1((
X
xi ) = K)
i∈N
ensures that K code-words are selected. Recall that definition of 1(.) depends on the semiring (equation (1.20)). The K-packing problem is defined on the min-max semiring. However, since we plan to
solve the min-max inference by sum-product reductions, it is important to note that the py reduction
of 1(.) for any y is 1(.) as defined for sum-product ring.
• Pairwise factors are only effective if both xi and xj are non-zero
f{i,j} (xi , xj ) = min 1(xi = 0 ∨ xj = 0), max 1(xi = 1 ∧ xj = 1), −Ai,j
where the tabular form is simply
∀i, j
(3.10)
92
CHAPTER 3. CONSTRAINT SATISFACTION
xj
xi
0
1
0
−∞
−∞
1
−∞
−Ai,j
Here the use of −Ai,j is to convert the initial max-min objective to min-max.12 Recall that G(A, y)
defines a graph based on the distance matrix A, s.t. two nodes are connected iff their distance is “not
larger than” y. This means G(−A, −y) corresponds to a graph in which the connected nodes have a
distance of “at least” y.
Proposition 3.6.1. The py -reduction of the K-packing factor-graph above for the distance matrix
A ∈ RN ×N defines a uniform distribution over the cliques of G(−A, −y) of size K.
Proof. Since py (x) is uniform over its domain it is enough to show that:
• Every clique of size K in G(−A, −y) corresponds to a unique assignment x∗ with py (x∗ ) > 0:
Given a clique C ⊆ V of size K in G(−A, −y) = (V, {(i, j) | Ai,j ≥ y}), define x∗ = {xi = ident(i ∈ C) |
i ∈ V}. It is easy to show that py (x∗ ) > 0. For this we need to show that all the constraint factors in
py are satisfied. The K-of-N factor is trivially satisfied as |C| = K. The py -reduction of the pairwise
factor of equation (3.10) becomes
f{i,j} (xi , xj ) = 1(xi = 0 ∨ xj = 0) + 1(xi = 1 ∧ xj = 1)1(−Ai,j ≤ −y)
∀i, j
(3.11)
where we replaced min and max with + and . operators of the sum-product semiring and thresholded
−Ai,j by −y. To see that all pairwise constraint factors are satisfied consider two cases: (I) nodes
i, j ∈ C, and therefore x∗i = x∗j = 1. This also means i and j are connected and definition of G(−A, −y),
this implies Ai,j ≥ y. Therefore the second term in factor above evaluates to one. (II) either i or j are
not in C, therefore the first term evaluates to one. Since both pairwise factors and the K-of-N factors
for x∗ are non-zero py (x∗ ) > 0.
• every assignment x∗ with py (x∗ ) > 0 corresponds to a unique clique of size K in G(−A, −y):
equation (3.11) implies xi = 1 ∧ xj = 1
⇒
Ai,j ≥ y. On the other hand, py (x) > 0 means K-of-N
factor is satisfied and therefore exactly K variables xi are nonzero. Therefore the index of these variables
identifies subset of nodes in G(−A, −y) that are connected (because Ai,j ≥ y), forming a clique.
Claim 3.6.2. The time-complexity of each iteration of sum-product BP for the factor-graph above is
– O(N K + |E|) for variable and factor synchronous update.
– O(N 2 K) for factor-synchronous (f-sync) message update.
– O(N 3 ) for completely asynchronous update.
12
The original objective is max-min because it aims to maximize the minimum distance between any two code-words.
3.6. CLIQUE PROBLEM, INDEPENDENT SET AND SPHERE PACKING
93
Figure 3.7: Using message passing to choose K = 30 out of N = 100 random points in the Euclidean plane to
maximize the minimum pairwise distance (with T = 500 iterations for PBP). Touching circles show the minimum
distance
Proof. The cost of f-sync calculation of factor-to-variable messages for the K-of-N factor is O(N K)
and the cost of factor-to-variable messages for pairwise factors is O(1). Since there are |E| such factor,
this cost evaluates to O(N K + |E|). Since each node is adjacent to |E(i, ·)| other nodes, the variableP
synchronous update of variable-to-factor messages is O( i∈V |E(i, ·)|) = O(|E|), which gives a total
time-complexity of O(N K + |E|).
In asynchronous update, the cost of factor-to-variable messages for the K-of-N factor is O(N 2 K)
as we need to calculate each message separately. Moreover, updating each variable-to-factor message
P
is O(|E(i, ·)), resulting a total of O( i∈V |E(i, ·)|2 ) = O(N 3 ). Since K ≤ N , when all the updates are
asynchronous, this cost subsumes the factor-to-variable cost of O(N 2 K).
A corollary is that the complexity of finding the approximate min-max solution by sum-product
reduction is O((N K + |E|) log(N )) using synchronous update. Figure 3.7 shows an example of the
solution found by message passing for K-packing with Euclidean distance.
3.6.2
Categorical variable factor-graph
Define the K-packing factor-graph as follows:Let x = {x1 , . . . , xK } be the set of K variables where
xi ∈ {1, . . . , N }. For every two distinct points 1 ≤ i < j ≤ K define the factor
f{i,j} (xi , xj ) = max − Axi ,xj , 1(xi 6= xj )
(3.12)
Here each variable represents a code-word and this factor ensures that code-words are distinct. Moreover
if xi and xj are distinct, f{i,j} (xi , xj ) = −Axi ,xj is the distance between the nodes that xi and xj
94
CHAPTER 3. CONSTRAINT SATISFACTION
represent.
The tabular form of this factor is
xj
xi
1
2
···
N −1
N
1
∞
−A1,2
···
−A1,N −1
−A1,N
2
−A2,1
..
.
∞
..
.
···
..
.
−A2,N −1
..
.
−A2,N
..
.
−AN −1,1
−AN −1,2
···
∞
−AN −1,N
−AN,1
−AN,2
···
−AN,N −1
∞
···
N −1
N
The following proposition relates this factor-graph to K-clique problem.
Proposition 3.6.3. The py -reduction of the K-packing factor-graph above for the distance matrix
A ∈ RN ×N defines a uniform distribution over the cliques of G(−A, −y) of size K.
Proof. Since py defines a uniform distribution over its support, it is enough to show that any clique of size
K over G(−A, −y) defines a unique set of assignments all of which have nonzero probability (py (x) > 0)
and any assignment x with py (x) > 0 defines a unique clique of size at least K on G(−A, −y). First
note that the basic difference between G(A, y) and G(−A, −y) is that in the former all nodes that are
connected have a distance of at most y while in the later, all nodes that have a distance of at least y
are connected to each other. Consider the py -reduction of the pairwise factor of equation (3.12)
f{i,j} (xi , xj ) = 1 max(−Axi ,xj , 1(xi 6= xj )) ≤ y
(3.13)
= 1(−Axi ,xj ≤ y ∧ xi 6= xj )
where, basically, we have replaced max from min-max semiring with ⊗ operator of the sum-product
semiring and thresholded Axi ,xj .
• Any clique of size K in G(−A, −y), defines K unique assignments, such that for any such assignment
x∗ , py (x∗ ) > 0:
For a clique C = {c1 , . . . , cK } ⊆ V of size K, define x∗i = cπ(i) , where π : {1, . . . , K} → {1, . . . , K}
is a permutation of nodes in clique C. Since there are K such permutations we may define as many
assignments x∗ . Now consider one such assignment. For every two nodes x∗i and x∗j , since they belong
to the clique C in G(−A, −y), they are connected and Axi ,xj ≥ y. This means that all the pairwise
factors defined by equation (3.13) have non-zero values and therefore py (x) > 0.
• Any assignment x∗ with py (x∗ ) > 0 corresponds to a unique clique of size K in G(−A, −y): Let
C = {x∗1 , . . . , x∗K }. Since py (x∗ ) > 0, all pairwise factors defined by equation (3.13) are non-zero.
Therefore ∀i, j 6= i Ax, xj ≥ y, which means all xi and xj are connected in G(−A, −y), forming a clique
of size K.
3.6. CLIQUE PROBLEM, INDEPENDENT SET AND SPHERE PACKING
95
To put simply, in acquiring the py -reduction, we set the values in the table form above to zero if their
value is less than y and set them to one otherwise. The resulting factor-graph, defines a distribution
py (x), s.t. py (x) > 0 means x defines a clique of size K in a graph G(−A, −y) which connects nodes
with distance larger than y.
Claim 3.6.4. Each iteration of BP for this factor-graph with pairwise factors is O(N 2 K 2 ), for synchronized update and O(N 2 K 3 ) for asynchronous update. Using the sum-product reduction of min-max
inference this suggests a O(N 2 K 2 log(N )) (for sync. update) and O(N 2 K 3 log(N )) (for asynchronous
update) procedure for K-packing problem.
Proof. (Claim 3.6.4) Since the factors are not sparse, the complexity of calculating a single factor-tovariable message is O(|XI |) = O(N 2 ), resulting in O(N 2 K 2 ) cost per iteration of variable-synchronous
update for BP. However if we update each message separately, since each message update costs O(N 2 K),
the total cost of BP per iteration is O(N 2 K 3 )
Since the diversity of pairwise distances is that of elements in A – i.e., |Y| = O(N 2 ) – the general cost
of finding an approximate min-max solution by message passing is O(N 2 K 2 log(N )) for sync. message
update and O(N 2 K 3 log(N )) for async. update.
The py -reduction of our second formulation was first proposed by [252] to find non-linear binary
codes. The authors consider the Hamming distance between all binary vectors of length n (i.e., N = 2n )
to obtain binary codes with known minimum distance y. As we saw, this method is O(N 2 K 2 ) =
O(22n K 2 ) –i.e., grows exponentially in the number of bits n. In the following section, we introduce a
factor-graph formulation specific to categorical variables with Hamming distance whose message passing
complexity is polynomial in n = log(N ). Using this formulation we are able find optimal binary and
ternary codes where both n and y are large.
3.6.3
Efficient Sphere packing with Hamming distance
Our factor-graph defines a distribution over the K binary vectors of length n such that the distance
between every pair of binary vectors is at least y.
13
To better relate this to K-clique problem, consider 2n binary code-words of length n as nodes of a
graph and connect two nodes iff their Hamming distance is at least y. Finding a K-clique in this graph is
equivalent to discovery of so-called nonlinear binary codes – a fundamental problem in information
theory (e.g., see [75, 193]). Here assuming y is an odd number, if at most
y−1
2
digits of a code-word
are corrupted (e.g., in communication), since every pair of codes are at least y digits apart, we can
still recover the uncorrupted code-word. The following is a collection of K = 12 ternary code-words
of length n = 16, obtained using the factor-graph that we discuss in this section, where every pair of
code-words are different in at least y = 11 digits.
13
For convenience we restrict this construction to the case of binary vectors. A similar procedure may be used to find
maximally distanced ternary and q-ary vectors, for arbitrary q.
96
CHAPTER 3. CONSTRAINT SATISFACTION
Figure 3.8: The factor-graph for sphere-packing with Hamming distances, where the distance factors are white
squares and z-factors are in black.
2
1
2
1
0
1
1
2
2
1
2
1
2
1
0
2
1
1
1
2
1
0
0
2
2
1
1
1
0
2
1
0
0
0
1
2
0
1
2
0
2
1
2
2
1
0
2
0
0
0
0
2
1
0
1
2
1
0
1
0
2
1
0
1
2
2
0
2
2
2
1
1
2
0
2
2
1
2
1
2
0
2
0
0
0
2
0
2
0
0
2
1
0
0
0
0
1
1
2
0
0
1
2
2
0
0
1
0
1
2
2
1
2
0
2
0
2
1
0
1
1
2
0
1
1
1
2
0
0
2
1
1
1
1
1
1
0
1
0
0
0
0
1
2
1
0
0
1
0
2
0
0
1
2
0
0
2
2
1
2
1
2
2
1
1
0
2
1
1
2
2
2
2
0
2
1
1
0
1
0
2
2
2
1
0
1
1
2
2
1
0
2
The construction of this factor-graph is more involved than our previous constructions. The basic
idea to avoid the exponential blow up is to have one variable per digit of each code-word (rather than
one variable per code-word). Then for each pair of code-words we define an auxiliary binary vector of
the same length, that indicates if the two code-words are different in each digit. Finally we define an
at-least-y-of-n constraint over each set of auxiliary vectors that ensures every two pair of code-words
are at least different in y digits. Figure 3.8 shows this factor-graph.
More specifically, let x = {x1:· , . . . , xK:· } be a set of binary vectors, where xi:· = {xi:1 , . . . , xi:n }
represents the ith binary vector or code-word. Additionally for each two code-words 1 ≤ i < j ≤ K,
define an auxiliary binary vector z i:j:· = {zi:j:1 , . . . , zi:j:n } of length n.
For each distinct pair of binary vectors xi:· and xj:· , and a particular digit 1 ≤ k ≤ n, the auxiliary
variable is constrained to be zi:j:k = 1 iff xi:k 6= xj:k . Then we define an at-least-y-of-n factor over z i:j:·
for every pair of code-words, to ensure that they differ in at least y digits.
3.6. CLIQUE PROBLEM, INDEPENDENT SET AND SPHERE PACKING
The factors are defined as follows
• Z-factors: For every 1 ≤ i < j ≤ K and 1 ≤ k ≤ n, define
f{i:k,j:k,i:j:k} (xi:k , xj:k , zi:j:k ) = 1
(xi:k 6= xj:k ) ∧ zi:j:k = 1 ∨ (xi:k = xj:k ) ∧ zi:j:k = 0
This factor depends on three binary variables, therefore we can explicitly define its tabular form containing 23 = 8 possible inputs. Here the only difference between binary and ternary (and q-ary) codes
in general is in the tabular form of this factor. For example for ternary codes, the tabular for of this
factor is a 3 × 3 × 2 array (zi:j:k is always binary).
• Distance-factors: For each z i:j:· define at-least-y-of-n factor (section 2.2.1):
fi:j:· (z i:j:· ) = 1(
X
zi:j:k ≥ y)
1≤k≤n
Claim 3.6.5. Each iteration of sum-product BP over this factor-graph is
– O(K 2 ny) for variable and factor synchronous update.
– O(K 2 n2 y) for variable-sync update.
– O(K 3 n + K 2 n2 y) for completely asynchronous BP update.
Proof. (Claim 3.6.5) We first consider the complexity of variable-to-factor updates: Each auxiliary
variable zi:j:k is connected to three factors and each xi:j is connected to O(K) z-factors. Since there are
O(nK), xi:j variables a variable-sync. update of variable-to-factor messages is O(nK 2 ), while async.
update is O(nK 3 ).
Next we consider two possibilities of factor-to-variable updates: We have O(nK 2 ) z-factors, and
factor-to-variable update for each of them is O(1), this cost is subsumed by the minimum variable-tofactor cost. The factor-graph also has O(K 2 ) distance factors, where the cost of each factor-to-variable
update is O(ny). Since |∂I| = K for a distance factor, a factor-sync. update is O(K 2 ny) in total, while
an async. update is O(K 2 n2 y).
Adding the cost of variable-to-factor and factor-to-variable in different scenarios we get the timecomplexities stated in the claim.
The following table reports some optimal binary codes (including codes with large number of bits
n) from [193], recovered using this factor-graph. We used Perturbed BP with variable-sync update and
T = 1000 iterations to find an assignment x∗ , z ∗ with p(x∗ , z ∗ ) > 0.
97
98
CHAPTER 3. CONSTRAINT SATISFACTION
n
3.7
K
y
n
K
y
n
K
y
n
K
y
8
4
5
11
4
7
14
4
9
16
6
9
17
4
11
19
6
11
20
8
11
20
4
13
23
6
13
24
8
13
23
4
15
26
6
15
27
8
15
28
10
15
28
5
16
26
4
17
29
6
17
29
4
19
33
6
19
34
8
19
36
12
19
32
4
21
36
6
21
38
8
21
39
10
21
35
4
23
39
6
23
41
8
23
39
4
25
43
6
23
46
10
25
47
12
25
41
4
27
46
6
27
48
8
27
50
10
27
44
4
29
49
6
29
52
8
29
53
10
29
Optimization variations of CSPs
This section briefly reviews the optimization variations of the CSPs, we have studied so far. The
optimization version of satisfiability is known as max-SAT or maximum weighted SAT, where each
clause has a weight, and the objective is to maximize the weighted sum of satisfied clauses. Here,
simply using factors fI (xI ) : XI → {0, −wI }, where wI is the positive weight of clause I, min-sum BP
will attempt to find the max-SAT solution. Note that here fI is not a “constraint” factor anymore (see
definition 1.3.2). Alternative approaches using variations of energetic survey propagation has also been
used to improve max-SAT results [63, 64].
A less studied optimization variation of satisfiability is
adversarial SAT which corresponds to min-max-sum inference [54].
For minimum coloring – a.k.a. chromatic number – and minimum clique-cover problem, since
the optimal value 1 ≤ K ∗ ≤ Kmax is bounded, by access to an oracle for the decision version (or an
incomplete solver [160] such as message passing), we can use binary search to find the minimum K in
O(τ log(Kmax )) time, where the decision problem has a O(τ ) time-complexity. In particular, since the
chromatic number is bounded by the maximum degree [51], approximating the chromatic number using
binary search and message passing gives a O(log(maxi (|E(i, ·)|)) maxi (|E(i, ·)|) |E|) time procedure.
The same binary-search approach can be used for minimum dominating-set, minimum set-cover,
maximum clique, maximum independent set and minimum vertex cover. However, these optimization
variations also allow a more efficient and direct approach. Note that both min-max variation and
the minimization (or maximization) variations of these problems use binary search. For example both
minimum clique-cover and K-clustering (see section 4.3) can be solved using binary-search over Kclique-cover decision problem. The parameter of interest for binary search is K in the former case, and
y in the later case, where y is a threshold distance that defines connectivity in G (see definition 3.6.1).
However, often both variations (i.e., min-max and minimization or maximization over K) also allow
direct message passing solutions.
For minimum dominating-set and set-cover, we replace the sum-product semiring with min-
3.7. OPTIMIZATION VARIATIONS OF CSPS
99
sum semiring and drop the K-of-N factor. Instead, a local factor fi:i (xi:i ) = −xi:i wi gives a weight to each
node i ∈ V. Here, the min-sum inference seeks a subset of nodes that form a dominating set and have
P
the largest sum of weights i∈D wi . Also note that by changing the semiring to min-sum, the identity
functions 1(.) change accordingly so that the leader factor and consistency constraints remain valid.
This gives an efficient O(|E|) synchronous procedure for minimum set-cover and minimum dominatingset. This problem and the resulting message passing solution are indeed a variation of K-medians and
affinity propagation respectively (see section 4.1).
The same idea applies to maximum clique and maximum independent set: as an alternative
to fixing or maximizing the “size” of a clique or an independent set, we may associate each node with a
weight wi ∀i ∈ V and seek a subset of nodes that form an independent set (or a clique) with maximum
weight. Sanghavi et al. [274] study the max-product message passing solution to this problem and
its relation to its LP-relaxation. In particular they show that starting from uniform messages, if BP
converges, it finds the solution to LP relaxation.
Here we review their factor-graph for maximum independent set using min-sum inference. Let
x = {x1 , . . . , xN } ∈ {0, 1}N be a set of binary variables, one for each node in V, where xi = 1 means
i ∈ D, the independent set.
• Local factors capture the cost (negative weight) for each node and is equal to −wi if xi = 1 and
zero otherwise
fi (xi )
min(−wi , 1(xi = 0)) ∀i ∈ V
=
• Pairwise factors ensure that if (i, j) ∈ E, then either xi = 0 or xj = 0
f{i,j} (x{i,j} )
=
1(xi = 0 ∨ xj = 0) ∀(i, j) ∈ E
It is easy to see that using a variable synchronous update, message passing can be performed very
efficiently in O(|E|).
Weigt and Zhou [309] (also see [333]) propose an interesting approach to minimum vertex cover
using energetic survey propagation. Here, again x = {x1 , . . . , xN } ∈ {0, 1}N has one binary variable
per node i ∈ V and a pairwise factor ensure that all edges are covered by at least one node in the cover
f{i,j} (x{i,j} )
=
1(xi = 1 ∨ xj = 1) ∀(i, j) ∈ E
Using the xor-and semiring, the resulting fixed points of warning propagation reveal minimal (but not
necessarily optimal) vertex covers. The authors then suggest using survey propagation, and decimation
to find the warning propagation fixed points with lowest energy (i.e., smallest size of cover).
Chapter 4
Clustering
Clustering of a set of data-points is a central problem in machine learning and data-mining [4, 102, 180,
184, 191, 230, 236, 248, 251, 266, 318, 322]. However many interesting clustering objectives, including
the problems that we consider in this section are NP-hard.
In this section we present message passing solutions to several well-known clustering objectives
including K-medians, K-clustering, K-centers and modularity optimization. Message passing has also
been used within Expectation Maximization to obtain some of the best results in learning stochastic
block models (a hidden variable model for clustering) [82]. The message passing solution to K-medians
and its generalization, clustering by shallow trees are proposed by other researchers. However, for
completeness, we review their factor-graphs in sections 4.1 and 4.2. expressing K-clustering and Kcenters problems as min-max inference problems on factor-graphs in sections 4.3 and 4.4.
4.1
K-medians
Given a symmetric matrix of pairwise distances A ∈ RN ×N between N data-points, and a number of
clusters K, K-medians seeks a partitioning of data-points into K clusters, each associated with a cluster
center, s.t. the sum of distances from each data-point to its cluster center is minimized.
This problem is NP-hard; however, there exists several approximation algorithms for metric distances
[15, 58]. Here we present the binary variable factor-graph [112] for a slightly different version of this
objective, proposed by Frey and Dueck [102]. The simplified form of min-sum BP messages in this
factor-graph is known as affinity propagation. Here, instead of fixing the number of clusters, K, the
objective is modified to incorporate each cluster’s cost to become a center. This cost is added to the
sum of distances between the data-points and their cluster centers and used to decide the number of
clusters at run-time.
Let A be the distance matrix of a directed graph G = (V, E), where (i, j) ∈
/ E ⇔ Ai,j = ∞. Moreover
let Ai,i denote the willingness of node i to be the center of a cluster. A simple heuristic is to set this
4.1. K-MEDIANS
101
value uniformly to mediani6=j Ai,j .
Define x = {xi:j , ∀ (i, j) ∈ E} as a set of binary variables – one per each directed edge (i, j) – where
xi:j ∈ {0, 1} indicates whether node i follows node j as its center. Here node i can follow itself as center.
The following factors define the cost and constraints of affinity propagation:
• Leader factors ensure that each node selects exactly one cluster center.
fE + (i,·) (xE(i,·) ) = 1((
X
xi:j ) = 1)
∀i ∈ V
(i,j)∈E + (i,·)
where as before E + (i, .) = {(i, j) ∈ E} ∪ {(i, i)} is the set of edges leaving node i.
• Consistency factors as defined by equation (3.8) ensure that if any node i ∈ E(., j) selects node j
as its center of cluster, node j also selects itself as the center
At this point we note that we used both of these factor-types for set-cover and dominating set
problem in section 3.5. The only addition (except for using a different semiring) is the following factors.
• Local factors take distances and the willingness to become a center into account
fi:j (x(i,j) ) = min Ai,j , 1(xi:j = 0)
∀(i, j) ∈ E ∪ {(i, i) | i ∈ V}
where in effect fi:j (xi:j ) is equal to Ai,j if xi:j = 1 and 0 otherwise
See equation (1.20) for definition of 1(.) in min-sum semiring and note the fact that ⊕ = min in this
case. This means extensions to other semirings (e.g., min-max) need only to consider a different 1(.)
and use their own ⊕ operator. However, direct application of min-max inference to this factor-graph
is problematic for another reason: since the number of clusters K is not enforced, as soon as node i
becomes center of a cluster, all nodes j with Aj,j ≤ Ai,i can become their own centers without increasing
the min-max value. In section 4.4, we resolve this issue by enforcing the number of clusters K and use
inference on min-max semiring to solve the corresponding problem known as K-center problem.
The complexity of min-sum BP with variable and factor synchronous message update is O(|E|) as
each leader and consistency factor allows efficient O(|E(., j)|) and O(|E(i, .)|) calculation of factor-tovariable messages respectively. Moreover, all variable-to-factor messages leaving node i can be calculated
simultaneously in O(|E(i, .)|) using variable-synchronous update of equation (2.13).
4.1.1
Facility location problem
A closely related problem to K-medians is the facility location problem, where a matrix A ∈ R|V1 |×|V2 |
specifies the pairwise distance between two parts of bipartite graph G = (V = (V1 , V2 ), E). The goal is
to select a sub-set of facilities D ⊂ V1 to minimize the sum of distances of each customer i ∈ V2 to its
associated facility in D.
The uncapacitated version of this problem (no restriction on |D|) can be solved as special K-median
problem where Ai,i = ∞ ∀i ∈ V2 . Message passing solution for min-sum variations of this problem are
102
CHAPTER 4. CLUSTERING
discussed in [187, 188]. We discuss the min-max facility location as a special case of k-center problem
in section 4.4.
4.2
Hierarchical clustering
By adding a dummy node ∗ and connecting all the cluster centers to this node with the cost Ai,∗ = Ai,i ,
we can think of the K-median clustering of previous section as finding a minimum cost tree of depth
two with the dummy node as its root. The shallow trees of Bailly-Bechet et al. [18] generalize this
notion by allowing more levels of hierarchy. The objective is to find a tree of maximum depth d, that
minimizes the sum over all its edges. An alternative approach to hierarchical clustering based on nested
application of affinity propagation is discussed in [111, 289].
Previously we presented the binary-variable model for affinity propagation. However, it is possible
to obtain identical message updates using categorical variables [102]. Here x = {x1 , . . . , xN }, where
xi ∈ Xi = {j | (i, j) ∈ E} selects one of the neighbouring nodes as the center of the cluster for node i.
In this case we can ignore the leader factors and change the consistency and local factors accordingly.
The idea in building hierarchies is to allow each node i to follow another node j even if i itself is
followed by a third node k (i.e., dropping the consistency factors). Also xi = i is forbidden. However,
this creates the risk of forming loops. The trick used by Bailly-Bechet et al. [18] is to add an auxiliary
“depth” variable zi ∈ {0, . . . , d} at each node. Here, the depth of the dummy node ∗ is zero and the
depth factors ensure that if i follows j then zi = zj + 1:
fi,j (xi , z i,j ) = 1 zi = zj + 1 ∨ xi 6= j
4.2.1
Spanning and Steiner trees
Although finding trees is not a clustering problem, since one could use the same techniques used for
clustering by shallow trees, we include it in this section.
Given a graph G = (V, E), a penalty
wi:j < 0 per edge (i, j) ∈ E and a prize wi > 0 per node i ∈ V, the prize-collecting Steiner tree’s
objective is to select a connected sub-graph G 0 = (V 0 ⊆ V, E 0 ⊆ E) with maximum sum of prizes
P
P
i∈V 0 wi +
(i,j)∈E 0 wi:j . Since the optimal sub-graph is always a tree, a construction similar to that of
shallow trees (with min-sum BP) finds high-quality solutions to depth-limited versions of this problem
in O(d|E|) [19, 20, 23, 31].
The main difference with factor-graph of shallow trees is that here the root node is a pre-determined
member of V and several tricks are used to find best (set of) root(s). However, since the |V 0 | may be
smaller than |V|, a different dummy node ∗ is introduced, with zero cost of connection to all nodes i ∈ V
s.t. xi = ∗ means node i is not a part of the Steiner tree (i ∈ V \ V 0 ).
Alternatively, if we do not introduce this dummy node such that V 0 = V and moreover, set the node
penalties to zero, the result is a depth limited spanning tree. Bayati et al. [24] show that if the
4.3. K-CLUSTERING PROBLEM
103
maximum depth is large enough (e.g., d = N ) and BP is convergent, it will find the minimum spanning
tree.
4.3
K-clustering problem
Given a symmetric matrix of pairwise distances A ∈ RN ×N between N data-points, and a number of
clusters K, K-clustering (a.k.a. min-max clustering) seeks a partitioning of data-points that minimizes
the maximum distance between all the pairs in the same partition.
We formulate this problem as min-max inference problem in a factor-graph. Let x = {x1 , . . . , xN }
with xi ∈ {1, . . . , K} be the set of variables, where xi = k means, point i belongs to cluster k. The
Potts factor
f{i,j} (xi , xj ) = min 1(xi 6= xj ), Ai,j
∀1 ≤ i < j ≤ N
(4.1)
is equal to Ai,j if two nodes are in the same cluster and −∞ otherwise. It is easy to see that using
min-max inference on this factor graph, the min-max solution x∗ (equation (1.26)) defines a clustering
that minimizes the maximum of all inter-cluster distances.
Recall that the y-neighborhood graph (definition 3.6.1) for a distance matrix A is the graph
G(A, y) = (V, E(A, y) = {(i, j) | Ai,j ≤ y}).
Claim 4.3.1. The py -reduction of the min-max clustering factor-graph above is identical to K-cliquecover factor-graph of section 3.4 for G(A, y).
Proof. The py -reduction of the K-clustering factor (equation (4.1)) is
f{i,j} (xi , xj ) = 1 min 1(xi 6= xj ), Ai,j ≤ y = 1(xi 6= xj ∨ Ai,j ≤ y)
Recall that the K-clique-cover factor-graph, defines a pairwise factor between any two nodes i and
j whenever (i, j) ∈
/ E(A, y), – i.e., whenever Ai,j ≤ y it does not define a factor. However, Ai,j ≤ y
means that the reduced constraint factor above is satisfied and therefore we only need to consider the
cases where Ai,j > y. This gives f{i,j} (xi , xj ) = 1(xi 6= xj ), which is the K-clique-cover factor between
two nodes i and j s.t. (i, j) ∈
/ E(A, y).
We use binary-search over sum-product reductions to approximate the min-max solution (see equation (1.25)). Considering the O(N 2 K) cost of message passing for K-clique-cover, this gives O(N 2 K log(N 2 )) =
O(N 2 K log(N )) cost for K-clustering, where the log(N 2 ) is the cost of binary search over the set of all
possible pairwise distance values in A.
104
CHAPTER 4. CLUSTERING
Figure 4.1 compares the performance of min-max clustering using message passing to that of Furthest
Point Clustering (FPC) [119] which is 2-optimal when the triangle inequality holds. Note that message
passing solutions are superior even when using Euclidean distance.
Figure 4.1: Min-max clustering of 100 points with varying numbers of clusters (x-axis). Each point is an average
over 10 random instances. The y-axis is the ratio of the min-max value obtained by sum-product reduction (and
using perturbed BP with T = 50 to find satisfying assignments ) divided by the min-max value of Furthest Point
Clustering (FPC). (left) Clustering of random points in 2D Euclidean space. The red line is the lower bound on
the optimal result based on the worst case guarantee of FPC. (right) Using symmetric random distance matrix
where Ai,j = Aj,i ∼ U (0, 1).
4.4
K-center problem
Given a pairwise distance matrix D ∈ RN ×N , the K-center problem seeks a partitioning of nodes, with
one center per partition such that the maximum distance from any node to the center of its partition
is minimized.
This problem is known to be NP-hard, even for Euclidean distance matrices [201], however Kcenter has some approximation algorithms that apply when the distance matrix satisfies the triangle
inequality [87, 135]. The method of Dyer and Frieze [87] is very similar to furthest point clustering and
extends to weighted K-center problem in which the distance from any point to all the other points is
scaled by its weight. The more general case of asymmetric distances does not allow any constant-factor
approximation (however o(log(N ))-approximation exists [66, 237]).
Here we define a factor-graph, whose min-max inference results in the optimal solution for K-center
problem. For this consider the graph G(V, E) induced by the distance matrix A s.t. V = {1, . . . , N }
and Ai,j = ∞ ⇔ (i, j) ∈
/ E.
Let x = {xi:j | (i, j) ∈ E} be the set of variables, where xi:j ∈ {0, 1} indicates whether j is the center
of cluster for i. Define the following factors:
4.4. K-CENTER PROBLEM
105
Figure 4.2: The factor-graph for K-center problem, where the local factors are black squares, the leader factors
are light grey, consistency factors are white and K-of-N factor is dark grey.
• Local factors:
f(i,j) (xi:j ) = min Ai,j , 1(xi:j = 0)
∀(i, j) ∈ E
• Leader, Consistency and at-most-K-of-N factors as defined for the sum-product case of induced
K-set-cover and K-independent set (section 3.5). Here we need to replace sum-product 1(.) with the
min-max version (equation (1.20)).
For variants of this problem such as the capacitated K-center, additional constraints on the maximum/minimum points in each group may be added as the at-least/at-most K-of-N factors.
Claim 4.4.1. The py -reduction of the K-center clustering factor-graph above is identical to K-set-cover
factor-graph of section 3.5 for G(A, y).
Proof. Leader, consistency and at-most-K-of-N factors are identical to the factors used for K-set-cover
(which is identical to their py -reduction), where the only difference is that here we have one variable
per each (i, j) ∈ E whereas in K-set-cover for G(A, y) we have one variable per (i, j) ∈ E | Ai,j ≤ y. By
considering the py -reduction of the local factors
fi:j (xi:j ) = 1 min Ai,j , 1(xi:j) = 0) ≤ y = 1(Ai,j ≤ y ∨ xi:j = 0)
we see that we can assume that for Aij > 0, xi:j = 0 and we can drop these variables from the factorgraph. After this we can also omit the py -reduction of the local factors as they have no effect. This
gives us the factor-graph of section 3.5 for set-cover.
Similar to the K-clustering problem, we use the sum-product reduction to find near-optimal solutions
106
CHAPTER 4. CLUSTERING
Figure 4.3: (left) K-center clustering of 50 random points in a 2D plane with various numbers of clusters
(x-axis). The y-axis is the ratio of the min-max value obtained by sum-product reduction (T = 500 for perturbed
BP) over the min-max value of 2-approximation of [87]. (right) Min-max K-facility location formulated as an
asymmetric K-center problem and solved using message passing. Yellow squares indicate 20 potential facility
locations and small blue circles indicate 50 customers. The task is to select 5 facilities (red squares) to minimize
the maximum distance from any customer to a facility. The radius of circles is the min-max value.
to this problem. The binary search seeks the optimal y ∈ Y (where Y is the collective range of all the
factors). Here, since only local factors take values other than ±∞, the search is over their range, which
is basically all the values in A. This adds an additional log(N ) multiplicative factor to the complexity
of K-set-cover (which depends on the message update).
We can significantly reduce the number of variables and the complexity by bounding the distance
to the center of the cluster y. Given an upper bound y, we may remove all the variables xi:j where
Ai,j > y from the factor-graph. Assuming that at most R nodes are at distance Ai,j ≤ y from every
node j, the complexity of min-max inference with synchronous update drops to O(N R2 log(N )). This
upper bound can be obtained for example by applying approximation algorithms.
Figure 4.3(left) compares the performance of message-passing and the 2-approximation of [87] when
triangle inequality holds. The min-max facility location problem can also be formulated as an asymmetric K-center problem where the distance to all customers is ∞ and the distance from a facility to
another facility is −∞ (figure 4.3(right)).
4.5
Modularity maximization
A widely used objective for clustering (or community mining) is Modularity maximization [229]. However, exact optimization of Modularity is NP-hard [45]. Modularity is closely related to fully connected
Potts graphical models [259, 332]. Many have proposed various other heuristics for modularity optimization [34, 67, 228, 259, 265]. Here after a brief review of the Potts model in section 4.5.1, we introduce
a factor-graph representation of this problem that has a large number of factors in section 4.5.2. We
then use the augmentation technique (section 2.2.2) to incrementally incorporate violated constraints
4.5. MODULARITY MAXIMIZATION
107
in these models.
4.5.1
The Potts model
Let G = (V, E) be an undirected graph, with M = |E|, N = |V| and the adjacency matrix B ∈ RN ×N ,
P
where Bi,j 6= 0 ⇔ (i, j) ∈ E. Let A be the normalized adjacency matrix where i<j A = 1. Also
def P
let AE(i,·) =
j Ai,j denote the normalized degree of node vi . Graph clustering using modularity
optimization seeks a partitioning of the nodes into unspecified number of clusters C = {C1 , . . . , CK },
maximizing
def
m(C) =
X
Ci ∈C
X
Ai,j − ζAE(i,·) AE(j,·)
(4.2)
i,j∈Ci
The first term of modularity is proportional to within-cluster edge-weights. The second term is proportional to the expected number of within cluster edge-weights for a null model with the same weighted
node degrees for each node i. Here the null model is a fully-connected graph. The resolution parameter ζ – which is by default set to one – influences the size of communities, where higher resolutions
motivates a larger number of clusters.
In the Potts model, each node i ∈ V is associated with a variable xi ∈ {1, . . . , Kmax }, where Kmax
is an upper bound on the number of clusters. Here, each pair of variables have a pairwise interaction
f{i,j} (x{i,j} )
=
ζ(A
E(i,·) AE(j,·) ) − Ai,j
0
xi = xj
(4.3)
xi 6= xj
where min-sum inference on this fully connected factor-graph gives an assignment of each node to a
cluster so as to maximize the modularity.
4.5.2
Clique model
Here, we introduce an alternative factor-graph for modularity optimization. Before introducing our
factor-graph representation, we suggest a procedure to stochastically approximate the null model using
a sparse set of interactions.
We generate a random sparse null model with M null < αM weighted edges (E null ), by randomly
p
sampling two nodes, each drawn independently from Pr(i) ∝ AE(i,·) , and connecting them with a
p
weight proportional to Bnull
∝ AE(i,·) AE(j,·) . If they have been already connected, this weight is
i,j
added to their current weight. We repeat this process αM times, however since some of the edges
are repeated, the total number of edges in the sparse null model may be less than αM . Finally the
108
CHAPTER 4. CLUSTERING
normalized edge-weight in the sparse null model is
def
Bnull
i,j
null .
k,l Bk,l
Anull
= P
i,j
It is easy to see that this generative process in expectation produces the fully connected null model.1
Factor-graph representation
Here we use the following binary-valued factor-graph formulation. Let x = {xi:j ∈ {0, 1} | (i, j) ∈
E ∪ E null } be a set of binary variables, and let L denote the cardinality of E ∪ E null . The variable xi:j
is equal to one means the corresponding edge, (i, j), is present in the final model, where our goal is
to define the factor-graph such that the final model consists of cliques. For this, define the factors as
follows:
• Local factor for each variable are equal to the difference between the weighted adjacency and the null
model if an edge is present (i.e., xi:j = 1)
null
fi:j (xi:j ) = min 1(xi:j = 0), Ai,j − Ai,j
(4.4)
By enforcing the formation of cliques, while minimizing the sum of local factors the negative sum of
local factors evaluates to modularity (equation (4.2)):
• For each three edges (i, j), (j, k), (i, k) ∈ E ∪ E null , i < j < k that form a triangle, define a clique
factor as
f{i:j,j:k,i:k} (xi:j , xj:k , xi:k ) = 1(xi:j + xj:k + xi:k 6= 2)
(4.5)
These factors ensure the formation of cliques – i.e., if two edges that are adjacent to the same node are
present (i.e., xi:i = 1 and xi:k = 1), the third edge in the triangle should also be present (i.e., xj:k = 1).
The computational challenge here is the large number of clique constraints. For the fully connected null
model, we need O(N 3 ) such factors and even using the sparse null model – assuming a random edge
3
3
L
L
3
probability a.k.a. Erdos-Reny graph – there are O( N
6 N ) = O( N 3 ) triangles in the graph (recall that
L = |E ∪ E null |).
Brandes et al. [45] first introduced an LP formulation with similar form of constraints. However,
since they include all the constraints from the beginning and the null model is fully connected, their
method is only applied to small toy problems.
1
The choice of using square root of weighted degrees for both sampling and weighting is to reduce the variance. One
may also use pure importance sampling (i.e., use the product of weighted degrees for sampling and set the edge-weights in
the null model uniformly), or uniform sampling of edges, where the edge-weights of the null model are set to the product
of weighted degrees.
4.5. MODULARITY MAXIMIZATION
109
Simplified message update and augmentation
Here, we give technical details as how to simplify the min-sum BP message update for this factor-graph.
The clique factor is satisfied only if either zero, one, or all three of the variables in its domain are
non-zero. Therefore, in order to derive message updates b
p{i:j,j:k,i:k}→i:j from the clique factor f{i:j,j:k,i:k}
to variable xi:j for a particular value of xi:j (e.g., xi:j = 0), we apply ⊕ operator (i.e., minimize) over
all the valid cases of incoming messages (e.g., when xi:k and xj:k in clique factor f{i:j,j:k,i:k} are zero).
This gives the simplified factor-to-variable messages
b
p{i:j,j:k,i:k}→i:j (0) = min{0, b
pj:k→{i:j,j:k,i:k} , b
pi:k→{i:j,j:k,i:k} }
b
p{i:j,j:k,i:k}→i:j (1) = min{0, b
pj:k→{i:j,j:k,i:k} + b
pi:k→{i:j,j:k,i:k} }
(4.6)
where for xi:j = 0, the minimization is over three feasible cases (a) xj:k = xi:k = 0, (b) xj:k = 1, xi:k = 0
and (c) xj:k = 0, xi:k = 1. For xi:j = 1, there are two feasible cases (a) xj:k = xi:k = 0 and (b) xj:k =
xi:k = 1.
Here we work with normalized messages, such that b
pI→i (0) = 0 2 and use b
pI→i to denote b
pI→i (1).
The same applies to the marginal b
pi:j , which is a scalar called bias. Here b
pi:j > 0 means b
pi:j (1) > b
pi:j (0)
and shows a preference for xi:j = 1. Normalizing clique-factor messages above we get the following form
of simplified factor-to-variable messages for clique constraints
b
p{i:j,j:k,i:k}→i:j = min{0, b
pj:k→{i:j,j:k,i:k} + b
pi:k→{i:j,j:k,i:k} }−
(4.7)
min{0, b
pj:k→{i:j,j:k,i:k} , b
pi:k→{i:j,j:k,i:k} }.
In order to deal with large number of factors in this factor-graph, we use the augmentation approach
of section 2.2.2. We start with no clique-factors, and run min-sum BP to obtain a solution (which may
even be unfeasible). We then find a set of clique constraints that are violated in the current solution and
augment the factor-graph with factors to enforce these constraints. In order to find violated constraints
in the current solution, we simply look at pairs of positively fixed edges (xi:j = 1 and xi:k = 1)
around each node i and if the third edge of the triangle is not positively fixed (xj:k = 0), we add
the corresponding clique factor (f{i:j,j:k,i:k} ) to the factor-graph. See algorithm 3 (in the appendix) for
details of our message-passing for Modularity maximization.
Experiments
We experimented with a set of classic benchmarks3 . Since the optimization criteria is modularity, we
compared our method only against best known “modularity optimization” heuristics: (a) FastModularity[67],
2
3
Note that this is different from the standard normalization for min-sum semiring in which minxi b
pI→i (xi ) = 0.
Obtained form Mark Newman’s website: http://www-personal.umich.edu/~mejn/netdata/
110
CHAPTER 4. CLUSTERING
Figure 4.4: (left) Clustering of power network (N = 4941) by message passing. Different clusters have different
colors and the nodes are scaled by their degree. (right) Clustering of politician blogs network (N = 1490) by
message passing and by meta-data – i.e., liberal or conservative.
Table 4.1: Comparison of different modularity optimization methods.
13.55%
0.506
.04
0.525
1.648
0.467
0.179
0.501
0.643
0.489
0.03
5635
17.12%
0.594
0.14
0.601
0.87
0.487
0.151
0.548
0.08
0.602
0.019
wkarate
n
34
78
562
12.34%
0.431
0
431
15.14%
0.401
0
0.444
0.557
0.421
0.095
0.410
0.085
0.443
0.027
netscience
n
1589
2742
NA
NA
NA
NA
53027
.0004%
0.941
2.01
0.907
8.459
0.889
0.303
0.926
0.154
0.948
0.218
dolphins
y
62
159
1892
14.02%
0.508
0.01
1269
6.50%
0.521
0.01
0.523
0.728
0.491
0.109
0.495
0.107
0.517
0.011
lesmis
n
77
254
2927
5.14%
0.531
0
1601
1.7%
0.534
0.01
0.529
1.31
0.483
0.081
0.472
0.073
0.566
0.011
celegansneural
n
297
2359
43957
16.70%
0.391
10.89
21380
3.16%
0.404
2.82
0.406
5.849
0.278
0.188
0.367
0.12
0.435
0.031
polblogs
y
1490
19090
NA
NA
NA
156753
.14%
0.411
32.75
0.427
67.674
0.425
0.33
0.427
0.305
0.426
0.099
karate
y
34
78
562
14.32%
0.355
17.54%
0.390
0
0.417
0.531
0.393
0.086
0.380
0.079
0.395
0.009
Time
3624
0.41
Modularity
.07
0.591
423
Modularity
0.511
27.85%
Time
5.68%
6554
0
Modularity
5461
615
Time
Time
441
115
NA
Modularity
Modularity
105
y
Time
Cost
y
football
L
Time
polbooks
L
Modularity
Louvian
Cost
FastGreedy
Edges
L-Eigenvector
Nodes
Spin-glass
Weighted?
message passing (sparse)
Problem
message passing (full)
4.5. MODULARITY MAXIMIZATION
(b) Louvain [34], (c) Spin-glass [259] and (d) Leading eigenvector [228].
111
4
table 4.1 summarizes our results (see also Figure 4.4). Here for each method and each data-set,
we report the time (in seconds) and the Modularity of the communities found by each method. The
table include the results of message passing for both full and sparse null models, where we used a
constant α = 20 to generate our stochastic sparse null model. For message passing, we also included
L = |E + E null | and the saving in the cost using augmentation. This column shows the percentage of
the number of all the constraints considered by the augmentation. For example, the cost of .14% for
the polblogs data-set shows that augmentation and sparse null model meant using .0014 times fewer
clique-factors, compared to the full factor-graph.
Overall, the results suggest that our method is comparable to state-of-the-art in terms both time
and quality of clustering. Although, we should note that the number of triangle constraints in large and
dense graphs increases very quickly, which deteriorates the performance of this approach despite using
the augmentation. Despite this fact, our results confirm the utility of augmentation by showing that it
is able to find feasible solutions using a very small portion of the constraints.
4
For message passing, we use λ = .1, max = median{|Ai,j − Anull
i,j |}(i,j)∈E∪E null and Tmax = 10. Here we do not perform
any decimation and directly fix the variables based on their bias b
pi:j > 0 ⇔ xi:j = 1.
Chapter 5
Permutations
5.1
Matching and permanent
The integration and maximization problems over unrestricted permutations define several important
combinatorial problems. Two notable examples of integration problems are permanent and determinant of a matrix. Determinant of matrix A ∈ RN ×N is defined as
det(A) =
X
x∈SN
sign(x)
N
Y
Ai,xi
i=1
where SN is the set of all permutations of N elements (a.k.a. symmetric group) and xi ∈ {1, . . . , N }
is the index of ith element in particular permutation x. Here the sign(.) classifies permutations as even
(sign(x) = 1) and odd (sign(x) = −1), where we can perform an even (odd) permutation by even (odd)
number of pairwise exchanges. The only difference in definition of permanent is removal of the sign
function
perm(A) =
N
X Y
Ai,xi
x∈SN i=1
Here, we see that both permanent and determinant are closely related with two easy combinatorial
problems on graphs – i.e., perfect matching and spanning sub-tree. While calculating the permanent
for A ∈ {0, 1}N ×N is #P-hard [297], the determinant can be obtained in O(N 3 ) [116].
The matrix-tree theorem states that the number of spanning trees in a graph with adjacency
matrix A is equal to det(L(i, i)) for an arbitrary 1 ≤ i ≤ N . Here L = A − D is the Laplacian of A,
P
where D is a diagonal matrix with degree of each node on the diagonal (i.e., Di,i = j Ai,j ) and L(i, i)
is the (N − 1) × (N − 1) sub-matrix of L in which row and column i are removed.
An intermediate step in representing the permutation as a graphical model is to use a bipartite
5.1. MATCHING AND PERMANENT
113
graph G = (V = (V1 , V2 ), E), where |V1 | = |V2 | = N and the edge-set E = {(i, j) | i ∈ V1 , j ∈ V2 }.
A perfect matching is a one to one mapping of elements of V1 to V2 and can be represented using
the corresponding edges E 0 ⊂ E. It is easy to see that any perfect matching E 0 ⊂ E identifies a
permutation x.
Here the maximum weighted matching (a.k.a. assignment problem) problem is to
QN
find a perfect matching x∗ = argx∈SN max
i=1 Ai,xi , while the bottleneck assignment problem
seeks x∗ = argx∈SN min maxN
i=1 Ai,xi . The factor-graph representation of the next section shows that
bipartite matching and bottleneck assignment problems correspond to the max-product (min-sum) and
min-max inference, and computation of permanent corresponds to sum-product inference over the same
factor-graph.
Interestingly min-sum and min-max inference in this setting are in P [121, 182] while sum-product
is in #P. Indeed the application of max-product BP to find the maximum weighted matching [22] (and
its generalization to maximum weighted b-matching [138]) is one of the few cases in which loopy BP
is guaranteed to be optimal. Although MCMC methods (section 2.6.1) can provide polynomial time
approximation schemes for permanent [150] (and many other combinatorial integration problems [149]),
they are found to be slow in practice [139]. This has motivated approximations using deterministic
variational techniques [62, 308] and in particular BP [139], which is guaranteed to provide a lower
bound on the permanent [301].
5.1.1
Factor-graph and complexity
Here we review the factor-graph of Bayati et al. [22] for maximum bipartite matching. Given the
bipartite graph G = ((V1 , V2 ), E) and the associated matrix A ∈ RN ×N with non-negative entries,
define two sets of variables x = {xi ∈ V2 | i ∈ V1 } and z = {zj ∈ V1 | j ∈ V2 }, where xi = j and zj = i
both mean node i is connected to node j. Obviously this representation is redundant and for (i, j) ∈ E
a pairwise factor should ensure xi and zi are consistent:
• Consistency factors ensure the consistency of x and z
f{i,j} (xi , zj ) = 1 (xi = j ∧ zj = i) ∨ (xi 6= j ∧ zj 6= i)
∀(i, j) ∈ E
• Local factors represent the cost of a matching
∀i ∈ V1
f{i} (xi ) = Ai,xi
where if i and j are connected in a matching, two local factors f{i} (xj ) =
p
p
Ai,j and f{j} (xi ) = Ai,j
account for Ai,j . Figure 5.1 shows this factor-graph.
It is easy to see that the joint form q(x, z) is equal to
NN
for any consistent assignment
P
to x, z and it is equal to 1 otherwise. Therefore sum-product and max-product (i.e., x,z q(x, z) and
⊕
i=1 Ai,xi
114
CHAPTER 5. PERMUTATIONS
Figure 5.1: The factor-graph for matching where the local factors are black and consistency factors are white
squares.
maxx,z q(x, z)) produce the permanent and max-weighted matching respectively.
The cost of each iteration of both max-product and sum-product BP in the factor-graph above is
O(N 2 ). Moreover, for max-product BP, if the optimal solution is unique, BP is guaranteed to converge
∗
to this solution after O( Ny ) iterations, where is the difference between the cost of first and second
best matchings and y ∗ is the cost of best matching [22].
An alternative is to use a binary variable model in which each edge of the bipartite graph is
associated with a binary variable and replace the consistency factors with degree constraint to ensure
that each node is matched to exactly one other node. However this model results in BP updates
equivalent to the one above (the simplification of updates discussed in [25] is exactly the updates of
binary variable model).
For min-max semiring, q(x, z) = maxN
i=1 Ai,xi for a consistent assignment to x, z and it evaluates to
min
1 = ∞. Therefore min-max inference here seeks an assignment that minimizes the maximum matching
cost – a.k.a. bottleneck assignment problem.
5.1.2
Arbitrary graphs
We can also use message passing to solve max-weighted matching in an arbitrary graph G = (V, E)
with adjacency matrix A. Zdeborová and Mézard [331] proposed a factor-graph for the related task of
counting of the perfect matchings in an arbitrary graph. For this, each edge (i, j) ∈ E is assigned to one
binary variable xi:j ∈ {0, 1} and the degree factors on each node restrict the number of non-zero values
to one
fE(i,.) (xE(i,.) )
=
1((
X
(i,j)∈E(i,.)
xi:j ) ≤ 1)
5.2. TRAVELING SALESMAN PROBLEM
115
where E(i, .) is the set of all the edges adjacent to node i.
Sanghavi [273] consider the problem of maximum weighted b-matching in arbitrary graphs by changing the degree factor to
fE(i,.) (xE(i,.) )
=
1((
X
xi:j ) ≤ b)
(i,j)∈E(i,.)
and also local factor fi:j (xi:j ) = xi:j Ai,j that takes the weights into account. They show that if the
solution to the corresponding LP relaxation is integral then BP converges to the optimal solution.
Moreover, BP does not converge if the LP solution is not integral.
Matchings in an arbitrary graph G is also related to the permanent and also (vertex disjoint) cycle
covers; a set of directed cycles in G that cover all of its nodes exactly once. The number of such cycle
covers in an un-weighted graph is equal to perm(A), which is in turn equal to the square of number of
perfect matchings [222, ch. 13].
In fact a directed cycle cover with maximum weight is equivalent to the maximum weighted bipartite
matching in the construction of the previous section. In section 5.2 below we will use message passing
to obtain a minimum weighted “undirected” cycle cover and further restrict these covers to obtain a
minimum weighted cover with a single cycle – i.e., a minimum tour for TSP.
5.2
Traveling salesman problem
A Traveling Salesman Problem (TSP) seeks the minimum length tour of N cities that visits each city
exactly once. TSP is NP-hard and for general distances, no constant factor approximation to this
problem is possible [238]. The best known exact solver, due to Held and Karp [128], uses dynamic
programming to reduce the cost of enumerating all orderings from O(N !) to O(N 2 2N ).
The devel-
opment of many (now) standard optimization techniques are closely linked with advances in solving
TSP. Important examples are simulated annealing [56, 166], mixed integer linear programming [118],
dynamic programming [28], ant colony optimization [85] and genetic algorithms [115, 120].
Since Dantzig et al. [77] manually applied the cutting plane method to 49-city problem, a combination
of more sophisticated cuts, used with branch-and-bound techniques [21, 233] has produced the state-ofthe-art TSP-solver, Concorde [13].
Other notable results on very large instances have been reported
by Lin-Kernigan heuristic [130] that continuously improves a solution by exchanging nodes in the tour.
For a readable historical background of the state-of-the-art in TSP and its applications, see [12].
The search over the optimal tour is a search over all permutations of N cities that contains no
sub-tours – that is the permutation/tour is constrained such that we dot not return to the starting
city without visiting all other cities. Producing the permutation with minimum cost that may include
sub-tours is called the (vertex disjoint) cycle-cover and is in P (see section 5.1.2).
We provide two approaches to model TSP: section 5.2.1 presents the first approach, which ignores the
116
CHAPTER 5. PERMUTATIONS
subtour constraints – i.e., finds cycle covers – and then “augment” the factor-graph with such constraints
when they become violated. This augmentation process is repeated until a feasible solution is found.
The second approach, presented in section 5.2.2, is to use the variables that represent the time-step in
which a node is visited. By having the same number of time-steps as cities, the subtour constraint is
automatically enforced. This second formulation, which is computationally more expensive, is closely
related to our factor-graph for sub-graph isomorphism (see section 5.3). This is because one can think
of the problem of finding a Hamiltonian cycle in G as finding a sub-graph of G that is isomorphic to
a loop of size |V|.
5.2.1
Augmentative approach
Let G = (V, E) denote a graph of our problem with the positively weighted symmetric adjacency matrix
A, s.t. Ai,j = 0 ⇔ (i, j) ∈
/ E. The objective is to select a subset of E that identifies shortest tour of N
cities. Let x = {xi:j | (i, j) ∈ E} be a set of M binary variables (i.e., xi:j ∈ {0, 1}), one for each edge in
the graph (i.e., M = |E|) where xi:j = 1 means (i, j) is in the tour. We use xi:j and xj:i to refer to the
same variable. Recall that for each node i, E(i, ·) denotes the edges adjacent to i. Define the factors of
the factor-graph as follows
• Local factors represent the cost associated with each edge
fi:j (xi:j ) = min 1(xi:j = 0), Ai,j
∀(i, j) ∈ E
(5.1)
where fi:j (xi:j ) is either Ai,j or zero.
Any valid tour satisfies the following necessary and sufficient constraints – a.k.a. Held-Karp constraints [127]:
• Degree factors ensure that exactly two edges that are adjacent to each vertex are in the tour
X
fE(i,·) (xE(i,·) ) = 1 (
xi:j ) = 2
(5.2)
(i,j)∈E(i,·)
• Subtour factors ensure that there are no short-circuits – i.e., there are no loops that contain strict
def
subsets of nodes. To enforce this, for each S ⊂ V, define E(S, .) = {(i, j) ∈ E | i ∈ S, j ∈
/ S} to be the
set of edges, with one end in S and the other end in V \ S. We need to have at least two edges leaving
each subset S. The following set of factors enforce these constraints
fE(S,·) (xE(S,·) ) = 1 (
X
xi:j ) ≥ 2
∀ S ⊂ V, S =
6 ∅
(5.3)
(i,j)∈E(S,·)
These three types of factors define a factor-graph, whose minimum energy configuration is the smallest
tour for TSP. Therefore we can use min-sum inference to obtain the optimal tour. Note that both
5.2. TRAVELING SALESMAN PROBLEM
117
subtour and degree constraints depend on large number of variables, however, due to sparsity they
allow efficient linear time calculation of factor-to-variable messages; see equation (5.3). The more
significant computational challenge is that the complete TSP factor-graph has O(2N ) subtour factors,
one for each subset of variables. In figure 5.2 we address this problem using factor-graph augmentation.
Simplified messages
In section 2.2.1, we introduced the K-of-N factors for min-sum inference. Both degree and subtour
factors are different variations of this types of factor. For simplicity we work with normalized message
b
pI→i:j = b
pI→i:j (1) − b
pI→i:j (0), which is equivalent to assuming b
pI→i:j (0) = 0 ∀I, i : j ∈ ∂I. The same
notation is used for variable-to-factor message, and marginal belief. As before, we refer to the normalized
marginal belief, b
pi:j = b
p(xi:j = 1) − b
p(xi:j = 0) as bias.
Recall that a degree constraint for node i (fE(i,·) ) depends on all the variables xi:j for edges (i, j)
that are adjacent to i. Here we review the factor-to-variable message for min-sum BP
b
pE(i,·)→i:j (xi:j )
=
X
min fE(i,·) (xE(i,·) )
x\i:j
b
pi:k→E(i,·) (xi:k )
(5.4)
(5.5)
(i,k)∈E(i,·)\(i,j)
We show that this message update simplifies to
b
pE(i,·)→i:j (1)
b
pE(i,·)→i:j (0)
=
=
min b
pi:k→E(i,·) | (i, k) ∈ E(i, ·) \ (i, j)
min b
pi:k→E(i,·) + b
pi:l→E(i,·)
∀i ∈ V
| (i, k), (i, l) ∈ E(i, ·) \ (i, j)
(5.6)
where for xi:j = 1, in order to satisfy degree constraint fE(i,·) (xE(i,·) ), only one other xi:k for (i, k) ∈
E(i, ·) \ (i, j) should be non-zero. On the other hand, we know that messages are normalized such
that b
pi:j→E(i,·) (0) = 0, which means they can be ignored in the summation of equation (5.4). For
xi:j = 0, in order to satisfy the constraint factor, two of the adjacent variables should have a non-zero
value. Therefore we seek two such incoming messages with minimum values. Let mink A denote the k th
smallest value in the set A – i.e., min A ≡ min1 A. We combine the updates above to get a “normalized
message”, b
pE(i,·)→i:j , which is simply the negative of the second largest incoming message (excluding
b
pE(i,·)→i:j ) to the degree factor fE(i,·) :
2
b
pE(i,·)→i:j = b
pE(i,·)→i:j (1) − b
pE(i,·)→i:j (0) = − min{b
pi:k→E(i,·) | (i, k) ∈ E(i, ·) \ (i, j)}
(5.7)
Following a similar procedure, factor-to-variable messages for subtour factors is given by
2
b
pE(S,·)→i:j = − max 0, min{b
pi:k→E(S,·) | (i, k) ∈ E(S, ·) \ (i, j)}
(5.8)
While we are searching for the minimum incoming message, if we encounter two messages with
118
CHAPTER 5. PERMUTATIONS
negative or zero values, we can safely assume b
pE(S,·)→i:j = 0, and stop the search. This results in
significant speedup in practice. Note that both equation (5.7) and equation (5.8) only need to calculate
the second smallest incoming message to their corresponding factors, less the current outgoing message.
In the asynchronous calculation of messages, this minimization should be repeated for each outgoing
message. However in a factor-synchronous update, by finding three smallest incoming messages to each
factor, we can calculate all the factor-to-variable messages at the same time.
iteration 1
iteration 3
iteration 5
iteration 11
Figure 5.2: The message passing results after each augmentation step for the complete graph of printing board
instance from [260]. The blue lines in each figure show the selected edges at the end of message passing. The pale
red lines show the edges with the bias that, although negative (b
pi:j < 0), were close to zero.
Augmentation
To deal with the exponentially large number of subtour factors, we use the augmentation procedure of
section 2.2.2. Starting with a factor-graph with no subtour factor, we find a solution x∗ using min-sum
BP. If the solution is feasible (has no subtours) we are done. Otherwise, we can find all subtours in O(N )
by finding connected components. We identify all the variables in each subtour as S ⊂ V and add a
subtour factor fE(S,·) to ensure that this constraint is satisfied in the next iteration of augmentation. Here
to speed up the message passing we reuse the messages from the previous augmentation step. Moreover
in obtaining x∗ from min-sum BP marginals b
p(xi ), we ensure that no degree constraint is violated
(i.e., each node has two neighbouring edges in x∗ ). Figure 5.2 shows iterations of augmentation over a
print board TSP instance from [260]. Algorithm 4 in the appendix gives details of our message passing
solution to TSP, which also uses several other minor tricks to speed up the BP message updates.
5.2. TRAVELING SALESMAN PROBLEM
nodes
102
1.08
101
100
10-1
10-2
103
nodes
103
4000
3500
3000
2500
1500
1000
2000
augmentation iters.
101
102
nodes
103
subtour factors.
103
100
102
nodes
103
102
101
100 1
10
102
nodes
103
103
101
100
10-1 1
10
102
100 1
10
103
101
augmentation iters.
nodes
3500
10-2
3000
10-1
nodes
10
2500
100
102
Random Distance2 Matrix
1500
optimality
101
4.5
4.0
3.5
3.0
2.5
2.0
1.5
1.0
2000
2
102
500
nodes
103
100
10-1 1
10
1000
102
nodes
1000
10-3 1
10
4000
10-2
103
101
103
101
10-1 1
10
102
nodes
102
100 1
10
103
10
800
10-1
102
nodes
Correlation Distance
2
600
100
1.8
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1.0
nodes
400
1
optimality
time (sec)
1000
1.02
102
10-3 1
10
800
1.06
1.04
0.98
103
1
10
200
nodes
10
103
103
Hamming Distance
2
1.00
102
102
100 1
10
augmentation iters.
1.10
10
400
nodes
103
10
600
1.0
103
200
102
nodes
optimality
time (sec)
10-3 1
10
10-3 1
10
augmentation iters.
1.1
10-2
time (sec)
1.2
500
10
-1
102
nodes
104
10
subtour factors.
optimality
100
101
100 1
10
103
Random 2D Euclidean
Distance
3
1.3
102
subtour factors.
1.4
101
102
nodes
subtour factors.
2
101
103
subtour factors.
1.5
102
100 1
10
1000
103
103
10
time (sec)
102
nodes
augmentation iters.
10-3 1
10
800
10-2
104
10
BP
NN
greedy
600
100
10-1
TSPLIB Instances
3
400
101
optimality
time (sec)
102
1.8
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1.0
200
103
119
102
nodes
103
102
101
100 1
10
102
nodes
103
Figure 5.3: Results of message passing for TSP on different benchmark problems. From left to right, the
plots show: (a) running time, (b) optimality ratio (compared to Concorde), (c) iterations of augmentation and
(d) number of subtours constraints – all as a function of number of nodes. The optimality is relative to the result
reported by Concorde. Note that all plots except optimality are log-log plots where a linear trend shows a monomial
relation (y = axm ) between the values on the x and y axis, where the slope shows the power m.
120
CHAPTER 5. PERMUTATIONS
Experiments
Here we evaluate our method over five benchmark datasets1 : (I) TSPLIB, which contains a variety of
real-world benchmark instances, the majority of which are 2D or 3D Euclidean or geographic distances.2
(II) Euclidean distance between random points in 2D. (III) Random (symmetric) distance matrices.
(IV) Hamming distance between random binary vectors with fixed length (20 bits). This appears in
applications such as data compression [153] and radiation hybrid mapping in genomics [29]. (V) Correlation distance between random vectors with 5 random features (e.g., using TSP for gene co-clustering
[69]). In producing random points and features as well as random distances (in (III)), we used uniform
distribution over [0, 1].
For each of these cases, we report the (a) run-time, (b) optimality, (c) number of iterations of
augmentation and (d) number of subtour factors at the final iteration. In all of the experiments, we
use Concorde [13] with its default settings to obtain the optimal solution.3 The results in figure 5.3
(2nd column from left) reports the optimality ratio – i.e., ratio of the tour found by message passing,
to the optimal tour. To demonstrate the non-triviality of these instance, we also report the optimality
ratio for two heuristics that have (1 + dlog2 (N )e)/2- optimality guarantees for metric instances [154]:
(a) nearest neighbour heuristic (O(N 2 )), which incrementally adds to any end of the current path the
closest city that does not form a loop; (b) greedy algorithm (O(N 2 log(N ))), which incrementally adds
a lowest cost edge to the current edge-set, while avoiding subtours.
All the plots in figure 5.3, except for the second column, are in log-log format. When using log-log
plot, a linear trend shows a monomial relation between x and y axes – i.e., y = axm . Here m indicates
the slope of the line in the plot and the intercept corresponds to log(a). By studying the slope of the
linear trend in the run-time (left column) in figure 5.3, we observe that, for almost all instances, message
passing seems to grow with N 3 (i.e., slope of ∼ 3). Exceptions are TSPLIB instances, which seem to
pose a greater challenge, and random distance matrices which seem to be easier for message passing. A
similar trend is suggested by the number of subtour factors and iterations of augmentation, which has
a slope of ∼ 1, suggesting a linear dependence on N . Again the exceptions are TSPLIB instances that
grow faster than N and random distance matrices that seem to grow sub-linearly.
Overall, we observe that augmentative message-passing is able to find near-optimal solutions in polynomial time. Although powerful branch-and-cut methods, such as Concorde, are able to exactly solve
instances with several thousands of variables, their general run-time on random benchmark instances
1
In all experiments, we used the full graph G = (V, E), which means each iteration of message passing is O(N 2 τ ), where
τ is the number of subtour factors. All experiments use Tmax = 200 iterations, max = median{Ai,j | i, j} and damping with
λ = .2. We used decimation, and fixed 10% of the remaining variables (out of N ) per iteration of decimation. Note that
here we are only fixing the top N variables with positive bias. The remaining M − N variables are automatically clamped
to zero. This increases the cost of message passing by an O(log(N )) multiplicative factor, however it often produces better
results.
2
Geographic distance is the distance on the surface of the earth as a large sphere.
3
For many larger instances, Concorde (with default setting and using CPLEX as LP solver) was not able to find the
optimal solution. Nevertheless we used the upper bound on the optimal produced by Concord in evaluating our method.
5.2. TRAVELING SALESMAN PROBLEM
121
remains exponential [13, p495], while our approximation on random instances appears to be O(N 3 ).
5.2.2
Using pairwise factors
Here we present an alternative factor-graph for finding permutations without subtours. This formulation
has O(N 2 ) factors and therefore the complete factor-graph remains tractable. However in practice, the
min-sum inference over this factor-graph is not as effective as the augmentation approach. Therefore,
here we use this factor-graph to solve min-max version of TSP, known as bottleneck TSP through
sum-product reductions.4
Given an asymmetric distance matrix A ∈ RN ×N , the task in the Bottleneck Traveling Salesman
Problem (BTSP) is to find a tour of all N points such that the maximum distance between two consecutive cities in the tour is minimized [156]. Any constant-factor approximation for arbitrary instances
of this problem is NP-hard [242].
Let x = {x1 , . . . , xN } denote the set of variables where xi ∈ Xi = {0, . . . , N − 1} represents the
time-step at which node i is visited. We assume modular arithmetic (module N ) on members of Xi –
e.g., N ≡ 0 mod N and 1 − 2 ≡ N − 1 mod N . For each pair xi and xj of variables, define the factor
f{i,j} (xi , xj ) = min 1(|xi − xj | > 1), max(Ai,j , 1(xi = xj − 1)), max(Aj,i , 1(xj = xi − 1))
(5.9)
where the tabular form is
xj
xi
0
1
···
N −2
N −1
0
∞
Ai,j
−∞
···
−∞
−∞
Dj,i
1
Aj,i
∞
Ai,j
···
−∞
−∞
−∞
2
..
.
−∞
..
.
Ai,j
..
.
∞
..
.
···
..
.
−∞
..
.
−∞
..
.
−∞
..
.
N −3
−∞
−∞
−∞
···
∞
Ai,j
−∞
N −2
−∞
−∞
−∞
···
Aj,i
∞
Ai,j
N −1
Ai,j
−∞
−∞
···
−∞
Aj,i
∞
⊗
Here, 1 = ∞ on diagonal enteries ensures that xi 6= xj . Moreover |xi − xj | > 1 means cities i and
⊕
j are not visited consecutively, so this factor has no effect ( 1 = −∞). However if two cities are visited
one after the other, depending on whether i was visited before j or vicee-versa, Ai,j or Aj,i represent
the distance between them. This factor can be easily converted to min-sum domains by replacing
4
A binary-variable formulation with similar symantics is proposed in [306]. However the authors applied their message
passing solution to an instance with only five cities.
122
CHAPTER 5. PERMUTATIONS
the identity values −∞, +∞ with −∞, 0 in the tabular form above. Alternatively we can replace the
identity, min and max operations in equation (5.9) with their corresponding min-sum functions.
Here we relate the min-max factor-graph above to a uniform distribution over Hamiltonian cycles
and use its sum-product reduction to find solutions to Bottleneck TSP. Recall G(A, y) is a graph in
which there is a connection between node i and j iff Ai,j ≤ y.
Proposition 5.2.1. For any distance matrix A ∈ RN ×N , the py -reduction of the BTSP factor-graph
above, defines a uniform distribution over the (directed) Hamiltonian cycles of G(A, y).
Proof. First note that py defines a uniform distribution over its support as its unnormalized value is
only zero or one. Here w.l.o.g we distinguish between two Hamiltonian cycles that have a different
starting point but otherwise represent the same tour. Consider the py -reduction of the pairwise factor
of equation (5.9)
f{i,j} (xi , xj ) = 1(|xi − xj | > 1) + 1(xi = xj − 1 ∧ Ai,j ≤ y)
+ 1(xi = xj + 1 ∧ Aj,i ≤ y)
(5.10)
(5.11)
• Every Hamiltonian cycle over G(A, y), defines a unique assignments x with py (x) > 0: Given the
Hamiltonian cycle H = h0 , h2 , . . . , hN −1 where hi ∈ {1, . . . , N } is the ith node in the path, for each
i define xi = j s.t. hj = i. Now we show that all pairwise factors of equation (5.10) are non-zero
for x. Consider two variables xi and xj . If they are not consecutive in the Hamiltonian cycle then
f{i,j} (xi , xj ) = 1(|xi − xj | > 1) > 0. Now w.l.o.g. assume i and j are consecutive and xi appears
before xj . This means (i, j) ∈ E and therefore Ai,j ≤ y, which in turn means f{i,j} (xi , xj ) = 1(xi =
xj − 1 ∧ Ai,j ≤ y) > 0 Since all pairwise factors are non-zero, py (x) > 0.
• Every x for which py (x) > 0, defines a unique Hamiltonian path over G(A, y): Given assignment
x, construct H = h0 , . . . , hN −1 where hi = j s.t.xj = i. Now we show that if p(x) > 0, H defines
a Hamiltonian path. If p(x) > 0, for every two variables xi and xi , one of the indicator functions of
equation (5.10) should evaluate to one. This means that first of all, xi 6= xj for i 6= j, which implies H
is well-defined and hi 6= hj for i 6= j. Since all xi ∈ {0, . . . , N − 1} values are distinct, for each xi = s
there are two variables xj = s − 1 and xk = s + 1 (recall that we are using modular arithmetic) for which
the pairwise factor of equation (5.10) is non-zero. This means Aj,i ≤ y and Ai,k ≤ y and therefore
(j, i), (i, k) ∈ E (the edge-set of G(A, y)). But by definition of H, hs = i, hs−1 = j and hs+1 = k are
consecutive nodes in H and therefore H is a Hamiltonian path.
This proposition implies that we can use the sum-product reduction of this factor-graph to solve
Hamiltonian cycle problems. The resulting factor-graph for Hamiltonian cycle problem is an special
case of our graph-on-graph technique, where the adjacency matrix of one graph is directly used to build
factors in a second graph. Here, as the tabular form of the factor above suggests, the second graph is a
simple loop of length N (see section 5.3).
5.2. TRAVELING SALESMAN PROBLEM
123
Figure 5.4: The min-max solution (using sum-product reduction) for Bottleneck TSP with different number of
cities (x-axis) for 2D Euclidean space (left) as well as asymmetric random distance matrices (right) with T = 5000
for Perturbed BP. The error-bars in all figures show one standard deviation over 10 random instances.
As expected, due to sparse form of this pairwise factor, we can perform BP updates efficiently.
Claim 5.2.2. The factor-to-variable BP messages for the sum-product reduction of factor of equation (5.9) can be obtained in O(N ).
Proof. The py -reduction of the min-max factors of equation (5.9) is given by:
f{i,j} (xi , xj ) = 1(f{i,j} (xi , xj ) ≤ y)
(5.12)
= 1(|xi − xj | > 1) + 1(xi = xj − 1 ∧ Ai,j ≤ y)
(5.13)
+ 1(xi = xj + 1 ∧ Aj,i ≤ y)
(5.14)
The matrix-form of this factor (depending on the order of Ai,j , Aj,i , y) takes several forms all of which
P
are band-limited. Assuming the variable-to-factor messages are normalized (i.e., xi b
pj→I (xi ) = 1) the
factor-to-variable message is given by
b
p{i,j}→i (xi ) = 1 − b
pj→{i,j} (xi )+
1(Ai,j ≤ y)(1 − b
pj→{i,j} (xi − 1))+
1(Aj,i ≤ y)(1 − b
pj→{i,j} (xi + 1))
Therefore the cost of calculating factor-to-variable message is that of normalizing the variable-tofactor message, which is O(N ).
Since there are N 2 pairwise factors, this gives O(N 3 ) time-complexity for each iteration of sumproduct BP in solving the Hamiltonian cycle problem (i.e., the sum-product reduction) and O(N 3 log(N ))
for bottleneck-TSP, where the log(N ) factor is the cost of binary search (see equation (1.25)).
124
CHAPTER 5. PERMUTATIONS
Figure 5.4 reports the average performance of message passing (over 10 instances) as well as a lower
bound on the optimal min-max value for tours of different length (N ). Here we report the results for
random points in 2D Euclidean space as well as asymmetric random distance matrices. For Euclidean
problems, the lower bound is the maximum over j of the distance of two closest neighbors to each
node j. For the asymmetric random distance matrices, the maximum is over all the minimum length
incoming edges and minimum length outgoing edges for each node.5
5.3
Graph matching problems
Consider two graphs G = (V, E) and G 0 = (V 0 , E 0 ), with (weighted) adjacency matrices A, A0 respectively. Here we enumerate several of the most important problems over permutations based on two
graphs [72]. Section 5.3.1 introduces the graphical model for the problems of (sub-graph) isomorphism
and monomorphism. Then we study the message passing solution to graph homomorphism, use it to
find symmetries in graphs. In section 5.3.5 we introduce a general factor-graph for graph alignment
based on the previous work of Bradde et al. [44] and show how it can model other problems such as
quadratic assignment problem and maximum common subgraph. The common idea in all these settings
is that a “variation” of the adjacency matrix A0 of graph G 0 is used as a pairwise factor over the edges
(or/and non-edges) of graph G (with adjacency A), defining a Markov network (a factor-graph with
pairwise factors).
5.3.1
Sub-graph isomorphism
The graph isomorphism problem asks whether G = (V, E) and G 0 = (V 0 , E 0 ) are identical up to a
permutation of nodes (written as G ∼
= G 0 ). That is, it seeks a one-to-one mapping π : V → V 0 such
that (i, j) ∈ E ⇔ (π(i), π(j)) ∈ E 0 . With some abuse of notation we also write π(G) = G 0 . Although,
there have been polynomial time solutions to special instances of graph isomorphism problem [95], the
general case has remained elusive. So much so that we do not know whether the problem is NP-complete
(e.g., see [222]).
A permutation π ∈ SN (recall SN is the symmetric group) such that π(G) ∼
= G is called an automorphism. The automorphisms of G, under composition form a group, called the automorphism
group Aut(G). The automorphism group also defines a natural notion of symmetry on the nodes of
graphs. Here, the orbit of each node, is the set of nodes that are mapped to i in any automorphism –
def
i.e., orbit(i) = {π(i) | π ∈ Aut(G)}. The orbits partition the set of nodes V into group of nodes that are
in a sense symmetric, which makes them a prime candidate for defining symmetry in complex networks
[198, 321].
5
If for one node the minimum length in-coming and out-going edges point to the same city, the second minimum length
in-coming and out-going edges are also considered in calculating a tighter bound.
5.3. GRAPH MATCHING PROBLEMS
125
Sub-graph isomorphism asks whether G 0 is isomorphic to a vertex induced subgraph of G. –
i.e., it seeks an injective mapping π : V 0 → V where
(π(i), π(j)) ∈ E ⇔ (i, j) ∈ E 0
∀i, j ∈ V 0
When dealing with sub-graph morphisms, we assume that the mapping is from the smaller graph to the
larger graph and therefore |V| ≤ |V 0 |.
The factor-graph for subgraph isomorphism is defined as follows: We have one variable per i ∈ V,
where the domain of each variable xi is V 0 – i.e., x = {xi ∈ V 0 | i ∈ V}. The factor-graph has two types
of pairwise factors:
• Edge factors: ensure that each edge in G = (V, E) is mapped to an edge in G 0 = (V 0 , E 0 )
f{i,j} (x{i,j} ) = 1((xi , xj ) ∈ E 0 ) ∀(i, j) ∈ E
(5.15)
where assuming the tabular form of this factor for sum-product semiring is simply the adjacency matrix
A0 ∈ {0, 1}|V
0 |×|V 0 |
of G 0 .
• Non-edge factors: ensure that each non-edge in G = (V, E) is mapped to a non-edge in G 0 = (V 0 , E 0 )
f{i,j} (x{i,j} ) = 1((xi , xj ) ∈
/ E 0 ∧ xi 6= xj )
∀i, j ∈ V, i 6= j, (i, j) ∈
/E
(5.16)
where again for sum-product semiring, the tabular form takes a simple form w.r.t. the binary valued
adjacency matrix of G 0 – i.e., 1 − A0 . Using sum-product semiring, this fully connected Markov network
defines a uniform distribution over the (subgraph) isomorphisms from G to G 0 . We could use sumproduct BP with decimation or perturbed BP to sample individual assignments x∗ ∼ p(x). Here, each
assignment x ≡ π is an (injective) mapping from V to V 0 , where xi = j 0 means node i ∈ V is mapped
to node j 0 ∈ V 0 .
In particular, for G = G 0 , the integral is equal to the cardinality of the automorphism group q(∅) =
|Aut(G)| and two nodes i, j ∈ V are in the same orbit iff the have the same marginals – i.e.,
p(xi ) = p(xj ) ⇔ i ∈ orbit(j)
This also suggests a procedure for finding (approximate) symmetries in graphs. We can use sum-product
BP to find marginals and group the nodes based on the similarity of their marginals. However, the cost
of message passing in this graphical model is an important barrier in practice.
Claim 5.3.1. Assuming |V| ≤ |E| ≤ |V|2 and |V 0 | ≤ |E 0 | ≤ |V 0 |2 and using variable synchronous update,
the time complexity of each iteration of sum-product BP for subgraph isomorphism is O(|V|2 |E 0 |).
Proof. First, we calculate the cost of sending sum-product BP messages through the edge and non-edge
factors. For the edge factors (A0 ), we can go through each row of the tabular form of the factor and multi-
126
CHAPTER 5. PERMUTATIONS
ply the non-zero entries with corresponding message – i.e., g{i,j} (xi , xj ) = b
pi→{i,j} (xi )f{i,j} (xi , xj ). Then
P
we add the values in each column, to obtain the outgoing message – i.e., b
p{i,j}→j (xj ) = xi g{i,j} (xi , xj ).
This procedure depends on the number of non-zero entries – i.e., O(|E 0 |). The procedure is similar for
non-edge factors. Therefore the overall cost of sending sum-product BP messages through all O(|V|2 )
factors is O(|E 0 | |V|2 ). Using variable synchronous update, calculating all variable-to-factor messages
takes O(|V|2 |V 0 |). Using the assumption of the claim, the overall cost is therefore O(|E 0 | |V|2 ).
However it is possible to improve this complexity by considering sparse mappings. For example, we
can restrict the domain of each variable xi to all the nodes j 0 ∈ V 0 that have the same degree with node
i, or furthermore to all the nodes that also have neighbours with same degree as the neighbours of node
i.
5.3.2
Subgraph monomorphism and supermorphism
Sub-graph monomorphism relaxes the constraint of the subgraph isomorphism to
(i, j) ∈ E ⇒ (π(i), π(j)) ∈ E 0
∀i, j ∈ V
(5.17)
where π : V 0 → V has to be injective – i.e., nodes in V 0 are mapped to distinct nodes in V. However, G 0 is
allowed to cover a “subset” of edges in an induced subgraph of G. We note here that previous graphical
models introduce in [43, 44] for isomorphism in fact define monomorphism. The only difference between
this factor-graph and that of sub-graph isomorphism is that the non-edge factors are replaced by the
following uniqueness factors.
• Uniqueness factors are inverse Potts factors that ensure disconnected nodes are mapped to different
nodes (– i.e., the mapping is injective)
/E
f{i,j} (x{i,j} ) = 1(xi 6= xj ) ∀i.j ∈ V, j 6= i, (i, j) ∈
(5.18)
Despite this difference, when G = G 0 – that is when we are interested in automorphisms – or more
generally when |E| = |E 0 | and |V| = |V 0 |, the distribution defined by the monomorphism factor-graph is
identical to that of isomorphism factor-graph.
Claim 5.3.2. Assuming |V| ≤ |E| ≤ |V|2 and |V 0 | ≤ |E 0 | ≤ |V 0 |2 , using variable synchronous update,
the time complexity of each iteration of sum-product BP for sub-graph monomorphism O(|E| |E 0 | +
|V 0 | |V|2 ).6
Proof. The complexity of sending sum-product BP messages through the edge factors is O(|E 0 |) (see
proof of Claim 5.3.1). However, the uniqueness factors are inverse Potts factors and allow O(|V 0 |)
Bradde et al. [44] suggest a trick to reduce this time-complexity to O((|E| |V|)|V 0 |), but unfortunately details are
omitted and we are unable to follow their route.
6
5.3. GRAPH MATCHING PROBLEMS
127
calculation of messages. This means the overall cost of sending messages through all the factors is
O(|E| |E 0 | + |V 0 | |V|2 ). The cost of calculating variable-to-factor messages using variable-synchronous
update is also O(|V|2 |V 0 |), which gives the overall complexity of O(|E| |E 0 | + |V 0 | |V|2 ) per BP iteration.
So far we have seen two variations of mapping a graph G to a subgraph of G 0 . In subgraph isomorphism, the image of G strictly agrees with a sub-graph of G 0 . In monomorphism, the image is contained
within a sub-graph. A third possibility is to ask for a mapping such that the image of G “contains” a
sub-graph G 0 :
(i, j) ∈ E ⇐ (π(i), π(j)) ∈ E 0
∀i, j ∈ V
(5.19)
where again π : V 0 → V has to be injective. Due to lack of a better name, we call this mapping
subgraph supermorphism.
The factor-graph for sub-graph supermorphism has two types of factors, both of which we have seen
before: 1) The non-edge factors between non-edges (equation (5.16)) ensure that the image contains
a sub-graph of G 0 while 2) uniqueness factors between the edges (i, j) ∈ E ensure that the mapping is
injective.
Claim 5.3.3. Assuming |V| ≤ |E| ≤ |V|2 and |V 0 | ≤ |E 0 | ≤ |V 0 |2 , using variable synchronous update, the
time complexity of each iteration of sum-product BP for subgraph supermorphism is O((|V 2 |−|E|) |E 0 | +
|V 0 | |V|2 )
Proof. The complexity of sending sum-product BP messages through the non-edge factors is O(|E 0 |)
(see proof of Claim 5.3.1) and each uniqueness factor require O(|V 0 |) computation. This means the
overall cost of sending messages through all the factors is O((|V 2 | − |E|) |E 0 | + |V 0 | |E|). The cost
of calculating variable-to-factor messages using variable-synchronous update is also O(|V|2 |V 0 |), which
gives the overall complexity of O((|V 2 | − |E|) |E 0 | + |V 0 | |V|2 ) per BP iteration.
5.3.3
Graph homomorphism
Graph homomorphism [129] further relaxes the constraint of the sub-graph monomorphism such that
a homomorphic mapping can map two distinct nodes in V to the same node in V 0 . However, equation (5.17) should hold, and therefore if two nodes in V 0 are adjacent, they should still be mapped to
distinct and adjacent nodes in V. Compared to other areas of graph theory, the study of rich structure
and surprising properties of homomorphisms is relatively recent. Study of graph homomorphisms covers
diverse areas including property testing [9], graph sequences and limits [37, 39, 40, 197] and constraint
satisfaction problems [223]. Many of these areas are interested in “counting” the number of homomorphisms [38] from a graph G to G 0 . As we see shortly, graph homomorphism reduces to many interesting
CSPs and therefore it is not hard to show that it is NP-complete. Moreover, counting the number of
128
CHAPTER 5. PERMUTATIONS
Figure 5.5: (left) The number of endomorphisms for all distinct graphs up to 8 nodes compared to the BP
integral. Here, larger disks represent graphs with smaller number of nodes. (right) Comparison of normalized
BP marginals and the exact marginals for endomorphism graphical model.
homomorphisms is #P-complete [86]. The set of all homomorphisms π of a graph G to itself –i.e., its
endomorphisms – under composition form endomorphism monoid (see definition 1.1.1 for monoid).
Here, the identity element simply maps each element to itself. Our interest in endomorphism is because
through the conjecture 5.3.4, we can use it (instead of automorphism) to approximate symmetries in
graphs.
The graphical model representation of homomorphism has been investigated in different contexts [50,
52], but to our knowledge, message passing has not been previously used for counting and finding homomorphisms. The Markov network for homomorphism resembles that of isomorphism and monomorphism: The variables are x = {xi | i ∈ V} where xi ∈ V 0 , and the factor-graph only contains edge-factors
of equation (5.15). Assuming |V 0 | ≤ |E 0 | and |V| ≤ |E|, it is easy to see that the complexity of variablesynchronous message passing is O(|E| |E 0 |), which makes this method very efficient for sparse graphs.
This graphical model can be extended to represent homomorphism in weighted graphs [38]. For this
define the edge-factors as
A0x
f{i,j} (x{i,j} ) = Ai,j i
,xj
∀i, j ∈ V, i 6= j, Ai,j > 0
For small graphs we can exactly count the number of homomorphisms and endomorphisms. To
evaluate the accuracy of message passing, we compared the BP estimates with the exact number of
endomorphisms, for all isomorphically distinct graphs with |V| < 9 (i.e., > 13, 000 instances); Figure 5.5
reports this result as well as the accuracy of BP marginals, suggesting that despite existence of short
loops in these graphs BP estimates are relatively accurate.
5.3. GRAPH MATCHING PROBLEMS
129
Reduction to other CSPs
The relation between homomorphism and other CSPs is well-known [223]. Here, we review this relation
between the factor-graphs of this section and other CSP factor-graphs we encountered in this thesis.
Coloring and clique cover: (see example 3.0.1 and section 3.4) K-coloring of G corresponds to finding
a homomorphism from G to KK , the complete graph of order K. Similarly the homomorphism from the
complement of G to Kk corresponds to K-clique-cover. The relation is also reflected in the corresponding
factor-graphs as the adjacency matrix for KK is the inverse Potts model, used in K-coloring.
Clique problem and independent-set: (see section 3.6) Recall that K-clique problem corresponds
to finding a clique of size K in G. This is a special case of sub-graph isomorphism from KK to G
which is in this case identical to sub-graph monomorphism and homomorphism from KK to G. The
(sum-product reduction of the) categorical-variable model of section 3.6.2 is a fully connected Markov
network with edge factors that we used for isomorphism, monomorphism and homomorphism equation (5.15). Similarly, the K-independent set problem is equivalent to finding homomorphisms from KK
to the complement of G. Independent set has an alternative relation with graph homomorphism: any
homomorphism from G to a graph with two connected nodes and a self-loop on one of the nodes defines
an independent set for G.
Hamiltonian cycle problem corresponds to sub-graph monomorphism from C|V| , the cycle of length |V|,
to G. Alternatively, we can formulate it as subgraph supermorphism from G to C|V| . The sum-product
reduction of our min-max formulation for bottleneck TSP in section 5.2.2 is indeed the factor-graph of
sub-graph supermorphism from G to C|V| .
5.3.4
Finding symmetries
One may characterize a graph G using the “number” of homomorphism from/to other graphs G 0 . This
characterization is behind the application of graph homomorphism in property testing and definition
of graph sequences. Let Hom(H, G) be the set of homomorphism from H to G – i.e., the set of all
assignments x where p(x) > 0. Let H1 , . . . , HM be the sequence of all graphs whose number of nodes is
at most |V|. Then, the Lovász vector of G which is defined as
v(G) = (|Hom(H1 , G)|, |Hom(H2 , G)|, . . . , |Hom(HM , G)|)
(5.20)
uniquely identifies G up to an isomorphism [196].
Here, rather than identifying a particular graph G within the set of all graphs, we are interested
in identifying a node i ∈ V of a single graph G within the set of all nodes V. Note that here both
identifications are up to an isomorphism. Our objective is equivalent to finding the orbits of G and our
approach in finding the orbits using graph homomorphism (rather than isomorphism) is founded on the
following conjecture.
130
CHAPTER 5. PERMUTATIONS
Figure 5.6: The table on the right shows the unnormalized endomorphism marginals for the graph on the left.
Here row i corresponds to q(xi ) (normalization of which gives p(xi )), and the number at row i and column j is
the number of times node i is mapped to j in an endomorphism. The total number of endomorphisms for this
graph is q(∅) = 78. Here, the orbits are {2, 3}, {5, 6}, {1, 7}, {4}. However, q(x1 ) = q(x7 ) = q(x4 ) – that is node
4 maps “to” other nodes with the same frequency as nodes 1 and 7. However, the mappings to node 4 ( i.e., the
4th column of the table) remains different from the mappings to 1 and 7. Here, as predicted by conjecture 5.3.4,
nodes with similar rows and columns belong to the same orbit.
Conjecture 5.3.4. Given the uniform distribution over the endomorphisms of G:
p(x) ∝
Y
1((xi , xj ) ∈ E)
(i,j)∈E
the necessary and sufficient condition for i and j to be in the same orbit is
p(xi ) = p(xj ) and ∀k ∈ V : p(xk = i) = p(xk = j)
⇔
orbit(i) = orbit(j)
Note that p(xi = k) is simply the relative frequency of mapping of node i to node k in an endomorphism. Therefore this conjecture simply states that for node i and j to be equivalent up to an
automorphism of G, it is necessary and sufficient for them to have the same frequency of mapping
to/from all other nodes of G. While it is trivial to prove necessity (⇐), we found it difficult to prove
the statement in the other direction.
Therefore, similar to other related conjectures [163, 204], we turn to computational verification.
For this, we experimented with distinct graphs with up to 9 nodes (i.e., > 286, 000 instances). Here, we
obtained the exact marginals p(xi ) ∀i ∈ V and constructed the orbits as suggested by conjecture 5.3.4.
We then obtained the exact orbits using the software of McKay and Piperno [205]. In all cases two
partitioning of the nodes to orbits were identical.
We also note that in conjecture 5.3.4 restricting the condition to p(xi ) = p(xj ) – i.e., similar
frequency of mappings “to” other nodes – is not sufficient. Figure 5.6 shows a counter-example for this
weaker condition.
Example 5.3.1. In figure 5.7(left) we used the homomorphism marginals to find approximate symmetries in a graph with visible symmetries, known as Thomassen graph. Here, after obtaining the
marginals we used Principle Component Analysis (PCA) with whitening [226] to extract three values
5.3. GRAPH MATCHING PROBLEMS
131
Figure 5.7: Coloring of nodes by reducing the dimensionality of marginals to three dimensions of RGB using
PCA with whitening. (left) The marginals are calculated using sum-product BP. (right) The marginals of the
Kronecker graph is obtained using Gibbs sampling and annealing. The graph in the middle is the product of two
graphs on the top and right – i.e., G = G1 × G2 . Two nodes in G are connected iff the corresponding nodes in G1
and G2 are connected. Note that the product of each two colors produces the same color in the product graph and
similarly colored nodes have similar neighbours.
for RGB colors. As the figure suggests, this approach is able to identify similar nodes with similar
colors.
For dense graphs, message passing is no longer accurate. In figure 5.7(right) we use Gibbs sampling
with annealing to estimate the endomorphism marginals in Kronecker product[190] of two random
graphs. Here again, we use PCA to color the nodes. Note that the algorithm is unaware of the product
format. The choice of the product graph is to easily observe the fact that the product of similarly
colored nodes produce similarly colored nodes in the product graph.
An alternative is to use spectral clustering on the matrix of marginals to obtain clusters. In a related
context Krzakala et al. [180] use the matrix of non-backtracking random-walks to find symmetric clusters
in stochastic block models. Note that the basic difference with our approach is the matrix used with
the spectral clustering. Other notable matrices that are used within this context are the Laplacian
matrix [299], the modularity matrix [228] and the Bethe hessian [272]. The relation between the
clustering (in its conventional sense) and orbits of a graph is better understood in the extreme case:
when clusters form isolated cliques, they are identical to orbits.
5.3.5
Graph alignment
The graph alignment problem can be seen as the optimization counterpart of the decision problems of
isomprphism, monomorphism and homomorphism. In the past, different methods have tried to optimize
132
CHAPTER 5. PERMUTATIONS
a variety of different objectives [53, 72]. In the context of message passing two distinct approaches have
been used: 1) Bayati et al. [26] propose a factor-graph for “sparse” graph alignment and show that it
scales to very large instances. Here the term sparsity both refers to the number of edges of G and G 0
and also to the restricted possibility of matching nodes of V to V 0 – i.e., each node in V can match only
a few predetermined nodes in V 0 . The factor-graph used by the authors resembles the binary version of
the maximum bipartite matching factor-graph of section 5.1.1. 2) Bradde et al. [44] used the min-sum
BP to minimize the number of misalignment in a factor-graph similar to that of graph monomorphism
above. Here we follow their route, with the distinction that we suggest using the min-sum semiring with
“sub-graph isomorphism” factor-graph and account for different matching costs using several tricks.
Here we consider a general objective function that evaluates the mapping π : V 0 → V ∪ {null}. We
then show how to optimize this objective using max-sum inference in a factor-graph.
• Node matching preference for matching node i ∈ V to j 0 ∈ V 0 : ϕ(i, j 0 ) : V × V 0 → R.
• Edge matching preference for matching (i, j) ∈ E to (i0 , j 0 ) ∈ E 0 : ς((i, j), (k 0 , l0 )) : E × E 0 → R.
• Node merging preference for mapping nodes i, j ∈ V to the same node k ∈ V 0 : ϑ(i, j, k 0 ) :
V × V × V 0 → R.
• Node deletion preference δ(i) : V → R, is the penalty for ignoring the node i ∈ V – i.e., mapping
it to the null node.
• Edge deletion preference is the preference for dropping (i, j) ∈ E: $(i, j) : E → R.
• Edge insertion preference is the preference for adding an edge (i0 , j 0 ) ∈ E 0 , when it is not matched
against any edge in E: υ(i0 , j 0 ) : E 0 → R.
We can define these preferences in such a way that the optimal solution is also a solution to an
interesting decision problem.
Example 5.3.2. The optimal alignments with the following parameters reproduce Hom(G, G 0 ):
• node matching ϕ(i, j 0 ) = 0 ∀ i ∈ V, j 0 ∈ V 0
• edge matching ς((i, j), (k 0 , l0 )) = 1
• node merging ϑ(i, j, k 0 ) = 0
• node deletion δ(i) = −∞
∀ (i, j) ∈ E, (k 0 , l0 ) ∈ E 0
∀ i, j ∈ V, k 0 ∈ V 0
∀i∈V
• edge deletion $(i, j) = −∞
∀ (i, j) ∈ E
• edge insertion υ(k 0 , l0 ) = 0 ∀ (k 0 , l0 ) ∈ E 0
Alternatively, using positive and uniform node and edge matching preferences, if we set the merge
cost to −∞ and allow node deletion at zero cost, the optimal solution will be the maximum common
subgraph. In particular, for maximum-edge common subgraph we use
• node matching ϕ(i, j 0 ) = 0 ∀ i ∈ V, j 0 ∈ V 0
• edge matching ς((i, j), (k 0 , l0 )) = 1
∀ (i, j) ∈ E, (k 0 , l0 ) ∈ E 0
5.3. GRAPH MATCHING PROBLEMS
133
• node merging ϑ(i, j, k 0 ) = −∞
∀ i, j ∈ V, k 0 ∈ V 0
• node deletion δ(i) = 0 ∀ i ∈ V
• edge deletion $(i, j) = 0 ∀ (i, j) ∈ E
• edge insertion υ(k 0 , l0 ) = 0 ∀ (k 0 , l0 ) ∈ E 0
Given two weighted adjacency matrices A and A0 (for G and G 0 respectively), where Ai,j is the
“flow” between the facilities i and j, while A0k0 ,l0 is the “distance” between the locations k 0 and l0 , the
quadratic assignment problem, seeks a one-to-one mapping π ∗ : V → V 0 of facilities to locations in
order to optimize the flow
π ∗ = argπ max
X
Ai,j A0π(i),π(j)
i,j∈V
Here w.l.o.g. we assume all weights are positive and set the alignment preferences so as to optimize the
quadratic assignment problem:
• node matching ϕ(i, j 0 ) = 0 ∀ i ∈ V, j 0 ∈ V 0
• edge matching ς((i, j), (k 0 , l0 )) = Ai,j A0k0 ,l0
• node merging ϑ(i, j, k 0 ) = −∞
• node deletion δ(i) = −∞
∀ (i, j) ∈ E, (k 0 , l0 ) ∈ E 0
∀ i, j ∈ V, k 0 ∈ V 0
∀i∈V
• edge deletion $(i, j) = −∞
• edge insertion υ(k 0 , l0 ) = −∞
∀ (i, j) ∈ E
∀ (k 0 , l0 ) ∈ E 0
Now we define the factors based on various alignment preferences. The factor-graph has one variable
per node i ∈ V: x = {xi | i ∈ V}, where xi ∈ V 0 ∪ {null}. Here, xi = null corresponds to ignoring
this node in the mapping. The alignment factor-graph has three type of factors:
• Local factors: take the node matching preferences and node deletion preference into account:
δ(i)
xi = null
fi (xi ) =
ϕ(i, x ) otherwise
i
∀i ∈ V
• Edge factors: are defined for each edge (i, j) ∈ E and partly account for edge matching, node
merging and edge deletion:
0
ϑ(i, j, x )
i
fi (xi , xj ) =
ς((i, j), (xi , xj ))
$(i, j)
xi = null ∨ xj = null
xi = xj
(xi , xj ) ∈ E 0
otherwise
∀(i, j) ∈ E
134
CHAPTER 5. PERMUTATIONS
• Non-edge factors: are defined for non-existing edge i, j 6= i ∈ V, (i, j) ∈
/ E and partly account for
node merging and edge insertion:
0
ϑ(i, j, x )
i
fi (xi , xj ) =
υ(xi , xj )
0
xi = null ∨ xj = null
xi = xj
(xi , xj ) ∈ E 0
∀i, j 6= i ∈ V, (i, j) ∈
/E
otherwise
This factor-graph in its general form is fully connected. The cost of max-sum message-passing
through each of these factors is O(|V 0 | log(V 0 |)), which means each iteration of variable synchronous
max-sum BP is O(|V|2 |V 0 | log(V 0 |)). However, if the matching candidates are limited (a.k.a. sparse
alignment), and the graphs G and G 0 are sparse, this cost can be significantly reduced in practice.
Example 5.3.3. Figure 5.8 shows a matching of E-coli metabolic network against a distorted version,
where 50% of edges were removed and the same number of random edges were added. Then we generated
10 matching candidate for each node of the original graph, including the correct match and 9 other
randomly selected nodes. We used graph alignment with the following preferences to match the original
graph against the distorted version7
• node matching ϕ(i, j 0 ) = 0 ∀ i ∈ V, j 0 ∈ V 0
• edge matching ς((i, j), (k 0 , l0 )) = 1 ∀ (i, j) ∈ E, (k 0 , l0 ) ∈ E 0
• node merging ϑ(i, j, k 0 ) = −∞
• node deletion δ(i) = −∞
∀ i, j ∈ V, k 0 ∈ V 0
∀i∈V
• edge deletion $(i, j) = 0 ∀ (i, j) ∈ E
• edge insertion υ(k 0 , l0 ) = 0 ∀ (k 0 , l0 ) ∈ E 0
We observed that message passing using our factor-graph was able to correctly match all the nodes
in two graphs.
7
We used T = 100 initial iterations with damping parameter λ = .2. After this, we used decimation and fixed ρ = 1%
of variables after each T = 50 iterations. The total run-time was less than 5 minutes.
5.3. GRAPH MATCHING PROBLEMS
135
Figure 5.8: Matching the E-coli metabolic network against a highly distorted version using message passing.
Here 50% of |E| = 4306 edges in the original network (left) are removed and the same number of random edges
are added, to produce the distorted network ( i.e., |E 0 | = |E|). Each node had 10 random candidates for matching
(including the correct choice) and message passing was able to identify the correct matching with 100% accuracy.
Conclusion
This thesis studied a general form of inference in graphical models with an emphasis on algebraic
abstractions. We organized an important subset of these inference problems under an inference hierarchy
and studied the settings under which distributive law allows efficient approximations in the form of
message passing. We investigated different methods to improve this approximation in loopy graphs
using 1) variational formulation and loop correction; 2) survey propagation; 3) hybrid techniques.
We then studied graphical modelling of combinatorial optimization problems under different modes of
inference.
As with any other inference and optimization framework, graphical modeling has its pros and cons.
The cons of using graphical models for combinatorial optimization are twofold a) implementing message
passing procedures, when compared to other standard techniques such as using Integer Programming
solvers, is more complex and time consuming. This is further complicated by b) the fact that there is
no standard guideline for designing a factor-graph representation, so as to minimize the computational
complexity or increase the quality of message passing solution. Indeed we used many tricks to efficiently
approximate the solution to our problems; example include simplification of BP messages through
alternative normalization, augmentation, variable and factor-synchronous message update, introduction
of auxiliary variables, using damping and decimation etc.
On the other hand, when dealing with large scale and difficult optimization problems, one has to
resort to conceptual and computational decomposition, and graphical modelling and message passing
techniques are the immediate candidates. Message passing is mass parallelizable, scalable and often
finds high-quality solutions. By providing factor-graphs for a diverse set of combinatorial problems,
this thesis also was an attempt to establish the universality of message passing. Of course some of
these problems better lend themselves to graphical modelling than some others, resulting in better
computational complexity and quality of results. Table 5.1 summarizes some important information
about the message passing solutions to combinatorial problems that are proposed or reviewed in this
thesis.
5.3. GRAPH MATCHING PROBLEMS
137
Table 5.1: Summary of message-passing solutions to combinatorial problems. The time complexity is for one
iteration of message passing. We report different costs for different update schedules for the same problem. See
the text for references. Here, N is the number of nodes and M is the number of constraints/factors.
Problem
Belief Propagation
Semiring ops.
*
not (min, max)
*
Perturbed BP
Survey Propagation
K-satisfiability
K-coloring
K-clique-cover
(+, ×)
*
(+, ×)
min set-cover
K-independent-set
max independent set
async.
async.
reduces to BP
async.
-
v-sync.
O(KM )
P
(f,v)-sync.
(+, ×)
O(K(N 2 − |E|))
P
O(KN 2 + i∈V |E(i, ·)|2 + |E(·, i)|2 )
v-sync.
O(KN + |E|)
f-sync.
(+, ×)
async.
a-sync.
async.
K-clique-cover
≡ to K-coloring on G c
-
(min, +)
O(|E|)
(f,v)-sync.
similar to K-median
(+, ×)
O(N 3 )
async.
binary var. model
f-sync.
≡ K-independent-set on G c
O(KN 2 )
O(KN + |E|)
(f,v)-sync.
(min, max)
O(log(N )(KN + |E|))
(f,v)-sync.
(+, ×) reduction ≡ K-independent-set
(min, max)
O(KN + |E|)
(f,v)-sync.
min-max BP
(+, ×)
(min, max)
(max, +)
O(K 3 N 2 )
async.
categorical var. model
O(K 2 N 2 )
v-sync.
≡ K-independent-set on G c
O(K 2 N 2 log(N ))
v-sync.
(+, ×) reduction ≡ K-independent-set
O(|E|)
(f,v)-sync.
≡ max independent-set
& min vertex cover
sphere-packing
reduces to Gibbs samp. & BP
v-sync.
v-sync.
& K-clique
K-packing
f-sync.
O(K 2 M )
2
i |E(i, ·)| )
relation to others
-
v-sync.
O(K|E|)
P
O(K i (N − |E(i, ·))|2 )
& K-clique
K-packing
async.
O(K
& K-set-cover
K-independent-set
Schedule
(+, ×)
(+, ×)
K-dominating-set
Complexity
P
P
O( i |Xi ||∂i|2 + I |XI ||∂I|)
P
P
O( i |Xi ||∂i| + I |XI ||∂I|)
P
P
O( i |Xi ||∂i| + I |XI |)
P
P
2
O( i |Xi ||∂i| + I |XI ||∂I|)
P
P
O( i |Xi ||∂i| + I |XI ||∂I|)
P
P
O( i 2|Xi | |∂i|2 + I 2|XI | |∂I|)
P
O(K 2 M + i |∂i|2 )
(+, ×)
(Hamming)
n: digits
y: min dist.
O(K 2 22n )
v-sync.
O(K 3 n + K 2 n2 y)
async.
O(K 2 n2 y)
v-sync.
O(K 2 ny)
(f,v)-sync.
K-medians
(min, +)
O(|E|)
f-sync.
facility location
(min, +)
O(|E|)
f-sync.
d-depth min span. tree
(min, +)
O(d|E|)
v-sync.
prize-coll. Steiner tree
(min, +)
O(d|E|)
v-sync.
-
a.k.a. affinity propagation
K-clustering
(min, max)
O(KN 2 log(N ))
v-sync.
(+, ×) reduction ≡ K-clique-cover
K-center
(min, max)
O(log(N )(KN + |E|))
f-sync.
(+, ×) reduction ≡ K-set-cover
O(KN + |E|)
Modularity max
(min, +)
min-max BP
-
clique model, using augmentation
O(Kmax N 2 )
max matching
(min, +)
O(N 2 )
Potts model
v-sync
-
-
-
v-sync
-
& cycle cover
bottleneck assignment
(min, max)
-
max b-matching
(min, +)
O(bN 2 )
TSP
(min, +)
O(N 2 τ ) or ∼ N 3
bottleneck TSP
(min, max)
O(N 3 log(N ))
(f,v)-sync
async
(+, ×) reduction ≡ Hamiltonian cycle
subgraph isomorphism
(+, ×)
O(|V|2 |E 0 |)
v-sync.
G → G0
subgraph monomorphism
(+, ×)
O(|V|2 |V 0 | + |E| |E 0 |)
v-sync.
G → G0
subgraph supermorphism
(+, ×)
O((|V 2 | − |E|) |E 0 | + |V 0 | |V|2 )
v-sync.
G → G0
homomorphism
(+, ×)
O(|E| |E 0 |)
v-sync.
G → G0
graph alignment
(max, +)
O(|V|2 |V 0 | log(|V 0 |))
v-sync.
G → G 0 with general costs
max common sub-graph
a variation of graph alignment
quadratic assignment
a variation of graph alignment
138
CHAPTER 5. PERMUTATIONS
Future work
Due to breadth of the models and problems that we covered in this thesis, our investigation lacks
the deserved depth in many cases. This is particularly pronounced in the experiments. Moreover, we
encountered many new questions and possibilities while preparing this thesis. Here, we enumerate some
of the topics that demand a more in depth investigation in the future work
• Many of the problems that we discussed also have efficient LP relaxations that some times come
with approximation guarantees. A comprehensive experimental comparison of message passing
and LP relaxations for these problems, both in terms of speed and accuracy is highly desirable.
• Our algebraic approach to inference suggests that all message passing procedures discussed here,
including survey propagation, are also applicable to the domain of complex numbers. Extensions
to this domain not only may allow new applications (e.g., using Fourier coefficients as factors) but
may also produce better solutions to many problems that we have studied here (e.g., in solving
CSPs).
• Our study of using graph homomorphism and its application to finding symmetries is a work in
progress. In particular its relation to other methods such as stochastic block models and spectral
techniques needs further investigation.
• Although some preliminary results on Ising model suggested that using sum-product reductions
for min-max inference performs much better than direct min-max message passing, an extensive
comparison of these two approaches to min-max inference is missing in our analysis.
• In section 3.7, we noted that several optimization counterparts to CSPs allow using binary-search
in addition to a direct optimization approach. While using binary search is more expensive for
these problems, we do not know which approach will performs better in practice.
References
[1] D. Achioptas and G. B. Sorkin. Optimal myopic algorithms for random 3-sat. In Foundations of Computer
Science, 2000. Proceedings. 41st Annual Symposium on, pages 590–600. IEEE, 2000. 79
[2] D. Achlioptas and A. Coja-Oghlan. Algorithmic barriers from phase transitions. arXiv:0803.2122, Mar.
2008. 79, 80
[3] D. Achlioptas, A. Naor, and Y. Peres. Rigorous location of phase transitions in hard optimization problems.
Nature, 435(7043):759–764, June 2005. 80
[4] Y. Ahn, J. Bagrow, and S. Lehmann. Link communities reveal multiscale complexity in networks. Nature,
2010. 100
[5] H. Aissi, C. Bazgan, and D. Vanderpooten. Min–max and min–max regret versions of combinatorial optimization problems: A survey. European J Operational Research, 197(2):427–438, 2009. 28
[6] S. Aji and R. McEliece. The Generalized distributive law and free energy minimization. In Allerton Conf,
2001. 2, 48
[7] S. M. Aji and R. J. McEliece. The generalized distributive law. Information Theory, IEEE Transactions
on, 46(2):325–343, 2000. 1
[8] S. M. Aji, G. B. Horn, and R. J. McEliece. On the Convergence of iterative decoding on graphs with a
single cycle. In Proc 1998 IEEE Int Symp Information Theory, 1998. 36
[9] N. Alon and A. Shapira. Homomorphisms in graph property testing. In Topics in discrete mathematics,
pages 281–313. Springer, 2006. 127
[10] D. J. Amit. Modeling brain function: The world of attractor neural networks. Cambridge University Press,
1992. 1, 68
[11] C. Andrieu, N. de Freitas, A. Doucet, and M. Jordan. An introduction to MCMC for machine learning.
Machine Learning, 2003. 66
[12] D. Applegate. The traveling salesman problem: a computational study. Princeton, 2006. 43, 115
[13] D. Applegate, R. Bixby, V. Chvatal, and W. Cook. Concorde TSP solver, 2006. 115, 120, 121
140
APPENDIX . REFERENCES
[14] S. Arora and B. Barak. Computational complexity: a modern approach. Cambridge University Press, 2009.
18
[15] S. Arora, P. Raghavan, and S. Rao. Approximation schemes for euclidean k-medians and related problems.
In Proceedings of the thirtieth annual ACM symposium on Theory of computing, pages 106–113. ACM, 1998.
100
[16] I. Averbakh. On the complexity of a class of combinatorial optimization problems with uncertainty. Mathematical Programming, 90(2):263–272, 2001. 28
[17] F. Bacchus. Representing and reasoning with probabilistic knowledge: a logical approach to probabilities.
MIT press, 1991. 2
[18] M. Bailly-Bechet, S. Bradde, et al. Clustering with shallow trees. J Statistical Mechanics, 2009. 102
[19] M. Bailly-Bechet, A. Braunstein, and R. Zecchina. A prize-collecting steiner tree approach for transduction
network inference. In Computational Methods in Systems Biology, pages 83–95. Springer, 2009. 102
[20] M. Bailly-Bechet, C. Borgs, et al. Finding undetected protein associations in cell signaling by belief propagation. PPNAS, 2011. 102
[21] E. Balas and P. Toth. Branch and bound methods. The traveling salesman problem, 1985. 115
[22] M. Bayati, D. Shah, and M. Sharma. Maximum weight matching via max-product belief propagation. In
ISIT, pages 1763–1767. IEEE, 2005. 36, 113, 114
[23] M. Bayati, C. Borgs, A. Braunstein, J. Chayes, A. Ramezanpour, and R. Zecchina. Statistical mechanics
of steiner trees. Physical review letters, 101(3):037208, 2008. 102
[24] M. Bayati, A. Braunstein, and R. Zecchina. A rigorous analysis of the cavity equations for the minimum
spanning tree. Journal of Mathematical Physics, 49(12):125206, 2008. 102
[25] M. Bayati, D. Shah, and M. Sharma. Max-product for maximum weight matching: Convergence, correctness, and lp duality. Information Theory, IEEE Transactions on, 54(3):1241–1251, 2008. 114
[26] M. Bayati, D. F. Gleich, A. Saberi, and Y. Wang. Message-passing algorithms for sparse network alignment.
ACM Transactions on Knowledge Discovery from Data (TKDD), 7(1):3, 2013. 132
[27] M. J. Beal, F. Falciani, Z. Ghahramani, C. Rangel, and D. L. Wild. A Bayesian approach to reconstructing
genetic regulatory networks with hidden factors. Bioinformatics, 21(3):349–356, 2005. 1
[28] R. Bellman. Dynamic programming treatment of the travelling salesman problem. JACM, 1962. 115
[29] A. Ben-Dor and B. Chor. On constructing radiation hybrid maps. J Computational Biology, 1997. 120
[30] H. A. Bethe. Statistical theory of superlattices. Proceedings of the Royal Society A Mathematical Physical
and Engineering Sciences, 150(871):552–575, 1935. 45
141
[31] I. Biazzo, A. Braunstein, and R. Zecchina. Performance of a cavity-method-based algorithm for the prizecollecting steiner tree problem on graphs. Physical Review E, 86(2):026706, 2012. 102
[32] B. Bidyuk and R. Dechter. Cutset sampling for Bayesian networks. Journal of Artificial Intelligence
Research, 28(3):1–48, 2007. 49
[33] S. Bistarelli, U. Montanari, F. Rossi, T. Schiex, G. Verfaillie, and H. Fargier. Semiring-based csps and
valued csps: Frameworks, properties, and comparison. Constraints, 4(3):199–240, 1999. 8, 24
[34] V. Blondel, J. Guillaume, et al. Fast unfolding of communities in large networks. J Statistical Mechanics,
2008. 106, 111
[35] H. L. Bodlaender. Classes of graphs with bounded tree-width, 1986. 48
[36] L. Bordeaux and E. Monfroy. Beyond np: Arc-consistency for quantified constraints. In Principles and
Practice of Constraint Programming-CP 2002, pages 371–386. Springer, 2002. 21
[37] C. Borgs, J. Chayes, L. Lovász, V. T. Sós, B. Szegedy, and K. Vesztergombi. Graph limits and parameter
testing. In Proceedings of the thirty-eighth annual ACM symposium on Theory of computing, pages 261–270.
ACM, 2006. 127
[38] C. Borgs, J. Chayes, L. Lovász, V. T. Sós, and K. Vesztergombi. Counting graph homomorphisms. In
Topics in discrete mathematics, pages 315–371. Springer, 2006. 127, 128
[39] C. Borgs, J. T. Chayes, L. Lovász, V. T. Sós, and K. Vesztergombi. Convergent sequences of dense graphs
i: Subgraph frequencies, metric properties and testing. Advances in Mathematics, 219(6):1801–1851, 2008.
127
[40] C. Borgs, J. T. Chayes, L. Lovász, V. T. Sós, and K. Vesztergombi. Convergent sequences of dense graphs
ii. multiway cuts and statistical physics. Annals of Mathematics, 176(1):151–219, 2012. 127
[41] E. Boros and P. L. Hammer. Pseudo-boolean optimization. Discrete applied mathematics, 123(1):155–225,
2002. 26
[42] S. Boyd and L. Vandenberghe. Convex Optimization, volume 25. Cambridge University Press, 2010. 46
[43] S. Bradde. Statistical physics of subgraph identification problem. PhD thesis, SISSA, 2010. 126
[44] S. Bradde, A. Braunstein, H. Mahmoudi, F. Tria, M. Weigt, and R. Zecchina. Aligning graphs and finding
substructures by a cavity approach. EPL (Europhysics Letters), 89(3):37009, 2010. 4, 124, 126, 132
[45] U. Brandes, D. Delling, et al. on clustering. IEEE KDE, 2008. 106, 108
[46] A. Braunstein and R. Zecchina. Survey propagation as local equilibrium equations.
mat/0312483, Dec. 2003. J. Stat. Mech., P06007 (2004). 58
arXiv:cond-
[47] A. Braunstein, M. Mezard, M. Weigt, and R. Zecchina. Constraint satisfaction by survey propagation.
Physics, (2):8, 2002. 56
142
APPENDIX . REFERENCES
[48] A. Braunstein, M. Mezard, and R. Zecchina. Survey propagation: an algorithm for satisfiability. Random
Structures and Algorithms, 27(2):19, 2002. 56, 66, 83, 84
[49] A. Braunstein, R. Mulet, A. Pagnani, M. Weigt, and R. Zecchina. Polynomial iterative algorithms for
coloring and analyzing random graphs. Physical Review E, 68(3):036702, 2003. 83
[50] G. R. Brightwell and P. Winkler. Graph homomorphisms and phase transitions. Journal of combinatorial
theory, series B, 77(2):221–262, 1999. 128
[51] R. L. Brooks. On colouring the nodes of a network. In Mathematical Proceedings of the Cambridge Philosophical Society, volume 37, pages 194–197. Cambridge Univ Press, 1941. 98
[52] A. Bulatov and M. Grohe. The complexity of partition functions. Theoretical Computer Science, 348(2):
148–186, 2005. 128
[53] H. Bunke. On a relation between graph edit distance and maximum common subgraph. Pattern Recognition
Letters, 18(8):689–694, 1997. 132
[54] M. Castellana and L. Zdeborová. Adversarial satisfiability problem. Journal of Statistical Mechanics:
Theory and Experiment, 2011(03):P03023, 2011. 98
[55] T. Castellani and A. Cavagna. Spin-glass theory for pedestrians. Journal of Statistical Mechanics: Theory
and Experiment, 2005(05):P05012, May 2005. 46
[56] V. Černỳ. Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm.
J Optimization theory and Applications, 1985. 115
[57] V. Chandrasekaran and N. Srebro. Complexity of inference in graphical models. UAI, 2008. 48
[58] M. Charikar, S. Guha, É. Tardos, and D. B. Shmoys. A constant-factor approximation algorithm for the
k-median problem. In Proceedings of the thirty-first annual ACM symposium on Theory of computing, pages
1–10. ACM, 1999. 100
[59] P. Cheeseman, B. Kanefsky, and W. M. Taylor. Where the really hard problems are. In Proceedings of
the 12th IJCAI - Volume 1, IJCAI’91, pages 331–337, San Francisco, CA, USA, 1991. Morgan Kaufmann
Publishers Inc. 79
[60] M. Chertkov. Loop series for discrete statistical models on graphs. {. . . } of Statistical Mechanics: Theory
and {. . . }, 2006. 49
[61] M. Chertkov and V. Y. Chernyak. Loop calculus in statistical physics and information science. Physical
Review E - Statistical, Nonlinear and Soft Matter Physics, 73(6 Pt 2):4, 2006. 49
[62] M. Chertkov and A. B. Yedidia. Approximating the permanent with fractional belief propagation. J. Mach.
Learn. Res., 14(1):2029–2066, Jan. 2013. 113
[63] H. L. Chieu and W. S. Lee. Relaxed survey propagation: A sum-product algorithm for max-sat. In AAAI,
pages 247–252, 2008. 98
143
[64] H. L. Chieu, W. S. Lee, and Y. W. Teh. Cooled and relaxed survey propagation for mrfs. In Advances in
Neural Information Processing Systems, pages 297–304, 2007. 98
[65] P. S. Churchland and T. J. Sejnowski. The computational brain. The MIT press, 1992. 1
[66] J. Chuzhoy, S. Guha, E. Halperin, S. Khanna, G. Kortsarz, R. Krauthgamer, and J. S. Naor. Asymmetric
k-center is log* n-hard to approximate. JACM, 2005. 104
[67] A. Clauset. Finding local community structure in networks. Physical Review E, 2005. 106, 109
[68] P. Clifford. Markov random fields in statistics. Disorder in physical systems, pages 19–32, 1990. 8
[69] S. Climer and W. Zhang. Take a walk and cluster genes: A TSP-based approach to optimal rearrangement
clustering. In ICML, 2004. 120
[70] S. Cocco, O. Dubois, J. Mandler, and R. Monasson. Rigorous decimation-based construction of ground
pure states for spin glass models on random lattices. arXiv:cond-mat/0206239, June 2002. 79
[71] A. Coja-Oghlan. On belief propagation guided decimation for random k-sat. In Proceedings of the TwentySecond Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’11, pages 957–966. SIAM, 2011.
81
[72] D. Conte, P. Foggia, C. Sansone, and M. Vento. Thirty years of graph matching in pattern recognition.
International journal of pattern recognition and artificial intelligence, 18(03):265–298, 2004. 124, 132
[73] M. Cooke, J. R. Hershey, and S. J. Rennie. Monaural speech separation and recognition challenge. Computer
Speech & Language, 24(1):1–15, 2010. 1
[74] G. Cooper. The computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence, 42, 1990. 2
[75] T. M. Cover and J. A. Thomas. Elements of Information Theory, volume 6 of Wiley Series in Telecommunications. Wiley, 1991. 95
[76] C. Crick and A. Pfeffer. Loopy belief propagation as a basis for communication in sensor networks. In
Proceedings of the Nineteenth conference on Uncertainty in Artificial Intelligence, pages 159–166. Morgan
Kaufmann Publishers Inc., 2002. 1
[77] G. Dantzig, R. Fulkerson, and S. Johnson. Solution of a large-scale traveling-salesman problem. J Operations
Research society of America, 1954. 42, 115
[78] A. Darwiche. Recursive conditioning. Artificial Intelligence, 126(1):5–41, 2001. 49
[79] B. A. Davey and H. A. Priestley. Introduction to lattices and order. Cambridge university press, 2002. 9
[80] C. P. De Campos. New complexity results for map in Bayesian networks. In IJCAI, volume 11, pages
2100–2106, 2011. 47
144
APPENDIX . REFERENCES
[81] N. De Freitas, P. Hø jen Sø rensen, M. Jordan, and S. Russell. Variational MCMC. In Proceedings of the,
pages 120–127. Morgan Kaufmann Publishers Inc., 2001. 68
[82] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborová. Asymptotic analysis of the stochastic block model
for modular networks and its algorithmic applications. Physical Review E, 84(6):066106, 2011. 100
[83] R. Dechter and D. Larkin. Hybrid processing of beliefs and constraints. In Proceedings of the Seventeenth
conference on Uncertainty in artificial intelligence, pages 112–119. Morgan Kaufmann Publishers Inc., 2001.
8
[84] D. L. Donoho, A. Maleki, and A. Montanari. Message-passing algorithms for compressed sensing. Proceedings of the National Academy of Sciences, 106(45):18914–18919, 2009. 1
[85] M. Dorigo and L. Gambardella. Ant colony system: a cooperative learning approach to the traveling
salesman problem. IEEE Evolutionary Computation, 1997. 115
[86] M. Dyer and C. Greenhill. The complexity of counting graph homomorphisms. Random Structures and
Algorithms, 17(3-4):260–289, 2000. 128
[87] M. E. Dyer and A. M. Frieze. A simple heuristic for the p-centre problem. Operations Research Letters, 3
(6):285–288, 1985. 29, 104, 106
[88] D. J. Earl and M. W. Deem. Parallel Tempering: Theory, Applications, and New Perspectives. Physical
Chemistry Chemical Physics, 7(23):21, 2005. 68
[89] F. Eaton and Z. Ghahramani. Choosing a variable to clamp. In International Conference on Artificial
Intelligence and Statistics, pages 145–152, 2009. 49
[90] F. Eaton and Z. Ghahramani. Model reductions for inference: Generality of pairwise, binary, and planar
factor graphs. Neural computation, 25(5):1213–1260, 2013. 25
[91] B. Eckmann and P. J. Hilton. Group-like structures in general categories (I) multiplications and comultiplications. Mathematische Annalen, 145(3):227–255, 1962. 38
[92] J. Edmonds and D. Fulkerson. Bottleneck extrema. J Combinatorial Theory, 1970. 29
[93] S. F. Edwards and P. W. Anderson. Theory of spin glasses. Journal of Physics F Metal Physics, 5(5):
965–974, 1975. 13
[94] G. Elidan and I. McGraw. Residual belief propagation: Informed scheduling for asynchronous message
passing. In Proceedings of the Twenty- {. . . }, 2006. 36
[95] D. Eppstein. Subgraph isomorphism in planar graphs and related problems. In Proceedings of the sixth
annual ACM-SIAM symposium on Discrete algorithms, pages 632–640. Society for Industrial and Applied
Mathematics, 1995. 124
[96] S. Ermon, C. P. Gomes, A. Sabharwal, and B. Selman. Optimization with parity constraints: From binary
codes to discrete integration. arXiv preprint arXiv:1309.6827, 2013. 25
145
[97] S. Ermon, C. Gomes, A. Sabharwal, and B. Selman. Low-density parity constraints for hashing-based
discrete integration. In Proceedings of The 31st International Conference on Machine Learning, pages
271–279, 2014. 25
[98] P. F. Felzenszwalb and D. P. Huttenlocher. Efficient belief propagation for early vision. International
journal of computer vision, 70(1):41–54, 2006. 1
[99] M. E. Fisher. On the dimer solution of planar ising models. Journal of Mathematical Physics, 7(10):
1776–1781, 1966. 26
[100] G. D. Forney Jr. Codes on graphs: normal realizations. Information Theory, IEEE Transactions on, 47(2):
520–548, 2001. 8
[101] W. T. Freeman, E. C. Pasztor, and O. T. Carmichael. Learning low-level vision. International journal of
computer vision, 40(1):25–47, 2000. 1
[102] B. Frey and D. Dueck. Clustering by passing messages between data points. Science, 2007. 100, 102
[103] Y. Fu and P. W. Anderson. Application of statistical mechanics to NP-complete problems in combinatorial
optimisation. Journal of Physics A: Mathematical and General, 19(9):1605–1620, June 1986. 2, 79
[104] D. Fulkerson. Flow networks and combinatorial operations research. The American Mathematical Monthly,
73(2):115–138, 1966. 29
[105] M. H. Gail, J. H. Lubin, and L. V. Rubinstein. Likelihood calculations for matched case-control studies
and survival studies with tied death times. Biometrika, 68(3):703–707, 1981. 40
[106] R. G. Gallager. Low-density parity-check codes. Information Theory, IRE Transactions on, 8(1):21–28,
1962. 1, 36
[107] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness,
volume 44 of A Series of Books in the Mathematical Sciences. W. H. Freeman, 1979. 76, 88
[108] A. E. Gelfand and A. F. M. Smith. Sampling-Based Approaches to Calculating Marginal Densities. Journal
of the American Statistical Association, 85(410):398–409, 1990. 49
[109] S. Geman and D. Geman. Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of
Images. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-6(6):721–741, 1984. 67
[110] S. Gershman, M. Hoffman, and D. Blei. Nonparametric variational inference. June 2012. 68
[111] I.-E. Givoni. Beyond Affinity Propagation: Message Passing Algorithms for Clustering. PhD thesis, University of Toronto, 2012. 102
[112] I. E. Givoni and B. J. Frey. A binary variable model for affinity propagation. Neural computation, 21(6):
1589–1600, 2009. 100
[113] A. Globerson and T. Jaakkola. Approximate inference using conditional entropy decompositions. In International Conference on Artificial Intelligence and Statistics, pages 130–138, 2007. 46
146
APPENDIX . REFERENCES
[114] A. Globerson and T. S. Jaakkola. Fixing max-product: Convergent message passing algorithms for map
lp-relaxations. In Advances in neural information processing systems, pages 553–560, 2008. 46
[115] D. E. Goldberg and R. Lingle. Alleles, loci, and the traveling salesman problem. In Genetic algorithms and
Applications, 1985. 115
[116] G. H. Golub and C. F. Van Loan. Matrix computations, volume 3. JHU Press, 2012. 112
[117] V. Gomez, J. M. Mooij, and H. J. Kappen. Truncating the loop series expansion for belief propagation.
Journal of Machine Learning Research, 8:1987–2016, Sept. 2007. 50
[118] R. E. Gomory et al. Outline of an algorithm for integer solutions to linear programs. Bulletin of the
American Mathematical Society, 64(5):275–278, 1958. 42, 115
[119] T. F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science,
38:293–306, 1985. 29, 104
[120] J. Grefenstette, R. Gopal, B. Rosmaita, and D. Van Gucht. Genetic algorithms for the traveling salesman
problem. In Genetic algorithms and Applications, 1985. 115
[121] O. Gross. The bottleneck assignment problem. Technical report, DTIC Document, 1959. 29, 113
[122] M. Grotschel and L. Lovász. Combinatorial optimization. Handbook of combinatorics, 2:1541–1597, 1995.
1
[123] R. Gupta, A. A. Diwan, and S. Sarawagi. Efficient inference with cardinality-based clique potentials. In
Proceedings of the 24th international conference on Machine learning, pages 329–336. ACM, 2007. 39
[124] W. K. Hastings. Monte Carlo sampling methods using Markov chains and their applications. Biometrika,
57(1):97–109, 1970. 67
[125] T. Hazan and T. Jaakkola. On the partition function and random maximum a-posteriori perturbations.
arXiv preprint arXiv:1206.6410, 2012. 25
[126] T. Hazan and A. Shashua. Convergent Message-Passing Algorithms for Inference over General Graphs with
Convex Free Energies. Science, 51(11):264–273, 2008. 46
[127] M. Held and R. Karp. The traveling-salesman problem and minimum spanning trees. Operations Research,
1970. 116
[128] M. Held and R. M. Karp. A dynamic programming approach to sequencing problems. Journal of the Society
for Industrial & Applied Mathematics, 10(1):196210, 1962. 115
[129] P. Hell and J. Nesetril. Graphs and homomorphisms. Oxford University Press, 2004. 127
[130] K. Helsgaun. General k-opt submoves for the lin–kernighan tsp heuristic. Mathematical Programming
Computation, 2009. 115
147
[131] J. R. Hershey, S. J. Rennie, P. A. Olsen, and T. T. Kristjansson. Super-human multi-talker speech recognition: A graphical modeling approach. Computer Speech & Language, 24(1):45–66, 2010. 1
[132] T. Heskes. Stable fixed points of loopy belief propagation are minima of the Bethe free energy. Advances
in Neural Information Processing Systems NIPS2002, 14:343–350, 2003. 45
[133] T. Heskes. Convexity arguments for efficient minimization of the Bethe and Kikuchi free energies. JAIR,
26, 2006. 46, 55
[134] T. Heskes. Convexity Arguments for Efficient Minimization of the Bethe and Kikuchi Free Energies, 2006.
46
[135] D. S. Hochbaum and D. B. Shmoys. A best possible heuristic for the k-center problem. Mathematics of
operations research, 10(2):180–184, 1985. 104
[136] D. S. Hochbaum and D. B. Shmoys. A unified approach to approximation algorithms for bottleneck problems. Journal of the ACM (JACM), 33(3):533–550, 1986. 29
[137] R. A. Howard and J. E. Matheson. Influence diagrams. Decision Analysis, 2(3):127–143, 2005. 8
[138] B. Huang and T. Jebara. Loopy belief propagation for bipartite maximum weight b-matching. In AI and
Statistics, 2007. 113
[139] B. Huang and T. Jebara. Approximating the permanent with belief propagation.
arXiv:0908.1769, 2009. 18, 113
arXiv preprint
[140] M. Ibrahimi, A. Javanmard, Y. Kanoria, and A. Montanari. Robust max-product belief propagation.
In Signals, Systems and Computers (ASILOMAR), 2011 Conference Record of the Forty Fifth Asilomar
Conference on, pages 43–49. IEEE, 2011. 28, 39
[141] A. Ihler. Particle belief propagation. {. . . } Conference on Artificial Intelligence and Statistics, 2009. 68
[142] A. Ihler and M. David. Nonparametric belief propagation. 2003. 68
[143] A. T. Ihler, J. W. Fisher, R. L. Moses, and A. S. Willsky. Nonparametric belief propagation for selflocalization of sensor networks. Selected Areas in Communications, IEEE Journal on, 23(4):809–819, 2005.
1
[144] M. Isard. PAMPAS: real-valued graphical models for computer vision, 2003. 68
[145] S. Janson, T. Luczak, and V. F. Kolchin. Random graphs. Bull. London Math. Soc, 33:363383, 2001. 81
[146] E. T. Jaynes. Probability theory: the logic of science. Cambridge university press, 2003. 2
[147] C. S. Jensen, A. Kong, and U. Kjaerulff. Blocking-Gibbs Sampling in Very Large Probabilistic Expert
Systems. International Journal of Human-Computer Studies, 42:647–666, 1995. 68
[148] F. V. Jensen, S. L. Lauritzen, and K. G. Olesen. Bayesian updating in causal probabilistic networks by
local computations. Computational Statistics Quarterly, 4(4):269–282, 1990. 48
148
APPENDIX . REFERENCES
[149] M. Jerrum and A. Sinclair. The markov chain monte carlo method: an approach to approximate counting
and integration. Approximation algorithms for NP-hard problems, pages 482–520, 1996. 113
[150] M. Jerrum, A. Sinclair, and E. Vigoda. A polynomial-time approximation algorithm for the permanent of
a matrix with nonnegative entries. Journal of the ACM (JACM), 51(4):671–697, 2004. 113
[151] M. R. Jerrum, L. G. Valiant, and V. V. Vazirani. Random generation of combinatorial structures from a
uniform distribution. Theoretical Computer Science, 43:169–188, 1986. 25
[152] J. Jiang, P. Rai, and H. Daume. Message-passing for approximate map inference with latent variables. In
Advances in Neural Information Processing Systems, pages 1197–1205, 2011. 47
[153] D. Johnson, S. Krishnan, J. Chhugani, S. Kumar, and S. Venkatasubramanian. Compressing large Boolean
matrices using reordering techniques. In VLDB, 2004. 120
[154] D. S. Johnson and L. A. McGeoch. The traveling salesman problem: A case study in local optimization.
Local search in combinatorial optimization, page 215310, 1997. 120
[155] V. Jojic, S. Gould, and D. Koller. Accelerated dual decomposition for map inference. In Proceedings of the
27th International Conference on Machine Learning (ICML-10), pages 503–510, 2010. 46
[156] S. N. Kabadi and A. P. Punnen. The bottleneck tsp. In The Traveling Salesman Problem and Its Variations,
pages 697–735. Springer, 2004. 121
[157] M. Kaminski, V. V. Lozin, and M. Milanic. Recent developments on graphs of bounded clique-width.
Discrete Applied Mathematics, 157(12):2747–2761, 2009. 48
[158] V. Kann. On the approximability of NP-complete optimization problems. PhD thesis, Royal Institute of
Technology Stockholm, 1992. 88
[159] R. M. Karp. Reducibility among combinatorial problems. Springer, 1972. 86
[160] H. A. Kautz, A. Sabharwal, and B. Selman. Incomplete algorithms., 2009. 32, 98
[161] M. Kearns. Graphical games. Algorithmic game theory, 3:159–180, 2007. 39
[162] M. Kearns, M. L. Littman, and S. Singh. Graphical models for game theory. In Proceedings of the Seventeenth conference on UAI, pages 253–260. Morgan Kaufmann Publishers Inc., 2001. 28
[163] P. J. Kelly et al. A congruence theorem for trees. Pacific J. Math, 7(0957):961–968, 1957. 130
[164] S. Khuller and Y. J. Sussmann. The capacitated k-center problem. SIAM Journal on Discrete Mathematics,
13(3):403–418, 2000. 29
[165] R. Kikuchi. A Theory of Cooperative Phenomena. Physical Review, 81(6):988–1003, 1951. 48, 49
[166] S. Kirkpatrick, M. Vecchi, et al. Optimization by simmulated annealing. Science, 1983. 115
[167] T. Kloks and H. L. Bodlaender. Only few graphs have bounded treewidth. RUUCS, 1992. 48
149
[168] P. Kohli and P. H. S. Torr. Graph Cuts for Minimizing Robust Higher Order Potentials. Computing, 24
(7):996–1000, 2008. 40
[169] P. Kohli, M. P. Kumar, and P. H. S. Torr. P3 & Beyond: Solving Energies with Higher Order Cliques.
2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1–8, 2007. 40
[170] P. Kohli, L. Ladický, and P. H. S. Torr. Robust Higher Order Potentials for Enforcing Label Consistency.
International Journal of Computer Vision, 82(3):302–324, Jan. 2009. 40
[171] D. Koller and N. Friedman. Probabilistic Graphical Models: Principles and Techniques, volume 2009 of
Adaptive Computation and Machine Learning. The MIT Press, 2009. 48
[172] D. Koller, U. Lerner, and D. Angelov. A general algorithm for approximate inference and its application
to hybrid bayes nets. pages 324–333, July 1999. 68
[173] V. Kolmogorov. Convergent tree-reweighted message passing for energy minimization. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 28(10):1568–1583, 2006. 36
[174] N. Komodakis. Beyond pairwise energies: Efficient optimization for higher-order MRFs. computer.org,
2009. 40
[175] N. Komodakis and N. Paragios. Mrf energy minimization and beyond via dual decomposition. IEEE
Transactions on Pattern {. . . }, 2010. 46
[176] N. Komodakis, N. Paragios, and G. Tziritas. Mrf optimization via dual decomposition: Message-passing
revisited. In Computer Vision, 2007. ICCV 2007. IEEE 11th International Conference on, pages 1–8. IEEE,
2007. 46
[177] L. Kroc, A. Sabharwal, and B. Selman. Survey Propagation Revisited. 23rd UAI, pages 217–226, 2002. 58
[178] A. Krogh, B. Larsson, G. Von Heijne, and E. L. Sonnhammer. Predicting transmembrane protein topology
with a hidden markov model: application to complete genomes. Journal of molecular biology, 305(3):
567–580, 2001. 1
[179] F. Krzakala, A. Montanari, F. Ricci-Tersenghi, G. Semerjian, and L. Zdeborova. Gibbs states and the set
of solutions of random constraint satisfaction problems. Proceedings of the National Academy of Sciences
of the United States of America, 104(25):1031810323, 2007. 79, 80, 81
[180] F. Krzakala, C. Moore, E. Mossel, J. Neeman, A. Sly, L. Zdeborová, and P. Zhang. Spectral redemption in
clustering sparse networks. Proceedings of the National Academy of Sciences, 110(52):20935–20940, 2013.
100, 131
[181] F. R. Kschischang and B. J. Frey. Factor graphs and the sum-product algorithm. Information Theory,
IEEE, 2001. 8
[182] H. W. Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):
83–97, 1955. 113
150
APPENDIX . REFERENCES
[183] X. Lan, S. Roth, D. Huttenlocher, and M. J. Black. Efficient Belief Propagation with Learned Higher-Order
Markov Random Fields. Computer VisionECCV 2006, 3952:269–282, 2006. 40
[184] A. Lancichinetti and S. Fortunato. Community detection algorithms: A comparative analysis. Physical
Review E, 80(5):056117, 2009. 100
[185] S. L. Lauritzen and F. V. Jensen. Local computation with valuations from a commutative semigroup.
Annals of Mathematics and Artificial Intelligence, 21(1):51–69, 1997. 2, 8
[186] S. L. Lauritzen and D. J. Spiegelhalter. Local computations with probabilities on graphical structures and
their application to expert systems. Journal of the Royal Statistical Society Series B Methodological, 50(2):
157–224, 1988. 48
[187] N. Lazic. Message Passing Algorithms for Facility Location Problems. PhD thesis, University of Toronto,
2011. 102
[188] N. Lazic, B. J. Frey, and P. Aarabi. Solving the uncapacitated facility location problem using message
passing algorithms. In International Conference on Artificial Intelligence and Statistics, pages 429–436,
2010. 102
[189] C. Lecoutre. A collection of csp benchmark instances., Oct. 2013. 78
[190] J. Leskovec, D. Chakrabarti, J. Kleinberg, C. Faloutsos, and Z. Ghahramani. Kronecker graphs: An
approach to modeling networks. The Journal of Machine Learning Research, 11:985–1042, 2010. 131
[191] J. Leskovec, K. Lang, and M. Mahoney. Empirical comparison of algorithms for network community
detection. In WWW, 2010. 100
[192] D. A. Levin, Y. Peres, and E. L. Wilmer. Markov chains and mixing times. American Mathematical Soc.,
2009. 71
[193] S. Litsyn, R. E, and S. N. The table of nonlinear binary codes. http://www.eng.tau.ac.il/~litsyn/
tableand/, 1999. [Online; accessed Jan 30 2014]. 95, 97
[194] M. L. Littman, S. M. Majercik, and T. Pitassi. Stochastic boolean satisfiability. Journal of Automated
Reasoning, 27(3):251–296, 2001. 17, 21
[195] Q. Liu and A. Ihler. Variational algorithms for marginal map. The Journal of Machine Learning Research,
14(1):3165–3200, 2013. 47
[196] L. Lovász. Operations with structures. Acta Mathematica Hungarica, 18(3):321–328, 1967. 129
[197] L. Lovász and B. Szegedy. Limits of dense graph sequences. Journal of Combinatorial Theory, Series B,
96(6):933–957, 2006. 127
[198] B. D. MacArthur, R. J. Sánchez-Garcı́a, and J. W. Anderson. Symmetry in complex networks. Discrete
Applied Mathematics, 156(18):3525–3531, 2008. 124
151
[199] E. N. Maneva, E. Mossel, and M. J. Wainwright. A new look at survey propagation and its generalizations.
arXiv:cs/0409012, Sept. 2004. 58
[200] C. D. Manning and H. Schütze. Foundations of statistical natural language processing. MIT press, 1999. 1
[201] S. Masuyama, T. Ibaraki, and T. Hasegawa. The computational complexity of the m-center problems on
the plane. IEICE TRANSACTIONS (1976-1990), 64(2):57–64, 1981. 104
[202] D. Maua and C. De Campos. Anytime marginal map inference. arXiv preprint arXiv:1206.6424, 2012. 47
[203] A. McCallum, A. Corrada-Emmanuel, and X. Wang. Topic and role discovery in social networks. Computer
Science Department Faculty Publication Series, page 3, 2005. 1
[204] B. D. McKay. Small graphs are reconstructible. Australasian Journal of Combinatorics, 15:123–126, 1997.
130
[205] B. D. McKay and A. Piperno. Practical graph isomorphism, {II}. Journal of Symbolic Computation, 60
(0):94 – 112, 2014. 130
[206] O. Meshi, A. Jaimovich, A. Globerson, and N. Friedman. Convexifying the Bethe free energy. AUAI Press,
June 2009. 46
[207] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller. Equations of state
calculations by fast computing machines. Journal of Medical Physics, 21(6):1087–1092, 1953. 67
[208] M. Mézard and A. Montanari. Information, physics, and computation. Oxford, 2009. 1, 56, 58, 64, 68, 71,
81, 84
[209] M. Mezard and G. Parisi. The bethe lattice spin glass revisited. arXiv:cond-mat/0009418, Sept. 2000. Eur.
Phys. J. B 20, 217 (2001). 79
[210] M. Mezard and G. Parisi. The cavity method at zero temperature. J. Stat. Phys 111 1, 2003. 79, 80
[211] M. Mezard, G. Parisi, and M. A. Virasoro. Spin Glass Theory and Beyond. Singapore: World Scientific,
1987. 1, 46, 80
[212] M. Mezard, G. Parisi, M. A. Virasoro, and D. J. Thouless. Spin glass theory and beyond. Physics Today,
41:109, 1988. 13
[213] M. Mezard, G. Parisi, and R. Zecchina. Analytic and algorithmic solution of random satisfiability problems.
Science, 297(5582):812–815, Aug. 2002. 2, 79
[214] T. Minka and Y. Qi. Tree-structured approximations by expectation propagation. In NIPS, 2003. 55
[215] D. Mitchell, B. Selman, and H. Levesque. Hard and easy distributions of sat problems. In AAAI, volume 92,
pages 459–465, 1992. 2, 79
[216] R. Monasson, R. Zecchina, S. Kirkpatrick, B. Selman, and L. Troyansky. Determining computational
complexity from characteristic phase transitions. Nature, 400(6740):133–137, 1999. 2
152
APPENDIX . REFERENCES
[217] A. Montanari and T. Rizzo. How to Compute Loop Corrections to Bethe Approximation. Journal of
Statistical Mechanics: Theory and Experiment, (10):16, 2005. 50
[218] A. Montanari, F. Ricci-Tersenghi, and G. Semerjian. Solving constraint satisfaction problems through belief
propagation-guided decimation. arXiv:0709.1667, Sept. 2007. 81
[219] A. Montanari, F. Ricci-Tersenghi, and G. Semerjian. Clusters of solutions and replica symmetry breaking
in random k-satisfiability. arXiv:0802.3627, Feb. 2008. J. Stat. Mech. P04004 (2008). 87
[220] J. Mooij. Understanding and improving belief propagation. PhD thesis, Radboud Universiteit Nijmegen,
2008. 50, 52, 55
[221] J. Mooij. libDAI: A free and open source C++ library for discrete approximate inference in graphical
models. JMLR, 2010. 56
[222] C. Moore and S. Mertens. The nature of computation. Oxford University Press, 2011. 1, 115, 124
[223] M.-L. Mugnier. Knowledge representation and reasonings based on graph homomorphism. In Conceptual
Structures: Logical, Linguistic, and Computational Issues, pages 172–192. Springer, 2000. 127, 129
[224] K. Murphy, S. Mian, et al. Modelling gene expression data using dynamic Bayesian networks. Technical
report, Technical report, Computer Science Division, University of California, Berkeley, CA, 1999. 1
[225] K. Murphy, Y. Weiss, and M. Jordan. Loopy-belief Propagation for Approximate Inference: An Empirical
Study. In 15, pages 467–475, 1999. 36
[226] K. P. Murphy. Machine learning: a probabilistic perspective. MIT press, 2012. 130
[227] R. M. Neal. Probabilistic inference using markov chain monte carlo methods. 1993. 66
[228] M. Newman. Finding community structure in networks using the eigenvectors of matrices. Physical review
E, 2006. 106, 111, 131
[229] M. Newman and M. Girvan. Finding and evaluating community structure in networks. Physical Review E,
2004. 106
[230] A. Y. Ng, M. I. Jordan, Y. Weiss, et al. On spectral clustering: Analysis and an algorithm. Advances in
neural information processing systems, 2:849–856, 2002. 100
[231] N. Noorshams and M. J. Wainwright. Belief propagation for continuous state spaces: Stochastic messagepassing with quantitative guarantees. Journal of Machine Learning Research, 14:2799–2835, 2013. 68
[232] L. E. Ortiz and M. Kearns. Nash propagation for loopy graphical games. In Advances in Neural Information
Processing Systems, pages 793–800, 2002. 39
[233] M. Padberg and G. Rinaldi. A branch-and-cut algorithm for the resolution of large-scale symmetric traveling
salesman problems. SIAM review, 33(1):60–100, 1991. 115
153
[234] R. Paget and I. D. Longstaff. Texture synthesis via a noncausal nonparametric multiscale Markov random
field. IEEE Transactions on Image Processing, 7(6):925–931, 1998. 40
[235] P. Pakzad and V. Anantharam. Estimation and Marginalization Using the Kikuchi Approximation Methods.
Neural Computation, 17(8):1836–1873, Aug. 2005. 49
[236] G. Palla, I. Derényi, I. Farkas, and T. Vicsek. Uncovering the overlapping community structure of complex
networks in nature and society. Nature, 2005. 100
[237] R. Panigrahy and S. Vishwanathan. An o(log*n) approximation algorithm for the asymmetric p-center
problem. J Algorithms, 1998. 29, 104
[238] C. H. Papadimitriou. The euclidean travelling salesman problem is np-complete. Theoretical Computer
Science, 4(3):237–244, 1977. 115
[239] G. Papandreou and A. L. Yuille. Perturb-and-map random fields: Using discrete optimization to learn and
sample from energy models. In Computer Vision (ICCV), 2011 IEEE International Conference on, pages
193–200. IEEE, 2011. 25
[240] G. Parisi. A backtracking survey propagation algorithm for k-satisfiability. Arxiv preprint condmat0308510,
page 9, 2003. 81
[241] J. D. Park and A. Darwiche. Complexity results and approximation strategies for map explanations. J.
Artif. Intell. Res.(JAIR), 21:101–133, 2004. 20, 39
[242] R. G. Parker and R. L. Rardin. Guaranteed performance heuristics for the bottleneck travelling salesman
problem. Operations Research Letters, 2(6):269–272, 1984. 29, 121
[243] J. Pearl. Bayesian networks: a model of self-activated memory for evidential reasoning. 1985. 8
[244] J. Pearl. Probabilistic reasoning in intelligent systems. 1988. 1
[245] J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, volume 88 of
Representation and Reasoning. Morgan Kaufmann, 1988. 2, 49
[246] A. Pelizzola. Cluster Variation Method in Statistical Physics and Probabilistic Graphical Models. Journal
of Physics A: Mathematical and General, 38(33):36, 2005. 48
[247] C. C. Pinter. A book of abstract algebra. Courier Dover Publications, 2012. 9, 23
[248] P. Pons and M. Latapy. Computing communities in large networks using random walks. In Computer and
Information Sciences-ISCIS 2005, pages 284–293. Springer, 2005. 100
[249] B. Potetz and T. S. Lee. Efficient belief propagation for higher-order cliques using linear constraint nodes.
Computer Vision and Image Understanding, 112(1):39–54, 2008. 39, 40
[250] C. Pralet, G. Verfaillie, and T. Schiex. An algebraic graphical model for decision with uncertainties,
feasibilities, and utilities. J. Artif. Intell. Res.(JAIR), 29:421–489, 2007. 8
154
APPENDIX . REFERENCES
[251] U. N. Raghavan, R. Albert, and S. Kumara. Near linear time algorithm to detect community structures in
large-scale networks. Physical Review E, 76(3):036106, 2007. 100
[252] A. Ramezanpour and R. Zecchina. Cavity approach to sphere packing in hamming space. Physical Review
E, 85(2):021106, 2012. 95
[253] S. Ravanbakhsh and R. Greiner. Perturbed message passing for constraint satisfaction problems. JMLR,
to appear, 2014. 4
[254] S. Ravanbakhsh, R. Greiner, and J. Y. Chun-nam. A generalized loop correction method for approximate
inference in graphical models. In Proceedings of the 29th International Conference on Machine Learning
(ICML-12), pages 543–550, 2012. 4, 50, 55
[255] S. Ravanbakhsh, P. Liu, R. Mandal, J. R. Grant, M. Wilson, R. Eisner, I. Sinelnikov, X. Hu, C. Luchinat,
R. Greiner, et al. Accurate, fully-automated nmr spectral profiling for metabolomics. arXiv preprint
arXiv:1409.1456, 2014. 1
[256] S. Ravanbakhsh, R. Rabbany, and R. Greiner. Revisiting algebra and complexity of inference in graphical
models. arXiv:1409.7410, 2014. 4
[257] S. Ravanbakhsh, R. Rabbany, and R. Greiner. Augmentative message passing for traveling salesman problem
and graph partitioning. NIPS, 2014. 4
[258] S. Ravanbakhsh, C. Srinivasa, B. Frey, and R. Greiner. Min-max problems on factor-graphs. Proceedings
of T 31st International Conference on Machine Learning, pp. 1035-1043, 2014. 4, 17
[259] J. Reichardt and S. Bornholdt. Detecting fuzzy community structures in complex networks with a potts
model. Physical Review Letters, 93(21):218701, 2004. 106, 111
[260] G. Reinelt. Tspliba traveling salesman problem library. ORSA journal on computing, 3(4):376–384, 1991.
118
[261] F. Ricci-Tersenghi and G. Semerjian. On the cavity method for decimated random constraint satisfaction
problems and the analysis of belief propagation guided decimation algorithms. arXiv:0904.3395, Apr. 2009.
J Stat. Mech. P09001 (2009). 81
[262] T. Rizzo, B. Wemmenhove, and H. J. Kappen. On Cavity Approximations for Graphical Models. Physical
Review E, 76(1):1–9, 2006. 50
[263] C. P. Robert and G. Casella. Monte Carlo Statistical Methods (Springer Texts in Statistics). Springer-Verlag
New York, Inc., Secaucus, NJ, USA, 2005. 66
[264] H. Rogers. Theory of recursive functions and effective computability. 1987. 13
[265] P. Ronhovde and Z. Nussinov. Local resolution-limit-free potts model for community detection. Physical
Review E, 81(4):046114, 2010. 106
[266] M. Rosvall and C. T. Bergstrom. Maps of random walks on complex networks reveal community structure.
PNAS, 2008. 100
155
[267] D. Roth. On the hardness of approximate reasoning. In Proceedings of the 13th international joint conference
on Artifical intelligence-Volume 1, pages 613–618. Morgan Kaufmann Publishers Inc., 1993. 2, 17
[268] S. Roth and M. J. Black. Fields of Experts. International Journal of Computer Vision, 82(2):205–229,
2009. 40
[269] C. Rother and P. Kohli. Minimizing sparse higher order energy functions of discrete variables. IEEE
Conference on Computer Vision and Pattern Recognition (2009), 0:1382–1389, 2009. 39
[270] O. Roussel and C. Lecoutre. Xml representation of constraint networks: Format xcsp 2.1. arXiv preprint
arXiv:0902.2362, 2009. 78
[271] N. Ruozzi and S. Tatikonda. Message-passing algorithms: Reparameterizations and splittings. Information
Theory, IEEE Transactions on, 59(9):5860–5881, 2013. 46
[272] A. Saade, F. Krzakala, and L. Zdeborová. Spectral clustering of graphs with the bethe hessian. arXiv
preprint arXiv:1406.1880, 2014. 131
[273] S. Sanghavi. Equivalence of lp relaxation and max-product for weighted matching in general graphs. In
Information Theory Workshop, 2007. ITW’07. IEEE, pages 242–247. IEEE, 2007. 115
[274] S. Sanghavi, D. Shah, and A. S. Willsky. Message passing for max-weight independent set. In Advances in
Neural Information Processing Systems, pages 1281–1288, 2007. 25, 99
[275] T. Schiex, H. Fargier, and G. Verfaillie. Valued constraint satisfaction problems: hard and easy problems.
In Proceedings of the 14th international joint conference on Artificial intelligence-Volume 1, pages 631–637.
Morgan Kaufmann Publishers Inc., 1995. 8
[276] A. Schrijver. Min-max results in combinatorial optimization. Springer, 1983. 28
[277] B. Selman, H. A. Kautz, and B. Cohen. Noise strategies for improving local search. In AAAI, volume 94,
pages 337–343, 1994. 79, 80
[278] P. P. Shenoy. Valuation-based systems: A framework for managing uncertainty in expert systems. In Fuzzy
logic for the management of uncertainty, pages 83–104. John Wiley & Sons, Inc., 1992. 8
[279] P. P. Shenoy and G. Shafer. Axioms for probability and belief-function propagation. In Uncertainty in
Artificial Intelligence, 1990. 2
[280] D. Sherrington and S. Kirkpatrick. Solvable model of a spin-glass. Physical review letters, 35(26):1792,
1975. 13
[281] S. E. Shimony. Finding maps for belief networks is np-hard. Artificial Intelligence, 68(2):399–410, 1994. 17
[282] L. Song, A. Gretton, and C. Guestrin. Nonparametric Tree Graphical Models. AISTATS 13, pages 765–772,
2010. 68
[283] L. Song, A. Gretton, D. Bickson, Y. Low, and C. Guestrin. Kernel Belief Propagation. Artificial Intelligence,
15(m):11–13, 2011. 68
156
APPENDIX . REFERENCES
[284] D. Sontag and T. S. Jaakkola. New outer bounds on the marginal polytope. In Advances in Neural
Information Processing Systems, pages 1393–1400, 2007. 42
[285] D. Sontag, T. Meltzer, A. Globerson, T. S. Jaakkola, and Y. Weiss. Tightening lp relaxations for map using
message passing. arXiv preprint arXiv:1206.3288, 2012. 46
[286] D. A. Sontag. Cutting plane algorithms for variational inference in graphical models. PhD thesis, Massachusetts Institute of Technology, 2007. 42
[287] K. Steiglitz and C. H. Papadimitriou. Combinatorial optimization: Algorithms and complexity. PrinticeHall, New Jersey, 1982. 1
[288] L. J. Stockmeyer. The polynomial-time hierarchy. Theoretical Computer Science, 3(1):1–22, 1976. 13
[289] J. X. J. W. P. Tan and L. Quan. Joint affinity propagation for multiple view segmentation. 102
[290] R. M. Tanner. A recursive approach to low complexity codes. Information Theory, IEEE Transactions on,
27(5):533–547, 1981. 1
[291] D. Tarlow, I. E. Givoni, and R. S. Zemel. Hop-map: Efficient message passing with high order potentials.
In International Conference on Artificial Intelligence and Statistics, pages 812–819, 2010. 39
[292] D. Tarlow, K. Swersky, R. S. Zemel, R. P. Adams, and B. J. Frey. Fast exact inference for recursive
cardinality models. arXiv preprint arXiv:1210.4899, 2012. 39, 40
[293] Y. W. Teh and M. Welling. The unified propagation and scaling algorithm. Advances in Neural Information
Processing {. . . }, 2002. 46
[294] S. Thrun. Particle filters in robotics. In Proceedings of the Eighteenth conference on Uncertainty in artificial
intelligence, pages 511–518. Morgan Kaufmann Publishers Inc., 2002. 1
[295] L. Tierney. Markov chains for exploring posterior distributions. Annals of Statistics, 22(4):1701–1728, 1994.
67
[296] S. Toda. Pp is as hard as the polynomial-time hierarchy. SIAM Journal on Computing, 20(5):865–877,
1991. 20
[297] L. G. Valiant. The complexity of computing the permanent. Theoretical computer science, 8(2):189–201,
1979. 18, 112
[298] L. G. Valiant and V. V. Vazirani. Np is as easy as detecting unique solutions. Theoretical Computer Science,
47:85–93, 1986. 24
[299] U. Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416, 2007. 131
[300] J. Von Neumann and O. Morgenstern. Theory of Games and Economic Behavior (60th Anniversary Commemorative Edition). Princeton university press, 2007. 39
157
[301] P. O. Vontobel. The bethe permanent of a nonnegative matrix. Information Theory, IEEE Transactions
on, 59(3):1866–1901, 2013. 113
[302] K. W. Wagner. The complexity of combinatorial problems with succinct input representation. Acta Informatica, 23(3):325–356, 1986. 13
[303] M. J. Wainwright and M. I. Jordan. Graphical Models, Exponential Families, and Variational Inference.
Foundations and Trends in Machine Learning, 1(12):1–305, 2007. 1, 43
[304] M. J. Wainwright, T. S. Jaakkola, and A. S. Willsky. A New Class of Upper Bounds on the Log Partition
Function. IEEE Transactions on Information Theory, 51(7):2313–2335, July 2005. 46
[305] M. J. Wainwright, T. S. Jaakkola, and A. S. Willsky. MAP Estimation Via Agreement on Trees: MessagePassing and Linear Programming. IEEE Transactions on Information Theory, 51(11):3697–3717, Nov.
2005. 46
[306] C. Wang, J. Lai, and W. Zheng. Message-passing for the traveling salesman problem. 121
[307] S. Wasserman. Social network analysis: Methods and applications, volume 8. Cambridge university press,
1994. 1
[308] Y. Watanabe and M. Chertkov. Belief propagation and loop calculus for the permanent of a non-negative
matrix. Journal of Physics A: Mathematical and Theoretical, 43(24):242002, 2010. 113
[309] M. Weigt and H. Zhou. Message passing for vertex covers. Physical Review E, 74(4):046110, 2006. 99
[310] Y. Weiss. Correctness of belief propagation in Gaussian graphical models of arbitrary topology. Neural
computation, (June), 2001. 36
[311] Y. Weiss, C. Yanover, and T. Meltzer. MAP Estimation, Linear Programming and Belief Propagation with
Convex Free Energies. Uncertainty in Artificial Intelligence, pages 416–425, 2007. 46
[312] A. Weller and T. Jebara. Clamping variables and approximate inference. In Advances in Neural Information
Processing Systems, pages 909–917, 2014. 49
[313] A. Weller and T. S. Jebara. On map inference by mwss on perfect graphs. arXiv preprint arXiv:1309.6872,
2013. 36
[314] M. Welling. On the choice of regions for generalized belief propagation. In Proceedings of the 20th conference
on Uncertainty in {. . . }, 2004. 49
[315] M. Welling. A Cluster-Cumulant Expansion at the Fixed Points of Belief Propagation. UAI, (x), 2012. 50
[316] M. Welling, M. Rosen-Zvi, and G. E. Hinton. Exponential family harmoniums with an application to
information retrieval. In Advances in neural information processing systems, pages 1481–1488, 2004. 1
[317] M. Welling, T. Minka, and Y. Teh. Structured Region Graphs: Morphing EP into GBP. UAI 2005, (1998):
609–616, 2005. 49
158
APPENDIX . REFERENCES
[318] S. White and P. Smyth. A spectral clustering approach to finding communities in graph. In SDM. SIAM,
2005. 100
[319] W. Wiegerinck, T. Heskes, et al. Fractional belief propagation. Advances in Neural Information Processing
Systems, pages 455–462, 2003. 46
[320] J.-Q. Xiao and H. Zhou. Partition function loop series for a general graphical model: free energy corrections
and message-passing equations. Journal of Physics A: Mathematical and Theoretical, 44(42):12, 2011. 50
[321] Y. Xiao, M. Xiong, W. Wang, and H. Wang. Emergence of symmetry in complex networks. Physical Review
E, 77(6):066108, 2008. 124
[322] X. Xu, N. Yuruk, Z. Feng, and T. A. Schweiger. Scan: a structural clustering algorithm for networks. In
SIGKDD, 2007. 100
[323] C. Yanover and Y. Weiss. Approximate inference and protein-folding. In Advances in neural information
processing systems, pages 1457–1464, 2002. 1
[324] C. Yanover, T. Meltzer, and Y. Weiss. Linear Programming Relaxations and Belief Propagation - an
Empirical Study. Journal of Machine Learning Research, 7:1887–1907, 2006. 46
[325] C. Yanover, T. Meltzer, and Y. Weiss. Linear programming relaxations and belief propagation–an empirical
study. JMLR, 2006. 42
[326] J. Yedidia and W. Freeman. Generalized belief propagation. Advances in neural information, 2001. 2
[327] J. Yedidia, W. Freeman, and Y. Weiss. Constructing Free-Energy Approximations and Generalized Belief
Propagation Algorithms. IEEE Transactions on Information Theory, 51(7):2282–2312, 2005. 45, 49
[328] J. S. Yedidia, W. T. Freeman, and Y. Weiss. Characterization of belief propagation and its generalizations.
ITIEEE, 51(TR-2001-15):1–16, 2001. 45, 48, 49
[329] A. L. Yuille. CCCP algorithms to minimize the Bethe and Kikuchi free energies: convergent alternatives
to belief propagation. Neural Computation, 14(7):1691–1722, 2002. 46
[330] L. Zdeborova and F. Krzakala. Phase transitions in the coloring of random graphs. Physical Review E, 76
(3):031131, Sept. 2007. 87
[331] L. Zdeborová and M. Mézard. The number of matchings in random graphs. Journal of Statistical Mechanics:
Theory and Experiment, 2006(05):P05003, 2006. 114
[332] P. Zhang and C. Moore. Scalable detection of statistically significant communities and hierarchies: messagepassing for modularity. arXiv preprint arXiv:1403.5787, 2014. 106
[333] H. Zhou. Vertex cover problem studied by cavity method: Analytics and population dynamics. The
European Physical Journal B-Condensed Matter and Complex Systems, 32(2):265–270, 2003. 99
[334] H. Zhou, C. Wang, J.-Q. Xiao, and Z. Bi. Partition function expansion on region graphs and messagepassing equations. Journal of Statistical Mechanics: Theory and Experiment, 2011(12):L12001, Dec. 2011.
50
Appendix
160
APPENDIX . REFERENCES
input : Graph G = (V, E) with normalized (weighted)adjacency A, maximum iterations Tmax ,
damping λ, threshold max .
output: A clustering C = {C1 , . . . , CK } of nodes.
construct the null model
b
pi:j ← 0 ∀(i, j) ∈ E ∪ E null
while true do // the augmentation loop
← 0, T ← 0
while < max and T < Tmax do // BP loop
←0
for (i, j) ∈ E ∪ E null do
b
p0i:j ← b
pi:j
b
pi:j ← (Ai,j − Anull
i,j )
for I ∈ ∂i : j do // update beliefs
calculate b
pI→i:j using equation (4.7)
b
pi:j ← b
pi:j + b
pI→i:j
end
← max{, |b
pi:j − b
p0i:j |}
for I ∈ ∂i : j do // update msgs.
e
pi:j→I ← b
pi:j − b
pI→i:j
b
pi:j→I ← λe
pi:j→I + (1 − λ)b
pi:j→I
end
end
T ←T +1
end
for i ∈ V do
for (i, j), (i, k) ∈ E ∪ E null do
if b
pi:j > 0 and b
pi:k > 0 and b
pi:k ≤ 0 then add the corresponding clique factor to the
factor-graph
end
end
if no factor was added then break out of the loop
else b
pi:j→I ← 0 ∀I, i : j ∈ I
end
C ← ConnectedComponents((V, {(i, j) ∈ E ∪ E null | b
pi:j > 0}))
Algorithm 3: Message Passing for Modularity Maximization.
161
input : Graph G = (V, E), weighted (symmetric) adjacency matrix A, maximum iterations
Tmax , damping λ, threshold max .
output: A subset T ⊂ E of the edges in the tour.
construct the initial factor-graph
initialize the messages for degree constraints b
pi:j→E(i,·) ← 0∀i ∈ V, j ∈ E(i, ·)
initialize b
pi:j ← Ai,j ∀(i, j) ∈ E
while true do // the augmentation loop
← 0, T ← 0
while < max and T < Tmax do // BP loop
←0
for each fI do // including
from this factor)
fE(S,·) , fE(i,·) (updates all the outgoing messages
find three lowest values in {b
pi:j→I | i : j ∈ ∂I}
for each i : j ∈ I do
calculate e
pI→i:j using equation (5.7)
I→i:j ← e
pI→i:j − b
pI→i:j
b
pI→i:j ← b
pI→i:j + λI→i:j
b
pi:j ← b
pi:j + I→i:j
← max(, |I→e |)
end
end
T ←T +1
end
T ← {(i, j) ∈ E | b
pi:j > 0}
// respecting degree constraints.
C ← ConnectedComponents((V, T ))
if |C| = 1 then return T
else augment the factor-graph with fE(S,·) ∀S ∈ C
initialize b
pE(S,·)→i:j ← 0 ∀S ∈ C, i : j ∈ E(S, ·)
end
Algorithm 4: Message Passing for TSP
| 0math.AC
|
Concentration behavior of the penalized least
squares estimator
arXiv:1511.08698v2 [math.ST] 19 Oct 2016
Penalized least squares behavior
Alan Muro and Sara van de Geer
{muro,geer}@stat.math.ethz.ch
Seminar für Statistik, ETH Zürich
Rämistrasse 101, 8092 Zürich
Abstract
Consider the standard nonparametric regression model and take as estimator the penalized least squares function. In this article, we study the trade-off
between closeness to the true function and complexity penalization of the estimator, where complexity is described by a seminorm on a class of functions.
First, we present an exponential concentration inequality revealing the concentration behavior of the trade-off of the penalized least squares estimator around
a nonrandom quantity, where such quantity depends on the problem under consideration. Then, under some conditions and for the proper choice of the tuning
parameter, we obtain bounds for this nonrandom quantity. We illustrate our results with some examples that include the smoothing splines estimator.
Keywords: Concentration inequalities, regularized least squares,
statistical trade-off.
1
Introduction
Let Y1 , ..., Yn be independent real-valued response variables satisfying
Yi = f 0 (xi ) + i , i = 1, ..., n,
where x1 , ..., xn are given covariates in some space X , f 0 is an unknown function in
a given space F, and 1 , ..., n are independent standard Gaussian random variables.
We assume that f 0 is “smooth” in some sense but that this degree of smoothness is
unknown. We will make this clear below.
1
To estimate f 0 , we consider the penalized least squares estimator given by
fˆ := arg min ||Y − f ||2n + λ2 I 2 (f ) ,
(1)
f ∈F
where λ > 0 is a tuning parameter and I a given seminorm on F. Here, for a vector
u ∈ Rn , we write ||u||2n := uT u/n, and we apply the same notation f for the vector
f = (f (x1 ), ..., f (xn ))T and the function f ∈ F. Moreover, to avoid digressions from
our main arguments, we assume that expression (1) exists and is unique.
For a function f ∈ F, define
τ 2 (f ) := ||f − f 0 ||2n + λ2 I 2 (f ).
(2)
This expression can be seen as a description of the trade-off of f . The term ||f − f 0 ||n
measures the closeness of f to the true function, while I(f ) quantifies its “smoothness”. As mentioned above, we only assume that f 0 is not too complex, but that this
degree of complexity is not known. This accounts for assuming that I(f 0 ) < ∞,
but that an upper bound for I(f 0 ) is unknown. Therefore, we choose as model class
F = {f : I(f ) < ∞}. It is important to see that taking as model class F0 = {f :
I(f ) ≤ M0 }, for some fixed M0 > 0, instead of F could lead to a model misspecification error if the unknown smoothness of f 0 is large.
Estimators with roughness penalization have been widely studied. Wahba (1990) and
Green and Silverman (1993) consider the smoothing
estimator, which correR 1 (m) splines
2
2
sponds to the solution of (1) when I (f ) = 0 |f (x)| dx, where f (m) denotes the
m-th derivative of f . Gu (2002) provides results for the more general penalized likelihood estimator with a general quadratic functional as complexity regularization. If
we assume that this functional is a seminorm and that the noise follows a Gaussian
distribution, then the penalized likelihood estimator reduces to the estimator in (1).
Upper bounds for the estimation error can be found in the literature (see e.g. van der
Vaart and Wellner (1996), del Barrio et al. (2007)). When no complexity regularization
term is included, the standard method used to derive these is roughly as follows: first,
the basic inequality
!
n
X
2
||fˆ − f 0 ||2n ≤ sup
i (f (xi ) − f 0 (xi ))
(3)
n f
i=1
is invoked. Then, an inequality for the right hand side of (3) is obtained for functions
f in {g ∈ F : ||g − f 0 ||n ≤ R} for some R > 0. Finally, upper bounds for the estimation error are obtained with high probability by using entropy computations. When
a penalty term is included, a similar approach can be used, but in this case the process
2
in (3) is studied in terms of both ||f − f 0 ||n and the smoothness of the functions considered, i.e., I(f ) and I(f 0 ).
A limitation of the approach mentioned above is that it does not allow us to obtain
lower bounds. Consistency results have been proved (e.g. van de Geer and Wegkamp
(1996)), but it is not clear how to use these to derive explicit bounds. Chatterjee (2014)
proposes a new approach to estimate the exact value of the error of least square estimators under convex constraints. It provides a concentration result for the estimation
error by relating it with the expected maxima of a Gaussian process. One can then
use this result to get both upper and lower bounds. In van de Geer and Wainwright
(2016), a more “direct” argument is employed to show that the error for a more general class of penalized least squares estimators is concentrated around its expectation.
Here, the penalty is only assumed to be convex. Moreover, the authors also consider
the approach from Chatterjee (2014) to derive a concentration result for the trade-off
τ (fˆ) for uniformly bounded function classes under general loss functions.
The goal of this paper is to contribute to the study of the “behavior” of τ (fˆ) from
a theoretical point of view. We consider the approach from Chatterjee (2014) and extend the ideas to the penalized least squares estimator with penalty based on a squared
seminorm without making assumptions on the function space. We present a concentration inequality showing that τ (fˆ) is concentrated around a nonrandom quantity R0
(defined below) in the nonparametric regime. Here, R0 depends on the sample size
and the problem under consideration. Then, we derive upper and lower bounds for
R0 in Theorem 2. These are obtained for the proper choice of λ and two additional
conditions, including an entropy assumption. Combining this result with Theorem 1,
one can obtain both upper and lower bounds for τ (fˆ) with high probability for a sufficiently large sample size. We illustrate our results with some examples in section 2.2
and observe that we are able to recover optimal rates of convergence for the estimation
error from the literature.
We now introduce further notation that will be used in the following sections. Denote the minimum of τ over F by
Rmin := min τ (f ),
f ∈F
and let this minimum be attained by fmin . Note that fmin can be seen as the unknown
noiseless counterpart of fˆ and that Rmin is a nonrandom unknown quantity. For two
vectors u, v ∈ Rn , let hu, vi denote the usual inner product. For R ≥ Rmin and λ > 0,
define
F(R) := {f ∈ F : τ (f ) ≤ R} ,
and
3
Mn (R) := sup h, f − f 0 i/n.
f ∈F (R)
Additionally, we write
M (R) := EMn (R),
Hn (R) := Mn (R) −
R2
,
2
H(R) := EHn (R).
Moreover, define the random quantity
R∗ := arg max Hn (R),
R≥Rmin
and the nonrandom quantity
R0 := arg max H(R).
R≥Rmin
From Lemma 1, it will follow that R∗ and R0 are unique.
For ease of exposition, we will use the following asymptotic notation throughout this
∞
paper: for two positive sequences {xn }∞
n=1 and {yn }n=1 , we write xn = O(yn ) if
lim sup |xn /yn | < ∞ as n → ∞ and xn = o(yn ) if xn /yn → 0 as n → ∞. Moreover,
we employ the notation xn yn if xn = O(yn ) and yn = O(xn ). In addition, we
make use of the stochastic order symbols OP and oP .
It is important to note that the quantities λ, τ (fˆ), τ (f 0 ), Rmin , M (R), H(R), R∗ , and
R0 depend on the sample size n. However, we omit this dependence in the notation to
simplify the exposition.
1.1
Organization of the paper
First, in section 2.1, we present the main results: Theorems 1 and 2. Note that the
former does not require further assumptions than those from section 1, while the latter
needs two extra conditions, which will be introduced after stating Theorem 1. Then, in
section 2.2, we illustrate the theory with some examples and, in section 2.3, we present
some concluding remarks. Finally, in section 3, we present all the proofs. We deferred
to the appendix results from the literature used in this last section.
2
2.1
Behavior of τ (fˆ)
Main results
The first theorem provides a concentration probability inequality for τ (fˆ) around R0 .
It can be seen as an extension of Theorem 1.1 from Chatterjee (2014) to the penalized
least squares estimator with a squared seminorm on F in the penalty term.
4
Theorem 1. For all λ > 0 and x > 0, we have
!
4 √
ˆ
τ (f )
x ( nR0 )2
P
− 1 ≥ x ≤ 3 exp −
.
R0
32(1 + x)2
√
Asymptotics. If R0 satisfies 1/( nR0 ) = o(1), then
τ (fˆ)
− 1 = oP (1).
R0
Therefore, the random fluctuations of τ (fˆ) around R0 are of negligible size in comparison with R0 . Moreover, this asymptotic result implies that the asymptotic distribution
of τ (fˆ)/R0 is degenerate. In Theorem 2, we will provide bounds for R0 under some
additional conditions and observe that R0 satisfies the condition from above.
Remark 1. Note that we only consider a square seminorm in the penalty term. A
generalization of our results to penalties of the form I q (·), q ≥ 2, is straightforward
but omitted for simplicity. The case q < 2 is not considered in our study since our
method of proof requires that the square root of the penalty term is convex, as can be
observed in the proof of Lemma 1.
Before stating the first condition required in Theorem 2, we will introduce the following definition: let S be some subset of a metric space (S, d). For δ > 0, the δ-covering
number N (δ, S, d) of S is the smallest value of N such that there exist s1 , ..., sN in S
such that
min d(s, sj ) ≤ δ, ∀ s ∈ S.
j=1,...,N
Moreover, for δ >0, the δ-packing number ND (δ, S, d) of S is the largest value of N
such that there exist s1 , ..., sN in S with
d(sk , sj ) > δ, ∀k 6= j.
Condition 1. Let α ∈ (0, 2). For all λ > 0 and R ≥ Rmin , we have
α
R/λ
, u > 0,
log N (u, F(R), || · ||n ) = O
u
and for some c0 ∈ (0, 2],
1
= O ((c0 λ)α ) .
log N (c0 R, F(R), || · ||n )
5
Condition 1 can be seen as a description of the richness of F(R). We refer the reader to
Kolmogorov and Tihomirov (1959) and Birman and Solomjak (1967) for an extensive
study on entropy bounds, and to van der Vaart and Wellner (1996) and van de Geer
(2000) for their application in empirical process theory.
Condition 2. Rmin τ (f 0 ).
Condition 2 relates the roughness penalty of f 0 with the minimum trade-off achieved
by functions in F. It implies that our choice of the penalty term is appropriate. In
other words, I(fmin ) and I(f 0 ) are not “too far away” from each other when the
tuning parameter is chosen properly. Therefore, aiming to mimic the trade-off of fmin
points us in the right direction as we would like to estimate f 0 .
The following result provides upper and lower bounds for R0 . Note that I(f 0 ) is
permitted to depend on the sample size. We assume that I(f 0 ) remains upper bounded
and bounded away from zero as the sample size increases. However, we allow these
bounds to be unknown.
Theorem 2. Assume Conditions 1 and 2, and that I(f 0 ) 1. For
1
λ n− 2+α
suitably chosen depending on α, one has
1
R0 n− 2+α .
Therefore, we obtain
1
τ (fˆ) n− 2+α
α
α
0 2+α
00 2+α
with probability at least 1 − 3 exp −c n
− 3 exp −c n
, where c0 , c00 denote
some positive constants not depending on the sample size.
Theorem 2 shows that one can obtain
√ bounds for R0 if we choose the tuning parameter
properly. Since the condition 1/( nR0 ) = o(1) is satisfied, one obtains then convergence in probability of the ratio τ (fˆ)/R0 to a constant. Moreover, the theorem also
provides bounds for the trade-off of fˆ. Note that these hold with probability tending
to one.
Remark 2. From Theorem 2, one obtains
||fˆ − f 0 ||2n + λ2 I 2 (fˆ) min ||f − f 0 ||2n + λ2 I 2 (f )
f ∈F
6
with high probability for a sufficiently large sample size. One can then say that the
trade-off of fˆ mimics that of its noiseless counterpart fmin , i.e., fˆ behaves as if there
were no noise in our observations. Therefore, our choice of λ allows us to control the
random part of the problem using the penalty term and “over-rule” the variance of the
noise.
Remark 3. From Theorem 2, one can observe that
1
||fˆ − f 0 ||n = OP (n− 2+α ),
I(fˆ) = OP (1).
Therefore, we are able to recover the rate of convergence for the estimation error of
fˆ. This will be illustrated in section 2.2. Furthermore, we observe that the degree of
smoothness of fˆ is bounded in probability. Although we have a lower bound for τ (fˆ),
this does neither imply (directly) a lower bound for ||fˆ − f 0 ||n , nor for I(fˆ).
2.2
Examples
In the following, we require that the assumptions in Theorem 2 hold. In each example,
we provide references to results from the literature where one can verify that Condition
1 is satisfied and refer the interested reader to these for further details. For the lower
bounds, one may first note that
N (u, F(R), || · ||n )
R
R
0
, || · ||n
≥ N u, f ∈ F : ||f − f ||n ≤ √ , I(f ) ≤ √
2
2λ
and additionally insert the results from Yang and Barron (1999), where the authors
show that often global and local entropies are of the same order for some f 0 .
R1
Example 1. Let X = [0, 1] and I 2 (f ) = 0 |f (m) (x)|2 dx for some m ∈ {2, 3, ...}.
Then fˆ is the smoothing spline estimator and can be explicitly computed (e.g. Green
and Silverman (1993)). Moreover, it can be shown that in this case Condition 1 holds
with α = 1/m under some conditions on the design matrix (Kolmogorov and Tihomirov (1959) and Example 2.1 in van de mGeer (1990)). Therefore from Theorem
2, we have that the standard choice λ n− 2m+1 yields
m
R0 n− 2m+1 .
Moreover, we obtain upper and lower bounds for τ (fˆ) with high probability for large
n. Then by Remark 3, we recover the optimal rate of convergence for ||fˆ−f 0 ||n (Stone
(1982)).
7
Now we consider the case where the design points have a larger dimension. Let
X = [0, 1]d with d ≥ 2 and define the d-dimensional P
index r = (r1 , ..., rd ), where
the values ri are non-negative integers. We write |r| = di=1 ri . Furthermore, denote
by Dr the differential operator defined by
Dr f (x) =
∂ |r|
f (x1 , ..., xd )
∂xr11 · · · ∂xrdd
and consider as roughness penalization
2
I (f ) =
XZ
|Dr f (x)|2 dx,
|r|=m
with m > d/2. In this case, Condition 1 holds with α = d/m (Birman
and Solomjak
m
− 2m+d
we obtain
(1967)) Then by Theorem 2, we have that for the choice λ n
m
R0 n− 2m+d .
Similarly as above, we are able to recover the optimal rate for the estimation error.
Example 2. In this example, define the total variation penalty as
T V (f ) =
n
X
|f (xi ) − f (xi−1 )|,
i=2
where x1 , ..., xn denote the design points.
Let X be real-valued and I 2 (f ) = (T V (f ))2 . In this case, Condition 1 is fulfilled
for α = 1 (Birman and Solomjak (1967)). The advantage of the total variation penalty
over that from Example 1 is that it can be used for unbounded X .
1
Let α = 1 in Theorem 2. For the choice λ n− 3 , we have
1
R 0 n− 3 ,
and we also obtain bounds for τ (fˆ) with high probability for large n. By Remark 3,
we also recover the optimal upper bound for the estimation error.
2.3
Conclusions
Theorem 1 derives a concentration result for τ (fˆ) around a nonrandom quantity rather
than just an upper bound. In particular, we√observe that the ratio τ (fˆ)/R0 convergences in probability to 1 if R0 satisfies 1/( nR0 ) = o(1). This condition holds in
the nonparametric setting for λ suitably chosen and I(f 0 ) 1, as shown in Theorem
2 and in our examples from section 2.2.
8
The strict concavity of H and Hn (Lemma 1) plays an important role in the derivation of both theorems. In our work, the proof of this property requires that the square
root of the penalty term is convex. Furthermore, the proof of both Theorems 1 and
2 rely on the fact that the noise vector is Gaussian. This can be seen in Lemma 6,
where we invoke a concentration result for functions of independent Gaussian random
variables, and in Lemma 4, where we employ a lower bound for the expected value of
the supremum of Gaussian processes to bound the function H.
3
Proofs
This section is divided in two parts. In section 3.1, we first state and prove the lemmas necessary to prove Theorem 1. These follow closely the proof of Theorem 1.1
from Chatterjee (2014), however, here we include a roughness penalization term for
functions in F. At the end of this section, we combine these lemmas to prove the
first theorem. In section 3.2, we first prove an additional result necessary to establish
Theorem 2. After this, we present the proof of the second theorem. Some results from
the literature used in our proofs are deferred to the appendix.
3.1
Proof of Theorem 1
Lemma 1. For all λ > 0, Hn (·) and H(·) are strictly concave functions.
Proof. Let λ > 0. Take any two values rs , rb such that Rmin ≤ rs ≤ rb and define
Fs,b := {ft = tvs + (1 − t)vb | t ∈ [0, 1], vs ∈ F(rs ), vb ∈ F(rb )}.
Take vt ∈ Fs,b and let r = trs + (1 − t)rb . By properties of a seminorm, we have that
I(vt ) < ∞, which implies that vt ∈ F. Moreover, we have that
p
||vt − f 0 ||2n + λ2 I 2 (vt )
p
p
≤ t ||vs − f 0 ||2n + λ2 I 2 (vs ) + (1 − t) ||vb − f 0 ||2n + λ2 I 2 (vb )
≤ tτ (vs ) + (1 − t)τ (vb ) ≤ trs + (1 − t)rb ,
p
where the first inequality uses the fact that the square root of the penalty term λ2 I 2 (·)
is convex. Therefore, vt ∈ F(r). Using these equations, we have
τ (vt ) =
Mn (r) = sup h, f − f 0 i/n ≥ sup h, f − f 0 i/n
f ∈F (r)
f ∈Fs,b
= sup h, tvs + (1 − t)vb − f 0 i/n
vs ,vb ∈F
τ (vs )≤rs
τ (vb )≤rb
= tMn (rs ) + (1 − t)Mn (rb ).
9
(4)
Therefore, Mn is concave for all λ > 0. Taking expected value in the equations in (4)
2
yields that M is concave for all λ > 0. Since g(r) := − r2 is strictly concave, then Hn
and H are strictly concaves and we have our result.
Lemma 2. For all λ > 0, we have that τ (fˆ) = R∗ .
Proof. Let f ∗ ∈ F(R∗ ) be such that
h, f ∗ − f 0 i/n = sup h, f − f 0 i/n.
f ∈F (R∗ )
We will show first that τ (f ∗ ) = R∗ . Suppose τ (f ∗ ) = R̃ for some Rmin ≤ R̃ < R∗ .
Note that then Mn (R̃) = Mn (R∗ ), and therefore, we have
!
R∗ R̃
Hn (R̃) = Hn (R∗ ) +
−
> Hn (R∗ ),
2
2
which is a contradiction by definition of R∗ . We must then have that τ (f ∗ ) = R∗ .
Now we will prove that τ (fˆ) = τ (f ∗ ). For all λ > 0 and for all f ∈ F, we have
||Y − f ||2n + λ2 I 2 (f )
(
(5)
= ||Y − f 0 ||2n − 2 h, f − f 0 i/n −
1
||f − f 0 ||2n + λ2 I 2 (f )
2
!)
p
≥ ||Y − f 0 ||2n − 2Hn ( ||f − f 0 ||2n + λ2 I 2 (f ))
≥ ||Y − f ∗ ||2n + λ2 I 2 (f ∗ ).
In consequence, by definition of fˆ and the inequalities in (5), both fˆ and f ∗ minimize
||Y − f ||2n + λ2 I 2 (f ).
and by uniqueness, it follows that τ (fˆ) = τ (f ∗ ).
Lemma 3. For x > 0, define
s1 := R0 − x, s2 := R0 + x, z := H(R0 ) − x2 /4.
Moreover, define the event e = {{Hn (s1 ) < z} ∧ {Hn (s2 ) < z} ∧ {Hn (R0 ) > z}}.
We have
n x4
c
P (e ) ≤ 3 exp −
.
32(R0 + x)2
10
Proof. From the proof of Theorem 1.1 in Chatterjee (2014), one can easily observe
that for all λ > 0 and any R we have
H(R0 ) − H(R) ≥
(R − R0 )2
.
2
Applying the inequality above to R = s1 and R = s2 , we have that
H(si ) +
x2
x2
≤ H(R0 ) − ,
4
4
i = 1, 2.
By Lemma 6, we have
4
x2
− nx
P (Hn (R0 ) ≤ z) = P Hn (R0 ) ≤ H(R0 ) −
≤ e 32(R0 )2 .
4
Therefore, using again Lemma 6 yields
4
x2
− nx
≤ e 32(s1 )2 ,
P (Hn (s1 ) ≥ z) ≤ P Hn (s1 ) ≥ H(s1 ) +
4
and
4
x2
− nx
P (Hn (s2 ) ≥ z) ≤ P Hn (s2 ) ≥ H(s2 ) +
≤ e 32(s2 )2 .
4
By the equations above, we obtain
−
P (ec ) = P (Hn (s1 ) ≥ z) + P (Hn (s2 ) ≥ z) + P (Hn (R0 ) ≤ z) ≤ 3e
n x4
32(s2 )2
.
Now we are ready to prove Theorem 1.
Proof of Theorem 1. Let λ > 0. First, we note that H is equal to −∞ when R tends
to infinity or R < Rmin . Then, by Lemma 1, we know that R0 is unique.
For x > 0, define the event
e := {{Hn (s1 ) < z} ∧ {Hn (s2 ) < z} ∧ {Hn (R0 ) > z}} ,
where s1 = R0 − x, s2 = R0 + x, and z = H(R0 ) − x2 /4. Therefore, we have
that s1 < R0 < s2 by construction. Moreover, we know that Hn (R∗ ) ≥ Hn (R0 ) by
definition of R∗ . Since Hn is strictly concave by Lemma 1, we must have that, in e,
s1 < R∗ < s2 .
Combining Lemma 2 with equation (6) yields that, in e,
τ (fˆ)
x
−1 <
.
R0
R0
11
(6)
Therefore, by Lemma 3, and letting y = x/R0 , we have
!
4 √
τ (fˆ)
n y 4 R02
y ( nR0 )2
P
− 1 ≥ y ≤ 3 exp −
≤ 3 exp −
.
R0
32(1 + y)2
32(1 + y)2
3.2
Proof of Theorem 2
For proving Theorem 2, we will need the following result. This lemma gives us bounds
for the unknown nonrandom quantity H(R). We note that these bounds can be written
as parabolas with maximums and maximizers depending on α, n, and λ.
Lemma 4. Assume Condition 1 and let α be as stated there. For some constants
C ≥ 1/2, 0 < c0 ≤ 2, c2 > c1 > 0 not depending on the sample size and for all λ > 0,
we have
R ≥ Rmin ,
g1 (R) < H(R) < g2 (R),
where gi (R) = − 21 (R − Ki )2 +
Ki2
,
2
i = 1, 2, with
1−α/2 √
1/2
1
K1 := K1 (n, λ) :=
,
2
nλα
1/2
√
4C c2
1
.
K2 := K2 (n, λ) :=
2 − α nλα
c0
c1
Proof of Lemma 4. This proof makes use of known results for upper and lower bounds
for the expected maxima of random processes, which can be found in the appendix.
We will indicate this below.
Let λ > 0 and R ≥ Rmin
F(R) and 1 , ..., n standard Gaussian random
Pn. For f ∈ √
variables, define Xf := i=1 i f (xi )/ n. Take any two functions f, f 0 ∈ F and note
that Xf − Xf 0 follows a Gaussian distribution with expected value E[Xf − Xf 0 ] = 0
and variance
V ar(Xf − Xf 0 ) = ||f ||2n + ||f 0 ||2n − 2Cov(Xf , Xf 0 ) = ||f − f 0 ||2n .
Therefore {Xf : f ∈ F(R)} is a sub-Gaussian process with respect to the metric
d(f, f 0 ) = ||f − f 0 ||n on its index set (see Appendix). Note that, if we define the
diameter of F(R) as diamn (F(R)) := supx,y∈F (R) ||x − y||n , then it is not difficult to
see that diamn (F(R)) ≤ 2R.
Now we proceed to obtain bounds for M (R). By Dudley’s entropy bound (see Lemma
7 in Appendix) and Condition 1, for some constants C ≥ 1/2 and c2 , we have
12
1
M (R) = √ E
n
"
#
2R
r
log N (u, F(R), || · ||n )
du
n
0
√ Z
C c2 2R R α/2
≤ √
du
uλ
n 0
1/2
√
4C c2
1
≤
R.
2 − α nλα
Z
sup Xf − Xf 0 ≤ C
f ∈F (R)
Moreover, by Sudakov lower bound (see Lemma 8 in Appendix) we have
r
1
log ND (, F(R), || · ||n )
sup
≤ M (R).
2 0<≤diamn (F (R))
n
For some 0 < c0 ≤ 2, let diamn (F(R)) = c0 R and take = c0 R in the last equation.
By Condition 1, for some constant c1 we have
c0 R
2
r
r
log ND (c0 R, F(R), || · ||n )
c0 R log N (c0 R, F(R), || · ||n )
≥
n
2
n
1/2
1−α/2 √
c
c1
1
≥ 0
R.
2
nλα
Then, by the equations above and the definition of H(R), we have, for all R ≥ Rmin ,
1−α/2 √
c0
2
c1
1
nλα
1/2
1/2
√
4C c2
1
R2
R2
< H(R) <
.
R−
R
−
2
2 − α nλα
2
Writing Ki R − R2 /2 = − 12 (R − Ki )2 + Ki2 /2 for i = 1, 2 completes the proof.
Now, we are ready to prove Theorem 2.
Proof of Theorem 2. By Condition 2 and I(f 0 ) 1, we know that there exist constants 0 < b1 < b2 not depending on n such that
b1 λ ≤ Rmin ≤ b2 λ.
Take
1
(7)
1
c0 n− 2+α ≤ λ ≤ c00 n− 2+α
(8)
2
2+α
1−α/2 √
c
c
with c0 , c00 constants satisfying 0 < c0 < c00 ≤ 0 2b2 1
, where α is as in Condition 1, c0 and c1 are as in Lemma 4, and b2 as in equation (7).
13
First, we will derive bounds for R0 . Let g1 and g2 be as in Lemma 4 and recall that
g1 (R) < H(R) < g2 (R) for R ≥ Rmin . Note that g1 (R) ≤ K12 /2 for all R ≥ Rmin
and that this upper bound is reached at R = K1 . Moreover, we know that the function
g2 attain negative values when R < 0 and when R > 2K2 . Then, by strict concavity
of H (Lemma 1), if Rmin ≤ K1 , we must have that R0 ≤ 2K2 .
Now, combining equation (7) and the choice in (8), we obtain that Rmin ≤ K1 .
Therefore, following the rationale from above and substituting equation (8) into the
definition of K2 , we have that there exist some constant a2 > 0 such that
1
R0 ≤ a2 n− 2+α
(9)
Furthermore, combining again equations (7) and (8), and recalling that Rmin ≤ R0
yields that, for some constant a1 > 0,
1
R0 ≥ a1 n− 2+α
(10)
Joining equations (9) and (10) gives us the first result in our theorem.
We proceed to obtain bounds for τ (fˆ). Let A1 , A2 be some constants such that
0 < A1 < a1 < a2 < A2 . We have
1
1
P τ (fˆ) ≤ A1 n− 2+α = P R0 − τ (fˆ) ≥ R0 − A1 n− 2+α
1
− 2+α
ˆ
≤ P R0 − τ (f ) ≥ (a1 − A1 )n
!
α
(a1 − A1 )4 n 2+α
≤ 3 exp −
,
32(a2 + a1 − A1 )
where in the first inequality we used equation (10), and in the second, Theorem 1 and
equation (9). Similarly, we have
1
P τ (fˆ) ≥ A2 n− 2+α
1
≤ P τ (fˆ) − R0 ≥ (A2 − a2 )n− 2+α
!
α
(A2 − a2 )4 n 2+α
≤ 3 exp −
,
32A2
where in the first inequality we use equation (9), and in the second, Theorem 1 and
again equation (9). Therefore, we have
14
1
1
P A1 n− 2+α ≤ τ (fˆ) ≤ A2 n− 2+α
1
1
− 2+α
− 2+α
ˆ
ˆ
= P τ (f ) ≤ A2 n
− P τ (f ) ≤ A1 n
!
!
α
α
(a1 − A1 )4 n 2+α
(A2 − a2 )4 n 2+α
≥ 1 − 3 exp −
− 3 exp −
32(a2 + a1 − A1 )
32A2
and the second result of the theorem follows.
References
Birman, M. Š. and M. Z. Solomjak (1967), Piecewise polynomial approximations of
functions of classes Wp α , Matematicheskii Sbornik (N.S.) 73 (115), 331–355.
Boucheron, S., G. Lugosi, and P. Massart (2013), Concentration inequalities: a nonasymptotic theory of independence, Oxford University Press.
Chatterjee, S. (Dec. 2014), A new perspective on least squares under convex constraint,
Annals of Statistics 42.6, 2340–2381.
del Barrio, E., P. Deheuvels, and S. van de Geer (2007), Lectures on Empirical Processes: theory and statistical applications, EMS series of lectures in mathematics,
European Mathematical Society.
Green, P. and B. Silverman (1993), Nonparametric regression and generalized linear
models: a roughness penalty approach, Chapman & Hall/CRC Monographs on Statistics & Applied Probability, Taylor & Francis.
Gu, C. (2002), Smoothing Spline ANOVA Models, IMA Volumes in Mathematics and
Its Applications, Springer.
Kolmogorov, A. N. and V. M. Tihomirov (1959), ε-entropy and ε-capacity of sets in
function spaces, Uspekhi Matematicheskikh Nauk 14.2 (86), 3–86.
Koltchinskii, V. (2011), Oracle inequalities in empirical risk minimization and sparse
recovery problems: École d’ Eté de Probabilités de Saint-Flour XXXVIII-2008, Ecole
d’Eté de Probabilités de Saint-Flour, Springer.
Stone, C. J. (Dec. 1982), Optimal Global Rates of Convergence for Nonparametric
Regression, Annals of Statistics 10.4, 1040–1053.
van de Geer, S. (2000), Empirical Processes in M-Estimation, Cambridge Series in
Statistical and Probabilistic Mathematics, Cambridge University Press.
van de Geer, S. (June 1990), Estimating a Regression Function, Annals of Statistics
18.2, 907–924.
van de Geer, S. and M. J. Wainwright (2016), On concentration for (regularized) empirical risk minimization, preprint, arXiv:1512.00677.
van de Geer, S. and M. Wegkamp (Dec. 1996), Consistency for the least squares estimator in nonparametric regression, Annals of Statistics 24.6, 2513–2523.
15
van der Vaart, A. and J. Wellner (1996), Weak convergence and Empirical Processes,
Springer Series in Statistics, Springer.
Wahba, G. (1990), Spline models for observational data, CBMS-NSF Regional Conference Series in Applied Mathematics, Society for Industrial and Applied Mathematics.
Yang, Y. and A. Barron (Oct. 1999), Information-theoretic determination of minimax
rates of convergence, Annals of Statistics 27.5, 1564–1599.
A
Appendix
Lemma 5. [Gaussian Concentration Inequality. See, e.g. Boucheron et al. (2013)]
Let X = (X1 , ..., Xn ) be a vector of n independent standard Gaussian random variables. Let f : Rn −→ R denote an L-Lipschitz function. Then, for all t > 0,
2 /(2L2 )
P (f (X) − Ef (X) ≥ t) ≤ e−t
.
The following lemma applies the Gaussian Concentration Inequality from above to
show that the quantities Hn and H are close with exponential probability, as exploited
by Chatterjee (2014).
Lemma 6. For all λ > 0, R ≥ Rmin , and t > 0, we have
2 /(2R2 )
P (|Hn (R) − H(R)| ≥ t) ≤ 2e−n t
.
Proof. Let λ > 0. In this proof, we will write
Mn (R) = Mn (R, ) = sup h, f − f 0 i/n.
f ∈F (R)
Let u and v be two n-dimensional standard Gaussian random vectors. By properties
of the supremum and by Cauchy-Schwarz inequality, we have
|Mn (R, u) − Mn (R, v)| =
sup hu, f − f 0 i/n − sup hv, f − f 0 i/n
f ∈F (R)
≤ sup
f ∈F (R)
hu, f − f 0 i/n − hv, f − f 0 i/n
f ∈F (R)
R
≤ sup ||u − v||n ||f − f 0 ||n ≤ √
n
f ∈F (R)
16
n
X
(ui − vi )2
i=1
!1/2
.
√
Therefore, Mn (R, ·) is (R/ n) - Lipschitz in its second argument. By the Gaussian
Concentration Inequality (Lemma 5), for every t > 0 and every R ≥ Rmin , we have
P (Mn (R, ) − M (R, ) ≥ t) ≤ e−n t
2 /(2R2 )
.
Now, take −Mn (R, ). Applying again the Gaussian Concentration Inequality yields
2 /(2R2 )
P (Mn (R, ) − M (R, ) ≤ −t) ≤ e−n t
.
Combining the last two equations yields the result of this lemma.
For the next lemma, we will need the following definition:
A stochastic process {Xt : t ∈ T } is called sub-Gaussian with respect to the semimetric d on its index set if
P (|Xs − Xt | > x) ≤ 2e
−
x2
2d2 (s,t)
for every s, t ∈ T, x > 0.
,
Lemma 7. [Dudley’s entropy bound. See, e.g. Koltchinskii (2011) ] If {Xt : t ∈ T }
is a sub-Gaussian process with respect to d, then the following bounds hold with some
numerical constant C > 0:
Z
D(T )
p
log N (, T, d) d
E sup Xt ≤ C
t∈T
0
and for all t0 ∈ T
Z
D(T )
E sup |Xt − Xt0 | ≤ C
t∈T
p
log N (, T, d) d
0
where D(T ) = D(T, d) denotes the diameter of the space T .
Lemma 8. [Sudakov Lower Bound. See, e.g. Boucheron et al. (2013) ] Let T be a
finite set and let (Xt )t∈T be a Gaussian vector with EXt = 0 ∀t. Then,
E sup Xt ≥
t∈T
p
1
min0 E[(Xt − Xt0 )2 ] log |T |.
2 t6=t
Moreover, let d be a pseudo-metric on T defined by d(t, t0 )2 = E[(Xt − Xt0 )2 ]. For all
> 0 smaller than the diameter of T , the lower bound from above can be rewritten as
1 p
E sup Xt ≥ log ND (, T, d).
2
t∈T
17
| 10math.ST
|
arXiv:1710.07576v1 [math.PR] 20 Oct 2017
A SHORT SURVEY ON BOUNDING THE UNION
PROBABILITY USING PARTIAL INFORMATION
JUN YANG, FADY ALAJAJI, AND GLEN TAKAHARA
Abstract. This is a short survey on existing upper and lower
bounds on the probability of the union of a finite number of events
using partial information given in terms of the individual or pairwise event probabilities (or their sums). New proofs for some of
the existing bounds are provided and new observations regarding
the existing Gallot–Kounias bound are given.
Contents
1. Introduction
2. Review of Existing Bounds
3. Observations on the GK Bound
References
2
5
14
16
Department of Statistical Sciences, University of Toronto, Canada
Department of Mathematics and Statistics, Queen’s University,
Canada
E-mail addresses: [email protected], [email protected],
[email protected].
1
2
A SHORT SURVEY
1. Introduction
Consider a finite family of events {A1 , . . . , AN } in a general probability space (Ω, F , P ), where N is a fixed positive integer. Note
that there are only finitely many Boolean atoms1 specified by the
the probability of the
Ai ’s [DC97]. We are interested
S in bounding
N
finite union of events, i.e., P
i=1 Ai , in terms of partial probabilistic event information such as knowing the individual event probabilities, {P (A1 ), . . . , P (AN )}, and the pairwise event probabilities {P (Ai ∩
Aj ), i 6= j}, or (linear) functions of the probabilities of individual and
pairwise events.
For example, the well-known union upper bound and the Bonferroni
inequality [GS96] are respectively given as follows:
!
N
N
[
X
(1)
P
Ai ≤
P (Ai ),
i=1
(2)
P
N
[
Ai
i=1
!
≥
N
X
i=1
P (Ai ) −
i=1
X
P (Ai ∩ Aj ).
i<j
We note that the union upper bound (1) is established in terms of only
PN
i=1 P (Ai ) so that each of the individual event probability P (Ai ) is
actually not needed. However, the Bonferroni lower bound (2) is estabP
P
lished using two terms, N
i=1 P (Ai ) and
i<j P (Ai ∩ Aj ). Therefore,
the union upper bound (1) and the Bonferroni inequality (2) are established based on different partial information on the event probabilities.
In order to distinguish the use of different partial information, we
assume that a vector θ = (θ1 , . . . , θm ) ∈ Rm represents partial probaS
bilistic information about the union N
i=1 Ai . Specifically, we assume
that for a given integer m ≥ 1, Θ denotes the range of a function of
N
P (Ai )’s and P (Ai ∩ Aj )’s, ηm : [0, 1]N +( 2 ) → Rm . Then θ equals to the
value of the function ηm for given A1 , . . . , AN . For example,
(3)
θ = (P (A1), P (A2 ), . . . , P (AN )) ,
or
(4)
θ=
N
X
i=1
1The
P (Ai ),
X
i<j
!
P (Ai ∩ Aj ) .
problem can be directly reduced to the finite probability space case. Thus,
we will consider finite probability spaces where ω ∈ Ω denotes an elementary outcome instead of an atom.
A SHORT SURVEY
3
Then, we can define a lower bound (and similarly an upper bound)
SN
on P
i=1 Ai that is established using the partial information represented by θ as follows.
S
N
A
Definition 1.1. A lower bound of P
i=1 i is a function of θ, ℓ(θ),
such that
!
N
[
(5)
P
Ai ≥ ℓ(θ),
i=1
for any set of events {Ai } that the value of ηm for given {Ai } equals to
θ.
Note that, for given θ, such as θ = (P (A1 ), . . . , P (AN )), there are
multiple functions of θ that are lower bounds, for example,
!
N
[
P
Ai ≥ θ1 = P (A1 ),
i=1
(6)
P
N
[
Ai
!
≥
Ai
!
≥ max θi = max P (Ai ).
i=1
P
N
[
i=1
P
i θi
N
i
=
P
i
P (Ai )
,
N
i
Therefore, we need to define an optimal lower bound in a general class
of lower bounds that are functions of θ.
S
N
Let LΘ denote the set of all lower bounds on P
i=1 Ai that are
functions of only θ.
Definition 1.2. We say that a lower bound ℓ⋆ ∈ LΘ is optimal in LΘ
if ℓ⋆ (θ) ≥ ℓ(θ) for all θ ∈ Θ and ℓ ∈ LΘ .
Definition 1.3. We say that a lower bound ℓ ∈ LΘ is achievable if for
every θ ∈ Θ,
!
N
[
Ai = ℓ(θ),
(7)
inf P
A1 ,...,AN
i=1
where the infimum ranges over all collections {A1 , . . . , AN }, Ai ∈ F ,
such that {A1 , . . . , AN } is represented by θ.
For bounds in LΘ , the following lemma shows that achievability is
equivalent to optimality.
4
A SHORT SURVEY
Lemma 1.4. A lower bound ℓ⋆ ∈ LΘ is optimal in LΘ if and only if
it is achievable.
Proof. Suppose that ℓ⋆ is achievable. Let θ ∈ Θ and ǫ > 0 be given,
and let ℓ be any lower bound in LΘ . By achievability there exist sets
A1 , . . . , AN in F represented by θ such that
!
N
[
ℓ⋆ (θ) > P
Ai − ǫ ≥ ℓ(θ) − ǫ.
(8)
i=1
Since this holds for any ǫ we have ℓ⋆ (θ) ≥ ℓ(θ). We prove the converse
by the contrapositive. Suppose that ℓ⋆ is not achievable. Then there
exists θ′ ∈ Θ such that
!
N
[
inf P
(9)
Ai > ℓ⋆ (θ′ ),
A1 ,...,AN
i=1
where the infimum ranges over all collections {A1 , . . . , AN }, Ai ∈ F ,
such that {A1 , . . . , AN } is represented by θ′ . Define ℓ by
c if θ = θ′
(10)
ℓ(θ) =
0 if θ 6= θ′ ,
where c satisfies
(11)
inf
A1 ,...,AN
P
N
[
i=1
Ai
!
> c > ℓ⋆ (θ′ ).
Then ℓ ∈ LΘ and is larger than ℓ⋆ at θ′ . Hence, ℓ⋆ is not optimal.
Using Lemma 1.4, we can therefore prove that a lower bound ℓ(θ)
is optimal if for any value of θ ∈ Θ, one can constructa collection of
SN
∗
= ℓ(θ). The
events {A∗i } that is represented by θ and P
i=1 Ai
optimal upper bound can also be defined similarly (using a supremum
in (7)) and proved by achievability. For example, one can easily verify
the following by a construction proof of achievability.
S
P
N
i P (Ai )
• P
is the optimal lower bound in the class
i=1 Ai ≥
N
P
forθ = (
i P (Ai )).
SN
≥ maxi P (Ai ) is the optimal lower bound in the
• P
i=1 Ai
class
. . . , P (AN )).
Sfor θ =
(P (A1 ), P
N
• P
≤ min{ i P (Ai ), 1} is the optimal upper bound
i=1 Ai
P
in the classes for both θ = ( i P (Ai )) and θ = (P (A1 ), . . . , P (AN )).
A SHORT SURVEY
5
Furthermore, we can prove that a lower bound is not optimal by
showing it is not achievable. For example, in order to show that the
in the class of
Bonferroni inequality (2) is not an optimal
Plower bound
P
lower bounds that are functions of θ =
P
(A
),
P
(A
∩
A
)
,
i
i
j
i
i<j
we only need to show it is not achievable. Note that for N > 3, the
lower bound (2) can have negative values. However, according to the
definition of achievability, the LHS of (7) can never be negative, which
means the lower bound (2) cannot be achievable. Therefore, the Bonferroni inequality (2) is not optimal.
Throughout the survey, we mainly focus on lower bounds using different partial probabilistic information. Upper bounds are presented
as remarks.
2. Review of Existing Bounds
P
PWe start from the class of lower bounds in terms of i P (Ai ) and
i<j P (Ai ∩ Aj ), for which the Dawson-Sankoff (DS) lower bound
[DS67] is known as optimal.
P Then we introduce some lower bounds in
terms of {P (Ai )} and { j P (Ai ∩ Aj )}, including the D. de Caen (DC)
bound [DC97] and the Kuai-Alajaji-Takahara (KAT) bound [KAT00a].
Next, a review of some lower bounds in terms of {P (Ai )} and {P (Ai ∩
Aj )} is given, including the algorithmic stepwise lower bound [KAT00b]
and the Gallot-Kounias (GK) bound [Gal66; Kou68; FLS10]. Finally,
some existing upper bounds are reviewed, including the Hunter upper
bound and the algorithmic greedy upper bound [KAT00b].
We first define the degree of an atom (or outcome in finite probability
space) ω ∈ F as follows.
Definition 2.1. For each atom ω ∈ F , let the degree of ω, denoted
by deg(ω), be the number of Ai ’s that contain ω.
S
Therefore, the degree of any atom in i Ai equals to an integer in
{1, . . . , N}.
P
P
2.1. Lower Bounds Using
P
(A
)
and
i
i
i<j P (Ai ∩ Aj ). ConP
P
sidering θ =
i P (Ai ),
i<j P (Ai ∩ Aj ) , we note that the Bonferroni inequality (2) is a lower bound in this class. However, we have
shown that (2)
P is not optimal,
P which means there exists another function of only i P (Ai ) and i<j P (Ai ∩ Aj ) that is a lower bound of
S
N
P
A
and always sharper than the Bonferroni inequality.
i
i=1
6
A SHORT SURVEY
Defining
(12)
a(k) := P
(
ω⊆
[
Ai , deg(ω) = k
i
)!
,
one can easily verify the following identities:
!
N
N
[
X
(13)
P
Ai =
a(k),
i=1
N
X
(14)
k=1
P (Ai ) =
i=1
X
(15)
ka(k),
k=1
P (Ai ∩ Aj ) =
i<j
X
N
X
N
X
k
k=2
P (Ai ∩ Aj ) = 2
i,j
X
2
a(k),
P (Ai ∩ Aj ) +
i<j
(16)
=2
=
N
X
P (Ai )
i=1
N
X
k(k − 1)
k=2
N
X
2
a(k) +
N
X
ka(k)
k=1
k 2 a(k).
k=1
Note that using the above equalities, one can derive a lower bound
simply via the Cauchy-Schwarz inequality:
!
!
!2
X
X
X
(17)
a(k)
k 2 a(k) ≥
ka(k) ,
k
k
k
where equality holds if and only if a(k) > 0 only for a particular k, i.e.,
all outcomes in the union has the same degree k. The resulting lower
bound can be written as
!
P
N
[
( i P (Ai ))2
.
(18)
P
Ai ≥ P
i,j P (Ai ∩ Aj )
i=1
P
P
P
Since i,j P (Ai ∩ Aj ) ≤ i,j P (Ai ) = N i P (Ai ),
P
P
( i P (Ai ))2
P (Ai )
P
;
≥ i
(19)
N
i,j P (Ai ∩ Aj )
A SHORT SURVEY
7
P
P (A )
hence the lower bound (18) is always sharper thanP i N i , which has
been shown to be optimal in the class of θ = ( i P (Ai )). This is
reasonable since Pthe lower bound (18) is established using more inP (A )
formation than i N i . However, it can be readily shown that the
lower bound (18) is not always sharper than the Bonferroni inequality
(2). Therefore, it cannot be theoptimal lower bound in the class of
P
P
θ=
i<j P (Ai ∩ Aj ) .
i P (Ai ),
2.1.1. Dawson-Sankoff (DS) Bound. The
is known as the
P DS bound P
optimal lower bound in terms of only i P (Ai ) and i<j P (Ai ∩ Aj ).
P
P
Denoting θ1 := i P (Ai ) and θ2 := i<j P (Ai ∩ Aj ), the DS bound
[DS67] can be written as
!
N
[
κθ12
(1 − κ)θ12
(20)
P
Ai ≥
+
,
(2
−
κ)θ
+
2θ
(1
−
κ)θ
+
2θ
1
2
1
2
i=1
2
2
− ⌊ 2θ
⌋ and ⌊x⌋ denotes the largest integer less than or
where κ = 2θ
θ1
θ1
equal to x.
We first show that the DS bound is the solution of a linear programming (LP) problem in the following lemma.
Lemma 2.2. The DS bound is the solution of the following LP problem.
X
a(k),
ℓDS := min
a(k)
s.t.
(21)
k
X
ka(k) =
P (Ai ),
i
k
X
X
2
k a(k) =
k
a(k) ≥ 0,
X
P (Ai ∩ Aj ),
i,j
k = 1, . . . , N.
Proof. For an LP problem, when a feasible solution exists and when
the objective function (which is linear) is bounded, the optimal value
of the objective function is always attained on the boundary of the
optimal level-set and it is attained on at least one of the vertices of
the polyhedron formed by the constraints (which is the set of feasible
solutions) [BT97]. Then, the lemma can be readily verified using this
fact that one of the optimal feasible points of the LP problem (21) is a
vertex. To obtain a vertex, one need to make N − 2 of the inequalities
a(k) ≥ 0 active, which means there are only two integers k1 and k2
8
A SHORT SURVEY
that 1 ≤ k1 < k2 ≤ N, satisfying
min a(k1 ) + a(k2 ),
k1 ,k2
s.t.
k1 a(k1 ) + k2 a(k2 ) =
(22)
X
P (Ai ),
i
k12 a(k1 )
+
k22 a(k2 )
=
X
P (Ai ∩ Aj ),
i,j
a(k1 ) ≥ 0,
a(k2 ) ≥ 0.
It can be easily
shown that the solution of the above problem is achieved
P
i,j P (Ai ∩Aj )
at k1 = ⌊ P P (Ai ) ⌋ and k2 = k1 + 1. Thus, the solution of (21) is
i
the DS bound.
The existing proof of the optimality of the DS bound can be seen,
e.g., in [GS96, p. 22]. We herein give an alternative and simpler proof
by proving it is achievable.
Lemma 2.3.The DS bound is optimal inthe class of lower bounds in
P
P
terms of θ =
i P (Ai ),
i<j P (Ai ∩ Aj ) .
Proof. We have shown that the DS bound is the solution of (21) and
can be written as ℓDS = a(k1 ) + a(k2 ), for some a(k1 ) ≥ 0, a(k2 ) ≥ 0,
and a(k) = 0, k 6= k1 , k 6= k2 . Recalling the definition of a(k), one can
construct two outcomes ω1 and ω2 in a finite probability space such
that
(23)
P (ω1) = a(k1 ),
P (ω2 ) = a(k2 ).
Then consider the following construction of collection of events {A∗i },
A∗i = {ω1 , ω2 },
(24)
A∗i
A∗i
= {ω2 },
if i ≤ k1 ,
if k1 < i ≤ k2 ,
= ∅,
otherwise.
S
Then we always have ℓDS = P ( i A∗i ). Therefore, the DS bound is
achievable, and hence optimal.
Note that since the DS bound is optimal, it is always sharper than
the lower bound in (18). Actually, this can be easily proved since the
lower bound in (18) is a lower bound of the objective function of (21)
by Cauchy-Schwarz inequality using the two constraints of (21).
A SHORT SURVEY
9
P
2.2. Lower Bounds Using {P (Ai )} and { j6=i P (Ai ∩ Aj )}. In this
P
section, we review the lower bounds in terms of {P (Ai )} and { j6=i P (Ai ∩
Aj )}, including the DC [DC97] and the KAT [KAT00a] bounds.
Similar to the definition of a(k), define
(25)
ai (k) := P ({ω ⊆ Ai : deg(ω) = k}) , i = 1, . . . , N, k = 1, . . . , N.
P
Then one can verify that i ai (k) = ka(k), i.e.,
P
ai (k)
.
(26)
a(k) = i
k
For simplicity, we denote
(27)
X
P (Ai ∩ Aj ),
αi := P (Ai ), βi :=
γi := αi + βi =
j6=i
X
P (Ai ∩ Aj ).
j
We examine lower bounds that
of θ = (α1 , . . . , αN , γ1 , . . . , γN ).
S are functions
N
One can verify that P
i=1 Ai , αi and γi can all be written as
linear functions of {ai (k)} as follows.
!
N
[
X
X X ai (k)
.
(28)
P
Ai =
a(k) =
k
i=1
i
k
(29)
αi = P (Ai ) =
X
k
ai (k),
k
γi =
X
j
P (Ai ∩ Aj ) =
X
kai (k).
k
2.2.1. D. de Caen (DC) bound. Similar to the lower bound in (18),
using the Cauchy-Schwarz inequality
!2
!
!
X
X
X ai (k)
ai (k)
kai (k) ≥
(30)
k
k
k
k
for i = 1, . . . , N, and summing over i, one can get the DC bound as
follows.
(31)
!
[
X X ai (k) X α2 X
P (Ai )2
i
P
=
≥
=: ℓDC .
P
Ai =
k
γi
j P (Ai ∩ Aj )
i
i
i
i
k
It is noted by D. de Caen [DC97] that the above lower bound can be
(but is not always) sharper than the DS bound.
10
A SHORT SURVEY
2.2.2. The Kuai-Alajaji-Takahara (KAT) bound. Now, we introduce
the KAT bound
# )
("
γi
γi
N
X
−
⌊
⌋
1
αi
αi
αi ,
(32)
ℓKAT :=
γi −
γi
γi
⌋
⌋)(⌊
⌋)
⌊
(1
+
⌊
αi
αi
αi
i=1
as the solution of an LP problem, which is given in the following Lemma.
Lemma 2.4. The KAT bound is the solution of the following LP problem
(33)
N X
N
X
ai (k)
min
k
{ai (k),i=1,...,N,k=1,...,N }
i=1 k=1
s.t.
N
X
ai (k) = αi ,
N
X
kai (k) = γi ,
i = 1, . . . , N,
k=1
k=1
ai (k) ≥ 0,
i = 1, . . . , N,
k = 1, . . . , N,
Proof. One can separate each i in the problem (33) and solve N suboptimization problems separately for each i:
min
{ai (k),k=1,...,N }
(34)
s.t.
N
X
ai (k)
k=1
N
X
k
ai (k) = αi ,
k=1
ai (k) ≥ 0,
N
X
kai (k) = γi ,
k=1
k = 1, . . . , N.
Each of the sub-problems can be solved using the same method as
solving the LP problem (21) for the DS bound. One can see [KAT00a]
for details. An alternative proof is given in [Kua99] by solving the dual
LP problem of (34).
It has been shown that the KAT bound is always sharper than both
the DC bound and the DS bound [KAT00a]. Furthermore, Dembo has
shown [Dem] that the KAT bound improves the DC bound by a factor
of at most 89 . In the following lemma, we give alternative and simpler
proofs of the above results.
Lemma 2.5. Comparing with the DC and DS bounds, the KAT bound
satisfies
9
(35)
max{ℓDC , ℓDS } ≤ ℓKAT ≤ ℓDC .
8
A SHORT SURVEY
11
Proof. First, substituting (26) in (21), one can get that the DS bound
is the solution of the following LP problem of {ai (k)}
ℓDS =
min
{ai (k),i=1,...,N,k=1,...,N }
(36)
s.t.
N X
N
X
ai (k) =
i=1 k=1
X
ai (k) ≥ 0,
X
i
N X
N
X
ai (k)
i=1 k=1
αi ,
k
N X
N
X
i=1 k=1
k = 1, . . . , N.
kai (k) =
X
γi ,
i
i
Since every feasible point of (33) is also a feasible point of (36). The
LP problem (36) is a relaxed problem of (33). Therefore, ℓKAT ≥ ℓDS .
Next, it is easy to show that ℓKAT ≥ ℓDC since based on the constraints of (33), one can get the DC bound as a lower bound of the
objective function of (33) using the Cauchy-Schwarz inequality. Therefore, ℓKAT is lower bounded by ℓDC .
Finally, we prove that ℓKAT ≤ 89 ℓDC . Note that the DC bound (31)
P α2
is given by ℓDC = i γii and that the solution of (34) can be written
as aik(k11 ) + aik(k22 ) where k2 = k1 + 1. It then suffices to prove for any
i = 1, . . . , N and integer k = 1, . . . , N − 1
(37)
9 αi2
ai (k) ai (k + 1)
,
+
≤
k
k+1
8 γi
where ai (k) + ai (k + 1) = αi and kai (k) + (k + 1)ai (k + 1) = γi .
Denoting x := ai (k) and y := ai (k + 1), one can get
2
y
x
+ k+1
[kx + (k + 1)y]
ai (k) ai (k + 1)
αi
k
/
+
=
k
k+1
γi
αi2
k
xy + k+1
xy + y 2
x2 + k+1
1
xy
k
(38)
=
= 1+
2
(x + y)
k(k + 1) (x + y)2
1
9
1
1
≤1+ = .
≤1+
4 k(k + 1)
8
8
The first equality holds when x = y = α2i and the second equality holds
when k = 1. Therefore, the inequality ℓKAT ≤ 89 ℓDC can be active.
S
N
A
Remark 2.6. Finally, we can derive an upper bound for P
i
i=1
P
using {P (Ai )} and { j P (Ai ∩ Aj )} by maximizing the LP problem
for the KAT bound.
12
A SHORT SURVEY
The following LP problem
(39)
max
{ai (k),i=1,...,N,k=1,...,N }
s.t.
N X
N
X
ai (k)
i=1 k=1
N
X
k
ai (k) = αi ,
k=1
ai (k) ≥ 0,
N
X
kai (k) = γi ,
i = 1, . . . , N,
k=1
i = 1, . . . , N,
k = 1, . . . , N,
gives the upper bound
!
[
X
1 X
P
Ai ≤
αi −
βi
N
i
i
i
(40)
X
1 X
P (Ai ∩ Aj ) =: ~NEW-1 .
=
P (Ai ) −
N j6=i
i
⊳
2.3. Lower Bounds Using {P (Ai )} and {P (Ai ∩ Aj )}. Lower and
SN
in terms of the individual event probaupper bounds on P
i=1 Ai
bilities P (Ai )’s and the pairwise event probabilities P (Ai ∩ Aj )’s can
be seen as special cases of the Boolean probability bounding problem
[Bor+14; Viz04], which can be solved numerically via a linear programming (LP) problem involving 2N variables. Unfortunately, the number
of variables for Boolean probability bounding problems increases exponentially with the number of events, N, which makes finding the
solution impractical. Therefore, some suboptimal numerical bounds
are proposed [Bor+14; Viz04; PG05; GS96] in order to reduce the complexity of the LP problem, for example, by using the dual basic feasible
solutions.
On the other hand, analytical/algorithmic bounds are particularly
important.
One can apply an existing bound using {P (Ai ), i ∈ I} and
P
{ j∈I P (Ai ∩ Aj )} as a base bound, and then optimize the bound by
choosing the optimal subset I of {1, . . . , N} algorithmically. Note that
the bound by optimization via a subset exploits the full information of
{P (Ai )} and {P (Ai ∩ Aj )}. Examples of bounds in this class includes
the stepwise algorithmic implementation of the Kounias lower bound
[KAT00b], the greedy algorithmic implementation of the Hunter upper bound [KAT00b]. Other analytical bounds, like the KAT bound,
are also investigated in other works (e.g., see [CS06; Hop06; Hop09;
KAT00b; BAL05; BAL07; MCS13]).
A SHORT SURVEY
13
P
The other class of bounds is established by {P (Ai )} and { j cj P (Ai ∩
Aj )}, where {cj } can be arbitrarily chosen from a continuous set for
c = (c1 , . . . , cN )T or computed using {P (Ai )} and {P (Ai ∩ Aj )}. Then
the resulting bound also exploits the full information of {P (Ai )} and
{P (Ai ∩Aj )}. Typical example of the bounds in this class is the GallotKounias (GK) bound [Gal66; Kou68] (see also [FLS10; MCS13]).
2.3.1. Kounias Lower Bound and Algorithmic Implementation. The
Kounias lower bound, which is a Bonferroni-type bound, can be written
as
!
(
)
N
[
X
X
(41)
P
Ai ≥ max
P (Ai ) −
P (Ai ∩ Aj ) ,
i=1
I
i∈I
i,j∈I,i<j
where I is a subset of the set of indices {1, . . . , N}. However, the
computational complexity of the Kounias lower bound is exponential
since there are exponential number of subsets of {1, . . . , N}.
In order to reduce the computational complexity, an algorithmic algorithm is proposed in [KAT00b] using a stepwise algorithm to find
a sub-optimal index set that maximizes the RHS of (41). We will refer to this algorithmic implementation of Kounias lower bound as the
stepwise lower bound.
2.3.2. Gallot-Kounias (GK) Bound.
RN ×1 and
(42)
P (A1 ∩ A1 ) P (A1 ∩ A2 )
P (A2 ∩ A1 ) P (A2 ∩ A2 )
Σ=
..
..
.
.
P (AN ∩ A1 ) P (AN ∩ A2 )
Let α = (P (A1 ), · · · , P (AN ))T ∈
. . . P (A1 ∩ AN )
. . . P (A2 ∩ AN )
∈ RN ×N ,
..
...
.
. . . P (AN ∩ AN )
the GK bound [Gal66; Kou68] is given as
!
N
[
(43)
P
Ai ≥ cT Σc,
i−1
where Σc = α. Kounias has shown in [Kou68, Lemma 1.1] that the
vector α is in the range of Σ, i.e., α is orthogonal to the null space of Σ.
As a result, if Σ is singular, one can choose subsets of {A1 , . . . , AN }
to compute the corresponding GK bound, which results in the same
bound if the rank of the corresponding Σ is the same.
Therefore, without loss of generality (WLOG), we assume herein Σ
is non-singular, then the solution of Σc = α is unique
(44)
c̃ = Σ−1 α.
14
A SHORT SURVEY
and the GK bound can be written as
!
N
[
−1
(45)
P
Ai ≥ c̃T Σc̃ = αT ΣT
α = αT Σ−1 α =: ℓGK ,
i−1
−1
where ΣT
= Σ−1 as Σ is symmetric. Furthermore, the GK bound
was recently revisited by [FLS10]. The authors in [FLS10] have shown
that the GK bound can be reformulated as
P
[ i ci P (Ai )]2
.
(46)
ℓGK = max P P
c∈RN
i
k ci ck P (Ai ∩ Ak )
2.4. Upper Bounds. There are only a few analytical/algorithmic upper bounds in the literature. In the following, we introduce the Hunter
bound and its algorithmic implementation by a greedy algorithm.
2.4.1. Hunter Upper Bound and Algorithmic Implementation. The Hunter
upper bound, which is a Bonferroni-type bound, can be written as
!
N
N
X
[
X
(47)
P
P (Ai ) − max
Ai ≤
P (Ai ∩ Aj ),
i=1
i=1
T0 ∈T
(i,j)∈T0
where T is the set of all trees spanning the N indices, i.e., the trees
that include all indices as nodes.
However, the computational complexity of finding the optimal spanning tree is exponential via an exhaustive search. In order to reduce
the complexity, one algorithmic algorithm is proposed in [KAT00b] using Kruskal’s greedy algorithm for finding a sub-optimal spanning tree
for a weighted graph. We will refer to this algorithmic implementation
of the Hunter upper bound as the greedy upper bound, ~Greedy .
3. Observations on the GK Bound
Finally, we conclude this survey with two observations on the GK
bound.
3.1. Applying the GK bound to subsets of events. We note that
many existing lower bounds, which do not fully explore available information, can be further improved algorithmically via optimization
over subsets, as in [BAL07; Hop06]. However, in this section, we prove
that the GK bound cannot be improved by applying it to subsets of
{A1 , . . . , AN }.
A SHORT SURVEY
15
Lemma 3.1. For any given M ≥ N, the GK bound is the solution of
the following problem:
min
(48)
kxk2
{x∈RM ×1 ,A∈RN×M }
s.t.
Ax = α,
AAT = Σ ∈ RN ×N .
Proof. We can always write x as
x = AT k1 + AT⊥ k2 ,
(49)
where k1 ∈ RN ×1 , k2 ∈ R(M −N )×1 , and A⊥ ∈ R(M −N )×M satisfies
T
AAT⊥ = 0N ×(M −N ) . Particularly, let LL
be the Cholesky decompo
sition of Σ, then A = L, 0N ×(M −N ) Q where Q is any orthogonal
matrix is the solution of (48).
Then
(50)
kxk2 = k1T AAT k1 + k2T A⊥ AT⊥ k2 = k1T Σk1 + k2T A⊥ AT⊥ k2 .
The first constraint Ax = α implies AAT k1 = α, i.e., k1 = Σ−1 α.
Therefore, the minimum of kxk2 is achieved at k2 = 0, so that
(51)
min kxk2 = k1T Σk1 = αT Σ−1 α.
k1 ,k2
Theorem 3.2. The GK bound cannot be improved via optimization
over subsets of {A1 , A2 , . . . , AN }.
T
a1
aT2
Proof. Denoting A =
... , the first constraint of (48) is equivalent
aTN
to N constraints aTi x = αi , i = 1, . . . , N, and the second constraint of
(48) is equivalent to N 2 constraints aTi aj = Σij , i = 1, . . . , N, j =
1, . . . , N.
By selecting a subset of {1, 2, . . . , N}, the resulting GK bound is the
solution of a relaxed problem of (48) with a subset of constraints on x
and ai , i = 1, . . . , N. Since the objective value of the relaxed problem
must be no more than the original problem (48), the GK bound using a
subset cannot be higher than the GK bound using full information.
3.2. Iterative implementation of the GK bound.
Theorem 3.3. The GK bound can be computed iteratively. More
specifically, for n = 1, . . . , N, denote ℓGK (n) as the GK bound using
the information of A1 , . . . , An : αn = (P (A1 ), . . . , P (An ))T , βn =
16
A SHORT SURVEY
(P (A1 ∩ An ), . . . , P (An−1 ∩ An ))T and Σn , the n × n upper left submatrix of Σ, then if αn − βnH Σ−1
n−1 βn > 0, we have
1
bn bTn bn
T
(52) ℓGK (n) = ℓGK (n − 1) +
αn
αn ,
bTn
1
αn − βnH Σ−1
n−1 βn
where bn = −Σ−1
n−1 βn .
The matrix Σn is not invertible if and only if αn − βnH Σ−1
n−1 βn = 0.
H −1
The last case αn − βn Σn−1 βn < 0 never happens, i.e., the following
inequality holds
(53)
P (An ) ≥ (P (A1 ∩ An ), . . . , P (An−1 ∩ An )) ·
−1
P (A1 ∩ An )
P (A1 ∩ A1 ) . . . P (A1 ∩ An−1 )
P (A2 ∩ A1 ) . . . P (A2 ∩ An−1 ) P (A2 ∩ An )
,
..
..
..
.
.
.
...
P (An−1 ∩ An )
P (An−1 ∩ A1 ) . . . P (AN ∩ An−1 )
for n = 1, . . . , N.
Proof. Note that
(54)
Σn =
Σn−1 βn
βnT αn
.
and by the matrix inverse lemma for a Hermitian matrix [HJ86], we
have
−1
1
bn bTn bn
Σn−1 0n
−1
+
(55)
Σn =
bTn
1
0Tn
0
αn − βnH Σ−1
n−1 βn
Substituting to ℓGK (n) = αn Σ−1
n αn , ℓGK (n) can be computed using
ℓGK (n − 1).
Since we have proved ℓGK (n) ≥ ℓGK (n − 1) in Theorem 3.2, the
inequality is directly from αn − βnH Σ−1
n−1 βn ≥ 0 in (52) of Theorem
3.3.
Recent works: We close this survey by referring the reader to [YAT14;
YAT15a; YAT15b; YAT16a; YAT16b] for the most recent works on this
topic.
References
[BAL05]
F. Behnamfar, F. Alajaji, and T. Linder. “Tight error bounds
for space-time orthogonal block codes under slow Rayleigh
flat fading”. IEEE Transactions on Communications 53.6
(2005), pp. 952–956.
REFERENCES
[BAL07]
17
F. Behnamfar, F. Alajaji, and T. Linder. “An efficient algorithmic lower bound for the error rate of linear block codes”.
IEEE Transactions on Communications 55.6 (2007), pp. 1093–
1098.
[Bor+14] E. Boros, A. Scozzari, F. Tardella, and P. Veneziani. “Polynomially Computable Bounds for the Probability of the
Union of Events”. Mathematics of Operations Research (2014).
eprint: http://dx.doi.org/10.1287/moor.2014.0657.
[BT97]
D. Bertsimas and J. Tsitsiklis. Introduction to Linear Optimization. Athena Scientific, 1997.
[CS06]
J. T. Chen and E. Seneta. “A Frechet-Optimal Strengthening of the Dawson-Sankoff Lower Bound”. Methodology and
Computing in Applied Probability 8.2 (2006), pp. 255–264.
[DC97]
D De Caen. “A lower bound on the probability of a union”.
Discrete Mathematics 169.1 (1997), pp. 217–220.
[Dem]
A. Dembo. “Unpublished notes”. Communicated by I. Sasson, 2000.
[DS67]
D. Dawson and D Sankoff. “An inequality for probabilities”. Proceedings of the American Mathematical Society
18.3 (1967), pp. 504–507.
[FLS10]
C. Feng, L. Li, and J. Shen. “Some inequalities in functional
analysis, combinatorics, and probability theory”. The Electronic Journal of Combinatorics 17.R58 (2010), p. 1.
[Gal66]
S. Gallot. “A bound for the maximum of a number of random variables”. Journal of Applied Probability 3.2 (1966),
pp. 556–558.
[GS96]
J. Galambos and I. Simonelli. Bonferroni-type inequalities
with applications. Springer series in statistics: Probability
and its applications. Springer, 1996. isbn: 9780387947761.
[HJ86]
R. A. Horn and C. R. Johnson, eds. Matrix Analysis. New
York, NY, USA: Cambridge University Press, 1986. isbn:
0-521-30586-1.
[Hop06]
F. M. Hoppe. “Improving probability bounds by optimization over subsets”. Discrete Mathematics 306.5 (2006), pp. 526–
530.
[Hop09]
F. M. Hoppe. “The effect of redundancy on probability
bounds”. Discrete Mathematics 309.1 (2009), pp. 123–127.
[KAT00a] H Kuai, F Alajaji, and G Takahara. “A lower bound on the
probability of a finite union of events”. Discrete Mathematics 215.1-3 (2000), pp. 147–158.
18
REFERENCES
[KAT00b] H. Kuai, F. Alajaji, and G. Takahara. “Tight error bounds
for nonuniform signaling over AWGN channels”. IEEE Transactions on Information Theory 46.7 (2000), pp. 2712–2718.
[Kou68]
E. G. Kounias. “Bounds for the Probability of a Union,
with Applications”. The Annals of Mathematical Statistics
39.6 (1968), pp. 2154–2158.
[Kua99]
H. Kuai. “Tight Bounds on The Probability of a Union with
Applications to Non-uniform Signaling over AWGN Channels”. MA thesis. Department of Mathematics and Statistics, Queen’s University, 1999.
[MCS13] Z. Mao, J. Cheng, and J. Shen. “A new lower bound on
error probability for nonuniform signals over AWGN channels”. In: Proc. Wireless Communications and Networking
Conference. IEEE. 2013, pp. 3005–3009.
[PG05]
A. Prékopa and L. Gao. “Bounding the probability of the
union of events by aggregation and disaggregation in linear programs”. Discrete Applied Mathematics 145.3 (2005),
pp. 444–454.
[Viz04]
B. Vizvári. “New upper bounds on the probability of events
based on graph structures”. RUTCOR Research Report (2004).
[YAT14] J. Yang, F. Alajaji, and G. Takahara. “New bounds on the
probability of a finite union of events”. In: Proc. IEEE International Symposium on Information Theory. 2014, pp. 1271–
1275.
[YAT15a] J. Yang, F. Alajaji, and G. Takahara. “On bounding the
union probability”. In: Proc. IEEE International Symposium on Information Theory. IEEE. 2015, pp. 1761–1765.
[YAT15b] J. Yang, F. Alajaji, and G. Takahara. “On Bounding the
Union Probability Using Partial Weighted Information”.
arXiv preprint arXiv:1506.08331 (2015).
[YAT16a] J. Yang, F. Alajaji, and G. Takahara. “Lower bounds on
the probability of a finite union of events”. SIAM Journal
on Discrete Mathematics 30.3 (2016), pp. 1437–1452.
[YAT16b] J. Yang, F. Alajaji, and G. Takahara. “On bounding the
union probability using partial weighted information”. Statistics & Probability Letters 116 (2016), pp. 38–44.
| 7cs.IT
|
arXiv:1709.00164v2 [q-bio.PE] 23 Sep 2017
Whole Genome Phylogenetic Tree Reconstruction Using Colored de Bruijn Graphs
Cole A. Lyman∗ , M. Stanley Fujimoto∗ , Anton Suvorov† , Paul M. Bodily∗ ,
Quinn Snell∗ , Keith A. Crandall‡ , Seth M. Bybee† and Mark J. Clement∗
∗ Computer Science Department
Brigham Young University,
Provo, Utah 84602 USA
† Department of Biology
Brigham Young University,
Provo, Utah 84602 USA
‡ Computational Biology Institute
George Washington University,
Washington, DC 20052 USA
Email: [email protected]
Abstract—We present kleuren, a novel assembly-free
method to reconstruct phylogenetic trees using the Colored
de Bruijn Graph. kleuren works by constructing the Colored de Bruijn Graph and then traversing it, finding bubble
structures in the graph that provide phylogenetic signal.
The bubbles are then aligned and concatenated to form a
supermatrix, from which a phylogenetic tree is inferred. We
introduce the algorithms that kleuren uses to accomplish
this task, and show its performance on reconstructing the
phylogenetic tree of 12 Drosophila species. kleuren reconstructed the established phylogenetic tree accurately and
is a viable tool for phylogenetic tree reconstruction using
whole genome sequences. Software package available at:
https://github.com/Colelyman/kleuren.
Keywords-phylogenetics; algorithm; whole genome sequence;
colored de Bruijn graph
I. I NTRODUCTION
Whole genome sequences are readily available and affordable like never before [1] due to the advent of highthroughput Next Generation Sequencing (NGS) which has
provided researchers with vast amounts of genomic sequencing data that has transformed the landscape of understanding
of genomes. The field of phylogenetics, which discovers the
evolutionary relationship between taxa, has been no exception to this transformation. Phylogenetics has responded to
the copious amounts of high throughput data with novel
alignment-free and assembly-free methods [2], [3] that are
better suited [4] to handle the large amounts of data more
efficiently than the traditional alignment-based phylogenetic
methods. The traditional approach to phylogenetic tree reconstruction requires a homology search throughout the
genomes of the taxa, a Multiple Sequence Alignment (MSA)
of the homologs, and a tree construction from the resulting
matrix. Each of these steps can be computationally expensive
and may introduce many unnecessary assumptions that can
be avoided by using an alignment-free and assembly-free
method.
Alignment-free and assembly-free methods [5]–[8] don’t
come without their disadvantages, one of which being that
many of these methods abstract away the source of the phylogenetic signal to a method akin to shared kmer-counting.
We propose an assembly-free whole genome phylogenetic
tree reconstruction method using the Colored de Bruijn
Graph (CdBG) [9], a data structure that is commonly used
for detecting variation and comparing genomes.
The CdBG is similar to a traditional de Bruijn Graph
(dBG) in that the substrings of a certain length, referred to as
kmers, of a sequence represent the vertices of the dBG and
an edge exists between two vertices if the suffix of the first
vertex is the prefix of the second vertex. The CdBG differs
from the traditional dBG in that each vertex is associated to
an unique color (or set of colors) which could be a differing
sample, species, or taxon.
We introduce the kleuren (Dutch for ”colors” in tribute
of Nicolaas Govert de Bruijn, the de Bruijn graph’s namesake) software package which implements our methods.
kleuren works by finding bubble regions [9], [10] of the
CdBG, which are where one or more colors diverge at a
node, which act as pseudo-homologous regions between the
taxa. The sequence for each taxon in each bubble is then
extracted and a MSA is performed, then the MSA’s from
each bubble are concatenated to form a supermatrix in which
a phylogenetic tree of evolution is constructed.
II. M ETHODS
A. Definitions
Given the alphabet Σ = {A, C, G, T } which are nucleotide codes, let a dBG G, be defined as G = (V, E)
where V = {v1 , v2 , . . . , vi , . . . , vs } is the set of vertices
and where vi is the ith unique sequence of length k of
G and where E = {e1 , e2 , . . . , ei , . . . , et } is the set of
edges and where ei = (vi , vi+1 ) is an edge connecting two
vertices such that the sequence of vi and vi+1 overlap by
(k − 1) characters. Let a CdBG, CG, be defined as CG =
{G1 , G2 , . . . , Gi , . . . , Gu } for u taxa where Gi = (Vi , Ei )
is the dBG of the ith taxon. We refer to each G ∈ CG as
a distinct color or taxon.
Furthermore, let a path, P = (v1 , . . . , vw ) in Gi be
defined as a sequence of vertices from Vi such that for all
subsequences (vj , vj+1 ) of P , the edge (vj , vj+1 ) ∈ Ei . Let
a bubble, B, in CG be defined as B = {P1 , . . . , Pz } such
that each P ∈ B is associated with one or more colors, that
the first and last vertices of ∀P ∈ B are identical, and that
2 ≤ z ≤ u (see Figure 1).
Finally, let K be defined as K = {V1 ∪ V2 ∪ . . . ∪ Vi ∪
. . . ∪ Vu } where Vi is the vertices (or the unique kmers) of
the ith dBG, Gi .
B. Software Architecture
We use the dbgfm software package [11] to construct
and represent the dBG’s of the individual taxa. kleuren
provides an interface to interact with the individual dBG’s
to create a CdBG, where each taxon is considered a color.
The dbgfm package uses the FM-Index [12], as a space
efficient representation of the dBG.
C. kleuren Algorithms
Algorithm 1 kleuren Algorithm
1: function KLEUREN (K, CG)
2:
bubbles ← [ ] . bubbles is initialized to an empty
list
3:
for all k ∈ K do
4:
if k is in c or more colors of CG then
5:
endV ertex ← FIND E NDV ERTEX(k, CG)
6:
for all color ∈ CG do
7:
path ← EXTEND PATH(k, endV ertex,
color)
8:
add path to bubble
9:
end for
10:
append bubble to bubbles
11:
end if
12:
end for
13:
alignments ← [ ]
14:
for all bubble ∈ bubbles do
15:
alignment ← multiple sequence alignment of
each path in bubble
16:
append alignment to alignments
17:
end for
18:
supermatrix ← concatenation of alignments
19: end function
1) Overall Algorithm: kleuren works by iterating over
the superset of vertices, K, and discovering vertices that
could form a bubble. A vertex, s, could form a bubble if
s is present in c or more colors of CG, where c is set by
the user as a command line parameter. Note that the lower
A. Bubble in a Colored de Bruijn Graph
Color 1
Color 2
Color 3
Path: ACTGTG
Path: ACTAGGTG
Path: ACTAGTG
B. Paths in the Bubble of Each Color
Figure 1. A. An example of a bubble in a Colored de Bruijn Graph
with 3 colors (i.e. 3 taxa), and where k = 3. The colors of the vertices
represent the following: gray- all colors contain the vertex, purple- Color
2 and Color 3 contain the vertex, yellow- Color 1 contains the vertex, redColor 2 contains the vertex, and blue- Color 3 contains the vertex. In this
example ACT is the startVertex and GTG is the endVertex which are both
contained in all of the colors. B. The extended paths of each color between
the startVertex and endVertex.
that c is, the more potential bubbles that may be found, but
kleuren will take longer to run because more vertices will
be considered as the starting vertex of a bubble. Let s be
considered as the starting vertex of the bubble, b; then the
end vertex, e, of b is found (see Section II-C2). After the
end vertex is found, the path, p, between s and e is found
for each color in CG (see Section II-C3). This process is
repeated until each vertex in K has been either considered
as a starting vertex of a bubble, or has been visited while
extending the path between a starting and ending vertex.
2) Finding the End Vertex: The end vertex is found by
traversing the path from the startV ertex until a vertex is
found that is in at least c colors. The endV ertex is then
used in the function to extend the path (see Section II-C3).
Algorithm 2 Find End Vertex Function
1: function FIND E NDV ERTEX (startV ertex, CG)
2:
endV ertex ← “ ” . endV ertex is initialized to
an empty string
3:
neighbors ← GET N EIGHBORS(startV ertex)
4:
while
!IS E MPTY(neighbors)
and
IS E MPTY (endV ertex) do
5:
for all neighbor ∈ neighbors do
6:
if k is in c or more colors of CG then
7:
endV ertex ← neighbor
8:
end if
9:
end for
10:
end while
11:
return endV ertex
12: end function
3) Extending the Path: The main functions that discover
the sequences found in a bubble are the Extend the Path
Functions (see Section II-C3). To extend the path between
the startV ertex and endV ertex we use a recursive function that traverses the dBG for a color in which every
possible path between the startV ertex and endV ertex is
explored up to the maxDepth (provided as a command line
parameter by the user). The maxDepth parameter allows
the user to specify how thorough kleuren will search for
a bubble; the higher the maxDepth the more bubbles that
kleuren will potentially find, but the longer kleuren
will take because at each depth there are exponentially more
potential paths to traverse.
Algorithm 3 Extend the Path Functions
1: function EXTEND PATH (startV ertex, endV ertex, color,
maxDepth)
2:
path ← “ ”
3:
visited ← {} . visited is initialized to the empty
set
4:
if RECURSIVE PATH(startV ertex, endV ertex, path,
color, visited, 0, maxDepth) then
5:
return path
6:
end if
7: end function
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
D. Data Acquisition
26:
To measure the effectiveness of our method we used
12 Drosophila species, obtained from FlyBase [13]. We
chose this group of species because there is a thoroughly
researched and established phylogenetic tree [14].
27:
28:
29:
30:
31:
function RECURSIVE PATH(currentV ertex, endV ertex,
path, color, visited, depth, maxDepth)
add currentV ertex to visited
if depth >= maxDepth then
return f alse
end if
if currentKmer == endKmer then
return true
end if
neighbors ← NEIGHBORS(currentV ertex, color)
for all neighbor ∈ neighbors do
if neighbor is in visited then
continue
end if
oldP ath ← path
append suffix of currentKmer to path
depth ← depth + 1
if !RECURSIVE PATH(neighbor, endV ertex, path,
color, visited, depth, maxDepth) then
path ← oldP ath
else
return true
end if
end for
end function
E. Tree Construction and Parameters
We used the DSK software package [15] to count the
kmers present in all of the Drosophila species. To find the
bubbles, we used the following parameters: k = 17 (kmer
size of 17) and c = 12 (all colors in the CG were required
to contain a vertex in order to search for a bubble starting at
that vertex) and ran 32 instances of kleuren concurrently
for 4 days to find 3, 277 bubbles. When all of the bubbles
in the CdBG had been identified, we used MAFFT [16]
to perform a MSA for each sequence in every bubble that
kleuren identified (see Figure 2 A.). Then each MSA was
concatenated to form a supermatrix (see Figure 2 B.) using
Biopython [17]. The phylogenetic tree was then inferred
from the supermatrix by Maximum Likelihood using IQTREE [18] (see Figure 2 C.).
Once the tree was constructed, we used the ETE 3
software package [19] to compare the tree to the established
one and Phylo.io [20] to visualize the trees.
F. Bubble Assumptions
Our method is based on the assumption that bubbles are
representative of homologous regions of the taxa genomes.
We propose that this assumption is reliable because it has
been shown that dBG’s are a suitable method to align
sequences [21]–[23], and by identifying the bubbles in the
CdBG we find the sections of the graph that contain the
most phylogenetic signal.
III. R ESULTS
kleuren constructed a tree (see Figure 3) consistent
with the established tree found in [14] (the Robinson-Foulds
distance [24] between the two trees is 0). Even though we
Color 1
Color 2
Color 3
Path: ACT--GTG
Path: ACTAGGTG
Path: ACTA-GTG
A. Multiple Sequence Alignment of the Sequences in
Bubble (Figure 1)
Color 1
Color 2
Color 3
Path: ACT--GTGATT-A...
Path: ACTAGGTGATTC-...
Path: ACTA-GTGATTCA...
B. Supermatrix of Multiple Sequence Alignments
concatenated
Color 3
Color 2
Color 1
C. Phylogenetic Tree
Figure 2. A. The Multiple Sequence Alignment (MSA) of the sequences
from the bubble presented in Figure 1. B. The MSA’s from each bubble
are concatenated into a supermatrix, from which a phylogenetic tree
is constructed. C. The resulting tree from the supermatrix inferred by
Maximum Likelihood.
ran many concurrent instances of kleuren for multiple
days (see Section II-E), not all of the kmers in K were
explored for potential bubbles; meaning that many more
bubbles could be found in this CdBG which would only
make the phylogeny more concrete.
Before this final successful run, there were a number
of unsuccessful attempts made to construct the tree. Initial
attempts were unsuccessful because K (the super-set of
kmers) that kleuren uses to find bubbles was semi-sorted
(segments of the file were sorted, but all of the kmers in
the file were not in lexicographic order) so the vertices that
kleuren used to search for bubbles were skewed towards
vertices that were lexicographically first. We remedied this
issue by shuffling the order of the kmer file so that there was
no lexicographic bias towards the bubbles that kleuren
finds.
A previous attempt resulted in a tree that had a 0.44 normalized Robinson-Fould’s distance from the established tree
occurred because there were too few bubbles, and therefore
there was not enough phylogenetic signal for the correct tree
to be constructed. To find more bubbles, we split up the kmer
file into parts so that multiple instances of kleuren could
find bubbles concurrently. We also discovered that there was
a high frequency of adenines (A) (a frequency around 40% in
comparison to the other nucleotides) in the final supermatrix
that could skew the final tree because nucleotides have
differing mutation rates. We thought this bias towards A
was due to the fact that in the recursiveP ath function
(see Algorithm 3) the neighbors may be sorted, so the
function would traverse the neighbor that started with an A
before traversing the other neighbors (see Algorithm 3, line:
18). Similar to the previous sorting problem, we shuffled
the order of the neighbors so that the first neighbor that
was traversed would not always be lexicographically first.
Despite this change, the final supermatrix that produced the
true tree still had a bias towards A (see Section V).
IV. C ONCLUSION
We introduced a novel method of constructing accurate
phylogenetic trees using a CdBG. Our method, kleuren,
uses whole genome sequences to construct a CdBG representation, then it traverses the CdBG to discover bubble
structures which become the basis for phylogenetic signal
between taxa and eventually produces a phylogenetic tree.
As the NGS era progresses, whole genome sequences are
becoming more prevalent for more non-model organisms,
in which phylogenies of these organisms have never been
constructed. kleuren is a viable method to relatively
quickly and accurately construct the phylogenies for these
newly sequenced organisms.
V. F UTURE W ORK
We plan to optimize kleuren so that it can find more
bubbles in a shorter amount of time. We will do this by
replacing the underlying data structure for how the CdBG is
represented. dbgfm, the current method used to represent
the dBG in kleuren, sacrifices time efficiency for memory
efficiency by storing the FM-Index entirely on disk, thus
slowing down queries into the dBG. When kleuren runs
faster, more bubbles will be found, and more phylogenetic
signal will be present so that a more accurate tree can be
constructed.
We also plan to investigate the reasons for the high
abundance of A’s in the supermatrix (see Section III) further,
and balance the frequency of nucleotides in the supermatrix.
Furthermore, we would like to look into how kleuren
performs when the CdBG is constructed using read sequencing data rather than assembled genomes.
ACKNOWLEDGMENT
This work was funded through the Utah NASA Space
Grant Consortium and EPSCoR and through the BYU
Graduate Research Fellowship.
The authors would like to thank Kristi Bresciano, Michael
Cormier, Justin B. Miller, Brandon Pickett, Nathan Schulzke,
and Sage Wright for their thoughts concerning the project.
The authors would also like to thank the Fulton Supercomputing Laboratory at Brigham Young University for
their work to maintain the super-computer on which these
experiments were run.
Figure 3. The phylogenetic tree of 12 Drosophila species constructed using kleuren. This tree resulted from using a kmer size of 17 and required all
species to contain a vertex in order for the algorithm to search for a bubble starting at that vertex; and this tree is consistent with the established tree for
these 12 species.
R EFERENCES
[1] S. C. Schuster, “Next-generation sequencing transforms
today’s biology,” Nature Methods, vol. 5, no. 1, pp. 16–
18, dec 2007. [Online]. Available: https://doi.org/10.1038/
nmeth1156
[2] H. Fan, A. R. Ives, Y. Surget-Groba, and C. H. Cannon,
“An assembly and alignment-free method of phylogeny
reconstruction from next-generation sequencing data,” BMC
Genomics, vol. 16, no. 1, jul 2015. [Online]. Available:
https://doi.org/10.1186/s12864-015-1647-5
[3] J. Qi, H. Luo, and B. Hao, “CVTree: a phylogenetic tree
reconstruction tool based on whole genomes,” Nucleic Acids
Research, vol. 32, no. Web Server, pp. W45–W47, jul 2004.
[Online]. Available: https://doi.org/10.1093/nar/gkh362
[4] C. X. Chan and M. A. Ragan, “Next-generation
phylogenomics,” Biology Direct, vol. 8, no. 1, jan 2013.
[Online]. Available: https://doi.org/10.1186/1745-6150-8-3
[5] B. Haubold, “Alignment-free phylogenetics and population
genetics,” Briefings in Bioinformatics, vol. 15, no. 3, pp.
407–418, nov 2013. [Online]. Available: https://doi.org/10.
1093/bib/bbt083
[6] O. Bonham-Carter, J. Steele, and D. Bastola, “Alignmentfree genetic sequence comparisons: a review of recent
approaches by word analysis,” Briefings in Bioinformatics,
vol. 15, no. 6, pp. 890–905, jul 2013. [Online]. Available:
https://doi.org/10.1093/bib/bbt052
[7] C. X. Chan, G. Bernard, O. Poirion, J. M. Hogan,
and M. A. Ragan, “Inferring phylogenies of evolving
sequences without multiple sequence alignment,” Scientific
Reports, vol. 4, no. 1, sep 2014. [Online]. Available:
https://doi.org/10.1038/srep06504
[8] H. Yi and L. Jin, “Co-phylog: an assembly-free phylogenomic
approach for closely related organisms,” Nucleic Acids
Research, vol. 41, no. 7, pp. e75–e75, jan 2013. [Online].
Available: https://doi.org/10.1093/nar/gkt003
[9] Z. Iqbal, M. Caccamo, I. Turner, P. Flicek, and G. McVean,
“De novo assembly and genotyping of variants using
colored de Bruijn graphs,” Nature Genetics, vol. 44,
no. 2, pp. 226–232, jan 2012. [Online]. Available: https:
//doi.org/10.1038/ng.1028
[10] G. Peng, P. Ji, and F. Zhao, “A novel codon-based
de Bruijn graph algorithm for gene construction from
unassembled transcriptomes,” Genome Biology, vol. 17,
no. 1, nov 2016. [Online]. Available: https://doi.org/10.1186/
s13059-016-1094-x
[11] R. Chikhi, A. Limasset, S. Jackman, J. T. Simpson, and
P. Medvedev, “On the representation of de bruijn graphs,” in
Research in Computational Molecular Biology: 18th Annual
International Conference, RECOMB 2014, Pittsburgh, PA,
USA, April 2-5, 2014, Proceedings, R. Sharan, Ed. Cham:
Springer International Publishing, 2014, pp. 35–55. [Online].
Available: https://doi.org/10.1007/978-3-319-05269-4 4
[12] P. Ferragina and G. Manzini, “Opportunistic data structures
with applications,” in Proceedings of the 41st Annual
Symposium on Foundations of Computer Science, ser. FOCS
’00. Washington, DC, USA: IEEE Computer Society, 2000,
pp. 390–. [Online]. Available: http://dl.acm.org/citation.cfm?
id=795666.796543
[13] L. S. Gramates, S. J. Marygold, G. dos Santos, J.-M.
Urbano, G. Antonazzo, B. B. Matthews, A. J. Rey, C. J.
Tabone, M. A. Crosby, D. B. Emmert, K. Falls, J. L.
Goodman, Y. Hu, L. Ponting, A. J. Schroeder, V. B. Strelets,
J. Thurmond, P. Zhou, and FlyBase Consortium, “FlyBase at
25: looking to the future,” Nucleic Acids Research, vol. 45,
no. D1, pp. D663–D671, oct 2016. [Online]. Available:
https://doi.org/10.1093/nar/gkw1016
[14] M. W. Hahn, M. V. Han, and S.-G. Han, “Gene family
evolution across 12 drosophila genomes,” PLoS Genetics,
vol. 3, no. 11, p. e197, 2007. [Online]. Available:
https://doi.org/10.1371/journal.pgen.0030197
[15] G. Rizk, D. Lavenier, and R. Chikhi, “DSK: k-mer
counting with very low memory usage,” Bioinformatics,
vol. 29, no. 5, pp. 652–653, jan 2013. [Online]. Available:
https://doi.org/10.1093/bioinformatics/btt020
[16] K. Katoh and D. M. Standley, “MAFFT multiple sequence
alignment software version 7: Improvements in performance
and usability,” Molecular Biology and Evolution, vol. 30,
no. 4, pp. 772–780, jan 2013. [Online]. Available: https:
//doi.org/10.1093/molbev/mst010
[17] P. J. A. Cock, T. Antao, J. T. Chang, B. A. Chapman,
C. J. Cox, A. Dalke, I. Friedberg, T. Hamelryck, F. Kauff,
B. Wilczynski, and M. J. L. de Hoon, “Biopython:
freely available Python tools for computational molecular
biology and bioinformatics,” Bioinformatics, vol. 25, no. 11,
pp. 1422–1423, mar 2009. [Online]. Available: https:
//doi.org/10.1093/bioinformatics/btp163
[18] L.-T. Nguyen, H. A. Schmidt, A. von Haeseler, and B. Q.
Minh, “IQ-TREE: A fast and effective stochastic algorithm
for estimating maximum-likelihood phylogenies,” Molecular
Biology and Evolution, vol. 32, no. 1, pp. 268–274, nov 2014.
[Online]. Available: https://doi.org/10.1093/molbev/msu300
[19] J. Huerta-Cepas, F. Serra, and P. Bork, “ETE 3:
Reconstruction, analysis, and visualization of phylogenomic
data,” Molecular Biology and Evolution, vol. 33,
no. 6, pp. 1635–1638, feb 2016. [Online]. Available:
https://doi.org/10.1093/molbev/msw046
[20] O. Robinson, D. Dylus, and C. Dessimoz, “Phylo.io:
Interactive viewing and comparison of large phylogenetic
trees on the web,” Molecular Biology and Evolution,
vol. 33, no. 8, pp. 2163–2166, apr 2016. [Online]. Available:
https://doi.org/10.1093/molbev/msw080
[21] B. Raphael, “A novel method for multiple alignment of
sequences with repeated and shuffled elements,” Genome
Research, vol. 14, no. 11, pp. 2336–2346, nov 2004.
[Online]. Available: https://doi.org/10.1101/gr.2657504
[22] I. Minkin, A. Patel, M. Kolmogorov, N. Vyahhi, and S. Pham,
“Sibelia: A scalable and comprehensive synteny block
generation tool for closely related microbial genomes,” in
Algorithms in Bioinformatics: 13th International Workshop,
WABI 2013, Sophia Antipolis, France, September 2-4,
2013. Proceedings. Berlin, Heidelberg: Springer Berlin
Heidelberg, 2013, pp. 215–229. [Online]. Available: http:
//dx.doi.org/10.1007/978-3-642-40453-5 17
[23] I. Minkin, S. Pham, and P. Medvedev, “TwoPaCo: an efficient
algorithm to build the compacted de Bruijn graph from many
complete genomes,” Bioinformatics, p. btw609, sep 2016.
[Online]. Available: https://doi.org/10.1093/bioinformatics/
btw609
[24] D. Robinson and L. Foulds, “Comparison of phylogenetic
trees,” Mathematical Biosciences, vol. 53, no. 1-2, pp.
131–147, feb 1981. [Online]. Available: https://doi.org/10.
1016/0025-5564(81)90043-2
| 8cs.DS
|
Sending Information Through Status Updates
Abdulrahman Baknina1, Omur Ozel2 , Jing Yang3, Sennur Ulukus1 , and Aylin Yener3
1
arXiv:1801.04907v1 [cs.IT] 15 Jan 2018
2
Department of Electrical and Computer Engineering, University of Maryland, College Park, MD 20742
Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213
3
Department of Electrical Engineering, The Pennsylvania State University, University Park, PA 16802
Abstract— We consider an energy harvesting transmitter sending status updates regarding a physical phenomenon it observes
to a receiver. Different from the existing literature, we consider
a scenario where the status updates carry information about
an independent message. The transmitter encodes this message
into the timings of the status updates. The receiver needs to
extract this encoded information, as well as update the status
of the observed phenomenon. The timings of the status updates,
therefore, determine both the age of information (AoI) and the
message rate (rate). We study the tradeoff between the achievable
message rate and the achievable average AoI. We propose several
achievable schemes and compare their rate-AoI performances.
I. I NTRODUCTION
We consider an energy harvesting transmitter sending status
updates to a receiver via status update packets. Each status
update packet requires a unit of energy; and the transmitter
harvests energy stochastically over time, one unit at a time, at
random times.1 In order to minimize the age of information
(AoI), the transmitter needs to send frequent and regular (over
time) status updates, however, the frequency and regularity of
the updates are constrained by the stochastic energy arrival
process, which is known only causally at the transmitter.
In this paper, different from the existing literature, we
consider the scenario where the timings of the status updates
also carry an independent message; see Fig. 1. In order to
obtain a tractable formulation, we consider an abstraction
where the physical channel is noiseless and the transmitter has
a battery of unit size. Intuitively, as will be clarified shortly,
there is a tradeoff between the AoI and the rate of the message.
Our goal in this paper is to characterize this tradeoff.
For this scenario, under causal (i.e., online) knowledge of
energy arrivals, [1] has determined that, in order to minimize
the long-term average AoI, the transmitter needs to apply a
threshold based policy: There exists a fixed and deterministic
threshold τ0 such that if an energy arrives sooner than τ0
seconds since the last update, the transmitter waits until τ0
and sends the update packet; on the other hand, if it has been
more than τ0 seconds since the last update, the transmitter
sends an update packet right away when an energy arrives.
On the other hand, again for this scenario, [2] has considered
the information-theoretic capacity of this energy harvesting
channel. The main information-theoretic challenge arises due
to having a state-dependent channel (where the state is the
This work was supported by NSF Grants CCF 14-22111/14-22347, CNS
15-26608/15-26165.
1 Energy requirements and energy harvests are normalized.
Ei
B=1
measurements
message (M )
Tx
Rx
Fig. 1. An energy harvesting transmitter with a finite-sized battery, that sends
status updates and independent information to a receiver.
energy availability), time-correlation introduced in the state
due to the existence of a battery at the transmitter where energy
can be saved and used later, and the unavailability of the state
information at the receiver. Reference [2] converts the problem
from regular channel uses to a timing channel and obtains the
capacity in terms of some auxiliary random variables using a
bits through queues approach as in [3].
Sending information necessarily requires the transmitter to
send out a packet after a random amount of time following
an energy arrival in [2], whereas minimizing AoI requires the
transmitter to apply a deterministic threshold based policy in
[1]. Note that in [1], the transmitter sends a packet either at
a deterministic time τ0 after an energy arrival, or right at the
time of an energy arrival, thus, it cannot send any rate with
the packet timings even though it minimizes the AoI. This is
the main source of the tension between AoI minimization and
information rate maximization; and is the subject of this paper.
In this paper, we first present a general tradeoff region between the achievable AoI and the achievable information rate.
We then consider the class of renewal policies in which the
system action depends only on the most recent transmission.
Within this class of policies, we first propose policies that
determine the next transmission instant as a function of the
time difference between the most recent energy arrival and the
most recent status update. We then consider simpler policies
which we call separable policies. These policies separate the
update decision and information transmission in an additive
manner: When an energy arrives, the transmitter decides when
to update, neglecting the information transmission; once the
transmitter decides to send an update, it then encodes the
message on top of that update timing. For all the policies,
we derive the average achievable AoI and the achievable rate.
We then compare the tradeoff regions of these policies. We
observe numerically that the first class of policies achieve
better tradeoff regions. We also observe that as the value of the
∆(t)
energy arrival
info. symbol
V1
Q1
Q2
V2
Q4
Q3
t
Z1 (τ1 )
T1
Fig. 2.
V1
Z2 (τ2 )
T2
T3
T4
T1
An example evolution of instantaneous AoI.
Fig. 3.
average energy arrival increases, policies perform similarly.
Related Work: Minimizing the AoI has been studied in many
different settings, including settings with no energy constraints
[4]–[13] and settings with energy constraints in offline and
online energy harvesting models [1], [14]–[17]. Energy harvesting communication systems have been extensively studied
in scheduling-theoretic and information-theoretic settings, for
example, offline scheduling in single-user and multi-user settings have been considered in [18]–[26], online scheduling has
been considered in [20], [27]–[32], and information-theoretic
limits have been considered in [2], [33]–[36].
II. S YSTEM M ODEL
We consider a noiseless binary energy harvesting channel
where the transmitter sends status updates and an independent
message simultaneously as in Fig. 1. The transmitter has a
unit size battery, i.e., B = 1. Energy arrivals are known
causally at the transmitter and are distributed according to an
i.i.d. Bernoulli distribution with parameter q, i.e., P[Ei = 1] =
1 − P[Ei = 0] = q. Hence, the inter-arrival times between the
energy arrivals, denoted as τi ∈ {1, 2, · · · }, are geometric with
parameter q. Each transmission costs unit energy; thus, when
the transmitter sends an update, its battery is depleted. The
timings of the transmitted updates determine the average AoI
and the message rate.
The instantaneous AoI is given by
∆(t) = t − u(t)
τ2
τ1
V2 Z3 (τ3 ) V3 Z4 (τ4 ) V4
(1)
T2
Sending information through a timing channel.
Then, the minimum AoI is given by
∗
∆ = inf ∆ = inf lim sup E
π∈Π
π∈Π n→∞
" Pn
Tj2
Pj=1
n
2 j=1 Tj
#
(4)
where Π is the set of all feasible policies. Since the transmitter
is equipped with a unit-sized battery and due to energy causality [18], we have Ti ≥ τi . Note that due to the memoryless
property of the geometric distribution, we assume without loss
of generality, that τi is the time from the instant of the previous
update and not the time from the instant of the previous energy
arrival.
To send information through the timings of the status updates, we consider the model studied in [2, Section V.A]. Thus,
here, we assume the knowledge of the energy arrival instants
causally at the transmitter and the receiver. The information
in the time duration Ti is carried by the random variable
Vi ∈ {0, 1, · · · } where we have here Ti = τi + Vi , see Fig. 3.
The achievable rate of this timing channel is [2],
R = lim inf
n
= lim inf
n
I(T n ; V n |τ n )
Pn
p(V n |τ n )
i=1 E[Vi ] + E[τi ]
sup
sup
p(V n |τ n )
H(V n |τ n )
i=1 E[Vi ] + E[τi ]
Pn
(5)
(6)
where the second equality follows since H(V n |τ n , T n ) = 0.
We denote the AoI-rate tradeoff region by the tuple
(AoI(r), r), where r is the achievable rate and AoI(r) is the
minimum achievable AoI given that a message rate of at least
r is achievable,
#
" Pn
2
j=1 Tj
P
(7)
AoI(r) = inf lim sup E
M n→∞
2 nj=1 Tj
where u(t) is the time stamp of the latest received status update
packet and t is the current time. An example evolution of the
AoI is shown in Fig. 2. The average long-term AoI is
#
" Pn
where M is defined as
)
(
j=1 Qj
(2)
∆ = lim sup E Pn
H(V n |τ n )
∞
T
n→∞
j
≥r
M= {Ti }i=1 Ti ≥ τi , lim inf sup Pn
j=1
n p(V n |τ n )
" Pn
#
i=1 E[Vi ] + E[τi ]
2
Tj
(8)
Pj=1
= lim sup E
(3)
n
2
n→∞
j=1 Tj
where V n denotes (V1 , · · · , Vn ) and similarly for τ n . An
where Ti is the duration between two updates, Qj = Tj2 /2 is alternate characterization for the tradeoff region can also be
the total accumulated age between two updates represented by done using the tuple (α, R(α)) where the achievable AoI is
the area (see Fig. 2), and the expectation is over the energy equal to α and R(α) is the maximum achievable information
arrivals and possible randomness in the transmission decisions. rate given that the AoI is no more than α.
III. ACHIEVABLE T RADEOFF R EGIONS
update in the channel use immediately after an energy arrival.
In this section, we consider several achievable schemes. All
considered achievable schemes belong to the class of renewal
policies. A renewal policy is a policy in which the action Ti at
time i is a function of only the current energy arrival instant
τi . The long-term average AoI under renewal policies is,
" Pn
#
2
E[Ti2 ]
j=1 Tj
Pn
(9)
∆ = lim sup E
=
2E[Ti ]
2 j=1 Tj
n→∞
A. Energy Timing Adaptive Transmission Policy (ETATP)
In this policy, the information which is carried in V is a
(random) function of the energy arrival realization τ . This is
the most general case under renewal policies. The optimal
tradeoff can be obtained by solving the following problem
which results from renewal reward theory [37, Theorem 3.6.1].
Since we use renewal policies and τi is i.i.d., hereafter, we drop
the subscript i in the random variables. Then, the maximum
achievable information rate in (6) reduces to,
H(V |τ )
p(v|τ ) E[V ] + E[τ ]
R = max
(10)
min
p(v|τ )
s.t.
max
p(v|τ ),m
s.t.
(11)
Next, we present our achievable schemes. In the first
scheme, information transmission is adapted to the timing of
energy arrivals: If it takes a long time for energy to arrive, the
transmitter tends to transmit less information and if energy arrives early, the transmitter tends to transmit more information.
This scheme fully adapts to the timings of the energy arrivals,
but this comes at the cost of high computational complexity.
We then relax the adaptation into just two regions, divided by
a threshold c: If energy arrives in less than c slots, we transmit
the information using a geometric distribution with parameter
pb , and if energy arrives in more than c slots, we transmit
the information using another geometric random variable with
parameter pa . The choice of a geometric random variable for
V here and hereafter is motivated by the fact that it maximizes
the information rate when the energy arrival timings are known
at the receiver; see [2, Section V.A].
In the previous schemes, the instantaneous information rate
depends on the timings of energy arrivals. We next relax this
assumption and assume that the instantaneous information rate
is fixed and independent of timings of energy arrivals. We
call such policies separable policies. In these policies, the
transmitter has two separate decision blocks: The first block
is for the status update which takes the decision depending on
the timing of the energy arrival, and the second block is for
encoding the desired message on top of the timings of these
updates. This is similar in spirit to super-position coding. In
the first separable policy, the update decision is a threshold
based function inspired by [1]: if the energy arrives before a
threshold τ0 , the update block decides to update at τ0 and
if the energy arrives after τ0 , the update block decides to
update immediately. The information block does not generate
the update immediately, but adds a geometric random variable
to carry the information in the timing on top of the timing
decided by the update block. In the second separable policy,
which we call zero-wait policy, the update block decides to
(12)
The maximum possible value for r is equal to r∗ =
H(V |τ )
maxp(v|τ ) E[V
]+E[τ ] . The solution of this problem can be found
by considering the following alternative problem which gives
the same tradeoff region
and the AoI in (9) reduces to
E[T 2 ]
E[(V + τ )2 ]
∆=
=
2E[T ]
2E[V + τ ]
E[(V + τ )2 ]
2E[V + τ ]
H(V |τ )
≥r
E[V ] + E[τ ]
H(V |τ )
m
E[(V + τ )2 ] ≤ 2αE[V + τ ]
E[V + τ ] = m
(13)
For a fixed m, problem (13) is concave in p(v|τ ) and can be
solved efficiently. Then, to obtain the entire tradeoff region,
we sweep over all possible values of the parameter α (which
are all possible values of the AoI). The solution for (13) is
found numerically by optimizing over all possible conditional
pmfs p(v|τ ) for each value of m. Then, we use line search to
search for the optimal m. All this, has to be repeated for all
possible values of the AoI α. Finding the optimal solution for
(13) has a high complexity, hence, we propose the following
policy which reduces this complexity significantly, and at the
same time adapts to the timing of the energy arrivals to the
extent possible within this set of policies.
B. Simplified ETATP
In this policy, we simplify the form of the dependence of
the transmission on the timings of energy arrivals significantly.
The transmitter waits until an energy arrives, if the energy
takes more than c slots since the last update, we transmit the
information using a geometric random variable with probability of success pb , otherwise the transmitter transmits the information using a geometric random variable with probability
of success pa , i.e., the transmitter chooses p(v|τ ) as follows
(
pb (1 − pb )v−1 , τ < c
p(v|τ ) =
, v = 1, 2, · · · (14)
pa (1 − pa )v−1 , τ ≥ c
In this case, pa , pb and c are the variables over which the
optimization is performed. The average achieved information
rate as a function of pa , pb and c can be obtained as,
R=
H2 (pb )
pb (1
− (1 − q)c ) +
H2 (pa )
pa (1
− q)c
E[τ ] + E[V ]
(15)
where E[V ] is equal to
E[V ] =
(1 − pa )
(1 − pb )
(1 − (1 − q)c ) +
(1 − q)c
pb
pa
(16)
Now, we can calculate the average AoI with this policy as,
E[(τ + V )2 ]
∆=
=
2E[τ + V ]
2−q
q2
+ E[V 2 ] + 2E[τ V ]
2E[τ ] + 2E[V ]
where we have E[V 2 ] as
2 + p2b − 3pb
2
(1 − (1 − q)c )
E[V ] =
p2b
2 + p2a − 3pa
+
(1 − q)c
p2a
(17)
Finally, we note that in this case E[V 2 ] is equal to,
(1 − pb ) 1
(1 − (1 − q)c+1 ) − (c + 1)(1 − q)c
pb
q
(1 − pa ) (1 − q)c+1
+
(19)
+ c(1 − q)c−1
pa
q
This schemes is simpler than the general class of ETATP;
still, we need to search for the optimal pa , pb and c. We reduce
this complexity further in the next policy.
C. Threshold Based Transmission Policy
We now present the first separable policy. In this policy, we
assume that T = Z(τ ) + V , where the information is still carried only in V ; see Fig. 2. Z(τ ) is the duration the transmitter
decides to wait in order to minimize the AoI, while V is the
duration the transmitter decides to wait to add information in
the timing of the update. Z(τ ) and V are independent which
implies that H(V |Z(τ )) = H(V |τ ) = H(V ). The duration
Z(τ ) is determined according to a threshold policy as follows,
Z(τ ) = τ U (τ − τ0 ) + τ0 U (τ0 − τ − 1)
(20)
The optimal value of τ0 is yet to be determined and is
an optimization variable. The optimal value of τ0 is to be
calculated and, thus, known both at the transmitter and the
receiver; hence, this threshold policy is a deterministic policy.
This ensures that we still have H(V n |τ n , T n ) = 0, which
is consistent with (6). We then choose V to be a geometric
random variable with parameter p. The tradeoff region can
then be written as,
min
T (τ ),p
s.t.
E[(Z(τ ) + V )2 ]
2E[Z(τ ) + V ]
Z(τ ) ≥ τ
H2 (p)/p
r≤
(1 − p)/p + E[Z(τ )]
(21)
where r is a fixed positive number. The feasible values of r are
H2 (p)/p
in [0, r∗] where r∗ is equal to r∗ = maxp∈[0,1] (1−p)/p+E[τ
].
This follows because the smallest value that Z(τ ) can take is
equal to τ . The optimization problem in this case becomes a
function of only τ0 and p.
We now need to calculate E[Z(τ )] and E[Z 2 (τ )]. We
calculate E[Z(τ )] as follows,
E[Z(τ )] =(1 − q)τ0 +
(1 − q)τ0 +1
+ τ0
q
E[V 2 ] =
(18)
and E[τ V ] as
E[τ V ] =
and we calculate E[Z 2 (τ )] as follows,
2 − 3q
2
E[Z (τ )] =
(1−q)τ0 + 2(τ0 +1)(1−q)τ0
q2
(1 − q)τ0 +1
+ 2(τ0 + 1)
+ τ02
(23)
q
(22)
2 + p2 − 3p
p2
(24)
Substituting these quantities in the above optimization problem
and solving for p and τ0 jointly gives the solution.
D. Zero-Wait Transmission Policy
This policy is similar to the threshold based policy, with
one difference: The update block does not wait after an energy
arrives, instead, it decides to update right away, i.e., Z(τ ) = τ .
Hence, the tradeoff region can be obtained by solving,
min
p
s.t.
E[(τ + V )2 ]
2E[τ + V ]
H2 (p)/p
r≤
(1 − p)/p + E[τ ]
(25)
We can then calculate E[(τ +V )2 ] = E[τ 2 +V 2 +2V τ ], where
V and τ are independent as the message is independent of
the energy arrivals. Since τ is geometric E[τ 2 ] = 2−q
q2 . This
optimization problem is a function of only a single variable p.
This problem is solved by line search over p ∈ [0, 1].
IV. N UMERICAL R ESULTS
Here, we compare the tradeoff regions resulting from the
proposed schemes. We plot these regions in Figs. 4-6 for
different values of average energy arrivals, namely, q = 0.2,
q = 0.5 and q = 0.7. For low values of q, as for q = 0.2
in Fig. 4, there is a significant gap between the performance
of ETATP and the simplified schemes. For this value of q, in
most of the region, simplified ETATP performs better than the
threshold and zero-wait policies. As the value of q increases as
shown in Fig. 5 and Fig. 6, the gap between the performance
of the different policies decreases significantly. In Fig. 5, the
threshold and zero-wait policies overlap. In Fig. 6, simplified
ETATP, threshold and zero-wait policies overlap. In all cases,
zero-wait policy performs the worst. This is consistent with
early results e.g., [4], early results in the context of energy
harvesting e.g., [14], [15], and recent results [1], [16], [17],
where updating as soon as one can is not optimum.
R EFERENCES
[1] X. Wu, J. Yang, and J. Wu. Optimal status update for age of information
minimization with an energy harvesting source. IEEE Trans. on Green
Communications and Networking, 2017. To appear. Also available online
at arXiv:1706.05773.
[2] K. Tutuncuoglu, O. Ozel, A. Yener, and S. Ulukus. The binary energy
harvesting channel with a unit-sized battery. IEEE Trans. Info. Theory,
63(7):4240 – 4256, April 2017.
[3] V. Anantharam and S. Verdu. Bits through queues. IEEE Trans. Info.
Theory, 42(1):4–18, January 1996.
[4] S. Kaul, R. Yates, and M. Gruteser. Real-time status: How often should
one update? In IEEE INFOCOM, March 2012.
0.45
0.9
0.4
0.8
0.35
0.7
0.3
0.6
0.25
0.5
0.2
0.4
0.15
0.3
0.1
0.2
0.05
0.1
0
4.2
4.4
Fig. 4.
4.6
4.8
5
5.2
5.4
5.6
5.8
The rate-AoI tradeoff region for q = 0.2.
Fig. 6.
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1.4
Fig. 5.
1.6
1.8
2
2.2
2.4
2.6
0
0.8
2.8
The rate-AoI tradeoff region for q = 0.5.
[5] S. Kaul, R. Yates, and M. Gruteser. Status updates through queues. In
CISS, March 2012.
[6] R. Yates and S. Kaul. Real-time status updating: Multiple sources. In
IEEE ISIT, July 2012.
[7] C. Kam, S. Kompella, and A. Ephremides. Age of information under
random updates. In IEEE ISIT, July 2013.
[8] M. Costa, M. Codreanu, and A. Ephremides. Age of information with
packet management. In IEEE ISIT, June 2014.
[9] Y. Sun, E. Uysal-Biyikoglu, R. Yates, C. E. Koksal, and N. B. Shroff.
Update or wait: How to keep your data fresh. In IEEE INFOCOM, April
2016.
[10] A. Kosta, N. Pappas, A. Ephremides, and V. Angelakis. Age and value
of information: Non-linear age case. Available at arXiv:1701.06927,
2017.
[11] A. M. Bedewy, Y. Sun, and N. B. Shroff. Age-optimal information
updates in multihop networks. Available at arXiv:1701.05711, 2017.
[12] P. Parag, A. Taghavi, and J. Chamberland. On real-time status updates
over symbol erasure channels. In IEEE WCNC, pages 1–6, March 2017.
[13] R. Yates, E. Najm, E. Soljanin, and J. Zhong. Timely updates over an
erasure channel. In IEEE ISIT, June 2017.
[14] R. Yates. Lazy is timely: Status updates by an energy harvesting source.
In IEEE ISIT, June 2015.
[15] B. T. Bacinoglu, E. T. Ceran, and E. Uysal-Biyikoglu. Age of information under energy replenishment constraints. In UCSD ITA, February
2015.
[16] A. Arafa and S. Ulukus. Age minimization in energy harvesting
communications: Energy-controlled delays. In Asilomar, October 2017.
[17] A. Arafa and S. Ulukus. Age-minimal transmission in energy harvesting
two-hop networks. In IEEE Globecom, December 2017.
[18] J. Yang and S. Ulukus. Optimal packet scheduling in an energy
harvesting communication system. IEEE Trans. Comm., 60(1):220–230,
January 2012.
1
1.2
1.4
1.6
1.8
The rate-AoI tradeoff region for q = 0.7.
[19] K. Tutuncuoglu and A. Yener. Optimum transmission policies for
battery limited energy harvesting nodes. IEEE Trans. Wireless Comm.,
11(3):1180–1189, March 2012.
[20] O. Ozel, K. Tutuncuoglu, J. Yang, S. Ulukus, and A. Yener. Transmission
with energy harvesting nodes in fading wireless channels: Optimal
policies. IEEE JSAC, 29(8):1732–1743, September 2011.
[21] C. K. Ho and R. Zhang. Optimal energy allocation for wireless
communications with energy harvesting constraints. IEEE Trans. Signal
Proc., 60(9):4808–4818, September 2012.
[22] J. Yang, O. Ozel, and S. Ulukus. Broadcasting with an energy harvesting
rechargeable transmitter. IEEE Trans. Wireless Comm., 11(2):571–583,
February 2012.
[23] J. Yang and S. Ulukus. Optimal packet scheduling in a multiple access
channel with energy harvesting transmitters. Journal of Comm. and
Networks, 14(2):140–150, April 2012.
[24] Z. Wang, V. Aggarwal, and X. Wang. Iterative dynamic water-filling for
fading multiple-access channels with energy harvesting. IEEE JSAC,
33(3):382–395, March 2015.
[25] K. Tutuncuoglu and A. Yener. Sum-rate optimal power policies for
energy harvesting transmitters in an interference channel. Journal of
Comm. Networks, 14(2):151–161, April 2012.
[26] D. Gunduz and B. Devillers. A general framework for the optimization
of energy harvesting communication systems with battery imperfections.
Journal of Comm. Networks, 14(2):130–139, April 2012.
[27] Q. Wang and M. Liu. When simplicity meets optimality: Efficient
transmission power control with stochastic energy harvesting. In IEEE
INFOCOM, April 2013.
[28] M. B. Khuzani and P. Mitran. On online energy harvesting in multiple
access communication systems. IEEE Trans. Info. Theory, 60(3):1883–
1898, February 2014.
[29] D. Shaviv and A. Ozgur. Universally near-optimal online power control
for energy harvesting nodes. IEEE JSAC, 34(12):3620–3631, December
2016.
[30] A. Baknina and S. Ulukus. Optimal and near-optimal online strategies
for energy harvesting broadcast channels. IEEE JSAC, 34(12):3696–
3708, December 2016.
[31] H. A. Inan and A. Ozgur. Online power control for the energy harvesting
multiple access channel. In WiOpt, May 2016.
[32] A. Baknina and S. Ulukus. Online policies for multiple access channel
with common energy harvesting source. In IEEE ISIT, July 2016.
[33] O. Ozel and S. Ulukus. Achieving AWGN capacity under stochastic energy harvesting. IEEE Trans. Info. Theory, 58(10):6471–6483, October
2012.
[34] W. Mao and B.Hassibi. Capacity analysis of discrete energy harvesting
channels. IEEE Trans. Info. Theory, 63(9):5850–5885, September 2017.
[35] D. Shaviv, P. Nguyen, and A. Ozgur. Capacity of the energy-harvesting
channel with a finite battery. IEEE Trans. Info. Theory, 62(11):6436–
6458, November 2016.
[36] V. Jog and V. Anantharam. A geometric analysis of the AWGN channel
with a (σ,ρ) -power constraint. IEEE Trans Info. Theory, 62(8):4413–
4438, August 2016.
[37] S. M. Ross. Stochastic Processes, volume 2. John Wiley & Sons New
York, 1996.
| 7cs.IT
|
1
Rohlin Distance and the Evolution of Influenza A virus: Weak
Attractors and Precursors
arXiv:1112.2608v1 [q-bio.PE] 12 Dec 2011
Raffaella Burioni1,∗ , Riccardo Scalco2 , Mario Casartelli1
1 Dipartimento di Fisica e INFN, Università di Parma, Parma, Italy
2 Department of Biochemistry, University of Zurich, Zurich, Switzerland
3 Dipartimento di Fisica e INFN, Università di Parma, Parma, Italy
∗ E-mail: Corresponding [email protected]
Abstract
The evolution of the hemagglutinin amino acids sequences of Influenza A virus is studied by a method
based on an informational metrics, originally introduced by Rohlin for partitions in abstract probability
spaces. This metrics does not require any previous functional or syntactic knowledge about the sequences
and it is sensitive to the correlated variations in the characters disposition. Its efficiency is improved by
algorithmic tools, designed to enhance the detection of the novelty and to reduce the noise of useless
mutations. We focus on the USA data from 1993/94 to 2010/2011 for A/H3N2 and on USA data from
2006/07 to 2010/2011 for A/H1N1 . We show that the clusterization of the distance matrix gives strong
evidence to a structure of domains in the sequence space, acting as weak attractors for the evolution,
in very good agreement with the epidemiological history of the virus. The structure proves very robust
with respect to the variations of the clusterization parameters, and extremely coherent when restricting
the observation window. The results suggest an efficient strategy in the vaccine forecast, based on the
presence of ”precursors” (or ”buds”) populating the most recent attractor.
Introduction
There is a long history in approaching DNA and RNA sequences as texts, with quantitative estimates
for various kinds of statistical properties and complexity indicators [1, 2]. The general idea behind this
approach is that the information encoded in the sequence is strictly related to the properties of the
corresponding biological structures and that good indicators should be able to recognize similar functions
in different sequences.
The sector devoted to estimate the relevance of mutations along a time ordered set of evolving sequences is particularly interesting when a sufficiently long record of samples is accessible, as for viral
RNA of rapidly evolving diseases [3–6]. We focus on a definite kind of statistical properties, precisely
metric properties.
The distance between sequences is a concept admitting several implementations. In the context of
evolving viral RNA, distances based on the sequence symbols are mostly of the Hamming type: for two
strings a and b of characters, the Hamming distance dH (a, b) is the number of sites with different symbols
[7–9]. Such distances are sensitive to local features only, since mutations occurring at different sites are
non correlated. In this framework, an interesting solution has been proposed in [8, 9], where, focusing
on particular locations of the sequence (the epitopes, whose role and peculiarities in Influenza Virus
evolution are well known), authors succeeded in extracting important features of strains evolution. In a
sense, the extra information introduced with the choice of the epitopes proved efficient in overcoming the
intrinsic uncorrelation of the Hamming metrics, leading to interesting results. Other relevant approaches,
based on sequences information only, rely on entropic distances [10–14] referred to the Shannon’s entropy
or to compression algorithms, and they are mainly addressed to the comparison of strings of different
length in inhomogeneous frameworks, a procedure motivated by the fact that in evolving sequences,
beside substitutions, there are frequent insertions and deletions [15].
There are few remarkable alternatives to these ”sequence based” type distances. Hemagglutination
2
inhibition (HI) assays [16], reporting the ability of ferret antibodies, raised against one viral strain, to
inhibit a second strain’s ability to agglutinate red blood cells, are currently used to define similarity
between antigens [17]. Certainly, the metrics extracted from HI tests is directly related to the real
antigenic similarity between strains, but it requires HI assay animal data, which are difficult to obtain
with high precision.
We intend to introduce an enhanced version of a different metrics, known as Rohlin distance [18, 19],
which is based on the sequences symbols and is expected to be sensitive to their global distributions
and correlations. It is also founded on the Shannon’s entropy but, differently from other informational
functionals, applies in a biologically homogeneous framework. Moreover it does not deal with the frequency probabilities of symbols on single sites, which are too poor as units and do not touch the global
structure (see however [11] for interesting improvements in this directions). In our approach, the basic
entities are indeed the partitions of a sequence into subsequences, as they are determined, starting from
a configuration (the list of amino acids), by a projection operation: precisely, we consider the partitions
defined by homogeneous segments. This aims at evidencing the ordered collection of connected subsequences of equal symbols. For instance, the alphabetical string a = AAABBACCCB would be divided
into five subsequences, and each subsequence would determine a segment. To the first segment AAA
there correspond the site subset with labels (1, 2, 3), to BB the subset (4, 5), etc. The natural length of
each segment is the number of its symbols. In the example, the lengths are 3, 2, 1, 3, 1. Once such lengths
are correctly normalized, this assignment is equivalent to the definition of a probability measure on the
subset algebra, proportional to the number of sites contained in each subset. Finally, partitions can be
represented by their bounds, the segment extremes, or more economically by the left extremes, allowing
for a simple and straightforward comparison between partitions.
In the partition space the Rohlin distance dR is then defined, for any couple of partitions (α, β), by
the mutual conditional Shannon entropy (see Materials and Methods for details):
dR (α, β) = H(α|β) + H(β|α) .
The conditional Shannon entropy H(α|β) represents the residual information needed to describe the
segment disposition of α when the disposition of β is known, or, in other terms, how the knowledge of β
may contribute to the knowledge of α. Therefore the symmetrized form above, defining dR (α, β), is the
total information required to distinguish α and β, seen as schemes of segments with their probabilities.
In absence of a bias, our choice, assigning equal weight to each site and leading to a measure proportional to the length, is the most natural. Other probability measures can be defined, but they cannot
depend on the configurations, since in that case the same set would have different measures for two configurations, and the conditional Shannon entropy would loose its meaning, inhibiting the very definition
of Rohlin distance.
The segmentation provided by the partitioning of the sequences into homogeneous subsequences entails many advantages: its definition is simple and universal; sequences are not too tightly fractioned, as
by single symbols; no a priori knowledge is required, along with the exigence of a “black box” analysis.
Moreover, even if segments have no intrinsic biological meaning (and this could appear as an inconvenience), alterations in their overall distributions, as those emerging in historical records, are by definition
compatible with the biologically efficient features proposed by evolution [20].
Thus, by using probabilities which arise from this geometrical and topological asset, the distance
dR (α, β) measures the information content carried by evolution, giving evidence to the emerging dissimilarities. Clearly, this content should be filtered, as far as possible, from the effects of the non evolving
part: to this end we introduce the so called reduction process π, a method designed to amplify the
relevant differences between partitions by dropping their common sub-partitions. This means that for
b = π(α, β) at amplified distance. From a practical point
any couple (α, β) there is a reduced couple (b
α, β)
of view, the reduction consists in erasing the common extremes of segments between two partitions. A
key point is that this method proves surprisingly effective also in filtering the noise of useless mutations.
Details are given in the Material and Methods section.
3
We shall deal here with an interesting example of a highly mutating sequence, the RNA of influenza
A virus, whose databases are particularly rich. More precisely, we consider the amino acids sequence
of the surface protein hemagglutinin for H3N2 subtype Influenza A virus [21], with human as host, in
1470 strains collected in USA from 1993/94 to 2010/2011 [22], and the analogous sequence for H1N1
subtype for 2506 strains collected in USA from 2006/07 to 2010/2011. As our method works on equal
length strings and it is based on an informational metrics with long range correlations, it is expected
to perform better when applied to the longest sequences. We therefore choose the full length (566) HA
sequences, which are a subsample of the available sequences. We also consider sequences identified by
a complete date, as the time when the sequence appears represents an important information in our
analysis. The restriction to the USA sequences is motivated by several facts. First, choosing sequences
from the temperate regions, we give relevance to the seasonal timing of the virus evolution, minimizing the
interference with a dephased development; second, the geographical bounds in the sampling ensure that
we are looking, season by season, to a reasonably stable population. Notwithstanding these restrictions,
for example for H3N2 we process 1470 sequences over 1986 in the northern hemisphere (more than 74
%) and over 2759 in the world (more than 53 %). In other words, we keep the statistical majority,
disregarding only possible noise. Analogous estimates hold for the processed H1N1 sequences.
After partitioning each sequence, we calculate the Rohlin distance between all the partition pairs,
and analyze the whole sequences sample by the hierarchical complete linkage clustering algorithm [23]
on the distance matrix. The procedure has strong analogies with the analysis presented in [9, 24] but in
a completely different metric space, namely the partition space instead of the configuration space. Our
analysis traces the evolution of topological properties of the sequences, while the virus escapes following
its antigenic drift. Interestingly, the structures arising in the sequence space according to this metrics
result to be quite meaningful; they individuate indeed well defined regions of the sequences space acting
as weak attractors [9,17], where the evolution of the virus takes place for definite periods. Moreover, the
attractors display precursors [9, 17], i.e. sequences populating the regions well before they are identified
as circulating strains, an information which appears to be relevant in the forecast problem, suggesting
an alternative strategy in vaccines formulation. Once again there are analogies but also differences with
the privileged regions of evolution presented in previous analysis [9]. Our attractors arise indeed from of
similitude related to the overall disposition of homogeneous segments, and not to the actual data lying
on every site (with the supplementary restriction to epitopes). This means that, more than an overlap of
results, the two approaches present complementary points of view, enforcing one another. It is therefore
quite remarkable that the vaccine predictions of the two methods agree very well, a fact whose origin is
not completely understood yet.
Results
Clustering and weak attractors in Rohlin metrics
Using the sequences set, we calculate the whole N × N matrices of Hamming and Rohlin distances, i.e.
b ik , whose entries are respectively dH (ai , ak ) on sequences, or
Hik , Rik and reduced Rohlin distance R
dR (αi , αk ) and dR (b
αi , α
bk ) on the corresponding partitions (possibly simplified by the reduction process).
The primitive (non-reduced) partitions result to be by far less informative than the reduced ones, so that
from now on we shall omit to report calculations on them.
This fact is a non trivial result in itself: in principle, by deleting the common bounds, the reduction
process could imply indeed the survival of a random set of unstable bounds created by the incoming
mutations. This happens, as expected, with artificially created random mutations (as reported in the
last subsection). Clearly, far from being more informative, such chaotic residual set would be useless for
our scope.
The empirical fact we observe is quite different: once the reduction process has cleaned the useless
4
common sub-partitions, the surviving bounds (defining the new reduced partitions), far from being “random”, carry the winning novelty in the adaptive strategy of the virus. In other terms, the effective new
disposition of segments is not random because it has been selected. Moreover, the mutual disposition has
intrinsically a long range character, which is captured by such a non local metrics as the Rohlin distance
is (details on the amplification are given in the mathematical section in Materials and Methods).
b ik . The used clustering tool is
The scenario emerges from a “clustering analysis” on the matrix R
the standard hierarchical complete linkage algorithm [23], where the number p of clusters is an external
parameter. An important point is therefore the choice of the optimal p. Since at fixed p every clusterization corresponds to a partition of the whole set into groups populated by n1 , n2 , ..., np sequences, with
n1 + n2 + ... + np = N , by defining the probabilities wP
i = ni /N, i = 1, 2, ..., p to each clusterization we
can associate the Shannon entropy H(w1 , ..., wp ) = − i wi log wi of its probability distribution. Since
at growing p the cluster populations cannot increase, the entropy is non-decreasing [25]. However, if it
is substantially stable, this means that the clusters are also stable (or equivalently, that newly added
probabilities are very small); if it grows, then clusters (and probabilities) are almost continuously splitted
into smaller and smaller ones. Now, observing the behaviors of Rohlin and Hamming entropies in Figure
1 for H3N2, we note a clear quasi-plateau for Rohlin in the interval 14 < p < 34, while Hamming is always
growing. For the latter, the interpretation is that the Hamming clusters, being an artificial product of
the procedure, split with a remarkable continuity.
As to Rohlin, the long plateau clearly indicates that the clusters are real structures in the sequences
space, keeping a definite individuality in a large observation range. The growth for low p is simply due to
the fact that, if the imposed number is too small, (p = 1, 2, 3...), the calculated clusters must contain the
real ones, so that at growing p the splitting is effective, up to the optimal number when calculated and real
clusters coincide. On the contrary, for p too large, calculated clusters are so numerous that also the real
ones begin to split much more effectively than during the peripheral loss registered in the plateau. The
plateau extremes may be therefore roughly related to the typical isolation length among real clusters, and
to their maximal diameter respectively. Interestingly, the optimal value p = 14, obtained from clustering
without additional hypotheses, is consistent with the number of different circulating strains identified by
the WHO HI tests.
Analogous results obtained for H1N1 are shown in Figure S1. In that case, the optimal value for the
clustering parameter is p = 8.
Since every sequence is marked by its sampling date, a natural question is the time distribution of
the resulting clusters. The upshots for H3N2 are summarized in Figure 2. The part above refers to the
b ik . Below, for comparison, to the clustering on Hik . There is no scale on the Y axis
clustering on R
because the ordinate only distinguishes among clusters (same ordinate ≡ same cluster): order and color
have no intrinsic meaning and they have been chosen with readability criteria only. The 11 different
polygonal symbols represent the 11 reference viruses, including the alternates, observed in the Northern
Hemisphere according to WHO HI tests. Their names are indicated in the plot and the details of the
sequences are given in Table 1. The reference sequences are not identified by a time X coordinate. They
are processed together with the dataset and they are positioned in the clusters they belong to after the
Rohlin clustering procedure. Vertical lines separate winter seasons and are conventionally set at July 31.
From the dR clustering of Figure 2, which is very stable by changing p in the plateau range of Figure
1, we can draw several indications. First, there is a clear long temporal extension of clusters, which are
densely populated for several winter seasons. Interestingly, they present precursors, that we term ”buds”,
and successors, i.e. a bunch of sequences representing viruses that appear, in time, before or after the
main part of the cluster [9]. The identification of buds will be explained in details in the next sections.
Let us consider, for instance, the 94/95 season. In that year, we observe a bud (Wuhan strain, the
yellow cluster) which is getting stronger in the following season, living jointly with the Johannesburg
strain, the red cluster. Then, it becomes the dominant strain in 95/96 and 96/97, while it may be
considered a successor in 97/98, when the Sidney (light green) is the dominant strain. Notice that the
5
same Sidney was a bud in 96/97.
The distribution of clusters suggests that the evolution in the sequence space takes place in preferential
regions, corresponding to each cluster, which can be populated well before and after the main season.
Such regions act as a kind of weak (i.e. non definitive) attractors. For example, as mentioned, two H3N2
virus strains circulated during 96/97 winter season: Wuhan successors and Sydney buds. Through HI
tests, WHO revealed the Wuhan reference virus as the circulating one, recommending it as a vaccine for
the season 97/98. It is crucial that, already during 96/97, our analysis shows the emergence of a bud,
the Sydney family strain, which is the actual virus circulated in 97/98 winter season. Rohlin attractors
correctly describe also the heterogeneity coming from ”outliers” sequenced by the WHO, which must not
be treated separately, as it happened in other Hamming based approaches [24]. These sequences naturally
fall into a cluster, confirming that Rohlin correctly takes into account the variability present in outliers.
A second point is that dR -clustering is consistent with epidemiological WHO-HI data. For example, the
subsequent strains A/Wuhan/359/95, A/Sydney/5/97 and A/Moscow/10/99, appeared during years from
95/96 to 02/03 according HI tests, are represented by three well defined clusters (A/Panama/2007/1999
is a Moscow alternate). Interestingly, their reference sequences belong to the correct clusters once they
are included in the data set, without any a priori information.
In the lower part of Figure 2 the same clustering procedure is shown referring to the Hamming
matrix Hik . A definite temporal extension of clusters is observed, in agreement with previous results
[9, 24]. However, the cluster temporal distribution obtained from Hik is quite unstable, confirming the
dependence on p evidenced in Figure 1. The p used here is the same of Rohlin, but the choice is completely
arbitrary because there is not a clear plateau for Hamming. This means that some appearing spots are
not true buds, as they results from the almost continuous splitting of Hamming clusters, which are not
stable under a change in p. Namely, a new cluster can be produced simply by raising p. Moreover in
some seasons (e.g. from 99/00 to 02/03) there is a contrast between WHO indication and the cluster
arrangement: some clusters are not represented by any reference sequence, while others are wrongly
doubly represented, showing a poorer correlation with HI analysis. For example, Sydney and Moscow
reference strains belong to the same cluster while they are expected to be in different ones.
Buds in Rohlin weak attractors and vaccine forecast
An interesting evidence can be drawn from the position of symbols in the attractors. The symbols above
the upper diagram in Figure 2 show in the first row the vaccine indicated by the WHO on the basis of
the HI tests in previous seasons, while in the second row they represent the indication we would suggest
on the basis of the following criterion: by looking at the dR -cluster distribution, when in a year there
are simultaneous strains (with statistically significant populations), we would indicate the newest one,
i.e. the bud, as emergent in the next year. This criterion, in other words, sees the novelty carried in
emergent buds as a feature enhancing the aggressiveness of the virus. In both the WHO or buds analysis,
the symbols are shown in green when the vaccine choice agrees with the circulating strain, in red when
they disagree.We used a two-colors symbol when more than one virus circulated in the same season and
the corresponding prediction agrees with one of the two circulating strain. Now, the second symbols row
indicates that the buds criterion is able to identify correctly the circulating strain every year, apart from
the season 01/02 because of the lack of sequences, while the WHO criterion fails in 3 cases over 17. Of
course, it will be extremely interesting to verify the criterion on the set of sequences for next season, as
soon as they will be available.
In the lower part of Figure 2, displaying the results of the same procedure on the Hamming distance,
we note that buds as early warning of new strains are not reliable because, as discussed above, the
instability in p does not allow for an unambiguous detection of their appearance, as they can be produced
by raising the value of p.
The bud criterion can be successfully applied also to the very interesting case of H1N1 and the results
are shown in Figure 3. The sampling for H1N1 is very inhomogeneous. From 2006 the statistics increases,
6
so we will limit our analysis to this period. From the clustering entropy analysis (shown in Figure S1), the
optimal p is 8. The very relevant cluster starting suddenly around April 2009 (red line in Figure 3) [26]
represents precisely the pandemic virus appeared in the 2009 season. In that case, the bud criterion
partially fails, as it recognizes correctly only the strains circulated at the beginning of the season. This is
reasonable, since our method is expected to be effective in simple antigenic drift, and not in the case of a
dramatic change, as in the 2009 pandemic case. The shift probably sets a completely new “direction” in
the sequence space. Notice that our method evidences the simultaneous occurrence of four well distinct
clusters in 2009, a feature missing in the Hamming analysis. The observation of multiple clusters signals
could be related to a typical instability of post-pandemic periods, as the one we are facing according to
the WHO [26]. We expect this structure to be present also in the new set of sequences for season 2011/12.
Restricting the time window
Another natural question, the relevance of the examined time window, is treated in Figure 4, displaying
the results one would obtain by stopping the data collection at five different years, i.e by applying the
clustering to the restricted sets of sequences available at those times. This procedure is intended to clarify
how the bud criterion works and to check that it is not an unpredictive a posteriori verification of the
vaccine choice, but a real working framework.
The time distribution in the X axis describes at various years the position of the H3N2 sequences,
exactly as in the upper Figure 2. However, to reproduce exactly the situation in which the WHO vaccine
prediction is made, we collect all the sequences available up to March of a given year and we perform our
clustering analysis only on that dataset. There is therefore an increasing number of sampled sequences,
in every horizontal time sector starting for the upper part down to the lowest. In details, we apply our
entropy criterion each time to the dataset restricted to the end of the winter season of a given year and,
from that dataset only, we choose the optimal p for that case. Interestingly, there is a clear plateau in
the entropy analysis for every restricted time window, allowing for a unambiguous choice of the number
of clusters. The Rohlin entropy analysis from the restricted time window is shown in Figure S2, together
with the corresponding one for Hamming. Notice that for the Hamming clustering, there is no clear
indication of the number of clusters p from the entropy analysis, nor any evidence of emergent buds.
In principle, leaving out part of the data, these clusterings could be different from the final one. It is
remarkable that the structure remains the same. Buds are clearly present, as if the evolution took place
in a well defined landscape, with preferential “antigenic” directions that are filled during the genetic drift,
and acting therefore as weak attractors. The symbols of the reference WHO strains are excluded from
the clustering in the time restricted window, as they would represent an a posteriori knowledge. They
are associated to each cluster by an inverse analysis, i.e. by calculating the reference WHO strain which
has minimum distance with the sequences belonging to the bud cluster. Details on the reverse analysis,
applied also to the whole dataset are given in the next subsection. Now, in Figure 4 a syringe indicates
the most accredited vaccine for the next year based on the bud criterion: even when the previous years
database is poor, the forecast is very good and the cluster corresponding to the syringe is exactly the
prevailing cluster observed the next year. By confirming the coherence of the procedure, this result
supports the bud emergence criterion for the prediction of the new prevailing strain.
Interestingly, our bud criterion, which does not include any additional information on epitopes positions, agrees very well with the dominant strain prediction discussed in [9]. This appears to indicates
that the ”black box” Rohlin distance analysis is able to grasp the biological information included in the
”epitopes” metrics, which is certainly correct but requires an additional input (the relevant positions).
We notice that our results for the vaccine choice always agree with the prediction of [9] when there is
a single circulating strain, while they are complementary when there are two. We do not have a clear
explanation for this interesting fact, at the moment.
7
Testing the method: reverse analysis and random permutations
The clustering procedure can be performed by a completely different method, which does not consider
distances between sequences themselves, as in the hierarchical method, but refers to the WHO different
sequences identified by the HI analysis. Precisely, the Rohlin distance has been calculated between each
of the 1470 sequences of the H3N2 with all the 11 WHO reference virus strains. The sequences are
temporally aligned along the X axis of Figure S3, while the reference strains have a conventional position
along the Y axis. In the diagram, each sequence is represented by a point whose X coordinate is its
sampling date, and whose Y coordinate corresponds to the nearest WHO reference strain. Surprisingly,
the final result of this new procedure is almost the same as the one showed in Figure 2 of the main text,
while it is obtained from a completely different analysis. The analogous plot on Hamming distance does
not preserve the cluster structure. Once again, the Rohlin distance approach proves robust and consistent
with the HI tests analysis.
As a further check of the robustness of our results, we consider a random permutation of the site
labels, simultaneously performed on all the sequences of the H3N2 dataset. This operation leaves the
Hamming distances invariant by definition. Since such a random mixing of the amino acids is biologically
meaningless, a natural request is that the conclusions drawn from a correct metrics should crash: this
is precisely what happens with the Rohlin distance. In other words, with Rohlin, only the partitions
corresponding to the real sequences seem to encode correctly the antigenic drift during the evolution,
evidencing a meaningful relation with the global structure of the sequences. Vice versa, the simple
global “mutations counting” completely fails to recognize the information deletion caused by the label
permutation. The results are presented in Figure S4.
Discussion
The mechanism underlying influenza A antigenic plasticity, that is, how the virus continually escapes
the immune system by producing variant strains that cause re-infection within a few years, remains an
outstanding evolutionary problem.
There are two main general pictures for this evolution. The first one is based on an almost continuous
slow drift from an ancestor sequence, with some large shifts occurring at certain stages of the evolution
[27]. The second one relies on a punctuated evolution, where ”antigenic” swarms of sequences populate
“basins” in the sequence space for several years, until the circulating swarm “jumps” to another basin,
reinfecting the population. Dynamical model for this type of evolution have been built [28], and the
evidence of a punctuated antigenic evolution has been put forward by several authors [17, 24, 29–31].
The picture emerging from the Rohlin metrics seems to support such a punctuated evolution with a
better fit of epidemiological data, giving also insights on the relevant distance between circulating strains
and vaccines. In fact, clusters result to be organized into well defined regions of the sequences space: the
virus appears to explore for several seasons the sequence space region corresponding to a certain Rohlin
width, until a jump takes it to another attractor, where the evolution starts again (a return to a previous
region is also possible). These regions constitute therefore weak attractors in the sense that they are
able to trap the virus for a finite time, and they can also be re-populated after years. In other words,
weak attractors seem to identify privileged antigenic directions from genetic data.
Interestingly, the clusters present “precursors”, that we termed buds, i.e. a small number of sequences
which explore in advance the next attractor when most of the strains still belong to the previous one.
All goes as if such precursors manage to experience the winning escape strategy, that will be followed
by the main swarm in subsequent years, and a clear correlation emerges between the bud, i.e. the
younger attractor appeared in a given year, and the circulating strain of the subsequent season. This
bud criterion, in parallel with HI analysis, could be helpful in the correct choice of the vaccines. The
picture emerging from Rohlin distance analysis appears to hold also by processing analogous data sets
8
as the A/H1N1 in USA. Interestingly, in H1N1 the bud criterion partially fails in 2009, as it recognizes
correctly the emerging bud only before the pandemic period, while it is not able to predict the clear new
cluster that appears suddenly in April 2009. The analysis correctly signals also the high instability of
the post-pandemic phase in 09/10.
In conclusion, some main points should be stressed: the first is that no a priori knowledge of biological
nature has been used or put into the data set. The indications we have derived from clustering on the
distance matrix constitute a genuine emergence. It seems plausible therefore that the same approach could
work in similar circumstances, i.e. when a homogeneous set of equal length arrays are at disposal. The
second point is the existence of structures in the sequence space, that can be described as weak attractors,
where the evolution of the viral species takes place with a discontinuous dynamics. Clearly, a clustering
algorithm is expected to recognize a chronological order within the distance matrix, whenever the distance
is a monotone function of time, but in that case one would also expect the progressive fragmentation of
clusters as the external parameter p grows. Such is, substantially, the behavior suggested by the analysis
on the Hamming matrix. On the contrary, the presence of precursors [9], which discontinuously anticipate
the onset of future attractors, and the stability of the attractors structure at varying p or sampling, are
quite non trivial facts, implying that the Rohlin attractors are not a conventional decomposition in the
sequence space; they possess instead a robust, intrinsic, “natural” meaning. It seems therefore that the
Rohlin distance on reduced couples is able to evidence a selected variety of admissible “antigenic states”,
preferentially explored through mutations, which remains hidden in other metric approaches. The third
point is that the “buds emergence criterion” could offer a valuable complementary tool for an optimal
strategy in the choice of vaccines. The matter is obviously delicate, and a long series of experimental
checks should explore and confirm such a possibility before practical utilization, but we think that an
effort in this direction is worthwhile.
Materials and Methods
Data
The main database of reference for H3N2 is constituted by 1470 full-length (566 aa) HA proteins of
the H3N2 subtype Influenza A virus isolated in USA from 1993/94 to 2010/11, excluding sequences
with an incomplete sampling date. Such set is enriched with the 11 reference sequences corresponding
to the reference viruses circulated in the same years according to WHO HI analysis [32]. The total is
N = 1481 sequences , written in the 20 amino acids alphabet. As for H1N1, the main database of
reference is constituted by N = 2506 full-length (566 aa) HA proteins of the H1N1 subtype Influenza A
virus isolated in USA from 2006/07 to 2010/2011, excluding sequences with an incomplete date. Such set
is enriched with the 4 reference sequences, corresponding to the reference viruses circulated in the same
years according to WHO HI analysis [32]. The total is N = 2510 sequences, written in the 20 amino
acids alphabet.
Sequences and Rohlin Metrics
Let K be a finite alphabet of characters (amino acids in our case). A sequence a ≡ (a1 , a2 , ..., aL ),
where ak ∈ K, may be thought as a function on the one dimensional array M of sites labeled 1, 2, ..., L.
This function defines the configuration or state on M. Every sequence a is therefore an element in a
configuration space C(M). A probability measure µ on the finite subset algebra M of M is given by
the normalized number of sites in every subset. This means that all sites are assumed to be equivalent.
For instance, if L = 12 and a subset E includes sites (2, 5, 10), then µ(E) = 0.25. Other measures are
possible, e.g. by assigning weights to the sites. However, such weights (and measures) cannot depend on
the configurations, otherwise the same subset of M could have more than one measure simultaneously,
9
and the functionals defined below would loose any meaning.
A partition α ≡ (A1 , ..., Am ) of M is an exhaustive collection of disjoint subsets (called “atoms”) of
M. The space Z(M) is the set of all possible partitions of M, where a partial order α ≤ β means that
β refines α. The “product” γ = α ∨ β ≡ αβ (a close analogous of the minimal common multiple) is the
minimal partition refining both factors α and β. The unit partition ν has only one atom, the whole set
M. Obvious properties such as ν ≤ α = αα ≤ αβ easily follow for every α and β.
The operation σ = α ∧ β is the maximal common factor (i.e. the most refined common sub-partition)
of α and β. Clearly, ν ≤ α ∧ β ≤ α, etc.
Every partition may be thought as an experiment where an elementary or atomic event Ak occurs
with probability µ(Ak ). Then, the meaning of the definitions above is that ν is the trivial experiment,
α ≤ β means that α is a sub-experiment of β, etc.
The Shannon’s entropy H(α) is defined as:
H(α) = −
m
X
µ(Ai ) ln µ(Ai ) .
(1)
i=1
If β = (B1 , ..., Bs ), is another partition, the conditional entropy H(α|β) is
H(α|β) = −
s
m X
X
µ(Ai ∩ Bk ) ln
i=1 k=1
µ(Ai ∩ Bk )
= H(αβ) − H(β).
µ(Bk )
(2)
These quantities give respectively the mean incertitude of an experiment α and the residual mean incertitude on α when the result of β is known [25]. Now, for all α and β in Z(M), the Rohlin distance
is:
dR (α, β) = H(α|β) + H(β|α) .
(3)
The useful formula
dR (α, β) = 2H(αβ) − H(α) − H(β)
(4)
follows from Eq. 2. Thus, dR (α, β) is a measure of the overall non-similarity between α and β, giving
account of the mutual correlations among the respective outcomes [18,19]. These concepts and definitions
hold true in all probability spaces. For discrete spaces (graphs or lattices), where the states or configurations are determined by the values assumed by sites j’s in a finite alphabet K, dR is therefore deeply
different from the well known Hamming distance dH between configurations a and b. This distance is
defined, up to a possible normalization factor, by:
X
dH (a, b) =
| bj − aj | .
(5)
j
where | x − y | is a distance in K if the alphabet is numerical, otherwise is 1 for aj 6= bj and 0 for
aj = bj (as in our case, since K is the alphabet of amino acids). Counting the sites with different symbols
regardless of their position, dH tells one nothing about correlations between mutations. It is important
to stress that the Hamming and Rohlin distances are not defined on the same objects, the former being
between configurations in C(M), the latter between partitions in Z(M).
In our particular case, where M is a one-dimensional finite lattice, and the states (or configurations)
are character sequences of length L, we shall work with partitions generated by homogeneous segments,
i.e. consecutive sites with the same value in K. Of course, in Z(M) there exist much more partitions,
e.g. those with non connected atoms. As an example with L = 12, consider the fictional configuration
a = {A A A T T C A A A F F B } The atoms (indicated by the site labels) of the corresponding
partition α = Φ(a) are {(1, 2, 3), (4, 5), (6), (7, 8, 9), (10, 11), (12)}. The map Φ : C → Z is univocal but
non invertible, since several configurations are mapped into the same partition. For instance, a mutation
10
from C to B as in {... T T C A A ...} → {... T T B A A ...}, does not affect the boundaries, and it
leaves the segment structure unchanged. Thus, by the correspondence Φ and the Rohlin distance, we
can evaluate “how different” are the states on M with regard to the correlated distribution of segments.
It is true that there is a loss of information due to the projection of many configurations into the same
partition; but a comparable loss takes place also for Hamming, since the single site contribute gives
account only for the “equal-or-not” distinction in K. Moreover, as noticed, sites in dH are always totally
uncorrelated.
The non-similarity between two partitions could be confused and weakened by the presence of a
tight common factor, that we would eliminate as far as possible, in order to amplify the Rohlin distance
giving evidence to the real emerging novelty. However, such a “reduction” operation (analogous to the
reduction to minimal terms for fractions) is not uniquely defined because partitions do not admit a
unique factorization into primes [33,34]. The role of prime (i.e. indecomposable) factors can be played by
dichotomic sub-partitions, which are still extremely redundant. Then, the key point consists in defining
for each partition a restricted family E(α) of “elementary” dichotomic factors, with the following features:
1. E(α) must be well defined for every α ∈ Z, or at least in the subset of Z actually under investigation;
2. E(α) does not contain more factors than the number m of atoms in α;
3. ∨m
ek = α.
k=1 α
Now, assuming that the elementary factors families E(α) and E(β) have been defined, the reduction
process consists in the following steps:
1. define the maximal common divisor σ = α ∧ β;
2. drop from E(α) and E(β) those factors which are not relatively prime with σ, and note the surviving
factors α
bk and βbj respectively (i.e. α
bk ∧ σ = βbj ∧ σ = ν);
3. define α
b = ∨k α
bk and βb = ∨j βbj .
In other words, we drop those dichotomic factors which are subfactors of the maximal common factor σ,
b are generated by the surviving factors. The amplification of non-similarity is a
and the reduced (b
α, β)
consequence of the following property:
b ≥ dR (α, β).
Proposition: dR (b
α, β)
b indeed, as
The proof is elementary recalling that, if σ = α ∧ β, we can write α = σ α
b and β = σ β:
mentioned, σ contains all the factors dropped during the reduction. Therefore, by formula (4) and the
b can be rephrased as
fact that σσ = σ, the thesis dR (α, β) ≤ dR (b
α, β)
b − H(σ α
b ≤ 2H(b
b − H(b
b
2H(αβ) − H(α) − H(β) ≡ 2H(σ α
bβ)
b) − H(σ β)
αβ)
α) − H(β)
(6)
Moving terms between the sides
b − 2H(b
b ≤ H(σ α
b − H(β)
b ,
2H(σ α
bβ)
αβ)
b) − H(b
α) + H(σ β)
(7)
and using formula (2) for the conditional entropy, the thesis reduces to
b ≤ H(σ|b
b .
2H(σ|b
αβ)
α) + H(σ|β)
But this is clearly true since
b ≤ H(σ|b
b ≤ H(σ|β)
b ,
H(σ|b
αβ)
α) and H(σ|b
αβ)
because the conditioning terms are greater in the left sides, q.e.d.
(8)
11
It is important to remark that this amplification regards the anti-similarity of the couple as a whole,
while for reduced partitions as single entities the complexity possibly decreases, as expected: this means
H(b
α) ≤ H(α), etc.
b defining this reduction process is many-to-one and idempoThe correspondence π : (α, β) → (b
α, β)
tent. It is a projection from Z × Z on the subset of irreducible pairs. The process, therefore, essentially
depends on the family of elementary factors, a choice which a priori can be implemented in many ways,
reflecting the kind of interest the observer has in the experiment. Details and procedures in abstract
probability spaces may be found in [33, 34]. Here we sketch an algorithmically easy recipe, fitting the
very special case of character strings.
By exploiting one-dimensionality, a partition into segments (connected subsequences) can be economically represented by the list of the left bounds of segments. In the example above, α is fully determined
by (1, 4, 6, 7, 10, 12).
This suggest a very convenient choice of the family of elementary factors: precisely, for every α =
{A1 , A2 , ..., Ak , ...AN }, the k-th dichotomic factor α
ek ∈ E(α) is
α
ek = {
k
[
j=1
Aj ,
N
[
Aj } ,
j=k+1
and therefore, in terms of labels, the example above for α gives α
e1 = (1, 4), α
e2 = (1, 6), etc. With such a
choice, the reduction process π described above consists in erasing all the common labels apart the first
one (label 1 is indeed the necessarily common bound for alignment). For instance, consider again a as
above, and a new configuration b = {A A H H H Q Q Q Q A Q Q} . The list for α is (1, 4, 6, 7, 10, 12),
the list for β = Φ(b) is (1, 3, 6, 10, 11), the list for σ = α ∧ β is (1, 6, 10). Then, the reduced α
b and
b
β are represented by (1, 4, 7, 12) and (1, 3, 11) respectively. Note that they do not correspond to any
new sequences, since the reduction is performed directly in Z(M), not in C(M). A graphic intuitive
representation of this reduction is given in Figure 5.
Acknowledgments
This work is partially supported by the MIUR FIRB grant RBFR08EKEV and by the INFN project
”Biological applications of theoretical physics methods”. The authors thank Roberto Burioni and Nicola
Clementi for useful discussion and suggestions.
References
1. Waterman M (1995) Introduction to computational biology. Chapman and Hall, CRC Press.
2. Pevzner P (2000) Computational Molecular Biology: An Algorithmic Approach. MIT Press, Cambridge MA.
3. Hampson A (2002) Influenza. C.W. Potter Eds., Elsevier, London.
4. Hoftand D, Belshe R (2004) The genetic archaeology of influenza. New England Journal of Medicine
351: 2550-2551.
5. Nelson M, Holmes E (2007) The evolution of epidemic influenza. Nat Rev Genet 8: 196-205.
6. Liao Y, Lee M, Ko C, Hsiung C (2008) Bioinformatics models for predicting antigenic variants of
influenza A/H3N2 virus. Bioinformatics 24: 505-512.
12
7. Hamming R (1950) Error detecting and error correcting codes. Bell System Technical Journal 29:
147-160.
8. Gupta V, Earl D, Deem M (2006) Quantifying influenza vaccine efficacy and antigenic distance.
Vaccine 24: 3881-3888.
9. He J, Deem M (2010) Low-dimensional clustering detects incipient dominant influenza strain clusters. Prot Eng, Des Sel 23: 935–946.
10. Otu H, Sayood K (2003) A new sequence distance measure for phylogenetic tree reconstruction.
Bioinformatics 19: 2122-2130.
11. Xia Z, Jin G, Zhu J, Zhou R (2009) Using a mutual information-based site transition network to
map the genetic evolution of influenza A/H3N2 virus. Bioinformatics 25: 2309-2317.
12. Ouyang Z, Zhu H, Wang J, She Z (2004) Multivariate entropy distance method for prokaryotic
gene identification. J Bioinform Comput Biol 2: 353-373.
13. Butte A, Kohane I (2000) Mutual information relevance networks: functional genomic clustering
using pairwise entropy measurements. Pac Symposium on Biocomputing 5: 415-426.
14. Rao S, Rodriguez A, Benson G (2005) Evaluating distance functions for clustering tandem repeats.
Genome Informatics 16: 3-12.
15. Baxevanis A, Ouellette B (2005) Bioinformatics: A Practical Guide to the Analysis of Genes and
Proteins. Wiley New York.
16. Hirst G (1943) Studies of antigenic differences among strains of influenza A by means of red cell
agglutination. J Exp Med 78: 407-423.
17. Smith D, Lapedes A, de Jong J, Bestebroer T, Rimmelzwaan G, et al. (2004) Mapping the antigenic
and genetic evolution of influenza virus. Science 305: 371-376.
18. Arnold V, Avez A (1967) Problèmes Ergodiques de la Mécanique Classique. Gauthier-Villars,
Paris.
19. Martin N, England J (1981) Mathematical Theory of Entropy. Addison-Wesley, Reading MA.
20. Sun J, Deem M (2007) Spontaneous emergence of modularity in a model of evolving individuals.
Physical Review Letters 99: 228107-11.
21. Wolf Y, Nikolskaya A, Cherry J, Viboud C, Koonin E, et al. (2010) Projection of seasonal influenza
severity from sequence and serological data. PLoS Curr 3;2: RRN1200.
22. Bao Y, Bolotov P, Dernovoy D, Kiryutin B, Zaslavsky L, et al. (2008) The influenza virus resource
at the national center for biotechnology information. J Virol 82: 596-601.
23. Hastie T, Tibshirani R, Friedman J (2009) The Elements of Statistical Learning. Springer, New
York.
24. Plotkin J, Dushoff J, Levin S (2002) Hemagglutinin sequence clusters and the antigenic evolution
of influenza A virus. Proc Natl Acad Sci USA 99: 6263-6268.
25. Khinchin A (1957) Mathematical Foundations of Information Theory. Dover, New York.
26. Morens D, Taubenberger J, Fauci A (2010) The 2009 H1N1 pandemic influenza virus: What next?
mBio 1(4): e00211.
13
27. Fitch W, Bush R, Bender C, Cox N (1997) Long term trends in the evolution of H(3) HA1 human
influenza type A. Proc Natl Acad Sci USA 94: 7712-7718.
28. Koelle K, Cobey S, Grenfell B, Pascual M (2006) Epochal evolution shapes the phylodynamics of
interpandemic influenza a (H3N2) in humans. Science 314: 1898-1903.
29. Pybus O, Rambaut A (2009) Evolutionary analysis of the dynamics of viral infectious disease. Nat
Rev Genet 10: 540-550.
30. Recker M, Pybus O, Nee S, Gupta S (2007) The generation of influenza outbreaks by a network
of host immune responses against a limited set of antigenic types. Proc Natl Acad Sci USA 104:
7711-7716.
31. Holmes E, Ghedin E, Miller N, Taylor J, Bao Y, et al. (2005) Whole-genome analysis of human
influenza A virus reveals multiple persistent lineages and reassortment among recent H3N2 viruses.
PLoS Biol 3(9): e300.
32. Data was obtained from the NIAID IRD online through the web site at http://www.fludb.org.
33. Agliari E, Casartelli M, Vivo E (2010) Metric characterization of cluster dynamics on the sierpinski
gasket. J Stat Mech : P09002.
34. Casartelli M, Dall’Asta L, Rastelli E, Regina S (2004) Metric features of a dipolar model. J Phys
A: Math Gen 37: 11731-11749.
14
Figure Legends
Figure 1. Looking for optimal p in clustering. Clustering entropy for Hamming and Rohlin
distance at different p values for H3N2. The plateau, in Rohlin, suggests an optimal and stable result
for the clustering.
15
Figure 2. Clusters and time evolution for H3N2. (upper part) Rohlin clusters time evolution for
H3N2. (lower part) Hamming clusters time evolution for H3N2. The reference WHO sequences are
shown by the corresponding symbols and names, and the details are indicated in Table 1. In the upper
part, we indicate the vaccine choice according to the WHO indication (up) and to buds criterion
(lower). Green and red colors indicate right and wrong choice with respect to the corresponding
analysis on the real circulating strain. A double color is used when more than one strain circulated in
that year and the corresponding prediction agrees with one of the circulating strain.
16
Figure 3. Clusters and time evolution for H1N1. (upper part) Rohlin clusters time evolution for
H1N1. (lower part) Hamming clusters time evolution for H1N1. The reference WHO sequences are
shown by the corresponding symbols, as indicated in Table 2. In the upper part, we indicate the vaccine
choice according to the WHO indication (up) and to buds criterion (lower). Green and red colors
indicate right and wrong choice. In this case, some of the symbols does not correspond to a specific HI
test, so they are indicated by a star and a pentagon. Notice the onset of the pandemic virus and the
failure of the bud criterion after that line.
17
Figure 4. Changing the time window. Rohlin analysis during years. Clusters structure is robust
under a sampling increase, and buds appeared during seasons correctly reveal the future circulating
strains, as indicated by the syringe.
18
Figure 5. Reduction. Two partitions α and β (derived from sequences as given in the text) and their
maximal common factor σ. Thick vertical lines individuate the atomic segments. Below, the reduced
b Colors in α and β remind the source configurations. There are no colors in σ, α
partitions α
b and β.
b and
b
β because these partitions do not originate from sequences of symbols but are directly defined in the
partition space.
19
Tables
Table 1. Symbols legend for Fig. 2:
#
3
D
2
7
F
8
I
A/Johannesburg/33/1994
A/Wuhan/359/1995
A/Sydney/5/1997
A/Panama/2007/1999
A/Moscow/10/1999
A/Wyoming/03/2003
A/New York/55/2004
A/Hiroshima/52/2005
A/Brisbane/10/2007
A/Uruguay/716/2007
A/Perth/16/2009
AY661180
AY661190
EF566075
DQ508865
DQ487341
CY034108
CY033638
EU283414
CY035022
EU716426
GQ293081
Table 2. Symbols legend for Fig. 3:
#
7
3
D
A/New Caledonia/20/1999
A/Solomon Islands/3/2006
A/Brisbane/59/2007
A/California/07/2009
CY033622
EU124177
CY058487
FJ969540
20
Supporting Information Legends
Figure S1: Looking for optimal p in clustering for H1N1. Clustering entropy for Rohlin and
Hamming at different p values for influenza A H1N1. The long plateau, in Rohlin, suggests a stable and
well defined value for the optimal p. Notice that Hamming is growing.
21
Figure S2: Looking for optimal p in clustering for H3N2 in the restricted time window.
Clustering entropy for Rohlin and Hamming at different p values for influenza A H3N2, as obtained by
considering only the sequences up to the end of the winter season of the year indicated in the plot. In
each time window, the long plateau, in Rohlin, suggests a stable and well defined value for the optimal
p. This figure is in correspondence with Fig. 4 of the main text.
22
Figure S3: Reverse analysis for Rohlin clusters. Sequences of minimum distance with the
corresponding WHO reference sequences, during years. The great similarity with Fig. 2 shows a strong
consistency between Rohlin and HI analysis.
Figure S4: Clustering on random permutations. Effect of random permutation of symbols on the
entropy of the clustering, as a function of p. R indicates the entropy of clustering with the Rohlin
distance and P stands for the entropy of clustering in the sample, obtained under a random
permutation of symbols in each sequence.
| 5cs.CE
|
Finding Largest Common Substructures of
Molecules in Quadratic Time?
Andre Droschinsky, Nils Kriege, and Petra Mutzel
arXiv:1610.08739v1 [cs.DS] 27 Oct 2016
Dept. of Computer Science, Technische Universität Dortmund, Germany
{andre.droschinsky,nils.kriege,petra.mutzel}@tu-dortmund.de
Abstract. Finding the common structural features of two molecules is
a fundamental task in cheminformatics. Most drugs are small molecules,
which can naturally be interpreted as graphs. Hence, the task is formalized as maximum common subgraph problem. Albeit the vast majority
of molecules yields outerplanar graphs this problem remains NP-hard.
We consider a variation of the problem of high practical relevance, where
the rings of molecules must not be broken, i.e., the block and bridge structure of the input graphs must be retained by the common subgraph. We
present an algorithm for finding a maximum common connected induced
subgraph of two given outerplanar graphs subject to this constraint. Our
approach runs in time O(∆n2 ) in outerplanar graphs on n vertices with
maximum degree ∆. This leads to a quadratic time complexity in molecular graphs, which have bounded degree. The experimental comparison
on synthetic and real-world datasets shows that our approach is highly
efficient in practice and outperforms comparable state-of-the-art algorithms.
1
Introduction
The maximum common subgraph problem arises in many application domains,
where it is necessary to elucidate common structural features of objects represented as graphs. In cheminformatics this problem has been extensively studied [11,4,12] and is often referred to as maximum or largest common substructure
problem. Two variants of the problem can be distinguished: The maximum common induced subgraph problem (MCIS) is to find isomorphic induced subgraphs
of two given graphs with the largest possible number of vertices. The maximum common edge subgraph problem (MCES) does not require that common
subgraphs are induced and aims at maximizing the number of edges. Both variants can be reduced to a maximum clique problem in the product graph of the
two input graphs [11]. In cheminformatics MCES is used more frequently since
it (i) reflects the notion of chemical similarity more adequately [11], and (ii)
can reduce the running time of product graph based algorithms [10]. Although
such algorithms still have exponential running time in the worst case, they are
commonly applied to molecular graphs in practice [11].
?
This work was supported by the German Research Foundation (DFG), priority programme “Algorithms for Big Data” (SPP 1736).
2
A. Droschinsky, N. Kriege, P. Mutzel
However, there are several restricted graph classes which render polynomial
time algorithms possible [1,2,14]. The seminal work in this direction is attributed
to J. Edmonds [9], who proposed a polynomial time algorithm for the maximum
common subtree problem. Here, the given graphs and the desired common subgraph must be trees. Recently, it was shown that this problem can be solved in
time O(∆n2 ) for (unrooted) trees on n vertices with maximum degree ∆ [3]. The
(induced) subgraph isomorphism problem (SI) is to decide if a pattern graph is
isomorphic to an (induced) subgraph of another graph and is generalized by
MCIS and MCES, respectively. Both variants of SI are NP-complete, even when
the pattern is a forest and the other graph a tree [5]; just as when the pattern is
a tree and the other is outerplanar [13]. On the other hand, when both graphs
are biconnected and outerplanar, induced SI can be solved in time O(n2 ) [13]
and SI in O(n3 ) [8]. These complexity results and the demand in cheminformatics lead to the consideration of MCES under the so-called block and bridge
preserving (BBP) constraint [12], which requires the common subgraph to retain
the local connectivity of the input graphs. BBP-MCES is not only computable
in polynomial-time, but also yields meaningful results for cheminformatics. A
polynomial-time algorithm was recently proposed for BBP-MCIS, which requires
time O(n6 ) in series-parallel and O(n5 ) in outerplanar graphs [6].
Most of the above mentioned polynomial time algorithms are either not applicable to molecular graphs or impractical due to high constants. A positive
exception is the BBP-MCES approach of [12], which has been shown to outperform state-of-the-art algorithms on molecular graphs in practice. This algorithm
is stated to have a running time of O(n2.5 ), but in fact leads to a running time
of Ω(n4 ) in the worst case [3].
Our contribution. We take up the concept of BBP and propose a novel BBPMCIS algorithm with running time O(∆n2 ) in outerplanar graphs with n vertices
and maximum degree ∆. We obtain this result by combining ideas of [3] for the
maximum common subtree problem with a new algorithm for biconnected MCIS
in biconnected outerplanar graphs. For this subproblem we develop a quadratic
time algorithm, which exploits the fact that the outerplanar embedding of a
biconnected outerplanar graph is unique. Moreover, the algorithm allows to list
all solutions in quadratic total time. Our approach supports to solve BBP-MCIS
w.r.t. a weight function on the mapped vertices and edges. The experiments show
that BBP-MCIS in almost all cases yields the same results as BBP-MCES for
molecular graphs under an adequate weight function. Our method outperforms
in terms of efficiency the BBP-MCES approach of [12] by orders magnitude.
2
Preliminaries
We consider simple undirected graphs. Let G = (V, E) be a graph, we refer to the
set of vertices V by V (G) or VG and to the set of edges by E(G) or EG . An edge
connecting two vertices u, v ∈ V is denoted by uv or vu. The order |G| of a graph
G is its number of vertices. Let V 0 ⊆ V , then the graph G[V 0 ] = (V 0 , E 0 ) with
Finding Largest Common Substructures of Molecules in Quadratic Time
3
E 0 = {uv ∈ E | u, v ∈ V 0 } is called induced subgraph. For U ⊆ V we write G \ U
for G[V \ U ]. A graph is connected if there is a path between any two vertices.
A connected component of a graph G is a maximal connected subgraph of G.
A graph G = (V, E) with |V | ≥ 3 is called biconnected if G \ {v} is connected
for any v ∈ V . A maximal biconnected subgraph of a graph G is called block. If
an edge uv is not contained in any block, the subgraph ({u, v}, {uv}) is called a
bridge. A vertex v of G is called cutvertex, if G \ {v} consists of more connected
components than G. A graph is planar if it admits a drawing on the plane such
that no two edges cross. The connected regions of the drawing enclosed by the
edges are called faces, the unbounded region is referred to as outer face. An edge
and a face are said to be incident if the edge touches the face. Two faces are
adjacent if they are incident with a common edge. A graph is called outerplanar
if it admits a drawing on the plane without crossings, in which every vertex
lies on the boundary of the outer face. A matching in a graph G = (V, E) is a
set of edges M ⊆ E, such that no two edges share a vertex. A matching M is
maximal if there is no other matching M 0 ) M and perfect, if 2|M | = |V |. A
weighted graph is a graph endowed with a function
w : E → R. A matching M in
P
a weighted graph has weight by W (M ) := e∈M w(e); it is a maximum weight
matching (MWM) if there is no matching M 0 of G with W (M 0 ) > W (M ).
An isomorphism between two graphs G and H is a bijection φ : V (G) →
V (H) such that uv ∈ E(G) ⇔ φ(u)φ(v) ∈ E(H). A common (induced) subgraph isomorphism is an isomorphism between (induced) subgraphs G0 ⊆ G
and H 0 ⊆ H. A subgraph G0 ⊆ G is block and bridge preserving (BBP) if (i)
each bridge in G0 is a bridge in G, (ii) any two edges in different blocks in
G0 are in different blocks in G. A common subgraph isomorphism φ is BBP
if both subgraphs are BBP, it is maximal if it cannot be extended. Molecular graphs are typically annotated with atom and bond types, which should be
preserved under isomorphisms. More general, we allow for a weight function
ω : (VG × VH ) ∪ (EG × EH ) → R≥0 ∪ {−∞}. The weight W(φ) of an isomorphism φ between G and H under ω is the sum of the weights ω(v, φ(v)) and
ω(uv, φ(v)φ(v)) for all vertices v and edges uv mapped by φ. A common subgraph isomorphism φ is maximum if its weight W(φ) is maximum. A maximum
isomorphism does not map any vertices or edges contributing weight −∞ and
we call these pairs forbidden. We further define [1..k] := {1, . . . , k} for k ∈ N.
3
Biconnected MCIS in Outerplanar Graphs
In this section we present an algorithm to determine the weight of a maximum
common biconnected induced subgraph isomorphism (2-MCIS) of two biconnected outerplanar graphs. First we show how to compute the maximal common
biconnected subgraph isomorphisms. Since these may contain forbidden vertex
and edge pairs, we then describe how to obtain the weight of a maximum solution
from them. Finally we show how to output one or all maximum solutions.
Outerplanar graphs are well-studied and have several characteristic properties, see [13] for further information. In particular, our algorithm exploits the fact
4
A. Droschinsky, N. Kriege, P. Mutzel
that biconnected outerplanar graphs have a unique outerplanar embedding in
the plane (up to the mirror image). In these embeddings, every edge is incident
to exactly two faces that are uniquely defined. We observe that the mapping is
determined by starting parameters, i.e., an edge of both input graphs together
with the mapping of their endpoints and incident faces.
We say a face is mapped by an isomorphism φ if all the vertices bordering
the face are mapped by φ. We distinguish four cases to describe the mapping
of an edge uv ∈ E(G) to an edge u0 v 0 ∈ E(H) by an isomorphism φ between
biconnected induced subgraphs. Assume the edge uv is incident to the faces A
and B in G and u0 v 0 is incident to A0 and B 0 in H, see Fig. 1(a). At least
one face incident to uv must be mapped by φ, since the common subgraph
must be biconnected. For the sake of simplicity of the case distinction, we also
associate the two other faces, regardless of whether they are mapped or not. The
isomorphism may map the endpoints of the edges in two different ways—just as
the two incident faces. We can distinguish the following four cases:
(1) u 7→ u0 , v 7→ v 0 , A 7→ A0 , B 7→ B 0 , (2) u 7→ v 0 , v 7→ u0 , A 7→ A0 , B 7→ B 0 ,
(3) u 7→ u0 , v 7→ v 0 , A 7→ B 0 , B 7→ A0 , (4) u 7→ v 0 , v 7→ u0 , A 7→ B 0 , B 7→ A0 .
Given an isomorphism φ between biconnected common induced subgraphs that
maps the two endpoints of an edge e, let the function type(e, φ) ∈ [1..4] determine
the type of the mapping as above. The following result is the key to obtain our
efficient algorithm.
Lemma 1. Let φ and φ0 be maximal isomorphisms between biconnected common
induced subgraphs of the biconnected outerplanar graphs G and H. Assume e ∈
E(G) is mapped to the same edge e0 ∈ E(H) by φ and φ0 , then
type(e, φ) = type(e, φ0 ) ⇐⇒ φ0 = φ.
Proof. It is obvious that the direction ⇐= is correct. We prove the implication
=⇒. Since the common subgraph is required to be biconnected, the isomorphisms
φ and φ0 both must map at least one face of G incident to the edge e to a face
of H incident to e0 . The two faces as well as the mapping of endpoints of the
two edges are uniquely determined by the type of the mapping. We consider the
mapping of the vertices on the cyclic border of these faces. Since the mapping
of the endpoints of e are fixed, the mapping of all vertices on the border of the
face is unambiguously determined. Since the common subgraph is required to be
biconnected, every extension of the mapping must include all the vertices of a
neighboring face. For this face, again, the mapping of the endpoints of the shared
edge implicates the mapping of all vertices on the cyclic border and the extension
is unambiguous. Therefore, the mapping can be successively extended to an
unmapped face. Consequently φ(u) = φ0 (u) holds for all u ∈ dom(φ) ∩ dom(φ0 ).
Since φ and φ0 are maximal it is not possible that one of them can be extended
and, hence, we must have dom(φ) = dom(φ0 ) and the result follows.
t
u
The proof of Lemma 1 constructively shows how to obtain a maximal solution
given two edges uv ∈ E(G), u0 v 0 ∈ E(H) and a type parameter t ∈ [1..4]. We
assume that this approach is realized by the procedure MaximalIso(uv, u0 v 0 , t),
Finding Largest Common Substructures of Molecules in Quadratic Time
5
Algorithm 1: 2-MCIS in outerplanar graphs
1
2
3
4
5
6
7
Input
: Biconnected outerplanar graphs G and H.
Output : Weight of a maximum common biconnected subgraph isomorphism.
Data
: Table D(e, f, t), e ∈ E(G), f ∈ E(H), t ∈ [1..4] storing the weight of
a 2-MCIS φ mapping e to f with type(e, φ) = t.
forall the uv ∈ E(G), u0 v 0 ∈ E(H) and t ∈ [1..4] do
if type t valid for uv and u0 v 0 and D(uv, u0 v 0 , t) undefined then
φ ← MaximalIso(uv, u0 v 0 , t)
(φ1 , . . . , φk ) ← SplitIso(φ)
forall the edges e ∈ E(G)
mapped to f ∈ E(H) by φ do
W (φe ) if e is mapped by the split iso. φe
D(e, f, type(e, φ)) ←
−∞
otherwise.
return maximum entry in D
which returns the unique maximal isomorphism that maps the two given edges
according to the specified type. The algorithm can be implemented by means
of a tree structure that encodes the neighboring relation between inner faces,
e.g., SP-trees as in [6,7] or weak dual graphs similar to the approach of [13].
The running time to compute a maximal solution φ then is O(|φ|) ⊆ O(n). Note
that for some edge pairs not all four types of mappings are possible. The type
t ∈ [1..4] is valid for a pair of edges if at least one incident face can be mapped
according to type t, i.e., the edges are incident to faces that are bordered by the
same number of vertices.
A maximal solution φ may map vertex and edge pairs that are forbidden
according to the weight function. In order to obtain the maximum weight, we
split φ into split isomorphisms φ1 , . . . , φk such that each (i) has non-negative
weight and (ii) again is an isomorphism between biconnected induced common
subgraphs. The split isomorphisms can be obtained in time O(|φ|) as follows. We
consider the graph G0 = G[dom(φ)]. For every forbidden edge uv that is incident
to two inner faces in G0 , we split the graph into G0i [V (Ci ) ∪ {u, v}], where Ci
is a connected component of G0 \ {u, v}, i ∈ [1..2]. In these graphs we delete
the forbidden vertices and edges and determine the blocks B1 , . . . , Bk . Then φ,
restricted to the vertices V (Bi ) of a block Bi , yields the split isomorphism φi
for i ∈ [1..k]. This approach is realized by the function SplitIso(φ) used in
the following. Every edge e ∈ E(G) is mapped by at most one of the resulting
isomorphisms, referred to by φe . Every 2-MCIS is a split isomorphism obtained
from some maximal solution.
Algorithm 1 uses a table D(e, f, t), e ∈ E(G), f ∈ E(H), t ∈ [1..4] storing
the weight of a 2-MCIS under the constraint that it maps e to f according to
type t. The size of the table is 4|E(G)||E(H)| ∈ O(nm), where n = |V (G)|
and m = |V (H)|. The algorithm starts with all pairs of edges and all valid
types of mappings between them. For each, the maximal isomorphism between
biconnected common induced subgraphs is computed by extending this initial
mapping. By splitting the maximal solution, multiple valid isomorphisms with
6
A. Droschinsky, N. Kriege, P. Mutzel
non-negative weight are obtained. These weights are then stored in D for all
pairs of edges contained in φ considering the type of the mapping. This includes
the −∞ weights occurring if there are forbidden vertices or edges. Keeping these
values allows to avoid generating the same isomorphism multiple times. The
main procedure loops over all pairs of edges and the four possible mappings for
each pair. Note that a mapping φ and its split isomorphisms are computed in
time O(|φ|) ⊆ O(n). Improved analysis gives the following result.
Theorem 1. Algorithm 1 computes the weight of a 2-MCIS between biconnected
outerplanar graphs G and H in time O(|G||H|).
Proof. We allocate the costs for a call of MaximalIso followed by SplitIso
to cells of the table D. A mapping φ containing k edges is computed in time
O(k) and as a result exactly k cells of the table D are filled with a value. The
value of a cell is computed at most once: Line 2 assures that an edge mapping
of a specific type is not used as initial mapping when the corresponding cell is
already filled. Every initial mapping that is extended must lead to an isomorphism containing only edge mappings associated with undefined cells according
to Lemma 1. Therefore the total costs of the algorithm can be allocated to cells
of D, such that each cell pays a constant amount. This proves that the total
running time is bounded by the size of the table, which is O(|G||H|).
t
u
We can easily modify the algorithm to enumerate all maximum isomorphisms
without affecting the total running time. First we run Algorithm 1 once to obtain
the maximum weight Wmax . Then we run a modified version of Algorithm 1 that
outputs every split isomorphism φi of size W (φi ) = Wmax as soon as it is found,
right after SplitIso(φ) is called in line 4.
4
Solving BBP-MCIS in Outerplanar Graphs
In the previous section we have presented an algorithm to compute a 2-MCIS
between two biconnected outerplanar graphs. In this section we will generalize
it to compute a BBP-MCIS between two outerplanar graphs G and H. In the
following we assume the isomorphisms to be BBP. We require the input graphs
to be connected. Otherwise we compute a BBP-MCIS for all pairs of connected
components and select an isomorphism of maximum weight.
We proceed as follows. First, we give insight into the BC-tree data structure,
which helps to partition the set S of all BBP common subgraph isomorphisms
between G and H into subsets w.r.t. certain conditions. Then we compute an
isomorphism of maximum weight in each of the subsets using a dynamic programming approach similar to the one used in [3] to solve the maximum common
subtree problem. Among the computed isomorphisms we output one with maximum weight, thus a BBP-MCIS.
The BC-tree data structure. Given a BBP-MCIS, we can observe that
bridges of G are mapped to bridges of H and that edges in one block of G
Finding Largest Common Substructures of Molecules in Quadratic Time
c1
7
c2
c2
c1
A
u
v
c1
b3
c2
u
B
c3
(a)
(b) Input graph G
v
c2
b1
c1
b2
c2
u
c3
c3
b4
(c) BC-tree BC
v
c3
b5
G
Fig. 1. A biconnected outerplanar graph (a) with an edge uv incident to the faces A
and B; a connected outerplanar graph (b) and its BC-tree (c). Block nodes have a gray
background, while bridge nodes are not filled. The solid black nodes are the cutvertices.
The corresponding subgraphs of G are shown above the block and bridge nodes.
can only be mapped to edges contained in exactly one block of H, such that the
mapped edges form a biconnected common subgraph. For a connected graph G
let CG denote the set of cutvertices, BlG the set of blocks and BrG the set of
bridges and BG := BlG ∪ BrG . The BC-tree BCG of G is the tree with nodes
BG ∪ CG and edges between nodes b ∈ BG and c ∈ CG iff c ∈ V (b). We refer
to the vertices of the BC-tree as B- and C-nodes and distinguish block nodes
from bridge nodes. An example of a graph G and its BC-tree BCG is shown in
Fig. 1. For any graph G, we define CC(V 0 , U ) as the connected component of
G[V 0 ] that includes at least one vertex of U . We allow only such sets U , where
the component is unambiguous. For example, in Fig. 1, CC(VG \ Vb2 , Vb4 ) is the
graph G[{c3 , u, v}].
S
Partitioning of all BBP isomorphisms S into S = x S x . First, we
define S 1 and S 2 . Let b ∈ BG be an arbitrary block or bridge in G. We define
S 1 to contain all isomorphisms φ where at least one edge in b is mapped by the
isomorphisms, i.e., | dom(φ)∩V (b)| ≥ 2. S 2 is defined to contain all isomorphisms
where exactly one vertex in b is mapped by the isomorphism. We can observe
that S1 and S2 are disjoint and all other isomorphisms between G and H do
not contain any vertices of b. Let N = {b1 , . . . , bk } ⊆ BG be the blocks and
bridges that share a cutvertex with b, i.e., bi ∈ N iff there is a node c ∈ CG with
bc and cbi edges in the BC-tree BCG . Any isomorphism φ that maps no vertex
of b, maps vertices of at most one node bi , because G[dom(φ)] is connected by
definition. For every bi we recursively define sets S x of isomorphisms as described
above that map only vertices of CC(VG \ Vb , Vbi ).
As example consider Fig. 1(c) and let b := b2 . S 1 consist of isomorphisms
which map at least one edge of b2 to an edge in H. The isomorphisms in S 2 map
exactly one vertex of V (b) to H. The recursion continues on N = {b1 , b3 , b4 }.
Three additional sets consist of isomorphisms which map at least one edge (and
three more for exactly one vertex) of V (bi ), i ∈ {1, 3, 4}, but no vertex of V (b2 ),
operating on CC(VG \Vb2 , Vbi ). The recursion for b := b4 continues with N = {b5 }
and two additional sets. Some of the sets S x are empty.
8
A. Droschinsky, N. Kriege, P. Mutzel
S
Partitioning of S x into S x = y Pxy . Before computing an isomorphism
of maximum weight in a set S x , we partition S x into subsets Px1 , Px2 , . . .. The
focus for the separation now is on the graph H. We distinguish two cases. If S x
is a set, where at least one edge of a certain block (bridge) b is mapped, then S x
is partitioned into |BlH | (|BrH |) subsets. The meaning is that for each B-node
b̄ ∈ BlH (b̄ ∈ BrH ) the mapped vertices of the B-node b ∈ B G are mapped only
to V (b̄). In terms of BBP this is block (bridge) preserving between b and b̄, as
intended. If S x is a set, where exactly one vertex of b is mapped, the subsets are
defined as follows. For each (v, v̄) ∈ V (b) × V (H), where ω(vv̄) 6= −∞ and v is
in the CC we operate on, we define a subset with the restriction φ(v) = v̄.
Computing a maximum isomorphism in a subset Pxy . We now describe
how to compute an isomorphism φ of maximum weight in a subset Pxy ⊆ S x .
The idea is to recursively extend mappings between some vertices of two single
bridges or two single blocks along all pairs of mapped cutvertices into other Bnodes determined by MWMs, while preserving bridges and blocks. Between the
computed isomorphisms we select one of maximum weight.
First, let Pxy be a subset, where at least one edge of a B-node b ∈ BG has
to be mapped to an edge of a B-node b̄ ∈ BH . If b and b̄ are bridges, the two
possible mappings V (b) → V (b̄) are considered. If both are blocks, all maximal
common biconnected subgraph isomorphisms between the blocks are considered
(cf. Alg. 1). We may have given a fixed mapping v 7→ v̄ (cf. (i) below). We
call a considered isomorphism valid, if it respects the possible fixed mapping
and contains only vertices of the CC we are operating on. We extend all the
valid isomorphisms φ along all pairs φ(c) = c̄, c 6= v of mapped cutvertices as
follows. Let Bc := {b1 , . . . bk }, be the B-nodes of BG , where bcbi is a path, and
B̄c := {b̄1 , . . . b̄l }, be the B-nodes of BH , where b̄c̄b̄j is a path, i ∈ [1..k], j ∈ [1..l].
For each pair (bi , b̄j ) ∈ Bc × B̄c we recursively calculate a BBP-MCIS ϕij under
the following restrictions: (i) The cutvertices must be mapped: c 7→ c̄. (ii) bi
and b̄j are both bridges or both blocks. (iii) At least one other vertex in the
block (bridge) bi must be mapped, but only to V (b̄j ). Restriction (iii) assures
that at least one vertex is added to the isomorphism. Therefore, the recursion to
compute ϕij is the method described in this paragraph. After computing ϕij for
each pair (bi , b̄j ), we construct a weighted bipartite graph with vertices Bc ] B̄c
for each pair of mapped cutvertices. The weight of each edge bi b̄j is determined
by the weight of a BBP-MCIS under the above restrictions, subtracted by ω(c, c̄)
for the appropriate cutvertices c and c̄. If there in no such restricted BBP-MCIS,
there is no edge. Computing a MWM on each of the bipartite graphs determines
the extension of φ. For each matching edge the corresponding computed isomorphisms are merged with φ. After extending all valid isomorphisms, we select one
of maximum weight.
Second, let Pxy be a subset, where exactly one vertex v of V (b) is mapped,
and let φ(v) = v̄. If v is no cutvertex, the only possible expansion is within V (b),
which is not allowed in this subset. Therefore this subset contains exactly one
isomorphism, v 7→ v̄. Next, assume v is a cutvertex. If v̄ is a cutvertex, we may
Finding Largest Common Substructures of Molecules in Quadratic Time
9
extend φ similar to the previous paragraph. In doing so, c := v, c̄ := v̄ and Bc
as before. The only difference is B̄c , which is defined by all B-nodes containing
v̄ = c̄. The reason is that we have not mapped any other vertices yet, therefore
we may expand in all directions in H. If v̄ is no cutvertex, then v̄ is contained in
exactly one b̄ ∈ B H . We are interested in BBP isomorphisms only. This means,
all vertices that are mapped to V (b̄) must be in the same block or bridge b0 ∈ BG .
Therefore, for each b0 ∈ BG , where bvb0 is a path and b0 and b̄ are of the same
type (bridge/block), we compute an isomorphism with fixed mapping v 7→ v̄,
where at least one edge of b0 is mapped to b̄. This falls back to the method
of the above paragraph as well. Among the computed isomorphisms we select
one of maximum weight. The appendix lists the pseudocode for computing a
BBP-MCIS as described above.
Time Complexity. The time to compute a BBP-MCIS essentially depends
on the time to compute the BC-trees, the biconnected isomorphisms between
the blocks of G and H, and the time to compute all the MWMs. The time to
compute a BC-tree is linear in the number of edges and vertices. Considering
all pairs of blocks and Theorem 1PweP
can bound the time for computing all the
biconnected isomorphisms by O( b b̄ |Vb ||Vb̄ |) ⊆ O(|G||H|). We only need to
compute MWMs for the pairs of cutvertices of the two graphs. It follows from
the result of [3, Theorem 7] for the maximum common subtree problem, that the
total time for this is O(|G||H|(min{∆G , ∆H } + log max{∆G , ∆H })), where ∆G
is the maximum degree of a C-node in BCG . This proves the following theorem.
Theorem 2. BBP-MCIS between two outerplanar graphs G and H can be solved
in time O(|G||H|∆(G, H)), where ∆(G, H) = 1 iff G or H is biconnected or both
are of bounded degree; otherwise ∆(G, H) = min{∆G , ∆H } + log max{∆G , ∆H }.
5
Experimental Evaluation
In this section we evaluate our BBP-MCIS algorithm experimentally and compare to the BBP-MCES approach of [12].1 Both algorithms were implemented
in C++ and compiled with GCC v.4.8.4 as 64-bit application. Running times
were measured on an Intel Core i7-3770 CPU using a single core. The available
memory of 16 GB was sufficient for all the computations.
We are interested in answering the following questions:
(H1) To what extent differs BBP-MCIS from BBP-MCES on molecular graphs?
(H2) How large is the difference in terms of running time on molecular graphs?
(H3) How is the running time affected by specific properties of the input graphs?
To answer (H1) and (H2) we extracted 29000 randomly chosen pairs of
outerplanar molecular graphs from a large chemical database.2 The molecules
1
2
We are grateful to Leander Schietgat for providing the implementation used in [12].
NCI Open Database, GI50, http://cactus.nci.nih.gov
10
A. Droschinsky, N. Kriege, P. Mutzel
Fig. 2. Running times in ms for 28 399 BBP-MCIS computations. Each black dot
represents a BBP-MCIS computation on two randomly chosen outerplanar molecular
graphs. It directly compares the running time of our algorithm (MCIS, x-axis) and
the implementation from [12] (MCES, y-axis). The running times of another 601 BBPMCIS computations did not fit into the borders.
in the database contain up to 104 vertices and 22 vertices on an average. The
weight function ω was set to 1 for each pair of vertices and edges with the same
label and −∞ otherwise. This matches the setting in [12].
To answer (H3) we compared the algorithms on randomly generated connected outerplanar graphs. Our graph generator takes several parameters as
input. With them we evaluated three different properties: the graph size, the
average ratio |E|/|V | of edges to vertices, and the average block size. For any
outerplanar graphs the ratio of edges to vertices is less than 2. While evaluating
the effect of one property, we preserved the other two. This procedure allows
to verify whether our theoretical findings are consistent with the running times
observed in practice. We set the weight function ω to 1 for each pair of vertices
and edges, which corresponds to uniformly labeled graphs.
(H1) While comparing the weight of the isomorphisms computed by the two
algorithms we observed a difference for only 0.40% of the 29 000 tested molecule
pairs. This suggests that BBP-MCIS yields a valid notion of similarity for outerplanar molecular graphs as it was shown for BBP-MCES [12].
(H2) Our algorithm computed the solutions on average 84 times faster. The dots
in Fig. 2 represent the computation times of the two algorithms. The results
are summarized in Table 1. Schietgat et al. [12] compared their BBP-MCES
Finding Largest Common Substructures of Molecules in Quadratic Time
11
Table 1. Upper half: Running times for our implementation (MCIS) and the implementation from [12] (MCES). Lower half: Relative differences in computation times.
Algorithm
MCIS
MCES
Average time Median time
1.97 ms
1.51 ms
207.08 ms
41.43 ms
95% less than Maximum time
5.28 ms
40.35 ms
871.48 ms
26 353.68 ms
Comparison Average factor Median factor Minimum factor Maximum factor
MCES / MCIS
83.8
25.6
1.8
28912.5
Table 2. Average time ± SD over 100 BBP-MCIS computations on random outerplanar graphs, varying one property (graph size, ratio of edges to vertices, block size BS).
Note the units of measurement; timeout—total time exceeds 3 days.
Size
10
20
40
80
160
MCIS 0.7 ± 0.3 ms 2.3 ± 0.8 ms 8.2 ± 1.6 ms 33.5 ± 3.6 ms 133.2 ± 10.1 ms
MCES 207 ± 118 ms
3.4 ± 6.0 s 38.6 ± 90.6 s 234.2 ± 420.9 s
timeout
|E|/|V |
0.98
1.10
1.24
1.46
1.78
MCIS 3.8 ± 0.3 ms 4.0 ± 1.1 ms 8.2 ± 1.6 ms 30.8 ± 4.0 ms 110.3 ± 11.6 ms
MCES 223 ± 16 ms
2.2 ± 2.6 s 38.6 ± 90.6 s 111.0 ± 213.8 s 216.1 ± 288.3 s
BS
MCIS
MCES
3
5
10
20
27 ± 6.4 ms 13.3 ± 2.4 ms 8.4 ± 1.7 ms
5.5 ± 1.4 ms
132 ± 14 ms 689 ± 548 ms 83.7 ± 118.7 s 30.4 ± 27.8 min
40
4.5 ± 0.9 ms
timeout
algorithm to a state-of-the-art algorithm for general MCIS. Their algorithm had
similar computation times on small graphs and was much faster on large graphs.
The maximum time of the general MCIS algorithm was more than 24 hours.
In contrast, our computation time never exceeded 41 ms. This clearly indicates
that our algorithm is orders of magnitude faster than the general approach.
(H3) We first varied the size of the input graphs, while preserving an average
ratio of edges to vertices of 1.24 and an average block size of 8. Based on Theorem 2 we expected the average time to increase by a factor of a bit more than
4, if we double the size. The results in Table 2 closely match this expectation.
Next, we evaluated different ratios of edges to vertices. The graph size was set
to 40 and the average block size to 8. A higher ratio results in a higher number
of faces in the blocks and consequently affects the time required by Alg. 1. In
particular, the table size and, thus, the running time is expected to show a
quadratic growth. The increase in running time exceeds our expectation. This
might be explained by the increasing size of the data structure used to represent
the faces of the blocks.
Finally, we evaluated different average block sizes. The graph size was set to
40 and the average ratio of edges to vertices to 1.24. Higher block sizes mean less
MWMs to compute, which are the most costly part in the BBP-MCIS computation. Therefore we expected the running time to decrease. The results shown
in Table 2 support this.
12
6
A. Droschinsky, N. Kriege, P. Mutzel
Conclusion
We have developed an algorithm, which computes a well-defined, chemical meaningful largest common substructure of outerplanar molecular graphs in a fraction of a second. Hence, our method makes the graph-based comparison in large
molecular datasets possible. As future work, we would like to extend our approach to more general graph classes with a focus on efficiency in practice.
References
1. Akutsu, T.: A polynomial time algorithm for finding a largest common subgraph
of almost trees of bounded degree. IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences E76-A(9) (1993)
2. Akutsu, T., Tamura, T.: A polynomial-time algorithm for computing the maximum common connected edge subgraph of outerplanar graphs of bounded degree.
Algorithms 6(1), 119–135 (2013)
3. Droschinsky, A., Kriege, N., Mutzel, P.: Faster Algorithms for the Maximum Common Subtree Isomorphism Problem. In: MFCS 2016. LIPIcs, vol. 58, pp. 34:1–34:14
(2016), arXiv:1602.07210
4. Ehrlich, H.C., Rarey, M.: Maximum common subgraph isomorphism algorithms
and their applications in molecular science: a review. Wiley Interdisciplinary Reviews: Computational Molecular Science 1(1), 68–79 (2011)
5. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory
of NP-Completeness. W. H. Freeman (1979)
6. Kriege, N., Kurpicz, F., Mutzel, P.: On maximum common subgraph problems in
series-parallel graphs. In: Jan, K., Miller, M., Froncek, D. (eds.) IWOCA 2014.
LNCS, vol. 8986, pp. 200–212. Springer (2014)
7. Kriege, N., Mutzel, P.: Finding maximum common biconnected subgraphs in seriesparallel graphs. In: Csuhaj-Varjú, E., Dietzfelbinger, M., Ésik, Z. (eds.) MFCS
2014, LNCS, vol. 8635, pp. 505–516. Springer (2014)
8. Lingas, A.: Subgraph isomorphism for biconnected outerplanar graphs in cubic
time. Theoretical Computer Science 63(3), 295 – 302 (1989)
9. Matula, D.W.: Subtree isomorphism in O(n5/2 ). In: B. Alspach, P.H., Miller, D.
(eds.) Algorithmic Aspects of Combinatorics, Annals of Discrete Mathematics,
vol. 2, pp. 91–106. Elsevier (1978)
10. Nicholson, V., Tsai, C.C., Johnson, M., Naim, M.: A subgraph isomorphism theorem for molecular graphs. In: Graph Theory and Topology in Chemistry, pp.
226–230. No. 51 in Stud. Phys. Theoret. Chem., Elsevier (1987)
11. Raymond, J.W., Willett, P.: Maximum common subgraph isomorphism algorithms
for the matching of chemical structures. Journal of Computer-Aided Molecular
Design 16(7), 521–533 (2002)
12. Schietgat, L., Ramon, J., Bruynooghe, M.: A polynomial-time maximum common
subgraph algorithm for outerplanar graphs and its application to chemoinformatics.
Annals of Mathematics and Artificial Intelligence 69(4), 343–376 (2013)
13. Syslo, M.M.: The subgraph isomorphism problem for outerplanar graphs. Theoretical Computer Science 17(1), 91 – 97 (1982)
14. Yamaguchi, A., Aoki, K.F., Mamitsuka, H.: Finding the maximum common subgraph of a partial k-tree and a graph with a polynomially bounded number of
spanning trees. Inf. Process. Lett. 92(2), 57–63 (2004)
Finding Largest Common Substructures of Molecules in Quadratic Time
A
13
Pseudocode
Algorithm 2: BBP-MCIS of outerplanar graphs
Input
1
2
3
1
2
3
4
5
6
7
8
: Connected outerplanar graphs G and H,
weight function ω : (VG × VH ) ∪ (EG × EH ) → R≥0 ∪ {−∞}.
Output : A BBP-MCIS of G and H.
Data
: BC-trees BCG and BCH with node sets B G , C G , B H , C H .
Select an arbitrary B-node (block or bridge) b ∈ B G .
φ ← SetSx(b, ∅)
. Initial recursion call.
[Output φ]
Procedure SetSx(b, X)
Input
: B-node b ∈ B G , excluded vertices X ⊆ V (G).
Output : Isomorphism of maximum weight on CC(VG \ X, Vb ).
ϕ ← (∅ → ∅)
. Initialize as empty mapping
forall the B-nodes b̄ ∈ B H , where b and b̄ are both bridges or both blocks do
φ ← BBP-Edge(b, b̄, X)
. Pxy ⊆ S x , at least one edge is mapped
if W(φ) > W(ϕ) then ϕ ← φ
forall the pairs (v, v̄) ∈ (V (b) \ X) × V (H) do
if ω(vv̄) 6= −∞ then
φ ← BBP-SingleVertex(b, X, v, v̄)
if W(φ) > W(ϕ) then ϕ ← φ
. Pxy ⊆ S x , single vertex
11
forall the paths bcb0 in BCG , where c ∈
/ X do
φ ← SetSx(b0 , V (b))
. No vertex of V (b) is mapped, recursion
if W(φ) > W(ϕ) then ϕ ← φ
12
return ϕ
9
10
14
1
2
3
4
5
6
7
8
9
10
11
1
2
3
4
5
6
7
8
9
A. Droschinsky, N. Kriege, P. Mutzel
Procedure BBP-Edge(b, b̄, X, v, v̄)
Input
: B-nodes b ∈ B G , b̄ ∈ B H , X ⊆ V (G), mapping v 7→ v̄ (optional).
Output : Maximum isomorphism ϕ, where at least one edge of b is mapped
to b̄; restricted to CC(VG \ X, Vb ) and ϕ(v) = v̄ (if given).
if exactly one of b, b̄ is a block then
return ∅ → ∅
. not block and bridge preserving
forall the valid isomorphisms ϕ : V (b) → V (b̄) do
forall the pairs (c, c̄) 6= (v, v̄) of cutvertices mapped by ϕ do
forall the pairs (bi , b̄j ) ∈ BG × BH where bcbi and b̄c̄b̄j are paths do
w(bi , b̄j ) ← W(BBP-Edge(bi , b̄j , X, c, c̄)) − ω(c, c̄)
Compute MWM M on bipartite graph with edge weights w(bi , b̄j )
forall the edges bi b̄j ∈ M do
Extend ϕ by BBP-Edge(bi , b̄j , X, c, c̄)
if W(φ) > W(ϕ) then ϕ ← φ
return ϕ
Procedure BBP-SingleVertex(b, X, v, v̄)
Input
: B-node b ∈ B G , excluded vertices X ⊆ V (G), mapping v 7→ v̄.
Output : Maximum isomorphism ϕ, where a single vertex of V (b) is
mapped to V (b̄); restricted to CC(VG \ X, Vb ), ϕ(v) = v̄.
ϕ ← (v 7→ v̄)
if v ∈
/ C G then
return ϕ
. No expansion possible
if v̄ ∈ CH then
forall the pairs (bi , b̄j ) ∈ BG × BH , where bvbi is a path and v̄ ∈ b̄j do
w(bi , b̄j ) ← W(BBP-Edge(bi , b̄j , X, v, v̄)) − ω(v, v̄)
Compute MWM M on bipartite graph with edge weights w(bi , b̄j )
forall the edges bi b̄j ∈ M do
Extend ϕ by BBP-Edge(bi , b̄j , X, v, v̄)
14
else
b̄ ← the unambiguous node of the set BH that contains the vertex v̄
forall the bi ∈ BG , where bvbi is a path do
φ ← BBP-Edge(bi , b̄, X, v, v̄)
if W(φ) > W(ϕ) then ϕ ← φ
15
return ϕ
10
11
12
13
| 8cs.DS
|
National Laboratory for Scientific Computing
arXiv:1501.05290v2 [cs.DB] 12 Feb 2015
Graduate Program in Computational Modeling
Managing large-scale scientific hypotheses as uncertain
and probabilistic data
By
Bernardo Gonçalves
PETRÓPOLIS, RJ - BRASIL
FEBRUARY - 2015
MANAGING LARGE-SCALE SCIENTIFIC HYPOTHESES AS
UNCERTAIN AND PROBABILISTIC DATA
Bernardo Gonçalves
THESIS SUBMITTED TO THE EXAMINING COMMITTEE IN PARTIAL
FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF SCIENCES IN COMPUTATIONAL MODELING.
Approved by:
Prof. Fabio Porto, D.Sc.
(Chair)
Prof. Pedro L. Dias, Ph.D.
Prof. Marco A. Casanova, Ph.D.
Prof. Ana Carolina Salgado, D.Sc.
PETRÓPOLIS, RJ - BRASIL
FEBRUARY - 2015
© 2015, Bernardo Nunes Gonçalves.
All rights reserved.
Gonçalves, Bernardo
G635m
Managing large-scale scientific hypotheses as uncertain and probabilistic data / Bernardo Gonçalves. Petropólis, RJ. : National Laboratory for
Scientific Computing, 2015.
xvii, 128p. : il.; 29 cm
Orientador: Fabio Porto
Thesis (D.Sc.) – National Laboratory for Scientific Computing, 2015.
1. Hypothesis management. 2. Predictive analytics. 3. Uncertain and
probabilistic data. 4. Causal reasoning. 5. Probabilistic database design.
Porto, Fabio. II. LNCC/MCT. III. Tı́tulo.
CDD – 629.8
I.
Dedicatory
To my parents Tania and Francisco,
and to my special love, Marcelle,
for being an island of certainty in an
uncertain world.
Acknowledgments
This thesis work has been supported by LNCC (Graduate Program in Computational Modeling), CNPq (grant 141838/2011-6), FAPERJ (grant ‘Nota 10’
E-26/100.286/2013) and IBM (Ph.D. Fellowship 2013/2014).
I would like to express my gratitude to my advisor Fabio Porto for the gift
of the challenging topic of this thesis, so special to me. I am grateful to him
for being an inspiring advisor, and for nicely influencing me towards database
research. I also thank my thesis committee for their attention and time devoted in
the assessment of my work. I am indebted to Ana Maria Moura for her generous
advice and support throughout my Ph.D. research, and to Frederico C. Silva and
Adolfo Simões for their support to my research project at DEXL/LNCC.
I would like to thank very much all researchers at LNCC who give lectures in
the graduate program, specially Prof. José Karam Filho for teaching me generously
about the roots of mathematical modeling. They have contributed significantly to
my education as a cross-disciplinary thinker and the shaping of my scientific and
mathematical skills. I thank my Ph.D. colleagues and friends at LNCC, specially
Eduardo Lima, Ramon Costa, Klaus Wehmuth, Raquel Lopes, Karine Guimarães
and Diego Paredes for their companionship and joy shared in the pursuit of their
Ph.D. theses. I would also like to gratefully recall my earlier professors at UFES,
José Gonçalves, Giancarlo Guizzardi, Rosane Caruso and Berilhes Garcia for shaping the most essential building blocks in my education.
Finally, I would like to thank my family for their support: my father Francisco, my example of simplicity and goodness; my mother, Tania, for her tenacity
and true love; Leo and Sche, my dear brother and sister, true union for life; my
grandmothers Zeca, Orizontina and Rosita for their love and prayers. I thank also
Marcel and Maris for the greatest gift, their daughter and my near future wife,
Marcelle – the best partner one could ever hope for.
“Originally, there was just experimental science, and
then there was theoretical science, with Kepler’s
Laws, Newton’s Laws of Motion, Maxwell’s equations,
and so on. Then, for many problems, the theoretical models grew too complicated to solve analytically,
and people had to start simulating. These simulations
have carried us through much of the last half of the
last century. At this point, these simulations are generating a whole lot of data, along with a huge increase
in data from the experimental sciences.”
— Jim Gray, 2007
Abstract of Thesis presented to LNCC/MCT in partial fulfillment of the requirements for the degree of Doctor of Sciences (D.Sc.)
MANAGING LARGE-SCALE SCIENTIFIC HYPOTHESES AS
UNCERTAIN AND PROBABILISTIC DATA
Bernardo Gonçalves
February - 2015
Advisor: Fabio Porto, D.Sc.
In view of the paradigm shift that makes science ever more data-driven, in this
thesis we propose a synthesis method for encoding and managing large-scale deterministic scientific hypotheses as uncertain and probabilistic data.
In the form of mathematical equations, hypotheses symmetrically relate aspects of the studied phenomena. For computing predictions, however, deterministic
hypotheses can be abstracted as functions. We build upon Simon’s notion of structural equations in order to efficiently extract the (so-called) causal ordering between
variables, implicit in a hypothesis structure (set of mathematical equations).
We show how to process the hypothesis predictive structure effectively through
original algorithms for encoding it into a set of functional dependencies (fd’s) and
then performing causal reasoning in terms of acyclic pseudo-transitive reasoning
over fd’s. Such reasoning reveals important causal dependencies implicit in the hypothesis predictive data and guide our synthesis of a probabilistic database. Like in
the field of graphical models in AI, such a probabilistic database should be normalized so that the uncertainty arisen from competing hypotheses is decomposed into
factors and propagated properly onto predictive data by recovering its joint probability distribution through a lossless join. That is motivated as a design-theoretic
principle for data-driven hypothesis management and predictive analytics.
The method is applicable to both quantitative and qualitative deterministic
hypotheses and demonstrated in realistic use cases from computational science.
Resumo da Tese apresentada ao LNCC/MCT como parte dos requisitos necessários
para a obtenção do grau de Doutor em Ciências (D.Sc.)
GERÊNCIA DE HIPÓTESES CIENTÍFICAS DE LARGA-ESCALA
COMO DADOS INCERTOS E PROBABILÍSTICOS
Bernardo Gonçalves
Fevereiro, 2015
Orientador: Fabio Porto, D.Sc.
Tendo em vista a mudança de paradigma que faz da ciência cada vez mais guiada
por dados, nesta tese propomos um método para codificação e gerência de hipóteses
cientı́ficas determinı́sticas de larga escala como dados incertos e probabilı́sticos.
Na forma de equações matemáticas, hipóteses relacionam simetricamente aspectos do fenômeno de estudo. Para computação de predições, no entanto, hipóteses determinı́sticas podem ser abstraı́das como funções. Levamos adiante a noção
de Simon de equações estruturais para extrair de forma eficiente a então chamada
ordenação causal implı́cita na estrutura de uma hipótese.
Mostramos como processar a estrutura preditiva de uma hipótese através de
algoritmos originais para sua codificação como um conjunto de dependências funcionais (df’s) e então realizamos inferência causal em termos de raciocı́nio acı́clico
pseudo-transitivo sobre df’s. Tal raciocı́nio revela importantes dependências causais implı́citas nos dados preditivos da hipótese, que conduzem nossa sı́ntese do
banco de dados probabilı́stico. Como na área de modelos gráficos (IA), o banco de
dados probabilı́stico deve ser normalizado de tal forma que a incerteza oriunda de
hipóteses alternativas seja decomposta em fatores e propagada propriamente recuperando sua distribuição de probabilidade conjunta via junção ‘lossless.’ Isso é
motivado como um princı́pio teórico de projeto para gerência e análise de hipóteses.
O método proposto é aplicável a hipóteses determinı́sticas quantitativas e
qualitativas e é demonstrado em casos realı́sticos de ciência computacional.
Table of Contents
1 Introduction
1
1.1
Problem Space and Specific Goals . . . . . . . . . . . . . . . . . . .
4
1.2
Thesis Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
1.3
Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . .
13
1.4
Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
2 Vision: Hypotheses as Data
15
2.1
Running Example . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.2
Hypothesis Encoding . . . . . . . . . . . . . . . . . . . . . . . . . .
20
2.3
Reasoning over FD’s . . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.4
Uncertainty Introduction . . . . . . . . . . . . . . . . . . . . . . . .
22
2.5
Predictive Analytics . . . . . . . . . . . . . . . . . . . . . . . . . .
25
2.6
Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
2.7
Summary: Key Points . . . . . . . . . . . . . . . . . . . . . . . . .
30
3 Hypothesis Encoding
32
3.1
Preliminaries: Structural Equations . . . . . . . . . . . . . . . . . .
32
3.2
The Problem of Causal Ordering . . . . . . . . . . . . . . . . . . .
36
3.3
Total Causal Mappings . . . . . . . . . . . . . . . . . . . . . . . . .
39
3.4
The Encoding Scheme . . . . . . . . . . . . . . . . . . . . . . . . .
43
3.5
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
3.6
Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
3.7
Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . .
48
4 Causal Reasoning over FD’s
4.1
Preliminaries: Armstrong’s Inference Rules . . . . . . . . . . . . . .
49
49
TABLE OF CONTENTS
x
4.2
Acyclic Pseudo-Transitive Reasoning . . . . . . . . . . . . . . . . .
50
4.3
Equivalence with Causal Ordering . . . . . . . . . . . . . . . . . . .
55
4.4
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
4.5
Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
4.6
Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . .
60
5 Probabilistic Database Synthesis
61
5.1
Preliminaries: U-Relations and Probabilistic WSA . . . . . . . . . .
61
5.2
Running Example . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
5.3
U-Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
5.4
U-Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
70
5.5
Design-Theoretic Properties . . . . . . . . . . . . . . . . . . . . . .
73
5.6
Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
76
5.7
Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . .
78
6 Applicability
79
6.1
The Physiome Project as a Testbed . . . . . . . . . . . . . . . . . .
79
6.2
Case Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
80
6.3
System Prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . .
89
6.4
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
95
6.5
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
98
6.6
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7 Conclusions
102
7.1
Revisiting the Research Questions . . . . . . . . . . . . . . . . . . . 102
7.2
Significance and Limitations . . . . . . . . . . . . . . . . . . . . . . 105
7.3
Open Problems and Future Work . . . . . . . . . . . . . . . . . . . 106
7.4
Final Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . 107
TABLE OF CONTENTS
Bibliography
xi
108
Appendix
A Detailed Proofs
116
A.1 Proofs of Hypothesis Encoding . . . . . . . . . . . . . . . . . . . . . 116
A.2 Proofs of Causal Reasoning . . . . . . . . . . . . . . . . . . . . . . 119
A.3 Proofs of Probabilistic DB Synthesis . . . . . . . . . . . . . . . . . 126
List of Figures
Figure
1.1
Multi-fold view of a deterministic scientific hypothesis. . . . . . . .
2
1.2
A view of the scientific method life cycle . . . . . . . . . . . . . . .
5
1.3
The usual data ingesture pipeline of simulation data management .
7
1.4
Design-theoretic pipeline for processing hypotheses as uncertain and
probabilistic data . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.1
Scientific hypotheses seen as alternative functions to predict data. .
16
2.2
Predictive analytics in a data-intensive hypothesis evaluation study.
18
2.3
Descriptive (textual) data of Example 1. . . . . . . . . . . . . . . .
19
2.4
‘Big’ fact table H1 loaded with simulation raw data . . . . . . . . .
19
2.5
‘Explanation’ relational table H0 . . . . . . . . . . . . . . . . . . . .
23
2.6
Result set of query Q1 on simulation trial dataset for hypothesis H1 . 24
2.7
U-relational predictive tables rendered by query using the fd’s. . . .
25
2.8
Analytics on predicted position s conditioned on observation. . . . .
26
3.1 “Directed causal graphs” associated with the two systems. . . . . . .
33
3.2
Running Simon’s Causal Ordering Algorithm (COA) . . . . . . . . .
35
3.3
Directed causal graph Gϕ induced by mapping ϕ for structure S . .
36
3.4
Bipartite graph G of structure S from Example 3. . . . . . . . . . .
37
3.5
Another hypothesis structure example. . . . . . . . . . . . . . . . .
39
3.6
Complete matching M for structure S from Example 3. . . . . . . .
42
3.7
Encoded fd set Σ (cf. Alg. 3) for the structure from Example 3. . . .
44
3.8
Performance of hypothesis encoding (in logscale). . . . . . . . . . .
46
4.1
Fd set Σ encoding (cf. Alg. 3) the structure of Fig. 3.2a and its
folding Σ# derived by Alg. 5. . . . . . . . . . . . . . . . . . . . . .
52
LIST OF FIGURES
4.2
xiii
Fd set Σ encoding the structure of Fig. 3.2a and the folding Υ(Σ)#
of its υ-projection. . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
4.3
Performance of acyclic causal reasoning over fd’s (logscale). . . . . .
59
5.1
U-relation generated by the repair-key operation. . . . . . . . . . .
63
5.2
Fd sets encoded from the given structures Sk (Ek , Vk ) for hypotheses
k = 1..3 from Example 6. . . . . . . . . . . . . . . . . . . . . . . . .
5.3
‘Big’ fact table H3 of hypothesis k = 3 from Example 6 loaded with
trial datasets identified by special attribute tid. . . . . . . . . . . .
5.4
65
65
‘Big’ fact table H3 of hypothesis k = 3 from Example 6 with u-factors
{b, d} and {p, r} emphasized (resp.) in colors green and red. . . . .
67
5.5
Fd set Ω3 (compare with Σ3 ) and its folding Ω#
3. . . . . . . . . . . .
68
5.6
U-factor projections rendered for hypothesis υ = 3. . . . . . . . . .
69
5.7
U-relations rendered for hypothesis υ = 3. . . . . . . . . . . . . . .
72
6.1
Plot of hemoglobin oxygen saturation hypotheses . . . . . . . . . .
81
6.2
Descriptive (textual) data of Example 8. . . . . . . . . . . . . . . .
81
6.3
Fd set Σ28 of hypothesis υ = 28. . . . . . . . . . . . . . . . . . . . .
82
6.4
Fd set Σ31 of hypothesis υ = 31. . . . . . . . . . . . . . . . . . . . .
82
6.5
Fd set Σ32 of hypothesis υ = 32. . . . . . . . . . . . . . . . . . . . .
82
6.6
Result set of hypothesis management query Q1. . . . . . . . . . . .
83
6.7
Results of analytical study on the hemoglobin phenomenon. . . . .
83
6.8
Plot of baroreflex hypothesis for Dahl SS Rat . . . . . . . . . . . .
84
6.9
Descriptive (textual) data of Example 9. . . . . . . . . . . . . . . .
84
6.10 Result set of hypothesis management query Q2. . . . . . . . . . . .
85
6.11 Results of analytical study on the baroreflex phenomenon. . . . . .
85
6.12 Fd set Σ1001 of hypothesis υ = 1001. . . . . . . . . . . . . . . . . . .
86
6.13 Plot of myogenic behavior hypothesis . . . . . . . . . . . . . . . . .
87
6.14 Descriptive (textual) data of Example 10. . . . . . . . . . . . . . . .
87
6.15 Fd set Σ60 of hypothesis υ = 60. . . . . . . . . . . . . . . . . . . . .
88
6.16 Fd set Σ89 of hypothesis υ = 89. . . . . . . . . . . . . . . . . . . . .
88
6.17 Result set of hypothesis management query Q3. . . . . . . . . . . .
89
6.18 Results of analytics on the vessel’s myogenic behavior phenomenon.
89
LIST OF FIGURES
xiv
6.19 Lynx-Hare population observed . . . . . . . . . . . . . . . . . . . .
91
6.20 Lynx-Hare population observed . . . . . . . . . . . . . . . . . . . .
91
6.21 Screenshots of this first prototype of the Υ-DB system. . . . . . . .
92
6.22 Descriptive (textual) data of Example 10. . . . . . . . . . . . . . . .
93
6.23 Fd set Σ1 of hypothesis υ = 1. . . . . . . . . . . . . . . . . . . . . .
93
6.24 Fd set Σ2 of hypothesis υ = 2. . . . . . . . . . . . . . . . . . . . . .
93
6.25 Fd set Σ3 of hypothesis υ = 3. . . . . . . . . . . . . . . . . . . . . .
94
6.26 Results of analytics on the US population phenomenon. . . . . . . .
95
6.27 Results of analytics on the Hudson’s Bay lynx population phenomenon 95
6.28 Performance behavior of Υ-DB on a Physiome testbed scenario. . .
96
6.29 Physiome hypotheses used in the experiments. . . . . . . . . . . . .
97
6.30 Example of Boolean Network hypothesis. . . . . . . . . . . . . . . . 100
6.31 Example of Boolean Network model . . . . . . . . . . . . . . . . . . 101
List of Tables
Table
1.1
Simulation data management vs. hypothesis data management. . .
5
Acronyms
• fd: functional dependency
• p-DB: probabilistic database
• p-WSA: probabilistic world set algebra
• MayBMS: U-relational database management system
• AI: Artificial Intelligence
• GM: graphical models (e.g., Bayesian Networks)
• ETL: extract, transform, load
• OLAP: On-Line Analytical Processing
• DW: Data Warehouse
• SEM: structural equation model
• COA: causal ordering algorithm
• Sk : structure of hypothesis k
• E: set of equations in a structure
• V: set of variables appearing in the equations in a structure
• V ars(f ): set of variables appearing in equation f
• ϕ : E → V: total causal mapping from equations to variables in a structure
• Cϕ : set of causal dependencies
• Gϕ : causal graph induced by ϕ
• Hk : ‘big’ fact table of hypothesis k
• H: set of hypothesis ‘big’ fact tables
• Yk` : U-relation synthesized for hypothesis k
• Y k : set of U-relations synthesized for hypothesis k
• H0 : relational ‘explanation’ table
LIST OF TABLES
• Y0 : U-relational ‘explanation’ table
• φ: phenomenon identifier
• υ: hypothesis identifier
• Σ, Γ, ∆, Ω: fd sets
• ΣB : subset of the closure of an fd set derived by pseudo-transitivity
• Σ+ : closure of an fd set
• Σ# : the folding of an fd set
• synthesis ‘4U’: synthesis for uncertainty
• tid: hypothesis trial identifier
• BCNF: Boyce-Codd Normal Form
• MathML: Mathematical Markup Language
• TCM: total causal mapping algorithm
• lhs, rhs: left-hand side, right-hand side
• sch(R): data columns of table R
• Vi Di : condition columns of a table
xvii
Chapter 1
Introduction
In view of the paradigm shift that makes science ever more data-driven [1], in this
thesis we demonstrate that large deterministic scientific hypotheses can be effectively encoded and managed as a kind of uncertain and probabilistic data.
Deterministic hypotheses can be formed as principles or ideas, then expressed
mathematically and implemented in a program that is run to give their decisive
form of data (see Fig. 1.1). Hypotheses can also be learned in large scale, as exhibited in the Eureqa project [2]. Examples of ‘structured deterministic hypotheses’
include tentative mathematical models in physics, engineering and economical sciences, or conjectured boolean networks in molecular biology and social sciences.
These are important reasoning devices, as they are solved to generate valuable
predictive data for decision making in science and increasingly in business as well.
In fact, we can refer nowadays to a broad, modern context of data science [3]
and big data [4] in which the complexity and scale of so-called ‘data-driven’ problems require proper data management tools for the predicted data to be analyzed
effectively. In this thesis, we pay attention to a quite general class of (tentative)
computational science models,1
and we look at them in an original way as a
distinguished kind of data source.
1
‘Computational science’ is (sic.) “a rapidly growing multidisciplinary field that uses advanced computing capabilities to understand and solve complex problems” [5]. We may refer
to non-stochastic, tentative computational science models throughout this text as ‘structured
deterministic hypotheses.’
2
Law of free fall
a(t) = −g
“If a body falls from rest, its velocity at any point
is proportional to the time it has been falling.”
v(t) = −gt + v0
s(t) = −(g/2)t2 + v0 t + s0
(i)
(ii)
for k = 0:n;
t = k * dt;
v = -g*t + v_0;
s = -(g/2)*t^2 + v_0*t + s_0;
t_plot(k) = t;
v_plot(k) = v;
s_plot(k) = s;
end
(iii)
FALL
t
0
1
2
3
4
···
v
0
−32
−64
−96
−128
···
s
5000
4984
4936
4856
4744
···
(iv)
Figure 1.1. Multi-fold view of a deterministic scientific hypothesis.
It is generally considered that computational science models, interpreted here
as hypotheses to explain real-world phenomena, are of strategic relevance [5]. They
are usually complex in that they may have hundreds to thousands of intertwined
(coupled) variables and be computed along space, time or frequency domains in
arbitrarily large scale. It is important to note the distinction between the structure
and data levels. Consider, say, Lotka-Volterra’s model, which essentially consists
in (Eqs. 1.1) two ordinary differential equations, complemented by seven subsidiary equations f1 (t), f2 (x0 ), f3 (y0 ), f4 (b), f5 (p), f6 (r), f7 (d) to set the values of
its domain variable t and (input) parameters x0 , y0 , b, p, r, d.
ẋ = x(b − py)
ẏ = y(rx − d)
(1.1)
In a sense, it can be said fairly simple, as it is characterized by a set E of equations
and a set V of variables, sized |E| = |V| = 9. Yet, at the data level this model (cf.
Chapter 2) can be made very large just by computing its predictions in a fine time
resolution and/or along an extended time window.
As we shall see shortly, the technical challenges associated with this thesis
involve (not only but) majorly the structure level where, e.g., such Lotka-Volterra
model can be abstracted as a deterministic structure S(E, V) with |S| = 18.2
2
The structure length |S| is a measure of how dense the hypothesis structure is, comprising
the total sum of the number of variables appearing in each equation.
3
We are really concerned here with models whose structure S is in the order of
|S| . 1M , and whose results (data!) shall be difficult to analyze by handicrafted
practice. Note that the data level of a model can be set as large as wanted (set
the domain resolution and/or extension accordingly), but it shall be necessarily
large when its structure is itself large. By ‘large-scale hypotheses’ then we mean
tentative deterministic models that are large at structure level.
Overall, such class of hypotheses can be said to qualify to at least four of
the five v’s associated to the notion of big data:3
value, because of their role
in advancing science and technology; volume, due to the large scale of modern
scientific problems; variety, because of their structural heterogeneity, even when
they refer to the same phenomena; and veracity, due to their uncertainty.
The idea of managing hypotheses ‘as data’ may sound intriguing and in fact
it raises a number of research questions of both conceptual and technical nature.4
We start by outlining below the conceptual research questions.
RQ1. How to define and encode hypotheses ‘as data’ ? What are the sources of
uncertainty that may be present and should be considered?
RQ2. How does hypotheses ‘as data’ relate with observational data or, likewise,
phenomena ‘as data’ from a database perspective?
RQ3. Does every piece of simulated data qualify as a scientific hypothesis? What
is the difference between managing ‘simulation’ data from managing ‘hypotheses’ as data?
RQ4. Is there available a proper (machine-readable) data format we can use to
automatically extract mathematically-expressed hypotheses from?
It has been a challenge of this thesis to provide reasonable answers to these questions, which are brought together into the vision of hypotheses ‘as data’ (we call
it the Υ-DB vision) and its use case that we present in Chapter 2, and experiment
with in realistic scenarios in Chapter 6.
3
The ‘v’ of velocity may appear in connection with machine learning hypotheses, which we
discuss in Chapter 6.
4
We shall keep record of those questions and revisit them in §7.1.
1.1. PROBLEM SPACE AND SPECIFIC GOALS
4
The Υ-DB vision formulates the problem of hypothesis encoding as a problem
of probabilistic database design. A number of technical questions arise then.
We introduce now technical context, materials and methods identified and
selected in this thesis as a basis to realize the Υ-DB vision in terms of probabilistic
database design. We shall outline in the sequel the technical research questions to
be answered by the core of the thesis.
1.1. Problem Space and Specific Goals
It has been a goal of this thesis to investigate the capabilities of probabilistic
databases to enable hypothesis data management as a particular case of simulation
data management. In the sequel, we first characterize the use case of hypothesis
data management and then formulate it in terms of probabilistic DB design.
1.1.1 Simulation data management
Simulation laboratories provide scientists and engineers with very large, possibly huge datasets that reconstruct phenomena of interest in high resolution. Notorious examples are the John Hopkins Turbulance Databases [6], and the Human
Brain Project (HBP) neuroscience simulation datasets [7]. A core motivation for
the delivery of such data is enabling new insights and discoveries through hypothesis testing against observations. Nonetheless, while the use case for exploratory
analytics is currently well understood and many of its challenges have already been
coped with so that high-resolution simulation data is increasingly more accessible
[8, 9], only very recently, as part of this thesis work, the use case of hypothesis
management has been taken into account for predictive analytics [10].
In fact, there is a pressing call for innovative technology to integrate (observed) data and (simulated) theories in a unified framework [11, 12, 13]. The
point has just been raised by leading neuroscientists in the context of the HBP,
who are incisive on the compelling argument that massive simulation databases
should be constrained by experimental data in corrective loops to test precise hypotheses [14, p. 28]. Fig. 1.2 shows a simplified view of the (data-driven) scientific
method life cycle. It distinguishes the phases of exploratory analytics (context of
1.1. PROBLEM SPACE AND SPECIFIC GOALS
5
Context of discovery
Context of justification
no
Phenomenon
observation
Hypothesis
formulation
Computational
simulation
Testing
against data
valid?
yes Publishing
results
Figure 1.2. A view of the scientific method life cycle. It highlights hypothesis
formulation and a backward transition to reformulation if predictions ‘disagree’
with observations.
discovery) and predictive analytics (context of justification), and highlights the
loop between hypothesis formulation and testing [15].
Simulation data, being generated and tuned from a combination of theoretical and empirical principles, has a distinctive feature to be considered when
compared to data generated by high-throughput technology in large-scale scientific experiments. It has a pronounced uncertainty component that motivates the
use case of hypothesis data management for predictive analytics [10]. Essential
aspects of hypothesis data management can be described in contrast to simulation
data management as follows — Table 1.1 summarizes our comparison.
• Sample data. Hypothesis management shall not deal with the same volume
of data as in simulation data management for exploratory analytics, but
only samples of it. This is aligned, for example, with the architectural design of CERN’s particle-physics experiment and simulation ATLAS, where
there are four tier/layers of data. The volume of data significantly decreases from (tier-0) the raw data to (tier-3) the data actually used for
analyses such as hypothesis testing [8, p. 71-2]. Samples of raw simula-
Table 1.1. Simulation data management vs. hypothesis data management.
Simulation data management
Exploratory analytics
Raw data
Extremely large (TB, PB)
Dimension-centered access pattern
Denormalized for faster retrieval
Batch-, incremental-only data updates
Hypothesis data management
Predictive analytics
Sample data
Very large (MB, GB)
Claim-centered access pattern
Normalized for uncertainty factors
Probability distribution updates
1.1. PROBLEM SPACE AND SPECIFIC GOALS
6
tion data are to be selected for comparative studies involving competing
hypotheses in the presence of evidence (sample observational data). This
principle is also aligned with how data is delivered at model repositiories.
Since observations are usually less available, only the fragment (sample)
of the simulation data that matches in coordinates the (sample) of observations is required out of simulation results for comparative analysis. For
instance, we show in §6.2.2 a predictive analytical study extracted from the
Virtual Physiological Rat Project (VPR1001-M) comparing sample simulation data (heart rates) from a baroreflex model with observations on a
Dahl SS rat strain.5 The simulation is originally set to produce predictions
in the time resolution of t∆ = 0.01. But since the observational sample is
only as fine as t∆ = 0.1, there is no gain in rendering a predicted sample
with t∆ ≥ 0.1 for hypothesis testing. Note that such a ‘sampling’ does not
incur in any additional uncertainty as typical of statistical sampling [16].
• Claim-centered access pattern. In simulation data management the access
pattern is dimension-centered (e.g., based on selected space-time coordinates) and the data is denormalized for faster retrieval, as typical of Data
Warehouses (DW’s) and OLAP applications.6
In particular, on account
of the so-called ‘big table’ approach, each state of the modeled physical
system is recorded in a large, single row of data. This is fairly reasonable
for an Extract-Transform-Load (ETL) data ingesture pipeline characterized by batch-, incremental-only updates (see Fig. 1.3). Such a setting is
in fact fit for exploratory analytics, as entire states of the simulated system
shall be accessed at once (e.g., providing data to a visualization system).
Altogether, data retrieval is critical and there is no risk of update anomalies. Hypothesis management, in contrast, should be centered on claims
identified within the hypothesis structure w.r.t. available data dependencies. Since the focus is on resolving uncertainty for decision making (which
5
http://virtualrat.org/computational-models/vpr1001/.
On-Line Analytical Processing, as distinguished from OLTP (On-Line Transaction Processing. The latter is meant for transaction processing of daily queries and updates in operational
systems, while the former is for analytical queries in Data Warehouses (DW’s) that gather a lot
of data collected from different sources for decision making.
6
1.1. PROBLEM SPACE AND SPECIFIC GOALS
D1
D2
7
ETL
...
Dp
sim
Sp
i=1
Ri
Figure 1.3.
S The usual data ingesture pipeline of simulation data management.
Datasets pi=1 Di generated bySsimulation trials on (hypothesis) models are
loaded each into a ‘big’ table pi=1 R. The uncertainty is then “buried” in the
database, which lacks a logical organization for enabling data-driven hypothesis
management and predictive analytics.
hypothesis is a best fit?), the data must be normalized based on uncertainty factors. This is key for the correctness of uncertainty modeling and
efficiency of probabilistic reasoning, say, in a probabilistic database [17,
p.30-1].
• Uncertainty modeling. In uncertain and probabilistic data management
[17], the uncertainty may come from two sources: incompleteness (missing data), and multiplicity (inconsistent data). Hypothesis management
on sample simulation data is concerned with the multiplicity of prediction records due to competing hypotheses targeted at the same studied
phenomenon. Such a multiplicity naturally gives rise to a probability distribution that may be initially uniform and eventually conditioned on observations. Conditioning is an applied Bayesian inference problem that
translates into database update for transforming the prior probability distribution into a posterior [10].
Overall, hypothesis data management is also OLAP-like, yet markedly different from simulation data management.
A key point that distinguishes hypothesis management is that a fact or unit of
data is defined by its predictive content. That is, every clear-cut predicted
fact (w.r.t.available data dependencies) is a claim. Accordingly, the data should
be decomposed and organized for a claim-centered access pattern.
1.1. PROBLEM SPACE AND SPECIFIC GOALS
8
conditioning
Sk
ETL
Dk1
Dk2 ... Dkp
U-intro
h
Sn
k=1
Hk
y
Sn
k=1
Sm
`=1
Yk`
Figure 1.4. Pipeline for processing hypotheses as uncertain and probabilistic
data. For each hypothesis k, its structure Sk is given
Spin a machine-readable
format, and all of its sample simulation data trials i=1 Dki are indicated their
target phenomenon, say φ, to be loaded into a ‘big table’ Hk . Then
Snthe synthesis
comes into play to read a base of possibly very many hypotheses k=1 Hk and
transform them into a probabilistic
database where each hypothesis is decomSm
`
posed into claim tables `=1 Yk . A probability distribution is computed for each
phenomenon φ, covering all the hypotheses and their trials targeted at φ. This
distribution is then updated into a posterior in the presence of observational data.
To anticipate Chapter 2, the synthesis method we have developed in this
thesis work for processing hypotheses as uncertain and probabilistic data comprises
a design-theoretic pipeline (see Fig. 1.4) that extends the one shown in Fig. 1.3.
1.1.2 Probabilistic database design
Probabilistic databases (p-DB’s) have evolved into mature technology in the
last decade with the emergence of new data models and query processing techniques
[17]. One of the state-of-the-art probabilistic data models is the U-relational representation system with its probabilistic world-set algebra (p-WSA) implemented in
MayBMS [18]. That is an elegant extension of the relational model we shall refer to
in this thesis for the management of large-scale uncertain and probabilistic data.
We look at U-relations from the point of view of p-DB design, for which no
formal design methodology has yet been proposed. Despite the advanced state of
probabilistic data management techniques, a lack of methods for the systematic
design of p-DBs may prevent wider adoption. The availability of design methods
has been considered one of the key success factors for the rapid growth of applications in the field of Graphical Models (GM’s) [19], considered to inform research
in p-DB’s [17, p. 14]. Analogously, we have proposed to distinguish methods for
p-DB design in three classes [10]: (i) subjective construction, (ii) learning from
1.1. PROBLEM SPACE AND SPECIFIC GOALS
9
data, and (iii) synthesis from other kind of formal specification.
The first is the less systematic, as the user has to model for the data and
correlations by steering all the p-DB construction process (MayBMS’ use cases [18],
e.g., are illustrated that way). The second comprises analytical techniques to extract the data and learn correlations from external sources, possibly unstructured,
into a p-DB under some ad-hoc schema. This is the prevalent one up to date,
motivated by information extraction and data integration applications [17, p. 103]. In this thesis we present a methodology of the third kind, as we extract data
dependencies from some previously existing formal specification (the hypothesis
mathematical structure) to synthesize a p-DB algorithmically. Such a type of construction method has been successful, e.g., for building Bayesian Networks [19]. To
our knowledge, this thesis is the first synthesis method for p-DB design (cf. §5.6).
We shall develop means to extract the specification of a hypothesis and encode
it into a U-relational DB for data-driven hypothesis management and analytics.
That is, we shall flatten deterministic hypotheses into U-relations.
The synthesis method that we have developed for p-DB’s relies on the extraction of functional dependencies (fd’s; cf. [20, 21, 22]) that are basic input to
algorithmic synthesis.7 For an example of fd, consider relation FALL in Fig. 1.1.
There holds an fd t → v s, meaning that values of attribute time t functionally
determine values of both attributes velocity v and position s. More precisely, let
µ and τ be any two tuples (rows) in an instance of relation (table) FALL. Then it
satisfies fd t → v s iff µ[t] = τ [t] implies µ[v s] = τ [v s]. In our illustrative relation
FALL, that fd is, in particular, a key constraint, which means that (values of) t
play the role of a key to (provide access to the values of) v and s in the relation.
A related concept which is also a major one for us is that of normalization
[20, 21, 22], viz., to ensure that the DB resulting from a design process bears some
desirable properties which are associated with some notion of normal form (ibid.).
For hypothesis management, the uncertainty has to be modeled and should be
normalized so that the uncertainty of one claim may not be undesirably mixed
7
In fact, it has been considered a critical failure in traditional DB design the lack of techniques
to obtain important information such as fd’s in the real world [23, p. 62].
1.1. PROBLEM SPACE AND SPECIFIC GOALS
10
with the uncertainty of another claim. It is expected to involve a processing of the
causal dependencies implicit in the given hypothesis structure. We shall introduce
in detail such concepts in context when necessary.
1.1.3 Structural equations
The flattening of the user mathematical models into hypothesis p-DB’s, nonetheless, is not straightforward. It has been a goal of this thesis to investigate proper
abstractions on mathematical models in order to (partly) capture their semantics,
viz., to an extent that is tailored for hypothesis management (as opposed to, say,
model solving). We shall abstract mathematical models into intermediary artifacts
that are amenable to be further encoded into fd’s.
In fact, given a system of equations with a set of variables appearing in
them, in a seminal article Simon introduced an asymmetrical, functional relation
among variables that establishes a (so-called) causal ordering [24]. That became
known as structural equation models (SEM’s) or just ‘structural equations’ (cf. also
[25]). Along these lines, our goal is to extract the causal ordering implicit in the
structure of a deterministic hypothesis into a set of fd’s that guides our synthesis
of U-relational DB’s. As we shall see throughout this text,
the causal ordering we capture and process through fd’s provides causal dependencies implicit in the predictive data that are very useful information to
decompose uncertainty for the sake of probabilistic modeling and reasoning.
1.1.4 Uncertainty Model
In uncertain and probabilistic data management, there are essentially two sources
of uncertainty: incompleteness (missing data), and multiplicity (inconsistent data).
The kind of uncertainty that is dealt with in this work is the multiplicity of hypothesis trial records identified to be targeted at the same phenomenon record.
That is, the uncertainty arises from the existance of competing hypotheses.
If multiple hypotheses and trials are inserted for the same phenomenon, the
system interprets it as defining a probability distribution.
1.1. PROBLEM SPACE AND SPECIFIC GOALS
11
Such a probability distribution (usually uniform) on the multiplicity of competing hypotheses is in accordance with probability theory under possible-worlds
semantics [17, Ch. 1]. It is modeled into the U-relational data model and its
p-WSA operators, and implemented into the MayBMS system as we shall see in
§5.1.8 The conf() aggregate operator, for instance, in spite of the name, performs
standard (non-Bayesian) probabilistic inference on such probability distribution.
Eventually, however, there is a need to condition the initial probability distribution in the presence of observations. For the conditioning, then, we shall adopt
Bayesian inference so that the prior probability distribution can be updated to a
posterior.
The informal discussion of this section opens the way for a number of technical research questions that we outline next.
RQ5. Is there an algorithm to, given a SEM, efficiently extract its causal ordering? What are the computational properties of this problem?
RQ6. What is the connection between SEM’s and fd’s? Can we devise an encoding scheme to ‘orient equations’ and then effectively transform one into
the other with guarantees? Once we do it, what design-theoretic properties
have such a set of fd’s?
RQ7. Is such fd set ready to be used for p-DB schema synthesis as an encoding
of the hypothesis causal structure? If not, what kind of further processing
we have to do? Can we perform it efficiently by reasoning directly on the
fd’s? How does it relate to the SEM’s causal ordering?
RQ8. Is the uncertainty decomposition required for predictive analytics reducible
to the structure level (fd processing), or do we need to process the simulated
data to identify additional uncertainty factors? Finally, what properties
are desirable for a p-DB schema targeted at hypothesis management? Are
they ensured by this synthesis method?
RQ9. Given all such a design-theoretic machinery to process hypotheses into
(U-)relational DB’s, what properties can we detect on the hypotheses back
8
Our own system of hypothesis management is to be delivered on top of the MayBMS backend.
1.2. THESIS STATEMENT
12
at the conceptual level? Do we have now technical means to speak of
hypotheses that are “good” in terms of principles of the philosophy of
science?
The core of this thesis is devoted to answer these questions, and we shall accomplish
it throughout Chapters 3, 4 and 5.
1.2. Thesis Statement
The statement of this thesis is that it is possible to effectively encode and
manage large deterministic scientific hypotheses as uncertain and probabilistic
data. Its key challenges are of both conceptual and technical nature. Conceptually, we provide core, non-obvious abstractions to define and encode hypotheses
as data. Technically, we provide a number of algorithms that compose a designtheoretic pipeline to encode hypotheses as uncertain and probabilistic data, and
verify their efficiency and correctness. The applicability and effectiveness of our
method is demonstrated in realistic case studies in computational science.
Besides, it is worthwhile highlighting some non-goals of this thesis.
N1. Although we perform some sort of information extraction [26] for the acquisition of hypotheses from some model repositories on the web, it is very
basic and ad-hoc in order to obtain a testbed for our method. That is,
we are not proposing means for the systematic extraction of hypotheses
from available sources. In fact we shall outline it in §7.3 as an important
direction of future work.
N2. We do not address solving computational models or numerical analytics
in any sense. In fact we rely on the numerical solvers (implemented into
tools that we use) as ‘transaction processing’ systems, load their computed
data into a relational ‘big’ fact table and then render it into U-relational
tables synthesized by our method. We do not deal with data visualization
either in any sense.
N3. The efficiency and scalability of query processing in p-DB’s, in particular Urelational’s MayBMS and its p-WSA (which we rely on) is not addressed or
1.3. THESIS CONTRIBUTIONS
13
evaluated in this thesis. In fact, the performance of U-relations and p-WSA
has been extensively evaluated and shown to be effective [27, 18]. All performance tests carried out in this thesis comprise our design-theoretic techniques for the encoding and synthesis of U-relational hypothesis databases.
N4. In terms of uncertainty and statistical analysis, we stick to (i) process some
well-defined forms of multiplicity in the data which constitute the model of
uncertainty dealt with in this work; then (ii) by relying on MayBMS we perform probabilistic inference; and (iii) eventually (at application level) we
perform Bayesian inference and so that a posterior probability distribution
is propagated through p-DB updates. We do not provide any additional
form of uncertainty management. Rather, we manage the data extracted
into the system (under user control) and process its uncertainty in terms
of the specific sources of uncertainty recognized in Υ-DB (cf. Chapter 2).
1.3. Thesis Contributions
The contributions of this thesis are outlined as follows.
1.3.1 Innovative Contributions
This thesis presents the vision of hypotheses as data (and its use case) socalled Υ-DB vision. It has been published in the vision track of VLDB 2014 [10],
for its (sic.) potentially high-impact visionary content. The innovative system of
Υ-DB has been described in a ‘system prototype demonstration’ paper [28].9
1.3.2 Technical Contributions
This thesis presents specific technical developments over the Υ-DB vision. In
short, it shows how to encode deterministic hypotheses as uncertain and probabilistic data. Our detailed technical contributions (cf. Chapters 3, 4, and 5)
are formulated into a formal method for the design of hypothesis p-DB’s which is
described in a technical report [29].10
The method, together with our realistic
testbed scenarios and performance evaluation, are yet to be published.
9
10
Preliminary version available at CoRR abs/1411.7419.
Preliminary version available at CoRR abs/1411.5196.
1.4. THESIS OUTLINE
14
1.4. Thesis Outline
The structure of the remainder of this thesis is outlined for reference.
Chapter 2. [Υ-DB Vision]. The research vision of hypotheses as (uncertain and
probabilistic) data, the characterization of its use case, key points and technical
challenges are presented.
Chapter 3. [Encoding]. The problem of encoding a hypothesis ‘as data’ given
its formal specification (set of mathematical equations) is presented and addressed
by an encoding scheme that transforms the equations into fd’s with guarantees in
terms of preserving the hypothesis causal structure.
Chapter 4. [Causal Reasoning]. It is presented a technique for causal reasonig
as acyclic pseudo-transitive reasoning over the encoded fd’s. It processes the hypothesis causal ordering to find the ‘first causes’ for each of its predictive variables.
Chapter 5. [p-DB Synthesis]. It is presented a technique to address the problem
of uncertainty introduction and propagation for the transformation of hypotheses
into U-relational databases. The synthesized U-database is shown to bear desirable
properties for hypothesis management and predictive analytics.
Chapter 6. [Applicability]. A discussion of applicability, the implementation
of the proposed techniques into a prototype system for test and demonstration of
the vision realization through realistic case studies are presented.
Chapter 7. [Conclusions]. Research questions are revisited, and the significance
and limitations of the thesis with directions to future work and final considerations
are discussed.
Chapter 2
Vision: Hypotheses as Data
High-throughput technology and large-scale scientific experiments provide
scientists with empirical data that has to be extracted, transformed and loaded
before it is ready for analysis [1]. In this vision we consider theoretical data, or
data generated by simulation from deterministic scientific hypotheses, which also
needs to be pre-processed to be analyzed.
Hypotheses as data. In view of the age of data-driven science, we consider
deterministic scientific hypotheses from a multi-fold point of view: formed as principles or learned in large scale,1 hypotheses are formulated mathematically and
coded in a program that is run to give their decisive form of data (see Fig. 1.1).
Uncertain data. The semantic structure of relation FALL (Fig. 1.1), item
(iv) can be expressed by the functional dependency (fd) t → v s. This is typical
semantics assigned to empirical data in the design of experiment databases. A
space-time dimension (like time t in our example) is used as a key to observables
(like velocity v and position s). In empirical uncertainty, it is such “physical”
dimension keys like t that may be violated, say, by alternative sensor readings.
Hypotheses, however, are tentative explanations of phenomena [15], which
characterizes a different kind of uncertain data. In order to manage such theoretical
uncertainty, we shall need two special attributes to compose, say, the epistemological dimension of keys to observables: φ, identifying the studied phenomena; and
υ, identifying the hypotheses aimed at explaining them. That is, we shall leverage
the semantics of relations like FALL to φ υ t → v s. This leap is a core abstraction
1
As exhibited, e.g., in the Eureqa project [2].
16
~
X
Y~
a
~ → Y~
υ1 : X
υ1 (a)
υ1 (b)
b
υ2 (b)
~ → Y~
υ2 : X
Given state
Predicted state
Figure 2.1. Deterministic scientific hypotheses seen as alternative functions to
predict data, giving rise to both theoretical and empirical sources of uncertainty.
in this vision of Υ-DB.
Predictive data. Scientific hypotheses are tested by way of their predictions [15]. In the form of mathematical equations, hypotheses symmetrically relate
aspects of the studied phenomenon. However, for computing predictions, deterministic hypotheses are applied asymmetrically as functions [30]. They take a given
valuation over input variables (parameters) to produce values of output variables
(predictions). By observing that, we shall seek a principled method to transform
the (symmetric) mathematical equations of a hypothesis into (asymmetric) fd’s.
By looking at deterministic hypotheses as alternative functions to predict
data (see Fig. 2.1), in this vision we shall deal with two sources of uncertainty.
Given a well-defined context with a set of alternative hypotheses aimed at explaining (providing predictions for) a selected phenomenon:
• Theoretical uncertainty,2
comprises selecting the best tentative model
(function) to produce (the best) data?
• Empirical uncertainty,3 comprises, for each candidate model, what is the
(parameter) input setting that calibrates it the best way for the selected
phenomenon?
Note that these two sources of uncertainty are intertwined in that one cannot
‘clean’ one without cleaning the other — neither theory nor parameters are directly
2
3
That is, multiplicity of hypothesis entries associated with a phenomenon.
That is, multiplicity of hypothesis trial entries associated with a phenomenon.
17
observable, but only their joint results (the predictions) [15]. In this thesis we aim
at providing means to support such kind of ‘integrated’ analytics.
Applications. Big computational science research programs such as the
Human Brain Project,4 or Cardiovascular Mathematics,5 are highly-demanding
applications challenged by such theoretical (big) data. Users need to analyze results
of hundreds to thousands of data-intensive simulation trials.
Besides, recent initiatives on web-based model repositories have been fostering large-scale model integration, sharing and reproducibility in the computational
sciences (e.g., [31, 32, 33]). They are growing reasonably fast on the web, (i)
promoting some MathML-based standard for model specification, but (ii) with limited integrity and lack of support for rating/ranking competing models. For those
two reasons, they provide a strong use case for our vision of hypothesis management. The Physiome project [33, 34], e.g., is planned to integrate several large
deterministic models of human physiology — a fairly simple model of the human
cardiovascular system, e.g., has about 600+ variables.
Also, there is a pressing call for deep predictive analytic tools to support users
assessing what-if scenarios in business enterprises [35]. Deep predictive analytics
are based on first principles (deterministic hypotheses) and go beyond descriptive
analytics or shallow predictive analytics such as statistical forecasting (ibid.).6
U-relations. All that ratifies that hypothesis management is a promising
class of applications for probabilistic DB’s. The vision of Υ-DB is currently set to
be delivered on top of U-relations and probabilistic world-set algebra (p-WSA) [18].
These were developed in the influential MayBMS project.7 As implied by some of
its design principles, viz., compositionality and the ability to introduce uncertainty,
MayBMS’ query language fits well to hypothesis management. We shall look at it,
as previously mentioned, from the point of view of a synthesis method for p-DB
design. We shall particularly make use of the repair key operation, which gives
4
http://www.humanbrainproject.eu/.
http://icerm.brown.edu/tw14-1-pdecm.
6
The concept of ‘deep’ predictive analytics is from Haas et al. [35], and is discussed in more
detail in §2.6.1.
7
Project website: http://maybms.sourceforge.net/. MayBMS is as a backend extension
of PostgreSQL. It offers all the traditional querying capabilities of the latter in addition to the
uncertain and probabilistic’s.
5
2.1. RUNNING EXAMPLE
18
Figure 2.2. Predictive analytics in a data-intensive hypothesis evaluation study:
hypotheses (simulated data) compete to explain a phenomenon (observed data).
rise to alternative worlds as maximal-subset repairs of an argument key.
Predictive analytics tool. In database research (e.g., [36]), uncertainty
is usually seen as an undesirable property that hinders data quality. We shall
refer to U-relations and p-WSA as implemented in MayBMS, nonetheless, to show
that the ability to introduce controlled uncertainty into an (otherwise complete)
simulation dataset can be a tool for ‘deep’ predictive analytics on a set of competing
or alternative hypotheses. Fig.2.2 shows such a scenario of hypotheses ‘as data’
compete to explain a phenomenon ‘as data.’
As a roadmap to most of the remainder of this chapter, we claim that if hypotheses can be encoded and identified (see §2.2), and their uncertainty quantified
by some probability distribution (see §2.4), then they can be rated/ranked and
browsed by the user under selectivity criteria. Furthermore, their probabilities can
be conditioned for possibly being re-ranked in the presence of evidence (see §2.5).
2.1. Running Example
Let us consider Example 1 for the presentation of the vision.
Example 1 A research is conducted on the effects of gravity on a falling object
in the Earth’s atmosphere. Scientists are uncertain about the precise object’s
2.1. RUNNING EXAMPLE
19
density and its predominant state as a fluid or a solid. Three hypotheses are then
considered as alternative explanations of the fall (see Fig. 2.3). Due to parameter
uncertainty, six simulation trials are run for H1 , and four for H2 and H3 each. 2
PHENOMENON
φ
1
Description
Effects of gravity on an object falling in
the Earth’s atmosphere.
HYPOTHESIS
υ
1
2
3
Name
Law of free fall
Stokes’ law
Velocity-squared law
Figure 2.3. Descriptive (textual) data of Example 1.
H1
tid
1
2
3
4
5
6
1
2
φ
1
1
1
1
1
1
1
1
υ
1
1
1
1
1
1
1
1
t
0
0
0
0
0
0
0.1
...
g
32.0
32.0
32.0
32.2
32.2
32.2
32.0
32.0
v0
0
10
20
0
10
20
0
0
s0
5000
5000
5000
5000
5000
5000
5000
5000
a
−32.0
−32.0
−32.0
−32.2
−32.2
−32.2
−32.0
−32.0
v
0
10
20
0
10
20
−3.2
...
s
5000
5000
5000
5000
5000
5000
4999.84
...
Figure 2.4. ‘Big’ fact table H1 of hypothesis υ = 1 loaded with simulation raw
data: trials on H1 are identified by tid.
The construction of Υ-DB, a Data Warehouse (DW), requires a simple user description of a research. That is, descriptive records of the phenomena and hypotheses
dimensions (see Fig. 2.3) are to be inserted first such that basic referential constraints are satisfied by their associated datasets (fact tables). For instance, each
one of the six trial datasets for hypothesis H1 shall reference its id υ = 1 as a
foreign key from table HYPOTHESIS further in their synthesized relations.
Fig. 2.4 shows the ‘big’ fact table H1 for hypothesis υ = 1 loaded with its trial
datasets for phenomenon φ = 1. Although table H1 is denormalized for faster data
retrieval as usual in DW’s, the extraction of the hypothesis equations allows to
render it automatically since all variables must appear in some equation. Now we
proceed to the hypothesis encoding and start to address research questions RQ1-4.
2.2. HYPOTHESIS ENCODING
20
2.2. Hypothesis Encoding
We aim at extracting, for each hypothesis, a set of fd’s from its mathematical
equations. Suppose we are given a set of equations of hypothesis H1 below, and
let us examine the set Σ1 of fd’s we target at.8
H1 . Law of free fall
a(t) = g
v(t) = −gt + v0
s(t) = −(g/2)t2 + v0 t + s0
Σ1 = { φ
φ
φ
gυ
g v0 t υ
g v0 s0 t υ
→
→
→
→
→
→
g,
v0 ,
s0 ,
a,
v,
s }.
In order to derive Σ1 from the equations of H1 , we focus on their implicit data dependencies and get rid of constants and possibly complex mathematical constructs.
Equation v(t) = −gt+v0 , e.g., written this way (roughly speaking), suggests that
v is a prediction variable functionally dependent on t (the physical dimension), g
and v0 (the parameters). Yet a dependency like g v0 t → v may hold for infinitely
many equations.9 In fact, we need a way to identify H1 ’s mathematical formulation precisely, i.e., an abstraction of its data-level semantics. This is achieved by
introducing hypothesis id υ as a special attribute in the fd (see Σ1 ).
This is a data representation of a deterministic scientific hypothesis. It is built
into an encoding scheme (see §3.4) that leverages the semantics of structural
equations.
The other special attribute, the phenomenon id φ, is supposed to be a key to the values of parameters, i.e., determination of parameters is an empirical, phenomenondependent task. The fd φ → g v0 s0 is to be (expectedly) violated when the user
8
9
Recall that a rigorous presentation of the method to encode fd set Σ1 is due by Chapter 3.
Think of, say, how many polynomials satisfy that dependency ‘signature.’
2.3. REASONING OVER FD’S
21
is uncertain about the values of parameters. The same rationale applies to derive
Σ2 = Σ3 from the equations of H2 , H3 below. These, n.b., vary in structure w.r.t.
H1 (e.g., they include parameter D, the object’s diameter).
H2 . Stokes’ law
H3 . Velocity-squared law
a(t) = 0
a(t) = 0
p
v(t) = − gD/ 4.6×10−4
p
s(t) = −t gD/ 4.6×10−4 +s0
v(t) = −gD2 / 3.29×10−6
s(t) = −(gD2 / 3.29×10−6 ) t+s0
Σ2 = Σ3 = { φ
φ
φ
φ
gDυ
g D s0 t υ
→
→
→
→
→
→
g,
D,
s0 ,
a,
v,
s }.
The key point here is that, if the hypothesis structure (set of equations) is
given in a machine-readable format for mathematics, then the method to extract
the hypothesis fd set from its equations can be carefully designed based on such hypothesis data representation abstraction. In fact, we shall explore W3C’s MathML
as a format for hypothesis specification.10
2.3. Reasoning over FD’s
Once each hypothesis fd set has been extracted, some reasoning is to be performed
to discover implicit data dependencies. In fact, dependency theory is equipped
with a formal system (cf. §4.1) for reasoning over fd sets like Σ1 and derive other
fd’s in its closure Σ+
1 . As we elaborate on in Chapter 4, we shall be particularly
concerned with the pseudo-transitivity inference rule. Applied over fd’s { φ →
g, g υ → a } ⊂ Σ1 , for instance, it gives us φ υ → a . This inference allows us
to observe that {g} is a ‘factor’ on the uncertainty of a, but {φ υ} should be a
dimensional key constraint for values of a.
In fact, note that derived fd’s like hφ υ, ai ∈ Σ+
1 , which should be a constraint
on values of a in H1 , are (expectedly) violated in the presence of uncertainty: ob10
http://www.w3.org/Math/.
2.4. UNCERTAINTY INTRODUCTION
22
serve in Fig. 2.4 the multiplicity {32.0, 32.2} of values of a under the same pair
(φ 7→ 1, υ 7→ 1), which should functionally determine them in H1 . For that reason
we admit a special attribute ‘trial id’ tid to be overimposed into H1 for a trivial
repair, provisionally, until uncertainty can be introduced in a controlled way by
synthesis ‘4U.’ It is meant to identify simulation trials and “pretend” certainty not
to lose the integrity of the data. It is under this imposed certainty that the raw
simulation trial data is safely loaded from files (see Fig. 2.4). Note, however, how
‘certainty’ is held at the expense of redundancy and, mostly important, opaqueness for predictive analytics (since tid isolates or hides the inconsistency w.r.t. to
the violated constraints). This is until the next stage of the Υ-DB construction
pipeline, when uncertainty is to be introduced in a controlled manner.
2.4. Uncertainty Introduction
Before we proceed to the uncertainty introduction procedure, note in relation
H1 (Fig. 2.4), that the predicted acceleration values a are such that an association between the hypothesis and a target phenomenon, viz., (φ 7→ 1, υ 7→ 1) is
established. In fact, as of the insertion of each hypothesis trial dataset, the user
must set for it a target phenomenon. This may be non-obvious but is quite convenient a design decision for the envisioned system of Υ-DB because hypotheses,
as (abstract) universal statements [15], can only be derived predictions from (be
empirically grounded) by assigning (callibrating) them onto some real-world phenomenon. This assignment is set at data entry time because in fact it only holds
at the data level.11
It is to be recorded in an ‘explanation’ table named H0 by
default (see Fig. 2.5, top), being provided with weights for establishing a prior
probability distribution which (by user choice) may or may not be uniform.
The data transformation of ‘certain’ to ‘uncertain’ relations then starts with
query Q0 , whose result set is materialized into U-relational table Y0 (see Fig. 2.5).
As we introduce in detail in §5.1, U-relations have in their schema a set of pairs
(Vi , Di ) of condition columns [18] to map each discrete random variable xi created by the repair-key operation to one of its possible values (e.g., x0 7→ 1). The
11
Hypotheses are ‘universal’ by definition [15]. They (must) qualify for a class of different
situated phenomena, while its predictive datasets must be very specific (for one specific situation).
2.4. UNCERTAINTY INTRODUCTION
H0
φ
1
1
1
υ
1
2
3
Conf
3
1
1
Y0
23
V 7→ D
x0 7→ 1
x0 7→ 2
x0 7→ 3
φ
1
1
1
υ
1
2
3
W
V 7→ D
x0 7→ 1
x0 7→ 2
x0 7→ 3
Pr
.6
.2
.2
Figure 2.5. ‘Explanation’ relational table H0 and its associated U-relational
table Y0 rendered by application of the repair-key operation.
world table W is internal to MayBMS’ and automatically stores their marginal
probabilities. The formal semantics the repair-key operation is given in §5.1.
Q0 . create table Y0 as select φ, υ from (repair key φ in H0 weight by Conf);
The possible-world semantics of p-DB’s (cf. §5.1) can be seen as a generalization of data cleaning. In the context of p-DB’s [17], data cleaning does not
have to be one-shot — which is more error-prone [37]. Rather, it can be carried
out gradually, viz., by keeping all mutually inconsistent tuples under a probability
distribution (ibid.) that can be updated in face of evidence until the probabilities
of some tuples eventually tend to zero to be eliminated. This motivates Remark 1.
Remark 1 Consider U-relational table Y0 (Fig. 2.5). Note that it abstracts the
goal of a data-intensive hypothesis evaluation study, or the scientific method itself [15], as the repair of each φ as a key. That is, in Υ-DB users can develop
their research directly upon data with support of query and update capabilities to
rate/rank their hypotheses υ w.r.t. each φ, until the relationship r(φ, υ) is repaired
to be a function f : Φ → Υ from each phenomenon φ to its best explanation υ. 2
Given a ‘big’ fact table such as H1 , we need to identify/group the correlated
input attributes under independent uncertainty units, viz., ‘u-factors,’ each one
associated with a random variable.12
We illustrate that by means of query Q1 ,
which materializes view Y1 [g] for (let g = Zi ) identified u-factor Zi ⊆ Z in H1 [φ, Z].
12
An attribute can be inferred ‘input’ (viz., a parameter) by means of fd reasoning (cf. §3.4).
2.4. UNCERTAINTY INTRODUCTION
H1 [tid, φ, Z]
tid
1
2
3
4
5
6
φ
1
1
1
1
1
1
g
32
32
32
32.2
32.2
32.2
24
v0
0
10
20
0
10
20
s0
5000
5000
5000
5000
5000
5000
Y1 [g]
W
V 7→ D
x1 7→ 1
x1 7→ 2
V →
7 D
···
x1 7→ 1
x1 7→ 2
φ
1
1
g
32
32.2
Pr
···
.5
.5
Figure 2.6. Result set of query Q1 on simulation trial dataset for hypothesis H1 .
Q1 . create table Y1 [g] as select U.φ, U.g from (repair key φ in
(select φ, g, count(*) as Fr from H1 group by φ, g)
weight by Fr) as U;
The result set of Q1 is stored in Y1 [g], see Fig. 2.6. Note that the possible values
of g are mapped to random variable x1 , and that table H1 is considered source
for a joint probability distribution (on the values of H1 ’s input parameters) which
may not be uniform: we count the frequency Fr of each possible value of a u-factor
Zi ⊆ Z (as done for g in Q1 ) and pass it as argument to the weight-by construct.
So far, we have presented informally the procedure of u-factorization. Now
we proceed to u-propagation — both are presented rigorously in Chapter 5. We
consider g υ → a ∈ Σ1 again in order to synthesize predictive U-relation Y1 [a].
Since a is functionally determined by υ and g only, and these are independent, we
propagate their uncertainty onto a into Y1 [a] by query Q2 .
Q2 . create table Y1 [a] as select H1 .φ, H1 .υ, H1 .a from H1 , Y0 , Y1 [g] as G
where H1 .φ=Y0 .φ and H1 .υ=Y0 .υ and G.φ=H1 .φ and G.g=H1 .g;
Query Q02 (not shown) then selects φ, υ and a from Yi [a] for each i = 1..3. The
result sets of Q2 and Q02 (resp. Y1 [a] and Y [a]) are shown in Fig. 2.7.
2.5. PREDICTIVE ANALYTICS
Y1 [a]
Y [a]
25
V0 7→ D0
x0 7→ 1
x0 7→ 1
V0 7→ D0
x0 7→ 1
x0 7→ 1
x0 7→ 2
x0 7→ 3
V1 7→ D1
x1 7→ 1
x1 7→ 2
V1 7→ D1
x1 7→ 1
x1 7→ 2
−
−
φ
1
1
φ
1
1
1
1
υ
1
1
υ
1
1
2
3
a
−32
−32.2
a
−32
−32.2
0
0
Figure 2.7. U-relational predictive tables rendered by query using the fd’s.
Compare relations H1 [a] and Y1 [a]. By accounting for the correlations captured in the fd g υ → a, we could propagate onto a the uncertainty coming from
the hypothesis and the only parameter a is sensible to, thus precisely situating
tuples of Y1 [a] in the space of possible worlds. The same is done for predictive attributes v and s. In the end, Υ-DB shall be ready for predictive analytics, i.e., with
all competing predictions as possible alternatives which are mutually inconsistent.
A key point here is that all the synthesis process is amenable to algorithm
design. Except for the user ‘research’ description, the Υ-DB construction is fully
automated based on the hypothesis structure (set of equations) and the raw hypothesis trial data.
2.5. Predictive Analytics
Users of Example 1, has to be able, say, to query phenomenon φ = 1 w.r.t.
predicted position s at specific values of time t by considering all hypotheses υ
admitted. That is illustrated by query Q3 , which creates integrative table Y [s];
and by query Q4 , which computes the confidence aggregate operation [18] for all s
tuples where t = 3 (Fig. 2.8 shows Q4 ’s result, apart from column Posterior).
The confidence on each hypothesis for the specific prediction of Q4 is split
due to parameter uncertainty such that they sum up back to its total confidence.
For H2 and H3 , e.g., we have {g D s0 t υ → s} ⊂ Γ, where Γ = Σ2 = Σ3 . Since g
and D are the parameter uncertainty factors of s (s0 is certain), with 2 possible
values (not shown) each, then there are only 2 × 2 = 4 possible s tuples for H2 and
H3 each. Considering all hypotheses υ for the same phenomenon φ, the confidence
2.5. PREDICTIVE ANALYTICS
26
values sum up to one in accordance with the laws of probability.
Q3 . create table Y [s] as select U.φ, U.υ, U.t, U.s from
(select φ, υ, t, s from Y1 [s] union all
select φ, υ, t, s from Y2 [s] union all
select φ, υ, t, s from Y3 [s]) as U, Y0
where U.φ=Y0 .φ and U.υ=Y0 .υ;
Q4 . select φ, υ, s, conf() as Prior from Y [s] where t=3
group by φ, υ, s order by Prior desc;
Y [s]
φ
1
1
1
1
1
1
1
1
1
1
1
1
1
1
υ
1
1
1
1
1
1
2
2
2
2
3
3
3
3
s
2188.36
2205.82
2320.51
2337.97
2452.66
2470.12
2930.59
2943.44
4991.92
4991.97
4778.87
4779.56
4944.72
4944.89
Prior
.1
.1
.1
.1
.1
.1
.05
.05
.05
.05
.05
.05
.05
.05
Posterior
.167
.168
.167
.165
.149
.145
.020
.019
.000
.000
.000
.000
.000
.000
Figure 2.8. Analytics on predicted position s conditioned on observation.
Users can make informed decisions in light of such confidence aggregates,
which are to be eventually conditioned in face of evidence (observed data). Example 2 features such kind of Bayesian conditioning for discrete random variables
mapped to the possible values of predictive attributes (like position s) whose domain are continuous.
Example 2 Suppose position s = 2250 feet is observed at t = 3 secs, with standard deviation σ = 20. Then, by applying Bayes’ theorem for normal mean with
a discrete prior [16], Prior is updated to Posterior (see Fig. 2.8). 2
2.6. RELATED WORK
27
The procedure uses normal density function (2.1), with (say) σ = 20, to get
the likelihood f (y | µk ) of each alternative prediction of s from Y [s] as mean µk
given y at observed s = 2250. Then it applies Bayes’ rule (2.2) to get the posterior
p(µk | y) [16].
1
1
2
e− 2σ2 (y−µk )
2
2πσ
P
p(µk | y) = f (y | µk ) p(µk ) / ni=1 f (y | µi ) p(µi )
f (y | µk ) = √
(2.1)
(2.2)
In the general case (cf. examples shown in Chapter 6), we actually have phenomenon ‘as data:’ a sample of independent observed values y1 , ..., yn (e.g., Brazil’s
population observed by census over the years). Then, the likelihood f (y1 , ..., yn | µk )
Q
for each competing trial µk , is computed as a product nj=1 f (yj | µkj ) of the single likelihoods f (yj | µkj ) [16]. Bayes’ rule is then settled by (2.3) to compute the
posterior p(µk | y1 , ..., yn ) given prior p(µk ).
Qn
p(µk | y1 , . . . , yn ) =
j=1 f (yj
m Y
n
X
| µkj ) p(µk )
(2.3)
f (yj | µij ) p(µi )
i=1 j=1
As a result, the prior probability distribution assigned to u-factors via repair key
is to be eventually conditioned on observed data. This is an applied Bayesian inference problem that translates into a p-DB update one to induce effects of posteriors
back to table W . In a first prototype of the Υ-DB system (cf. 6.3), we accomplish it by performing Bayesian inference at application level and then applying
p-WSA’s update (a variant of SQL’s update) into MayBMS. This solution is good
enough to let us complete use case demonstrations of Υ-DB.13
2.6. Related Work
The vision of managing hypotheses as data has some roots in Porto and
Spaccapietra [38], who motivated a conceptual data model to support (the socalled) in silico science by means of a scientific model management system. We
13
Cf. Chapter 6, and §6.3 in particular.
2.6. RELATED WORK
28
discuss now the work we understand to be mostly related to our vision of datadriven hypothesis management and analytics.
2.6.1 Models-and-data
Haas et al. [35] provide an original long-term perspective on the evolution of
database technology. They characterize the data typically managed by traditional
DB systems as a record about the past, not a conclusion or an insight or a solution
(ibid.). In the context of scientific databases, e.g., their position is suggestive
that DB technology has been designed for empirical data, not the theoretical data
generated by simulation from domain-specific principles or scientific hypotheses.
They recognize current DB technology to have raised the art of scalable
‘descriptive’ analytics to a very high level; but point out, however, that nowadays
(sic.) what enterprises really need is ‘prescriptive’ analytics to identify optimal
business, policy, investment, and engineering decisions in the face of uncertainty.
Such analytics, in turn, shall rest on deep ‘predictive’ analytics that go beyond mere
statistical forecasting and are imbued with an understanding of the fundamental
mechanisms that govern a system’s behavior, allowing what-if analyses [35]. In
sum, there is a pressing call for deep predictive analytic tools in business enterprises
as much as in science’s.
In comparison with the Υ-DB vision, Haas et al. are proposing a long-term
models-and-data research program to pursue data management technology for deep
predictive analytics. They discuss strategies to extend query engines for model
execution within a (p-)DB. Along these lines, query optimization is understood as
a more general problem with connections to algebraic solvers.
Our framework in turn essentially comprises an abstraction and technique
for the encoding of hypotheses as data. It can be understood (in comparison) as
putting models strictly into a flattened data perspective. For that reason it has
been directly applicable by building upon recent work on p-DBs [17]. In principle,
it can be integrated into, say, the OLAP layer of the models-and-data project.
2.6.2 Scientific simulation data
As previsouly mentioned, science’s ETL is distinguished by its unfrequent, incremental-only updates and by having large raw files as data sources [8]. Challenges for
2.6. RELATED WORK
29
enabling an efficient access to high-resolution, raw simulation data have been documented from both supercomputing,[6] and database research viewpoints;[39] and
pointed as key to the use case of exploratory analytics. The extreme scale of the
raw data has motivated such non-conventional approaches for data exploration,
viz., the ‘immersive’ query processing (move the program to the data) [6, 40], or
‘in situ’ query processing in the raw files [41, 42]. Both exploit the spatial structure
of the data in their indexing schemes.
That line of research is motivated for equipping scientist end-users for an
immediate interaction with their very large simulation datasets.14
The NoDB
approach, in particular, argues to eliminate such ETL phase (viz., the loading) for
a direct access to data ‘in situ’ in the raw data files [42]. In fact, data exploration
is a fundamental use case of data-driven science.
Nonetheless, being generated from first principles or learned deterministic
hypotheses, simulation data has a pronounced uncertainty component that motivates a another use case, viz., the case of hypothesis management and predictive
analytics [35, 10]. As we have motivated in §1.1, the latter requires probabilistic
DB design for enabling uncertainty decomposition (factorization).
Hypothesis management shall not deal with the same volume of data as in
simulation data management for exploratory analytics, but samples of it (cf. Table
1.1 for a comparison). For instance, in CERN’s particle-physics experiment ATLAS
there are four tier/layers of data management. The volume of data significantly
decreases from the (tier-0) raw data to the (tier-3) data actually used for analyses
such as hypothesis testing [8, p. 71-2].
Overall, the overhead incurred in loading samples of raw simulation trial
datasets into a p-DB is justified for enabling a principled hypothesis evaluation
and rating/ranking according to the scientific method.
2.6.3 Hypothesis encoding
Our framework is comparable with Bioinformatics’ initiatives that address
hypothesis encoding into the RDF data model [43]: (i) the Robot Scientist [44] is a
knowledge-base system (KBS) for automated generation and testing of hypotheses
14
Sometimes phrased ‘here is my files, here is my queries, where are my results?’ [41].
2.7. SUMMARY: KEY POINTS
30
about what genes encode enzymes in the yeast organism; (ii) HyBrow [45] is a KBS
for scientists to test their hypotheses about events of the galactose metabolism
also of the yeast organism; and (iii) SWAN [46] is a KBS for scientists to share
hypotheses on possible causes of the Alzheimer disease.
The Robot Scientist relies on rule-based logic programming analytics to automatically generate and test RDF-encoded hypotheses of the kind ‘gene G has
function A’ against RDF-encoded empirical data [44]. HyBrow is likewise, but
hypotheses are formulated by the user about biological events [45]. SWAN in
turn disfavors analytic techniques for hypothesis evaluation and focus on descriptive aspects: hypotheses are high-level natural language statements retrieved from
publications. Each ‘hypothesis’ is associated with lower-level ‘claims’ (both RDFencoded) that are meant to support it on the basis of some empirical evidence
(RDF-encoded gene/protein data). In particular, SWAN [46] differs from the former in that each hypothesis is unstructured, being then more related to efforts on
the retrieval of textual claims from the narrative fabric of scientific reports [47].
All of them though, consist in some ad-hoc RDF encoding of sequence and
genome analysis hypotheses under varying levels of structure (viz., from ‘gene G
has function A’ statements to free text). Our framework in turn consists in the
U-relational encoding of hypotheses from mathematical equations, which is (to our
knowledge) the first work on hypothesis relational encoding.
Finally, as for hypothesis evaluation and comparison analytics, the Υ-DB
vision is distinguished in terms of its Bayesian inference approach. The latter has
been pointed out as a major direction for the improvement of the Bioinformatics’
initiatives just mentioned (cf. [43, p. 13]), and is in fact an influential model of
decision making for hypothesis evaluation [15, p. 220].
2.7. Summary: Key Points
We outline some key points in the Υ-DB vision:
• ‘Structured deterministic hypotheses’ are encoded as theoretical data and
distinguished from empirical data by the introduction of an epistemological
dimension into their semantic structure.
2.7. SUMMARY: KEY POINTS
31
• Two sources of uncertainty are considered: theoretical uncertainty, originating from competing hypotheses; and empirical uncertainty, derived from
alternative simulation trials on each hypothesis for the same phenomenon.
• A method to extract the structure of a hypothesis can be carefully designed
based on a hypothesis data representation and shall be reducible in terms of
machine-readable format for mathematical modeling, viz., W3C’s MathML,
which we shall adopt as a standard for hypothesis specification.
• We have seen that the controlled introduction of uncertainty into simulation data is amenable to algorithm design and then reducible to a designtheoretic synthesis method for the construction of U-relational DB’s.
• Simulation data can be modeled as hypothesis data whenever it is associated with a target phenomenon. As the same phenomenon may happen
to be associated with many such hypotheses, the research activity can be
modeled as a data cleaning problem in p-DB’s.
Essentially, the vision of Υ-DB comprises a design-theoretic pipeline (Fig.
1.4). For the insertion of a hypothesis k, we shall be given a MathML-compliant
structure Sk together with its simulation trial datasets Dk` in raw files (e.g., .mat,
.csv). Then we apply an Extract-Transform-Load (ETL) automatic procedure to
generate the hypothesis ‘big’ fact table Hk under the trial id’s.
The extracted equations are firstly encoded into fd’s. Then, at any time, as
many hypotheses may have been inserted into the system, the uncertainty introduction (U-intro) procedure can be applied to process the encoded fd’s and synthesize
the ‘uncertain’ U-relations that are to be eventually conditioned on observations.
Note, in Fig. 1.4, that the ETL procedure is operated in a ‘local’ view for
each hypothesis k, while the U-intro procedure and the conditioning are operated
in the ‘global’ view of all available hypotheses k = 1..n. The pipeline opens up
four main tracks of technical research challenges from the ETL stage on, viz., (i)
hypothesis encoding and (ii) causal reasoning over fd’s, (iii) p-DB synthesis and
(iv) conditioning. We address in the sequel the three first track of challenges in
depth. The problem of conditioning is outlined for further work in §7.3.
Chapter 3
Hypothesis Encoding
In this chapter we address the problem of hypothesis encoding. In §3.1 we
introduce notation and basic concepts of structural equations and the problem of
causal ordering. In §3.2 we study the problem of extracting the causal ordering
implicit in the structure of a deterministic hypothesis and show that Simon’s classical approach [24, 48] is intractable. In §3.3 then we build upon a less notorious
approach of Nayak’s [49] and borrow an efficient algorithm for it that fits very
well our use case for hypothesis encoding. In §3.4 we develop an encoding scheme
that builds upon the idea of structural equations through an original abstraction of
hypotheses ‘as data.’ In §3.5 we present experiments that attest how the encoding
scheme works in practice for large hypotheses. In §3.6 we discuss related work. In
§3.7 we summarize the results of this chapter.
3.1. Preliminaries: Structural Equations
Given a system of mathematical equations involving a set of variables, to build a
structural equation model (SEM) is, essentially, to establish a one-to-one mapping
between equations and variables [24]. That shall enable further detecting the
hidden asymmetry between variables, i.e., their causal ordering. For instance,
Einstein’s famous equation E = m c2 states the equivalence of mass and energy,
summarizing a theory that can be imputed two different asymmetries (for different
applications), say, given a fixed amount of mass m = m0 (and recall c is a constant),
predict the particle’s relativistic rest energy E; or given the particle’s rest energy,
predict its mass or potential for nuclear fission.
3.1. PRELIMINARIES: STRUCTURAL EQUATIONS
F = 10 [N ],
m = 5 [kg],
F = ma
a = 90 [m/s2 ],
m = 5 [kg],
F = ma
33
F
m
a
F
m
a
Figure 3.1. “Directed causal graphs” associated with the two systems.
To stress the point, consider Newton’s second law F = m a in such a scalar
setting. The modeler can either use it to compute (predict), say, acceleration
values given an amount of mass and different force intensities, or to predict force
intensities given a fixed acceleration (e.g., for testing an engineered dynamometer).
The point here is that Newton’s equation is not enough to derive predictions. That
is, it has a number of variables |V| = 3, which is larger than |E| = 1. It must be
completed with two more equations in order to qualify as an (applied) hypothesis
‘as data.’ Although usually it is interpreted an asymmetry towards a, technically,
there is nothing in its semantics to suggest so.1
Compare the two systems given
in Fig. 3.1.2 In sum, the causal ordering of any system of equations is not to be
guessed, as it can be inferred. In this chapter we rely on previous work (mostly
AI’s work, viz., [24, 49, 48]) and adapt it for the encoding of hypotheses into fd’s.
Def. 1 A structure is a pair S(E, V), where E is a set of equations over set V of
variables, |E| ≤ |V|, such that:
(a) In any subset of k equations of the structure, at least k different variables
appear;
(b) In any subset of k equations in which r variables appear, k ≤ r, if the
values of any (r − k) variables are chosen arbitrarily, then the values of the
1
2
As the equality construct ‘=’ is used as a predicate, not an assignment operator.
We shall introduce the notion of ‘directed causal graphs’ shortly.
3.1. PRELIMINARIES: STRUCTURAL EQUATIONS
34
remaining k variables can be determined uniquely — finding these unique
values is a matter of solving the equations.
Def. 2 Let S(E, V) be a structure. We say that S is self-contained or complete
if |E| = |V|.
In short, we are interested in systems of equations that are ‘structural’ (Def. 1) and
‘complete’ (Def. 2), viz., that has as many equations as variables and no subset of
equations has fewer variables than equations.3
Complete structures can be solved for unique sets of values of their variables.
In this work, however, we are not concerned with solving sets of mathematical
equations at all, but with processing their causal ordering in view of U-relational
DB design. Simon’s concept of causal ordering has its roots in econometrics studies
(cf. [24]) and has been taken further in AI with a flavor of Graphical Models (GMs)
[50, 25, 48]. In this thesis we translate the problem of causal ordering into the
language of data dependencies, viz., into fd’s.
Def. 3 Let S be a structure. We say that S is minimal if it is complete and there
is no complete structure S 0 ⊂ S.
Def. 4 The structure matrix AS of a structure S(E, V), with f1 , f2 , . . . , fn ∈ E
and x1 , x2 , . . . , xm ∈ V, is a n × m matrix of 1’s and 0’s in which entry aij is
non-zero if variable xj appears in equation fi , and zero otherwise.
Elementary row operations (e.g., row multiplication by a constant) on the
structure matrix may hinder the structure’s causal ordering and then are not valid
in general [24]. This also emphasizes that the problem of causal ordering is not
about solving the system of mathematical equations of a structure, but identifying
its hidden asymmetries.
Def. 5 Let S(E, V) be a complete structure. Then a total causal mapping over
S is a bijection ϕ : E → V such that, for all f ∈ E, if ϕ(f ) = x then x ∈ V ars(f ).
3
Also, we expect the systems of equations given as input to be ‘independent’ in the sense of
Linear Algebra. In our context, that means systems that can only have non-redundant equations.
In that case, if some subset of equations has fewer variables than equations, then the system must
be ‘overconstrained.’
3.1. PRELIMINARIES: STRUCTURAL EQUATIONS
35
x1 x 2 x3 x4 x5 x6 x7
x1 x2 x3 x4 x5 x6 x 7
f1
f2
f3
f4
f5
f6
f7
1
0
0
1
1
0
0
0
1
0
1
0
0
0
0
0
1
1
1
0
0
0
0
0
1
1
1
0
0
0
0
1
1
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
1
(a) Structure matrix as given.
→
f1
f2
f3
f4
f5
f6
f7
1
0
0
1
1
0
0
0
1
0
1
0
0
0
0
0
1
1
1
0
0
0
0
0
1
1
1
0
0
0
0
1
1
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
1
(b) COA execution in 3 recursive steps.
Figure 3.2. Running Simon’s Causal Ordering Algorithm (COA) on a given
structure (Fig. 3.2a). Minimal subsets detected in a recursive step k, highlighted
in different shades of gray, have their diagonal elements colored (Fig. 3.2b).
Simon has informally described an algorithm (cf. [24]) that, given a complete structure S(E, V), can be used to compute a partial causal mapping ϕp from partitions
on the set of equations to same-cardinality partitions on the set of variables. As
shown by Dash and Druzdzel [48], the causal mapping returned by Simon’s (socalled) Causal Ordering Algorithm (COA) is not total when S has variables that
are strongly coupled (because they can only be determined simultaneously). They
also have shown that any total mapping ϕ over S must be consistent with COA’s
partial mapping ϕp [48]. The latter is made partial by design (merge strongly
coupled variables into partitions or clusters) in order to force its induced causal
graph Gϕp to be acyclic. Algorithm 1, COAt , is a variant of Simon’s COA adapted
to illustrate our use case. It returns a total causal mapping ϕ, instead of a partial
causal mapping. We illustrate it through Example 3 and Fig. 3.2.
Example 3 Consider structure S(E, V) whose matrix is shown in Fig. 3.2a. Note
that S is complete, since |E| = |V| = 7, but not minimal. The set of all minimal
subsets S 0 ⊂ S is Sc = { {f1 }, {f2 }, {f3 } }. By eliminating the variables identified at
recursive step k, a smaller structure T ⊂ S is derived. Compare the partial causal
mapping eventually returned by COA, ϕp = { h{f1 }, {x1 }i, h{f2 }, {x2 }i, h{f3 }, {x3 }i,
h{f4 , f5 }, {x4 , x5 }i, h{f6 }, {x6 }i, h{f7 }, {x7 }i }, to the total causal mapping returned by COAt , ϕ = {hf1 , x1 i, hf2 , x2 i, hf3 , x3 i, hf4 , x4 i, hf5 , x5 i, hf6 , x6 i, hf7 , x7 i}.
3.2. THE PROBLEM OF CAUSAL ORDERING
36
Algorithm 1 COAt as a variant of Simon’s COA.
1: procedure COAt (S : structure over E and V)
Require: S given is complete, i.e., |E| = |V|
Ensure: Returns total causal mapping ϕ : E → V
2:
ϕ ← ∅, Sc ← ∅
3:
for all minimal S 0 ⊂ S do
4:
Sk ← Sk ∪ S 0
. store minimal structures S 0 found in S
0
0
5:
V ← S (V)
6:
for all f ∈ S 0 (E) do
7:
x ← any xa ∈ V 0
8:
ϕ ← ϕ ∪ hf, xi
9:
V 0 ← V 0 \ {x}
S
10:
T ← S \ S 0 ∈Sk S 0
11:
if T 6= ∅ then
12:
return ϕ ∪ COAt (T )
13:
return ϕ
Since x4 and x5 are strongly coupled (see Fig.3.2b), COAt maps them arbitrarily
(e.g., it could be f4 7→ x5 , f5 7→ x4 instead). Such total mapping ϕ renders a cycle
in the directed causal graph Gϕ induced by ϕ (see Fig.3.3). 2
x1
x2
x4
x6
x3
x5
x7
Figure 3.3. Directed causal graph Gϕ induced by mapping ϕ for structure S. An
edge connects a node xi towards a node xj , with xi , xj ∈ V, iff xi appears in the
equation f ∈ E such that ϕ(f ) = xj .
3.2. The Problem of Causal Ordering
The serious issue with Alg. 1, COA(t ), is that finding all minimal structures
in a given structure (cf. line 3) is a hard problem that can only be addressed
heuristically as a problem of co-clustering (also called biclustering [51, 52]) in
3.2. THE PROBLEM OF CAUSAL ORDERING
37
Boolean matrices. Simon’s approach, however, as we shall see next, is not the only
way to cope with the problem of causal ordering.
In fact, in order to study the computational properties of SEM’s and the
problem of causal ordering, we observe that any structure S(E, V) satisfying Def.
1 can be modeled straightforwardly as a bipartite graph G = (V1 ∪ V2 , E), where
the set E of equations and the set V of variables are the disjoint vertex sets, i.e.,
V1 7→ E, V2 7→ V, and E 7→ S is the edge set connecting equations to the variables
appearing in them. Fig. 3.4 shows the bipartite graph G corresponding to the
structure given in Example 3 — for a comprehensive text on graph concepts and
its related algorithmic problems, cf. Even [53].
f1
x1
f2
x2
f3
x3
f4
x4
f5
x5
f6
x6
f7
x7
Figure 3.4. Bipartite graph G of structure S from Example 3.
A biclique (or complete bipartite graph) is a bipartite graph G = (A ∪ B, E)
such that for every two vertices a ∈ A, b ∈ B, we have (a, b) ∈ E [53]. Note that
for balanced bicliques, i.e., when |A| = |B| = K, the degree deg(u) of any vertex
u ∈ A ∪ B must be deg(u) = |A| = |B| = K.
Recent approaches to co-clustering problems (e.g., [54]) have come with the
notion of pseudo-biclique (also called ‘quasi-biclique’), which is a relaxation of the
biclique concept to allow some less rigid notion of connectivity than the ‘complete
connectivity’ required in a biclique. Now, recall that Simon’s COA(t ) needs to
find, at each recursive step, all minimal subsets S 0 ⊆ S. Theorem 1 situates this
particular computational task in terms of its complexity, which for |E 0 | = |V 0 | ≥ 2
is equivalent to find, at each recursive step, the minimal-size ‘pseudo-bicliques’
3.2. THE PROBLEM OF CAUSAL ORDERING
38
(i.e., with the least K ≥ 2) in its corresponding bipartite graph (e.g., see Fig. 3.4).
Here we take, in Def. 6, a specific notion of pseudo-biclique.
Def. 6 Let G = (A ∪ B, E) be a bipartite graph. We say that G is a K-balanced
pseudo-biclique if |A| = |B| = K with |E| ≥ 2K and, for all vertices u ∈ A ∪ B,
deg(u) ≥ 2.
Now we state (originally) the balanced pseudo-biclique problem (BPBP) as a decision problem as follows.
(BPBP). Given a bipartite graph G = (V1 ∪ V2 , E) and a positive integer
K ≥ 2, does G contain a K-balanced pseudo-biclique?
Lemma 1 The balanced pseudo-biclique problem (BPBP) is NP-Complete.
Proof 1 We show (by restriction [55]) that the BPBP is a generalization of the
balanced biclique problem (BBP), referred ‘balanced complete bipartite subgraph’
problem [55, GT24, p. 196], which is shown to be NP-Complete by means of a
transformation from ‘clique’ [56, p. 446]. The restriction from BPBP to BBP
(special case) is made by requiring (cf. Def. 6) either (a) |E| = K 2 or (b) deg(u) =
K, for K ≥ 2,4 which are equivalent ways of enforcing the inquired K-balanced
pseudo-biclique to be a K-balanced biclique. 2
We introduce another hypothesis structure (see Fig. 3.5) to illustrate the correspondence between the pseudo-biclique property and COA’s algorithmic approach
as elaborated in the proof of Theorem 1.
Theorem 1 Let S(E, V) be a complete structure. Then the extraction of its causal
ordering by Simon’s COA(S) is NP-Hard.
Proof 2 We show that, at each recursive step k of COA, to find all non-trivial
minimal subsets (i.e., |E 0 | ≥ 2) translates into an optimization problem associated
with the decision problem BPBP, which we know by Lemma 1 to be NP-Complete.
See Appendix §A.1.1. 2
4
Note that clearly, for any positive integer K ≥ 2, we have for K 2 ≥ 2K.
3.3. TOTAL CAUSAL MAPPINGS
x1 x2 x 3 x4
f1
f2
f3
f4
1
1
0
1
0
1
1
1
1
0
1
1
0
0
0
1
(a) COA execution in 2 recursive steps.
39
f1
x1
f2
x2
f3
x3
f4
x4
(b) Bipartite graph.
Figure 3.5. Another hypothesis structure example.
Nonetheless, the problem of causal ordering can be solved efficiently by means
of a different, less notorious approach due to Nayak [49], which we introduce and
build upon next.
3.3. Total Causal Mappings
The problem of causal ordering can be solved in polynomial time by (i)
finding any total causal mapping ϕ : E → V over structure S given (cf. Def. 5);
and then (ii) by computing the transitive closure Cϕ+ of the set Cϕ (cf. Eq. 3.1) of
direct causal dependencies induced by ϕ.
Cϕ = { (xa , xb ) | there exists f ∈ E such that ϕ(f ) = xb and xa ∈ V ars(f ) } (3.1)
Def. 7 Let S(E, V) be a structure with variables xa , xb ∈ V, and ϕ a total causal
mapping over S inducing set of direct causal dependencies Cϕ and its transitive
closure Cϕ+ . We say that (xb , xa ) is a direct causal dependency in S if (xb , xa ) ∈
Cϕ , and that (xb , xa ) is a causal dependency in S if (xb , xa ) ∈ Cϕ+ .
In other words, (xa , xb ) is in Cϕ iff xb direct and causally depends on xa ,
given the causal asymmetries induced by ϕ. Those notions open up an approach
to causal reasoning that fits very well to our use case, which is aimed at encoding
hypothesis structures into fd sets and then performing (symbolic) causal reasoning
in terms of acyclic pseudo-transitive reasoning over fd’s (cf. Chapter 4).5
For it
to be effective, nonetheless, we shall need to ensure some properties of total causal
mappings first.
5
Note that it differs from AI research (e.g., [48]) geared for reasoning over GM’s.
3.3. TOTAL CAUSAL MAPPINGS
40
For a given structure S, there may be multiple total causal mappings over
S (recall Example 3). But the causal ordering of S must be unique (see Fig. 3.3).
Therefore, a question that arises is whether the transitive closure Cϕ+ is the same
for any total causal mapping ϕ over S. Proposition 1, originally from Nayak [49],
ensures that is the case.
Proposition 1 Let S(E, V) be a structure, and ϕ1 : E → V and ϕ2 : E → V be
any two total causal mappings over S. Then Cϕ+1 = Cϕ+2 .
Proof 3 The proof is based on an argument from Nayak [49], which we present in
arguably much clearer way (see Appendix, §A.1.2). Intuitively, it shows that if ϕ1
and ϕ2 differ on the variable an equation f is mapped to, then such variables, viz.,
ϕ1 (f ) and ϕ2 (f ), must be causally dependent on each other (strongly coupled). 2
Another issue is concerned with the precise conditions under which total
causal mappings exist (i.e., whether or not all variables in the equations can be
causally determined). In fact, by Proposition 2, based on Nayak [49] apud. Hall
[53, p. 135-7], we know that the existence condition holds iff the given structure
is complete.
Before proceeding to it, let us refer to Even [53] to briefly introduce the
additional graph-theoretic concepts which are necessary here. A matching in a
graph is a subset of edges such that no two edges in the matching share a common
node. A matching is said maximum if no edge can be added to the matching
(without hindering the matching property). Finally, a matching in a graph is said
‘perfect’ if every vertex is an end-point of some edge in the matching — in a
bipartite graph, a perfect matching is said a complete matching.
Proposition 2 Let S(E, V) be a structure. Then a total causal mapping ϕ : E →
V over S exists iff S is complete.
Proof 4 We observe that a total causal mapping ϕ : E → V over S corresponds
exactly to a complete matching M in a bipartite graph B = (V1 ∪ V2 , E), where
V1 7→ E, V2 7→ V, and E 7→ S. In fact, by Even apud. Hall’s theorem (cf. [53,
135-7]), we know that B has a complete matching iff (a) for every subset of vertices
3.3. TOTAL CAUSAL MAPPINGS
41
F ⊆ V1 , we have |F | ≤ |E(F )|, where E(F ) is the set of all vertices connected
to the vertices in F by edges in E; and (b) |V1 | = |V2 |. By Def. 1 (no subset of
equations has fewer variables than equations), and Def. 2 (number of equations
is the same as number of variables), it is easy to see that conditions (a) and (b)
above hold iff S is a complete structure. 2
The problem of finding a maximum matching is a well-studied algorithmic
problem. In this thesis we adopt the Hopcroft-Karp algorithm [57], which is known
p
to be polynomial-time, bounded by O( |V1 | + |V2 | |E|).6 That is, we handle the
problem of total causal mapping by (see Alg. 2) translating it to the problem of
maximum matching in a bipartite graph (in linear time) and then applying the
Hopcroft-Karp algorithm to get the matching and finally translate it back to the
total causal mapping, as suggested by the proof of Proposition 2.
Algorithm 2 Find a total causal mapping for a given structure.
1: procedure TCM(S : structure over E and V)
Require: S given is a complete structure, i.e., |E| = |V|
Ensure: Returns a total causal mapping ϕ
2:
B(V1 ∪ V2 , E) ← ∅
3:
ϕ←∅
4:
for all hf, Xi ∈ S do . translates the structure S to a bipartite graph B
5:
V1 ← V1 ∪ {f }
6:
for all x ∈ X do
7:
V2 ← V2 ∪ {x}
8:
E ← E ∪ {(f, x)}
11:
M ← Hopcroft-Karp(B)
. solves the maximum matching problem
for all (f, x) ∈ M do . translates the matching to a total causal mapping
ϕ ← ϕ ∪ {hf, xi}
12:
return ϕ
9:
10:
Fig. 3.6 shows the complete matching found by the Hopcroft-Karp algorithm for
the structure given in Example 3.
6
The Hopcroft-Karp algorithm solves maximum matching in a bipartite graph efficiently as
a problem of finding maximum flow in a network (cf. [53, p. 135-7], or [58, p. 664-9]).
3.3. TOTAL CAUSAL MAPPINGS
42
f1
x1
f2
x2
f3
x3
f4
x4
f5
x5
f6
x6
f7
x7
Figure 3.6. Complete matching M for structure S from Example 3.
Corollary 1 summarizes the results we have so far.
Corollary 1 Let S(E, V) be a complete structure. Then a total causal mapping
ϕ : E → V over S can be found by (Alg. 2) TCM in time that is bounded by
p
O( |E| |S|).
Proof 5 Let B = (V1 ∪ V2 , E) be the bipartite graph corresponding to complete
structure S given to TCM, where V1 7→ E, V2 7→ V, and E 7→ S. The translation of
S into B is done by a scan over it. This scan is of length |S| = |E|. Note that number |E| of edges rendered is precisely the length |S| of structure, where the denser
the structure, the greater |S| is. The re-translation of the matching computed by
internal procedure Hopcroft-Karp, in turn, is done at expense of |E| = |V| ≤ |S|.
Thus, it is easy to see that TCM is dominated by the maximum matching algorithm
p
p
Hopcroft-Karp, which is known to be O( |V1 | + |V2 | |E|), i.e., O( |E| + |V| |S|).
p
√ p
Since S is assumed complete, we have |E| = |V| then |E| + |V| = 2 |E|.
p
Therefore, TCM must have running time at most O( |E| |S|). 2
Remark 2 Let S(E, V) be a complete structure. Then we know (cf. Proposition
2) that a total causal mapping over S exists. Let it be defined ϕ , TCM(S). Then
the causal ordering implicit in S shall be correctly extracted (cf. Proposition 1)
by processing the causal dependencies induced by ϕ, as we show in Chapter 4. 2
Now we are ready to accomplish the hypothesis encoding into fd’s, as we
show next.
3.4. THE ENCODING SCHEME
43
3.4. The Encoding Scheme
We shall encode variables as relational attributes and map equations onto
fd’s through total causal mappings. Let Z be a set of attribute symbols such that
Z ' V, where S(E, V) is a complete structure; and let φ, υ ∈
/ Z be two special
attribute symbols kept to identify (resp.) phenomena and hypotheses. We are
explicitly distinguishing symbols in Z, assigned by the user into structure S, from
epistemological symbols φ and υ. Then we consider a sense of Simon’s into the
nature of scientific modeling and interventions [24], summarized in Def. 8.
Def. 8 Let S(E, V) be a structure and x ∈ V be a variable. We say that x is
exogenous if there exists an equation f ∈ E such that V ars(f ) = {x}. In this
case f can be written f (x) = 0, and must be mapped to x in any total causal
mapping ϕ over S. We say that x is endogenous otherwise.
Remark 3 introduces an interpretation of Def. 8 with a data dependency flavor.
Remark 3 The values of exogenous variables (attributes) are to be determined
empirically, outside of the system (proposed structure S). Such values are, therefore, dependent on the phenomenon id φ only. The values of endogenous variables
(attributes) are in turn to be determined theoretically, within the system. They
are dependent on the hypothesis id υ and shall be dependent on the phenomenon
id φ as well indirectly. 2
As introduced in §2.2, the encoding scheme we are presenting here is not
obvious. It goes beyond Simon’s structural equations to abstract the data-level
semantics of mathematical deterministic hypotheses. Whereas Simon’s structural
equations are able to represent only linear equations, our encoding scheme can
represent non-linear equations and arbitrarily complex mathematical operators by
means of its data representation of deterministic hypotheses.
For instance, take non-linear equation y = a x2 and suppose that, considering
the context of its complete system of equations, (Alg. 2) TCM maps it onto variable
y. Then, by an abstraction of the equation semantics, we shall encode it into fd
3.4. THE ENCODING SCHEME
44
Σ={ φ
φ
φ
x1 x2 x3 x5 υ
x1 x3 x4 υ
x4 υ
x5 υ
→
→
→
→
→
→
→
x1 ,
x2 ,
x3 ,
x4 ,
x5 ,
x6 ,
x7 }.
Figure 3.7. Encoded fd set Σ (cf. Alg. 3) for the structure from Example 3.
a x υ → y. That is, the hypothesis identifier υ captures the data-level semantics of
the hypothesis equation.7
We encode complete structures into fd sets by means of (Alg. 3) h-encode.
Fig. 4.1 presents an fd set defined Σ , h-encode(S), encoding the same structure
S from Example 3.
Algorithm 3 Hypothesis encoding.
1: procedure h-encode(S : structure over E and V, D : domain variables)
Require: S given is a complete structure, i.e., |E| = |V|
Ensure: Returns a non-redundant fd set Σ
2:
Σ←∅
3:
ϕ ← TCM(S)
4:
for all hf, xi ∈ ϕt do
5:
Z ← X \ {x}, where hf, Xi ∈ S
6:
if Z = ∅ or Z ⊆ D then
. x is exogenous
7:
if x ∈
/ D then
. supress φ-fd for dimensions like time t
8:
Σ ← Σ ∪ hZ ∪ {φ}, xi
9:
10:
11:
else
Σ ← Σ ∪ hZ ∪ {υ}, xi
. x is endogenous
return Σ
Now we study the design-theoretic properties of the encoded fd sets. We
shall make use of the concept of ‘canonical’ fd sets (also called ‘minimal’ [20, p.
390]), see Def. 9.
Def. 9 Let Σ be an fd set. We say that Σ is canonical if:
7
Note that, without the hypothesis id, infinitely many equations fit the pattern a x → y.
3.5. EXPERIMENTS
45
(a) each fd in Σ has the form X → A, where |A| = 1;
(b) For no hX, Ai ∈ Σ we have (Σ − {hX, Ai})+ = Σ+ ;
(c) for each fd X → A in Σ, there is no Y ⊂ X such that (Σ\{X → A}∪{Y →
A})+ = Σ+ .
For an fd set satisfying such properties (Def. 9) individually, we say that it is (a)
singleton-rhs, (b) non-redundant and (c) left-reduced. It is said to have an attribute
A in X that is ‘extraneous’ w.r.t. Σ if it is not left-reduced (Def. 9-c) [22, p. 74].
Finally, an fd X → Y in Σ is said trivial if Y ⊆ X. Note that the presence of a
trivial fd in a an fd set is sufficient to make it redundant.
Theorem 2 Let Σ be an fd set defined Σ , h-encode(S) for some complete structure S. Then Σ is non-redundant and singleton-rhs but may not be left-reduced
(then may not be canonical).
Proof 6 We show that properties (a-b) of Def. 9 must hold for Σ produced by
(Alg. 3) h-encode, but property (c) may not hold (i.e., encoded fd set Σ may not
be left-reduced). See Appendix, §A.1.3. 2
We draw attention to the significance of Theorem 2, as it sheds light on a
connection between Simon’s complete structures [24] and fd sets [20]. In fact, we
continue to elaborate on that connection in next chapter to handle causal ordering
processing symbolically by causal reasoning over fd’s.
3.5. Experiments
Fig. 3.8 shows the results of experiments we have carried out in order to study
how effective the procedure of hypothesis encoding is in practice, in particular
its behavior for hypotheses whose structure S has been randomly generated over
orders of magnitude |S| ≈ 2k , to have length up to |S| ≈ 220 . 1M . The largest
structure considered, with |S| ≈ 1M , has been generated to have exactly |E| =
2.5K.
3.6. RELATED WORK
46
We have executed ten runs for each tested order of magnitude, and then
taken its mean running time in ms.8
The plot is shown in Fig. 3.8 in logscale
base 2. In fact a a near-, sub-quadratic slope is expected for the curve structure
length |S| × time.
These scalability results are compatible with the computational complexity
p
of h-encode, which is (cf. Corollary 1) bounded by O( |E| |S|).9
time [ms]
214
211
28
25
27
29
211
213
215
217
structure length |S|
219
221
Figure 3.8. Performance of hypothesis encoding (in logscale).
3.6. Related Work
Modeling physical and socio-economical systems as a set of equations is a traditional modeling approach, and a very large bulk of models exist up to date. Simon’s
early work on structural equations and causal ordering comprises a specific notion
of causality aimed at further contributing to the potential of such modeling approach (cf. [59]). It is meant for identifying influences among variables (or their
values) that are implicit in the system model for enabling informed interventions.
These may apply either to the system (phenomenon) under study, or to the model
itself (say, when its predictions are not approximating observations very well).
Significant research effort has been devoted to causal modeling and reasoning
in the past decades in both statistics and AI (cf. [25, 19]). The notion of causality
used can be traced back to the early work of Simon’s and others in Econometrics.
8
The experiments were performed on a 2.3GHZ/4GB Intel Core i5 running Mac OS X 10.6.8.
Note that, for any arbitrary structure S(E, V), we have |E| ≤ |S| ≤ |E|2 . So, in worst case
(the densest structure possible)
we have |S| = |E|2 and then can establish a time bound function
p
2
of |E| only, viz., O(|E| |E|).
9
3.6. RELATED WORK
47
Nonetheless, there are two important differences to be emphasized:
• Such work is majorly devoted to deal with (statistical) qualitative hypotheses, not (deterministic) quantitative hypotheses;
• The causal model is assumed as given or is derived from data, instead of
being converted or synthesized from a set of equations.
These are both core differences that also apply to our work in comparison
to the bulk of existing work in probabilistic DB’s. Our main point here, though,
is to clarify the technical context and state of the art of the problem of causal
ordering. A few works have been concerned with extracting a causal model out
of some previous existing formal specification such as a set of equations. This is
a reason why causal ordering has been an yet barely studied problem from the
computational point of view.
Dash and Druzdzel revisit the problem and re-motivate it in light of modern
applications [48]. First, they provide a formal description of how Simon’s COA
gives a summary of the causal dependencies implicit in a given SEM. That is,
in clustering the strongly coupled variables into a causal graph, COA provides a
condensed representation of the causal model implicit in the given SEM. They
show then that any valid total causal mapping produced for a given SEM must be
consistent with COA’s partial causal mapping.
Yet, the serious problem is that the algorithm turns out to be intractable. In
fact, no formal study of COA’s computational properties can yet be found in the
literature. In this thesis we have obtained the (negative) hardness result that it is
intractable, which turns out to be compatible with Nayak’s intuition (sic.) that it
is a worst-case exponential time algorithm (cf. [60, p. 37]).
Inspired on Serrano and Gossard’s work on constraint modeling and reasoning [61], Nayak reports an approach that is provably quite effective to process the
causal ordering: extract a total causal mapping and then compute the transitive
closure of the direct causal dependencies. In this thesis we build upon it to perform causal reasoning in terms of a form of transitive reasoning. Such approach
fits very well to our use case, viz., the synthesis of p-DB’s from fd’s. As we show
3.7. SUMMARY OF RESULTS
48
in Chapter 4, we process the causal ordering of a hypothesis structure (abstracted
as a SEM) in terms of acyclic causal reasoning over fd’s and prove its correctness.
This is enabled by the encoding scheme presented in this chapter.
3.7. Summary of Results
In this chapter we have studied and developed an encoding scheme to process
the mathematical structure of a deterministic hypothesis into a set of fd’s towards
the encoding of hypotheses ‘as data.’ Then we have studied the design-theoretic
properties held by such an encoded fd set. We list the results achieved as follows.
• By Theorem 1, we know (an original hardness result) that Simon’s approach to process the causal ordering of a structure is intractable;
• By building upon on the work of Simon [24] and Nayak [49] (cf. Propositions 1 and 2), we have framed an approach to efficiently extract the basic
information (a total causal mapping) for processing the causal ordering
implicit in the mathematical structure of a deterministic hypothesis;
• By Corollary 1, we know how to process the complete structure of a hypothp
esis into a total causal mapping in time that is bounded by O( |E| |S|).
That is, the machinery of hypothesis encoding is provably suitable for very
large hypothesis structures.
• By Theorem 2, which studies the design-theoretic properties of the encoded
fd sets, we have unraveled the connection between Simon’s complete structures and fd sets to further explore it in next chapter.
• We have performed experiments (cf. Fig. 3.8) to study how effective the
approach is in practice, or how it scales for hypotheses whose structure S
is randomly generated to have length up to the order of |S| . 1M .10
10
The tests were up to this order only because of the hardware limitations of our experimental
settings. In theory (cf. complexity time bounds), larger structures can be handled very efficiently.
Chapter 4
Causal Reasoning over FD’s
In this chapter we present a technique to address the problem of causal ordering processing in order to enable the synthesis of U-relational DB’s. In §4.1 we
introduce Armstrong’s classical inference system to reason over fd’s. In §4.2 we
develop the core concept and algorithm of the folding of an fd set, as a method for
acyclic causal reasoning over fd’s. In §4.3 we show its connections (equivalence)
with causal reasoning. In §4.4 we present experiments on how the method behaves
in practice. §4.5 we discuss related work. In §4.6 we conclude the chapter.
4.1. Preliminaries: Armstrong’s Inference Rules
As usual notational conventions from the DB literature [20, 21], we write X, Y, Z
to denote sets of relational attributes and A, B, C to denote singleton attribute
sets. Also, we write XY as shorthand for X ∪ Y .
Functional dependency theory relies on Armstrong’s inference rules (or axioms) of (R0) reflexivity, (R1) augmentation and (R2) transitivity, which forms
a sound and complete inference system for reasoning over fd’s [20]. From R0-R2
one can derive additional rules, viz., (R3) decomposition, (R4) union and (R5)
pseudo-transitivity.
R0. If Y ⊆ X, then X → Y ;
R1. If X → Y , then XZ → Y Z;
R2. If X → Y and Y → W, then X → W ;
R3. If X → Y Z, then X → Y and X → Z;
R4. If X → Y and X → Z, then X → Y Z;
4.2. ACYCLIC PSEUDO-TRANSITIVE REASONING
50
R5. If X → Y and Y Z → W, then XZ → W .
Given an fd set Σ, one can obtain Σ+ , the closure of Σ, by a finite application of
rules R0-R5. We are concerned with reasoning over an fd set in order to process
its implicit causal ordering. The latter, as we shall see in §4.3, can be performed
in terms of (pseudo-)transitive reasoning. Note that R2 is a particular case of R5
when Z = ∅, then we shall refer to R5 reasoning and understand R2 included. The
next definition opens up a way to compute Σ+ very efficiently.
Let Σ be an fd set on attributes U , with X ⊆ U . Then X + , the attribute
closure of X w.r.t. Σ, is the set of attributes A such that hX, Ai ∈ Σ+ . Bernstein
has long given algorithm (Alg. 4) XClosure to compute X + . It is polynomial time
in |Σ| · |U | (cf. [62]), where Σ and U are (resp.) the given fd set and the attribute
set over which it is defined. A tighter time bound (linear time in |Σ| · |U |) is
achievable as discussed further in Remark 4.
Algorithm 4 Attribute closure X + (cf. [20, p. 388]).
procedure XClosure(Σ : fd set, X : attribute set)
Require: Σ is an fd set, X is a non-empty attribute set
Ensure: X + is the attribute closure of X w.r.t. Σ
2:
size ← 0
3:
Λ←∅
4:
X+ ← X
5:
while size < |X + | do
6:
size ← |X + |
7:
Σ←Σ\∆
8:
for all hY, Zi ∈ Σ do
9:
if Y ⊆ X + then
10:
∆ ← ∆ ∪ {hY, Zi}
11:
X+ ← X+ ∪ Z
12:
return X +
1:
. consumes fd
4.2. Acyclic Pseudo-Transitive Reasoning
As discussed in the previous chapter, we shall process the causal ordering in terms
of computing the transitive closure of each endogenous variable (predictive attribute). Before we proceed to that, we shall develop some machinery to reason
over fd’s in terms of Armstrong’s rule R5 (pseudo-transitivity). We shall then
4.2. ACYCLIC PSEUDO-TRANSITIVE REASONING
51
demonstrate the correspondence between this kind of reasoning with causal reasoning shortly in the sequel.
Def. 10 Let Σ be a set of fd’s on attributes U. Then Σ. , the pseudo-transitive
closure of Σ, is the minimal set Σ. ⊇ Σ such that X → Y is in ΣB, with XY ⊆ U ,
iff it can be derived from a finite (possibly empty) application of rule R5 over fd’s
.
in Σ. In that case, we may write X →
− Y and omit ‘w.r.t. Σ’ if it can be understood
from the context.
We are in fact interested in a very specific proper subset of ΣB, say, a kernel
of fd’s in ΣB that gives a “compact” representation of the causal ordering implicit
in Σ. Note that, to characterize such special subset we shall need to be careful
w.r.t. the presence of cycles in the causal ordering.
Def. 11 Let Σ be a set of fd’s on attributes U, and hX,Ai ∈ ΣB with XA ⊆ U .
#
We say that X → A is folded (w.r.t. Σ), and write X −
→ A, if it is non-trivial
and for no Y ⊂ U with Y + X, we have Y → X and X 6→ Y in Σ+ .
The intuition of Def. 11 is that an fd is folded when there is no sense in going
on with pseudo-transitive reasoning over it anymore (nothing new is discovered).
Given an fd X → A in fd set Σ, we shall be able to find some folded fd Z → A
by applying (R5) pseudo-transitivity as much as possible while ruling out cyclic or
trivial fd’s in some clever way.
Def. 12 Let Σ be an fd set on attributes U , and hX, Ai ∈ Σ be an fd with XA ⊆ U .
Then,
(a) A#, the (attribute) folding of A (w.r.t.Σ) is an attribute set Z ⊂ U such
#
that Z −
→ A;
(b) Accordingly, Σ#, the folding of Σ, is a proper subset Σ# ⊂ ΣB such that
#
an fd hZ, Ai ∈ ΣB is in Σ# iff X −
→ A for some Z ⊂ U .
Example 4 (continued). Fig. 4.1 shows an fd set Σ (left) and its folding Σ#
(right). Note that the folding can be obtained by computing the attribute folding
4.2. ACYCLIC PSEUDO-TRANSITIVE REASONING
Σ={ φ
φ
φ
x1 x2 x3 x 5 υ
x1 x3 x4 υ
x4 υ
x5 υ
→
→
→
→
→
→
→
52
Σ# = { φ
φ
φ
φ υ x5
φ υ x4
φ υ x5
φ υ x4
x1 ,
x2 ,
x3 ,
x4 ,
x5 ,
x6 ,
x7 }.
→
→
→
→
→
→
→
x1 ,
x2 ,
x3 ,
x4 ,
x5 ,
x6 ,
x7 }.
Figure 4.1. Fd set Σ encoding (cf. Alg. 3) the structure of Fig. 3.2a and its
folding Σ# derived by Alg. 5.
for A in each fd X → A in Σ. We illustrate below some reasoning steps to partially
compute an attribute folding considering the subset of fd’s in Σ with φ ∈
/ X.
→
→
→
→
→
1.
2.
3.
4.
5.
x 1 x2 x3 x 5 υ
x 1 x3 x 4 υ
x5 υ
x 1 x3 x4 υ
x 1 x2 x 3 x5 υ
x4
x5
x7
x7
x7
6. ∴
x1 x2 x3 x 4 υ → x7
[given]
[given]
[given]
[R5 over (2), (3)]
[R5 over (1), (4)]
[R5 over (2), (5)] .
Note that (6) is still amenable to further application of R5, say over (1), to
derive (7) x1 x2 x3 x5 υ → x7 . However, even though (1) and (6) have (resp.) the
.
.
form (1) X → A and (6) Y → B with Y →
− X, we have X →
− Y as well which
characterizes a cycle that fetches nothing into Y .1 In fact, if we consider only the
fd’s 1-3 given, then (6) itself satisfies Def. 11 and then is folded. The same holds,
e.g., for (1) by an empty application of R5. 2
Lemma 2 Let S(E, V) be a complete structure, ϕ a total causal mapping over S
and Σ an fd set encoded through ϕ given S. If hX, Ai ∈ Σ, then A#, the attribute
folding of A (w.r.t. Σ) exists and is unique.
Proof 7 See Appendix, §A.2.1. 2
1
Note, when at the step deriving (6) by R5 over (2), (5), that such cycle was not yet formed.
4.2. ACYCLIC PSEUDO-TRANSITIVE REASONING
53
We give an original algorithm (Alg. 5) to compute the folding of an fd set. At its
core there lies (Alg. 6) AFolding, which can be understood as a non-obvious variant
of XClosure (cf. Alg. 4) designed for acyclic pseudo-transitivity reasoning. In order
to compute the folding of attribute A in fd hX, Ai ∈ Σ, algorithm AFolding can be
seen as backtracing the causal ordering implicit in Σ towards A. Analogously, in
terms of the directed graph Gϕ induced by the causal ordering (see Fig. 3.3), that
would comprise graph traversal to identify the nodes x ∈ V that have xa ∈ V in
their reachability, i.e., x
xa . Rather, AFolding’s processing of the causal ordering
is fully symbolic based on Armstrong’s rewrite rule R5.
Example 5 Cyclicity in an fd set Σ may have the effect of making its folding Σ#
to degenerate to Σ itself. For instance, consider Σ = {A → B, B → A}. Note that
Σ is canonical, and AFolding (w.r.t. Σ) is B given A, and A given B. That is,
Σ# = Σ. 2
Algorithm 5 Folding of an fd set.
1: procedure folding(Σ : fd set)
Require: Σ given encodes complete structure S
Ensure: Returns fd set Σ# , the folding of Σ
2:
Σ# ← ∅
3:
for all hX, Ai ∈ Σ do
4:
Z ← AFolding(Σ, A)
5:
Σ# ← Σ# ∪ hZ, Ai
6:
return Σ#
4.2. ACYCLIC PSEUDO-TRANSITIVE REASONING
54
Algorithm 6 Folding of an attribute w.r.t. an fd set.
1: procedure AFolding(Σ : fd set, A : attribute)
Require: Σ is parsimonious
Ensure: Returns A#, the attribute folding of A (w.r.t. Σ)
2:
∆←∅
. consumed fd’s
3:
Λ←∅
. consumed attrs.
?
4:
A ← A . stores attrs. A is found to be ‘causally dependent’ on (cf. §4.3)
5:
size ← 0
6:
while size < |A? | do
. halts when A(i+1) = A(i)
7:
size ← |A? |
8:
Σ←Σ\∆
9:
for all hY, Bi ∈ Σ do
10:
if B ∈ A? then
11:
∆ ← ∆ ∪ {hY, Bi}
. consumes fd
?
?
12:
A ← A ∪Y
13:
Λ← Λ∪B
. consumes attr.
14:
for all C ∈ Y do
15:
if C ∈ Λ and B ∈ X for hX, Ci ∈ ∆ then
. cyclic fd
16:
Λ ← Λ \ B . reingests it to simulate cyclic app. of R5
17:
return A? \ Λ
Theorem 3 Let S(E, V) be a complete structure, and Σ an fd set encoded given
S. Now, let hX, Ai ∈ Σ. Then AFolding(Σ, A) correctly computes A#, the attribute
folding of A (w.r.t. Σ) in time O(|S|2 ).
Proof 8 For the proof roadmap, note that AFolding is monotone (size of A? can
only increases) and terminates precisely when A(i+1) = A(i) , where A(i) denotes
the attributes in A? at step i of the outer loop. The folding A# of A at step i
is A(i) \ Λ(i) . We shall prove by induction, given attribute A from fd X → A in
parsimonious Σ, that A? \ Λ returned by AFolding(Σ, A) is the unique attribute
folding A# of A. See Appendix, §A.2.2. 2
Remark 4 Let Σ be an arbitrary fd set on attribute set U . Beeri and Bernstein
gave a straightforward optimization to (Alg. 4) XClosure to make it linear in |Σ|·|U |
(cf. [63, p. 43-5]), where |Σ| · |U | is the maximum length for a string encoding all
the fd’s. Note that the actual length of such string in our case is exactly |S|. The
4.3. EQUIVALENCE WITH CAUSAL ORDERING
55
optimization mentioned applies likewise to (Alg. 6) AFolding.2 That is, AFolding
can be implemented to run in linear time in |S|. 2
Corollary 2 Let S(E, V) be a complete structure, and Σ an fd set encoded given
S. Then algorithm folding(Σ) correctly computes Σ#, the folding of Σ in time that
is f (|S|) Θ(|E|), where f (|S|) is the time complexity of (Alg. 6) AFolding.
Proof 9 See Appendix, §A.2.3. 2
Finally, it shall be convenient to come with a notion of parsimonious fd sets
(see Def. 13), which is suggestive of a distinguishing feature of such mathematical
information systems in comparison with arbitrary information systems.
Def. 13 Let Σ be set of fd’s on attributes U . Then, we say that Σ is parsimonious if it is canonical and, for all fd’s hX, Ai ∈ Σ with XA ⊆ U , there is no
Y ⊂ U such that Y 6= X and hY, Ai ∈ Σ.
Proposition 3 then shall be useful further in connection with the concept of
the folding.
Proposition 3 Let S(E, V) be a complete structure, ϕ a total causal mapping
over S and Σ an fd set encoded through ϕ given S. Let Σ# be the folding of Σ,
then Σ# is parsimonious.
Proof 10 See Appendix, §A.2.4. 2
4.3. Equivalence with Causal Ordering
Now we show the equivalence of acyclic pseudo-transitive reasoning with causal
ordering processing. We start with Theorem 4, which establishes the equivalence
between the notion of causal dependency and the fd encoding scheme presented in
Chapter 3.
2
We omit its tedious exposure here. In short, it shall require one more auxiliary data structure
to keep track, for each fd not yet consumed, of how many attributes not yet consumed appear in
its rhs.
4.3. EQUIVALENCE WITH CAUSAL ORDERING
56
Theorem 4 Let S(E, V) be a complete structure, ϕ a total causal mapping over
S and Σ an fd set encoded through ϕ given S. Then xa , xb ∈ V are such that
xb is causally dependent on xa , i.e., (xa , xb ) ∈ Cϕ+ iff there is some non-trivial fd
hX, Bi ∈ ΣB with A ∈ X, where B 7→ xb and A 7→ xa .
Proof 11 We prove the statement by induction. We consider first the ‘if’ direction, and then its ‘only if’ converse. See Appendix §A.2.5. 2
Def. 14 then gives useful terminology for a neat concept towards our goal in
this chapter.
Def. 14 Let S(E, V) be a structure with variables xa , xb ∈ V, and ϕ a total causal
mapping over S inducing set of direct causal dependencies Cϕ and its transitive
closure Cϕ+ . We say that xa is a first cause of xb in S if (xa , xb ) ∈ Cϕ+ and, for no
x ∈ V, we have (x, xa ) ∈ Cϕ+ .
Proposition 4 connects the notion of first cause with those of exogenous and
endogenous variables introduced in Chapter 3.
Proposition 4 Let S(E, V) be a structure with variable x ∈ V. Then x can only
be a first cause of some y ∈ V if x is exogenous. Accordingly, any variable y ∈ V
can only have some first cause x ∈ V if it is endogenous.
Proof 12 Straightforward from definitions, see Appendix §A.2.6. 2
Note that exogenous variables are encoded into fd’s X → A with φ ∈ X.
Since the values of such variables are assigned “outside” the system (cf. Remark
3), they are devoid of indirect causal dependencies and then have no uncertainty
except for their own. Thus, we have not to be concerned at all with processing
the causal (uncertainty) chaining towards them. Our goal is rather find the first
causes of the endogenous variables (predictive attributes).
We shall need then the terminology of Def. 15, and then we introduce Lemma
3 paving the way to our goal.
Def. 15 Let S be a structure, and Σ be a set of fd’s encoded over it. Then Υ(Σ),
the υ-projection of Σ, is the subset of fd’s X → A such that υ ∈ X. Accordingly,
Φ(Σ), the φ-projection of Σ, is the subset of fd’s X → A such that υ ∈
/ X.
4.3. EQUIVALENCE WITH CAUSAL ORDERING
Σ={ φ
φ
φ
x1 x2 x3 x 5 υ
x1 x3 x4 υ
x4 υ
x5 υ
→
→
→
→
→
→
→
x1 ,
x2 ,
x3 ,
x4 ,
x5 ,
x6 ,
x7 }.
Υ(Σ)# = { x1
x1
x1
x1
57
x2
x2
x2
x2
x3 υ x 5
x3 υ x4
x3 υ x5
x3 υ x4
→
→
→
→
x4 ,
x5 ,
x6 ,
x7 }.
Figure 4.2. Fd set Σ encoding the structure of Fig. 3.2a and the folding Υ(Σ)#
of its υ-projection.
Fig. 4.2 illustrates the υ-projection of an fd set and the folding applied over
such fd subset in order to compute the first causes of endogenous variables.
Lemma 3 Let S(E, V) be a complete structure, ϕ a total causal mapping over S
and Σ an fd set encoded through ϕ given S. Then a variable xa ∈ V can only be
a first cause of some variable xb ∈ V, where hX, Bi ∈ Σ, and B 7→ xb , A 7→ xa , if
either (i) A ∈ X or (ii) A ∈
/ X but there is hZ, Ci ∈ ΣB with A ∈ Z and C ∈ X.
Proof 13 We prove the statement by construction out of Theorem 4, see Appendix
§A.2.7. 2
Finally, Theorem 5 further clarifies the purpose of the folding and its meaning
in terms of causal ordering.
Theorem 5 Let S(E, V) be a complete structure, ϕ a total causal mapping over
S and Σ an fd set encoded through ϕ given S. Now, let B be an attribute that
encodes some variable xb ∈ V. If hX, Bi ∈ Υ(Σ)#,3 then every first cause xa of xb
(if any) is encoded by some attribute A ∈ X.
Proof 14 We show that the existance of a missing first cause xc of xb for folded
#
X −
→ B, where B 7→ xb and C 7→ xc but C ∈
/ X leads to a contradiction. See
Appendix §A.2.8. 2
3
Note that the folding is taken w.r.t. the υ-projection of Σ, then xb where B 7→ xb is an
endogenous variable.
4.4. EXPERIMENTS
58
Remark 5 Observe that, on the one hand, the goal of computing the transitive
closure Cϕ+ of a set of induced causal dependencies Cϕ is to derive the entire causal
ordering of a given structure. The goal of folding, on the other hand, is not to
discover all variables (attributes) a given variable (attribute) is causally dependent
on, but only all of its first causes (if any). 2
In particular, the results just shown comprise a method to compute, for each
endogenous variable (predictive attribute), all of its first causes. This is a core goal
of the reasoning device developed in this chapter in order to enable the automatic
synthesis of hypotheses as uncertain and probabilistic (U-relational) data.
4.4. Experiments
Fig. 4.3 shows the results of experiments we have carried out in order to study
how effective the causal reasoning over fd’s is in practice, in particular its behavior
for hypotheses whose structure S has been randomly generated over orders of
magnitude |S| ≈ 2k , to have length up to |S| ≈ 220 . 1M . The largest structure
considered, with |S| ≈ 1M , has been generated to have exactly |E| = 2.5K. Like
in the experiments of the previous chapter, we have executed ten runs for each
tested order of magnitude, and then taken its mean running time in ms.4
The plot is shown in Fig. 3.8 is in logscale base 2. Notice the linear rate
of growth across orders of magnitude (base 2) from 1K to 1M sized structures.
For a growth factor of 2 in structure length (doubled), the time required by causal
reasoning grows a factor of 2 (doubled as well). These scalability results are compatible with the computational complexity of folding, which is bounded by O(|S|2 ).
Yet, that is a bit overestimated time bound as we see in the plot of Fig. 4.3.
4.5. Related Work
The concept of fd set folding and the design of (Alg. 6) AFolding as a not quite
obvious variant of XClosure, is an original approach to the problem of processing
the causal ordering of a hypothesis via acyclic pseudo-transitive reasoning over
fd’s. To the best of our knowledge, such a specific form of fd reasoning was an
4
The experiments were performed on a 2.3GHZ/4GB Intel Core i5 running Mac OS X 10.6.8.
4.5. RELATED WORK
59
221
time [ms]
folding
216
211
26
27
29
211
213
215
217
219
221
structure length |S|
Figure 4.3. Performance of acyclic causal reasoning over fd’s (logscale).
yet unexplored problem in the database research literature (reasoning over fd’s is
extensively covered in Maier [22]).
Recent years have seen the emergence of some foundational work in causality
in databases [64]. It is motivated for improving DB usability in terms of providing
users with explanations to query answers (and non-answers). Essentially, the idea
is borrowed from AI work on causality (cf. §3.6) to identify causal ordering between
tuples. Given a query and its result set, the system should be able to explain to the
user what tuples ‘caused’ that answer, or why possibly expected tuples are missing.
That requires causal chain of tuples for a given query, which can be computationally
expensive as the database instance can be very large. For conjunctive queries, the
causality is said to be computed very efficiently [65].
A more specific problem addressed by Kanagal et. al is the so-called sensitivity analysis [66], which is aimed at establishing a more refined connection between
the query answer (output) and elements of the DB instance (input) for supporting
user interventions. Instead of providing the user with causes and non-causes, the
goal is to enable the user to know how changes in the input affect the output. This
line of work is strongly related to the vision of ‘reverse data management’ [67].
Causal reasoning in the presence of constraints (viz., fd’s) is an yet unexplored topic, though called for as worth of future work by Meliou et al. [64, p. 3].
The fd’s are rich information that can be exploited for the sake of explanation and
sensitivity analysis. Once they are available, it is intuitive that the search space of
such problems shall be significantly reduced.
4.6. SUMMARY OF RESULTS
60
In fact, our encoding of equations into fd’s captures the causal chain from
exogenous (input) to endogenous (output) tuples in at schema level. Nonetheless
our form of causal reasoning over fd’s is geared for hypothesis management and
analytics, from a uncertainty management point of view. A concrete connection
to causality in DB’s is not yet established.
4.6. Summary of Results
In this chapter we have studied and developed a technique for acyclic causal
reasoning over fd’s. We list the results achieved as follows.
• We have developed principled concepts and a core algorithm, viz., (Alg.
5) the folding, in order to perform acyclic pseudo-transitive reasoning over
fd’s. This is towards an efficient method for causal reasoning, yet elegant as a database formalism for the systematic construction of hypothesis
probabilistic DB’s.
• We have given a reasonably tight time bound for the behavior of such
reasoning device in terms of the structure given as input. We have established (cf. Theorem 3, Corollary 2) the time bound of O(|S|2 ) for the
folding algorithm.
• We have shown the correctness of the folding algorithm in connection with
causal reasoning (cf. Theorem 4, Theorem 5).
• We have defined the core notion of first causes (cf. Def. 14, Proposition 4), which is meant to guide the procedure of U-intro (Chapter 5) by
precisely capturing the uncertainty factors on endogenous variables (predictive attributes). This is similar, yet markedly different from computing
the transitive closure of causal dependencies (cf. Remark 5).
• We have performed experiments (cf. Fig. 4.3) to study how effective the
approach of causal reasoning over fd’s is in practice, or how it scales for
hypotheses whose structure S is randomly generated to have length up
to the order of |S| . 1M . The experiments show that the time bound,
though already effective for very large structures, are a bit overestimated.
Chapter 5
Probabilistic Database Synthesis
In this chapter we present a technique to synthesize hypothesis U-relations. At
this stage of the pipeline, relational schema H is loaded with datasets computed
from the hypotheses under alternative trials (input settings). The challenge is how
to model or design its probabilistic version (i.e., render the U-relations Y ) so that
it is suitable for data-driven hypothesis management and analytics.
In §5.1 we introduce U-relational DB’s. In §5.2 we present a running example
to illustrate the uncertainty introduction procedure (U-intro in the pipeline, cf. Fig.
1.4). In §5.3 we present the technique to factorize the uncertainty present in the
‘big’ fact table in terms of the well-defined uncertainty factors. In §5.4 then we
show how to propagate such uncertainty into the predictive attributes properly,
i.e., based on their first causes detected as shown in Chapter 4. In §5.6, we discuss
related work. Finally, in §5.7 we conclude the chapter.
1
5.1. Preliminaries: U-Relations and Probabilistic WSA
Three remarkable features of U-relations are: expressiveness (being closed under
positive relational algebra queries); succinctness (efficient storage of a very large
number of possible worlds through vertical decompositions to support attributelevel uncertainty); and efficient query processing (including confidence computation) [18].
A U-relational database or U-database is a finite set of structures,
1
n
W = { hR11 , ..., Rm
, p[1] i, ..., h R1n , ..., Rm
, p[n] i },
1
We postpone the presentation of experiments on p-DB synthesis (U-intro as a whole) to §6.4.
5.1. PRELIMINARIES: U-RELATIONS AND PROBABILISTIC WSA
i
of relations R1i , ..., Rm
and numbers 0 < p[i] ≤ 1 such that
62
P
1≤ i ≤n
p[i] = 1. An
i
, p[i] ∈ W is a possible world, with p[i] being its probability [18].
element R1i , ..., Rm
Probabilistic world-set algebra (p-WSA) consists of the operations of relational algebra, an operation for computing tuple confidence conf, and the repairkey operation for introducing uncertainty — by giving rise to alternative worlds as
maximal-subset repairs of an argument key [18].
Let R` [U ] be a relation, and XA ⊆ U . For each possible world hR1 , ..., Rm ,
pi ∈ W , let A ∈ U contain only numerical values greater than zero and let R`
satisfy the fd (U \ A) → U . Then, repair-key is:
Jrepair-keyX@A (R` )K(W) :=
n
o
h R1 , ..., R` , Rm , R̂` [U \ A], p̂ i ,
where hR1 , ..., R` , Rm , pi ∈ W , R̂` is a maximal repair of fd X → U in R` , and
Y
t.B
P
.
p̂ = p ·
s ∈ R` : s.X = t.X s.B
t∈R̂`
U-relations (cf. Fig. 5.1) have in their schema a set of pairs (Vi , Di ) of condition
columns (cf. [18]) to map each discrete random variable xi to one of its possible
values (e.g., x1 7→ 1). The world table W stores their marginal probabilities (cf.
the notion of pc-tables [17, Ch. 2]). For an illustration of the data transformation
from certain to uncertain relations, consider query (5.1) in p-WSA’s extension of
relational algebra, whose result set is materialized into U-relation Y0 as shown in
(Fig. 5.1).
Y0 := πφ,υ (repair-keyφ@Conf (H0 ) ).
(5.1)
Also, let R[ Vi Di | sch(R) ], S[ Vj Dj | sch(S) ] be two U-relations, where R. Vi Di is
the union of all pairs of condition columns Vi Di in R, then operations of selection
J σψ (R) K, projection J πZ (R) K, and product JR × SK issued in relational algebra
are rewritten in positive relational algebra on U-relations:
Jσψ (R)K := σψ (R[Vi Di | sch(R)]);
J πZ (R) K := π Vi Di Z (R);
JR×SK := π(R.Vi Di ∪ S.Vi Di )→V D ∪ sch(R) ∪ sch(S) (R ./R.Vi Di is consistent with S.Vj Dj S).
5.2. RUNNING EXAMPLE
63
H0
Y0
V 7→ D
x0 7→ 1
x0 7→ 2
x0 7→ 3
φ
1
1
1
φ
1
1
1
υ
1
2
3
υ
1
2
3
Conf
2
2
1
W
V 7→ D
x0 7→ 1
x0 7→ 2
x0 7→ 3
Pr
.4
.4
.2
Figure 5.1. U-relation generated by the repair-key operation.
If R and S have k and ` pairs of condition columns each, then JR × SK
returns a U-relation with k + ` such pairs. If k = 0 or ` = 0 (or both), then R or
S (or both) are classical relations, but the rewrite rules above apply accordingly.
All that rewriting is parsimonious translation (sic. [18]): the number of algebraic
operations does not increase and each of the operations selection, projection and
product/join remains of the same kind. Query plans are hardly more complicated
than the input queries. In fact, it has been verified hat off-the-shelf relational
database query optimizers do well in practice.
For a comprehensive overview of U-relations and p-WSA we refer the reader
to [18]. In this thesis we look at U-relations from the point of view of p-DB design,
for which no methodology has yet been proposed. We are concerned in particular
with hypothesis management applications [10].
5.2. Running Example
Before proceeding, we consider Example 6, which is fairly representative
to illustrate how to deal with correlations in the predictive data of deterministic
hypotheses for the sake of suitable data-driven analytics.
Example 6 We explore three slightly different theoretical models in population
dynamics with applications in Ecology, Epidemics, Economics, etc: (5.2) Malthus’
model, (5.3) the logistic equation and (5.4) the Lotka-Volterra model. In practice,
such equations are meant to be extracted from MathML-compliant XML files (cf.
Chapter 6). For now, consider that the ordinary differential equation notation ‘ẋ’
5.2. RUNNING EXAMPLE
64
is read ‘variable x is a function of time t given initial condition x0 .’
ẋ = bx
(5.2)
ẋ = b(1 − x/K)x
(5.3)
ẋ = x(b − py)
ẏ = y(rx − d)
(5.4)
The models are completed (by the user) with additional equations to provide the
values of exogenous variables (or ‘input parameters’),2 e.g., x0 = 200, b = 10, such
that we have SEM’s (resp.) Sk (Ek , Vk ) for k = 1..3,
• E1 = { f1 (t), f2 (x0 ), f3 (b), f4 (x, t, x0 , b) };
• E2 = {f1 (t), f2 (x0 ), f3 (K), f4 (b), f5 (x, t, x0 , K, b)};
• E3 = { f1 (t), f2 (x0 ), f3 (b), f4 (p), f5 (y0 ), f6 (d), f7 (r),
f8 (x, t, x0 , b, p, y), f9 (y, t, y0 , d, r, x) }.
Fig. 5.2 shows the fd sets encoded from structures Sk given above.3
We also
consider trial datasets for hypothesis υ = 3 (viz., the Lotka-Volterra model), which
are loaded into the ‘big’ fact table relation H3 ∈ H as shown in Fig. 5.3. We admit
a special attribute ‘trial id’ tid to keep hypothesis trials identified until uncertainty
is introduced in a controlled way by p-DB synthesis (U-intro stage, cf. Fig. 1.4).
2
2
Given S(E, V), it is actually a task of the encoding algorithm (viz., sub-procedure TCM) to
infer, for each variable x ∈ V, whether it is exogenous or endogenous by means of the total causal
mapping.
3
Recall that domain variables like time t are informed to the encoding algorithm to suppress
an fd φ → t.
5.2. RUNNING EXAMPLE
65
Σ1 = { φ → x0 ,
φ → b,
x0 b t υ → x }.
Σ2 = { φ
φ
φ
x0 K b t υ
→
→
→
→
Σ3 = { φ
φ
φ
φ
φ
φ
x0 b p t υ y
y0 d r t υ x
x0 ,
K,
b,
x }.
→
→
→
→
→
→
→
→
x0 ,
b,
p,
y0 ,
d,
r,
x,
y }.
Figure 5.2. Fd sets encoded from the given structures Sk (Ek , Vk ) for hypotheses
k = 1..3 from Example 6.
H3 tid
1
1
2
2
3
3
4
4
5
5
6
6
6
6
6
6
φ
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
υ
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
t
0
...
0
...
0
...
0
...
0
...
0
5
10
15
20
...
x0
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
b
.5
.5
.5
.5
.4
.4
.4
.4
.397
.397
.397
.397
.397
.397
.397
.397
p y0
.02 4
.02 4
.018 4
.018 4
.02 4
.02 4
.018 4
.018 4
.02 4
.02 4
.018 4
.018 4
.018 4
.018 4
.018 4
.018 4
d
.75
.75
.75
.75
.8
.8
.8
.8
.786
.786
.786
.786
.786
.786
.786
.786
r
.02
.02
.023
.023
.02
.02
.023
.023
.02
.02
.023
.023
.023
.023
.023
.023
x
30
...
30
...
30
...
30
...
30
...
30
50.1
13.8
79.3
12.6
...
y
4
...
4
...
4
...
4
...
4
...
4
62.9
8.65
8.23
30.7
...
Figure 5.3. ‘Big’ fact table H3 of hypothesis k = 3 from Example 6 loaded with
trial datasets identified by special attribute tid.
Given the ‘big’ fact table H3 , p-DB synthesis has two main parts: process
the ‘empirical’ uncertainty present in the ‘big’ fact table and synthesize it out
(decompose it) into independent u-factors (u-factorization); and then propagate it
precisely into the predictive data (u-propagation).
5.3. U-FACTORIZATION
66
5.3. U-Factorization
As we have seen in §5.1, the repair-key operation allows one to create a discrete
random variable in order to repair an argument key in a given relation. Our goal
here is to devise a technique to perform such operation in a principled way for
hypothesis management. It is a basic design principle to have exactly one random
variable for each distinct uncertainty factor (‘u-factor’ for short), which requires
carefully identifying the actual sources of uncertainty present in relations H.
The multiplicity of (competing) hypotheses is itself a standard one, viz., the
theoretical u-factor. Consider an ‘explanation’ table like H0 in Fig. 5.1, which
stores (as foreign keys) all hypotheses available and their target phenomena. We
can take such H0 as explanation table for the three hypotheses of Example 6. Then
a discrete random variable V0 is defined into Y0 [ V0 D0 | φ υ ] by query formula (5.1).
U-relation Y0 is considered standard in p-DB synthesis, as the repair of φ as a key
in (standard) H0 .
Hypotheses, nonetheless, are (abstract) ‘universal statements’ [15]. In order
to produce a (concrete) valuation over their endogenous attributes (predictions),
one has to inquire into some particular ‘situated’ phenomenon φ and tentatively
assign a valuation over the exogenous attributes, which can be eventually tuned
for a target φ. The multiplicity of such (competing) empirical estimations for a
hypothesis k leads to Problem 1, viz., learning empirical u-factors for each Hk ∈ H.
Problem 1 Let Σk be an fd set encoded given hypothesis structure Sk , and Hk its
‘big’ fact table relation loaded with trial data. Now, let Z be the set of attributes
encoding exogenous variables in Hk , then the problem of u-factor learning is:
(1) to infer in Hk ‘casual’ fd’s φ Bi → Bj , φ Bj → Bi not in Σk (strong input
correlations), where Bi , Bj ∈ Z;
(2) to form maximal groups G1 , ..., Gn ⊆ Z of attributes such that for all
Bi , Bj ∈ Ga , the casual fd’s φ Bi → Bj and φ Bj → Bi hold in Hk ;
(3) to pick, for each group Ga , any A ∈ Ga as a pivot representative and
insert φ A → B into an fd set Ωk for all B ∈ (Ga \ A).
5.3. U-FACTORIZATION
H3 tid
1
1
2
2
3
3
4
4
5
5
6
6
6
6
6
6
φ
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
67
υ
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
t
0
...
0
...
0
...
0
...
0
...
0
5
10
15
20
...
x0
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
30
b
.5
.5
.5
.5
.4
.4
.4
.4
.397
.397
.397
.397
.397
.397
.397
.397
p y0
.02 4
.02 4
.018 4
.018 4
.02 4
.02 4
.018 4
.018 4
.02 4
.02 4
.018 4
.018 4
.018 4
.018 4
.018 4
.018 4
d
.75
.75
.75
.75
.8
.8
.8
.8
.786
.786
.786
.786
.786
.786
.786
.786
r
.02
.02
.023
.023
.02
.02
.023
.023
.02
.02
.023
.023
.023
.023
.023
.023
x
30
...
30
...
30
...
30
...
30
...
30
50.1
13.8
79.3
12.6
...
y
4
...
4
...
4
...
4
...
4
...
4
62.9
8.65
8.23
30.7
...
Figure 5.4. ‘Big’ fact table H3 of hypothesis k = 3 from Example 6 with u-factors
{b, d} and {p, r} emphasized (resp.) in colors green and red.
U-factor learning is meant to process only the attributes Z ⊂ U from Hk [U ] that
are inferred exogenous in the given hypothesis, i.e., for all A ∈ Z, there is an fd
hX, Ai ∈ Φ(Σk ), where the latter is the φ-projection of Σk . Such attributes are
then ‘officially’ unrelated. In fact, by ‘casual’ fd’s we mean correlations that, for a
set of experimental trials, may occasionally show up in the trial input data; e.g.,
x0 ↔ y0 hold in H3 , but not because x0 and y0 are related in principle (theory).
Fig. 5.4 helps to illustrate Problem 1 through the ‘big’ fact table. We
emphasize u-factors {b, d} and {p, r} in colors green and red. Observe that values
of b are strongly correlated (one-to-one) with values of d for φ = 1, just like p and
r. Note also that {x0 , y0 } can be seen as a certain factor. From the user point
of view, this is a record that reflects a common practice in computational science
known as (parameter) sensibility analysis.
Problem 1 is dominated by the (problem of) discovery of fd’s in a relation,
which is not really a new problem (e.g., see [68]). We then keep focus on the
synthesis method as a whole and omit our detailed u-factor-learning algorithm in
5.3. U-FACTORIZATION
Ω3 = { φ x0
φb
φp
x0 b p t υ y
y0 d r t υ x
68
→
→
→
→
→
y0 ,
d,
r,
x,
y }.
Ω#
3 = { φ x0
φb
φp
φ x0 b p t υ y
φ x0 b p t υ x
→
→
→
→
→
y0 ,
d,
r,
x,
y }.
Figure 5.5. Fd set Ω3 (compare with Σ3 ) and its folding Ω#
3.
particular.4
Its output, fd set Ωk , is then filled in (completed) with the υ-
projection Υ(Σk ). For illustration consider hypothesis υ = 3 and its trial input
data recorded in H3 in Fig. 5.3. We show its resulting fd set Ω3 in Fig. 5.5 (left),
together with its folding Ω#
k (right). The latter is then input to (Alg. 7) merge
to get the final information necessary for the actual synthesis of U-relations, as
captured in Def. 16. For an illustration of the merging of fd’s with equivalent
left-hand sides, note in Fig. 5.5 (right) that φ x0 b p t υ y ↔ φ x0 b p t υ x holds in
+
(Ω#
3 ) .
Def. 16 Let Sk and Hk be the complete structure and ‘big’ fact table of hypothesis
k, and Σk an fd set defined Σk , h-encode(Sk ). Now, let Ωk , u-factor-learning( Hk ,
#
Φ(Σk ) ) ∪ Υ(Σk ), and Ω#
k be the folding of Ωk . Finally, define Γk , merge( Ωk ).
Then we say that Γk is the u-factorization of Sk over Hk .
Algorithm 7 Merge fd’s with equivalent left-hand sides.
1: procedure merge(Σ : fd set)
2:
Ω←∅
3:
for all hX, Ci ∈ Σ do
4:
if there is hZ, W i ∈ Ω such that X ↔ Z holds in Σ+ then
5:
Ω ← Ω \ hZ, W i
6:
S ←X \Z
7:
Ω ← Ω ∪ hZ, W SCi
. merges equivalent keys
8:
else
9:
Ω ← Ω ∪ hX, Ci
10:
4
return Ω
In short, we make use of relational algebra group-by operation and build a pruned lattice of
attribute groups having the same number of rows under the grouping (similarly to [68]).
5.3. U-FACTORIZATION
69
Remark 6 Let Γk be the u-factorization of structure Sk over ‘big’ fact table Hk .
Then every fd in Γk encodes a clear-cut claim, either empirical, in Φ(Γk ), or theoretical, in Υ(Γk ). That is ensured by the merge algorithm, which groups fd’s in
Ω#
k with equivalent left-hand sides. 2
We are now able to employ a notion of u-factor decomposition formulated in
Def. 17 into query formula (5.5) in p-WSA’s extension of relational algebra.
Def. 17 Let Sk be the complete structure of hypothesis k, and Hk [U ] its ‘big’ fact
table such that Γk is the u-factorization of Sk over Hk . Now, let Ga ⊂ U be a set
of attributes Ga = A G such that, for all B ∈ G, an fd φ A → B exists in Φ(Γk ).
Then we define U-relation Yki [Vi Di | φ A G ] by query formula (5.5), and say that
Yki is a u-factor projection of Hk .
Yki := π φ A G (repair-key φ @ count ( γ φ, A, G, count(∗) (Hk ) ) )
(5.5)
where γ is relational algebra’s grouping operator.
Y13
V 7→ D φ x0 y0
x2 7→ 1 2 30 4
Y23 V 7→ D
x3 7→ 1
x3 7→ 2
x3 7→ 3
φ b
d
2 .5
.5
2 .4
.8
2 .397 .786
Y33 V 7→ D φ p
r
x4 7→ 1 2 .020 .020
x4 7→ 2 2 .018 .023
Figure 5.6. U-factor projections rendered for hypothesis υ = 3.
The synthesis of u-factor projections, in particular the application of repair-key (cf.
Eq. 5.5), has an important consequence for the u-factorization Γk of Hk , viz., the
introduction of new fd’s into Γ0k defined as follows (see Def. 18). We shall consider
it (rather than Γk ) to study design-theoretic properties of synthesized Y k in §5.5.
Def. 18 Let Sk and Hk [U ] be (resp.) the complete structure and ‘big’ fact table
of hypothesis k, and Γk be the u-factorization of Sk over Hk . Now, let Γ0k ,
S
i
i ∈ I {φ → Ai Gi } ∪ Γk , where I indexes all u-factor projections Yk [ Vi Di | φ Ai Gi ]
of Hk . We say that Γ0k is the repaired factorization of Sk over Hk .
5.4. U-PROPAGATION
70
5.4. U-Propagation
U-propagation is a central part of U-intro and the pipeline itself. Recall that all
the machinery developed so far, from hypothesis encoding to causal reasoning to
u-factorization is for enabling predictive analytics. Let us briefly reconstruct it.
For hypothesis structure Sk (E, V), take any endogenous variable xc ∈ V encoded by attribute C 7→ xc . There should be exactly one fd hX, Ci ∈ Υ(Σk )#.
By Theorem 5, for every first cause xb of xc there is B ∈ X where B 7→ xb with
xb ∈ V. Now, observe that when Ωk is rendered by u-factor learning, it is filled
partly with fd’s from Υ(Σk ), and partly with fd’s processed from Φ(Σk ). This is
to summarize exogenous variables into clear-cut independent u-factors. It means
that, after u-factor learning, each first cause xb encoded by B shall be represented
by some pivot attribute Ai which is, if not Ai = B itself, then occasionally strongly
correlated to it (i.e., φ Ai → B and φ B → Ai hold in Hk ).
Further then Ωk is subject to folding such that, if B ∈ X for hX, Ci ∈ Υ(Ωk ),
now we have Ai ∈ Z for hZ, Ci ∈ Υ(Ω#
k ). This processing from fd X → C
(where X contains the first causes) into Z → C (where Z contains only their pivot
representatives) is meant for enabling an economical representation of uncertainty.
Our running example is small, but such a principle is quite relevant for large-scale
hypotheses (say, when |S| ≈ 1M ). The correctness of such u-factor summarization
shall be ensured by Proposition 3, which let us know that Ω#
k is parsimonious then
(by Def. 13) canonical, therefore (by Def. 9) left-reduced.
Now, all fd’s in Υ(Ω#
k ) have form Z → C and we are almost ready for upropagation. Note that, as a result of u-factorization, each pivot attribute Ai ∈ Z
is associated with random variable Vi from U-relation Yki [ Vi Di | φ Ai Gi ]. Then
we shall use each Ai ∈ Z (from Z → C) as a surrogate to Yki in order to propagate
factorized uncertainty into ‘predictive’ U-relations Ykj [ Vj Dj | S T ] by a join formula. Attribute sets S and T are defined after merging fd’s in Ω#
k with equivalent
lhs to get Γk and pass it as argument for synthesis. We let S = Z \ W such that
S contains the domain variables only (e.g., φ, υ, t). The pivot attributes in Z
shall not be included in the data columns of Ykj , but leave their trace through the
condition columns Vj Dj that annotate sch(Ykj ) as a repair of the key S → T .
5.4. U-PROPAGATION
71
All that (cf. Def. 19) is abstracted into general p-WSA query formula (5.6),
and employed in (Alg. 8 ) synthesize to accomplish u-propagation (Part II).
Def. 19 Let Sk be the complete structure of hypothesis k, and Hk [U ] its ‘big’ fact
table such that Γk is the u-factorization of Sk over Hk . Now, let Z → T be an fd
in Υ(Γk ). Then we define U-relation Ykj [ Vj Dj | S T ] by query formula (5.6), and
say that Ykj is a predictive projection of Hk , where:
Ykj := π S, T ( συ=k (Y0 ) ./ (./ i ∈ I Yki ) ./ Hk )
(5.6)
(a) Yki [ Vi Di | φ Ai Gi ] is a u-factor projection of Hk ;
(a) we have i ∈ I if Ai ∈ Z;
(c) we take S = Z \ W , where W is the set of all pivot attributes representing
first causes in Z.
Algorithm 8 p-DB synthesis applied over folding fd set.
1: procedure synthesize(Sk : structure, Hk : ‘big’ table, Y0 : explan. table)
Require: Sk is complete
Ensure: U-relations Y k returned are a BCNF, lossless decomposition of Hk
2:
Σk ← h-encode(Sk )
3:
Ωk ← u-factor-learning( Φ(Σk ), Hk ) ∪ Υ(Σk )
4:
Γk ← merge( folding(Ωk ) )
Part I: U-factorization
5:
for all hφ A, Gi ∈ Φ(Γk ) do
. scans over the u-factors of hypothesis k
i
6:
Yk ← π φ, A, G (repair-keyφ@count ( γ φ, A, G, count(∗) (Hk ) ) )
7:
Y k ← Y k ∪ Yki
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Part II: U-propagation
for all hZ, T i ∈ Υ(Γk ) do
. scans over the claims of hypothesis k
W ←∅
. prepares to keep track of u-factor pivot attributes
i
for all Yk [ φ Ai Gi ] ∈ Y k do
if A ∈ Z then
. A is a first cause of all B ∈ T
I = I ∪ {i}
. indexes the u-factor projection
W ←W ∪A
. keeps track of u-factor’s pivot attribute
S ←Z \W
. removes u-factor pivot attributes
j
i
Yk ← π S, T ( συ=k (Y0 ) ./ (./ i ∈ I Yk ) ./ Hk )
Y k ← Y k ∪ Ykj
return Y k
5.4. U-PROPAGATION
72
Fig. 5.7 shows the rendered U-relations for hypothesis k = 3 whose ‘big’
fact table is shown in Fig. 5.3. Note that tid = 6 in H3 corresponds now to
θ = { x1 7→ 3, x2 7→ 1, x3 7→ 3, x4 7→ 2 }, where θ defines a particular world in
W whose probability is Pr(θ) ≈ .055. This value is derived from the marginal
probabilities stored in world table W (see Fig. 5.7) as a result of the application
of formulas Eq. 5.1 and Eq. 5.5.
Remark 7 Observe that, although (Alg. 8) synthesize operates locally for each
hypothesis k, the effects of p-DB synthesis (U-intro) in the pipeline are global on
account of the (global) ‘explanation’ relation H0 (then U-relation Y0 ), e.g., see Fig.
5.7. In fact, the probability of each tuple (row), say, in U-relation Ykj with φ = p
for hypothesis υ = k, is distributed among all the hypotheses ` 6= k that are keyed
in Y0 under φ = p, i.e., all hypotheses that compete at φ = p. 2
W
Y0 V 7→ D
x1 7→ 1
x1 7→ 2
x1 7→ 3
Y13
Y43
V 7→ D φ x0 y0
x2 7→ 1 1 30 4
V1 7→ D1
x1 7→ 3
x1 7→ 3
...
x1 7→ 3
x1 7→ 3
x1 7→ 3
x1 7→ 3
x1 7→ 3
x1 7→ 3
V2 7→ D2
x2 7→ 1
x2 7→ 1
...
x2 7→ 1
x2 7→ 1
x2 7→ 1
x2 7→ 1
x2 7→ 1
x2 7→ 1
φ
1
1
1
υ
1
2
3
Y23 V 7→ D
x3 7→ 1
x3 7→ 2
x3 7→ 3
V3 7→ D3
x3 7→ 1
x3 7→ 1
...
x3 7→ 3
x3 7→ 3
x3 7→ 3
x3 7→ 3
x3 7→ 3
x3 7→ 3
V 7→ D
x1 7→ 1
x1 7→ 2
x1 7→ 3
x2 7→ 1
x3 7→ 1
x3 7→ 2
x3 7→ 3
x4 7→ 1
x4 7→ 2
φ b
d
1 .5
.5
1 .4
.8
1 .397 .786
V4 7→ D4
x4 7→ 1
x4 7→ 1
...
x4 7→ 2
x4 7→ 2
x4 7→ 2
x4 7→ 2
x4 7→ 2
x4 7→ 2
φ
1
1
1
1
1
1
1
1
1
Pr
.4
.4
.2
1
.33
.33
.33
.5
.5
Y33 V 7→ D φ p
r
x4 7→ 1 1 .020 .020
x4 7→ 2 1 .018 .023
υ
3
3
3
3
3
3
3
3
3
t
1900
...
...
1900
1901
1902
1903
1904
...
y
4
...
...
4
4.12
5.78
11.7
31.1
...
Figure 5.7. U-relations rendered for hypothesis υ = 3.
x
30
...
...
30
41.5
56.7
72.8
75.9
...
5.5. DESIGN-THEORETIC PROPERTIES
73
U-relations rendered by p-DB synthesis are ready for querying. Typical
queries comprise the conf() aggregate operation, inquiring the probability (or confidence) for each tuple to ‘true’ in the probability space captured by the hypothesis
competition. We illustrate queries in Chapter 6.
5.5. Design-Theoretic Properties
For the U-intro procedure to be meaningful, we have to study design-theoretic
properties of u-factor projections and prediction projections synthesized out of ‘big’
fact table Hk for the sake of predictive analytics. In particular, for the projections
to be claim-centered, we submit that they should satisfy Boyce-Codd normal form
(BCNF, cf. Def. 21) w.r.t. the repaired factorization Γ0k of Hk ; and for them to
be a correct decomposition of the uncertainty present in Hk , their join should be
lossless (preserve the data in Hk , cf. Def. 22) w.r.t. Γ0k .
Note that in this study we consider repaired factorization Γ0k (not u-factorization Γk ), since it is the one which actually holds in Y k after key repairing.
5.5.1 Claim-Centered Decomposition
As emphasized through Remark 6, every fd in u-factorization Γk is a claim (cf.
Remark 6), then the same holds for repaired factorization Γ0k . Thus, for a claimcentered decomposition of ‘big’ fact table Hk , it is desirable that U-relational
schema Y k that it satisfies BCNF w.r.t. Γ0k . BCNF (‘do not represent the same
fact twice’ [21, p. 251]) is our notion of ‘good design’ for uncertainty decomposition
in view of predictive analytics. This is to avoid the uncertainty of one claim to be
undesirably mixed with the uncertainty of another claim.
Def. 20 Let R[U ] be a relation scheme over set U of attributes, and Σ a set of
fd’s. Then the projection of Σ onto R[U ], written πU (Σ), is the subset Σ0 ⊆ Σ of
fd’s X → Z such that XZ ⊆ U .
Def. 21 Let R[U ] be a relation scheme over set U of attributes, and Σ a set of fd’s
on U . We say that:
(a) R is in BCNF if, for all hX, Ai ∈ Σ+ with A 6∈ X and XA ⊆ U , we have
X → U (i.e., X is a superkey for R);
5.5. DESIGN-THEORETIC PROPERTIES
74
(b) A schema R is in BCNF if all of its schemes R1 , ..., Rn ∈ R are in BCNF.
Example 7 To illustrate the concept of BCNF, let us consider canonical fd set
Σ = {A → B, B → C } over attributes U = {A, B, C}, and a tentative schema
containing a single relation R[ABC]. This relation is not in BCNF because, for
one, B → C violates it (C * B but B is not a superkey for R). 2
Observe also that an overdecomposed schema may (trivially) satisfy BCNF.
For example, let Σ = {A → B, A → C} then by Def. 21 both schemas R =
R1 [ABC] and R0 = {R1 [AB], R2 [AC]} are in BCNF w.r.t. Σ. The second, however, breaks data into two tables making their access more difficult than necessary
since both B and C brings in information about A. That is, if the schema were to
be synthesized over the fd’s in Σ, then it would be desirable to apply (R4) union or
merge them before. Our point is that, if we target at a BCNF-satisfying schema,
then it is also desirable for it to be the minimal-cardinality schema in BCNF.
Theorem 6 guarantees the BCNF property w.r.t. Γ0k by design for every
schema Y k rendered by (Alg. 8) synthesize over Γk .
Theorem 6 Let Sk and Hk be (resp.) the complete structure and ‘big’ fact table
of hypothesis k, and let Γ0k be the repaired factorization of Sk over Hk , and Y0 the
‘explanation’ table where hypothesis k is recorded. Now, let Y k be a U-relational
schema defined Y k , synthesize(Sk , Hk , Y0 ). Then Y k is in BCNF w.r.t. Γ0k and
is minimal-cardinality.
Proof 15 We exploit the fact that the projection of (Γ0k )+ onto u-factor projections
and predictive projections define a disjoint partition of (Γ0k )+ into its φ-projection
Φ(Γ0k )+ and υ-projection Υ(Γ0k )+ . Since we know the form of fd’s in each of them,
the search space for BCNF violations is significantly reduced. The minimality of
|Y k | in turn comes from (Alg. 7) merge. See Appendix, §A.3.1. 2
5.5.2 Correctness of Uncertainty Decomposition
Recall from the preliminaries (cf. §5.1) that the U-relational equivalent of the
relational product operation (main sub-operation of the join operation) has been
introduced. Now, we provide the classical definition of a lossless join [20], i.e.,
5.5. DESIGN-THEORETIC PROPERTIES
75
when a decomposition of data from a relation into two or more relations is known
to preserve the data in its original form by an application of the join.
Def. 22 Let R[U ] be a (U-)relational schema synthesized into collection R =
Sn
i=1 Ri and let Σ be an fd set on attributes U . We say that R has a lossless join
w.r.t. Σ if for every instance r of R[U ] satisfying Σ, we have r = ./ ni=1 πRi (r).
The lossless join property is of interest to ensure that our decomposition of
the data from the ‘big’ fact table into u-factor projections preserves the data so
that their join to ‘annotate’ the predictive projections when propagated by means
of the U-relational join operation is correct. Theorem 7 guarantees that is the case.
Theorem 7 Let Sk be the complete structure of hypothesis k, and Hk [U ] its ‘big’
fact table such that Γ0k is the repaired factorization of Sk over Hk and Y0 is the
‘explanation’ table where hypothesis k is recorded. Now, let Y k be a U-relational
schema defined Y k , synthesize(Sk , Hk , Y0 ). Then,
i
(a) the join ./ m
i=1 Yk [ Vi Di | φ Ai Gi ] of any subset of the u-factor projections
of Hk is lossless w.r.t. Γ0k .
(b) any predictive projection Ykj [ Vj Dj | S T ], result of a join of the theoretical
u-factor Y0 [ V0 D0 | φ υ ] with the ‘big’ fact table Hk [U ] and in turn with
u-factor projections Yki [ Vi Di | φ Ai Gi ], is lossless w.r.t. Γ0k .
Proof 16 We make use of a lemma from Ullman [20, p. 397], and then the proof
comes straightforwardly. See Appendix, §A.3.2. 2
Remark 8 The significance of Theorem 6 lies in that it guarantees the decomposition of uncertainty based on the causal ordering processing is in fact claim-centered
as desirable for predictive analytics. Theorem 7 in turn is significant as it ensures
that all the empirical uncertainty implicit in a hypothesis ‘big’ fact table Hk can
be decomposed into u-factor projections that are (a) independent (not strongly
correlated, cf. Problem 1), and (b) can be fully recovered by a join that is lossless
w.r.t. repaired factorization Γ0k of structure Sk over Hk . This is essential to make
sure that in u-propagation the composition of the required u-factors recovers the
uncertainty associated with the predictive data. Since repaired factorization Γ0k is
5.6. RELATED WORK
76
known to be a correct processing of the causal ordering (cf. results of Chapter 4),
altogether Theorem 7 guarantees that the first causes are joined together correctly
towards the predictive variables influenced by them. 2
As we have seen, the p-DB synthesis technique presented here is essentially
targeted at design-theoretic properties. It is also motivated by computational
performance, as uncertainty decomposition is desirable also to speed up probabilistic inference [17, p. 30-1]. In fact, the U-intro procedure is fully grounded
in U-relations and p-WSA as implemented in the MayBMS system. Its computational performance is dominated by U-relational query processing. We present
experimental studies on the U-intro procedure in §6.4, as they are designed from a
applicability point of view. The goal is to provide some reference computational
measures for prospective users.
5.6. Related Work
Informed on research on Graphical Models (GM) [69], Suciu et al. provide a
striking motivation for work on probabilistic database design [17, p.30-1]. In GM
design, probability distributions on large sets of random variables are decomposed
into factors of simpler probability functions, over small sets of these variables. The
factors can be identified, e.g., by using a set of axioms (the so-called ‘graphoids’) for
reasoning about the probabilistic independence of variables [70]. The same design
principle (sic.) applies to p-DB’s [17]: the data should be decomposed into its simplest components so that only key constraints hold in a table (i.e., it is in BCNF).
Attribute- and tuple-level correlations should guide the table decomposition into
simpler tables. Ideally, the original table with its probability distribution can be
recovered as a query (a view) from the decomposed tables. We have followed such
principle in our claim-centered decomposition for predictive analytics.
In fact, a connection between database normalization theory and factor decomposition in Graphical Models (GM) has been discussed by Verma and Pearl
[70], but has not been explored since then. To date, there is no formal design theory
for p-DB’s [17]. A step in that direction is taken by Sarma et al. [71]. Their initiative revisits dependency theory in view of reformulating fd’s for uncertain schema
5.6. RELATED WORK
77
design [71]. Our work takes a different direction. We refer to classical dependency
theory and U-relational operations (viz, its uncertainty-introduction operator) to
construct p-DB’s systematically from scratch. We have focused on the extraction
and processing of fd’s towards a factorized U-relational schema. The synthesized
schema is ensured to be in BCNF and have a lossless join.
Despite some major differences, our synthesis method builds upon the classical theory of relational schema design by synthesis [62]. Classical design by synthesis [62] was once criticized due to its too strong ‘uniqueness’ of fd’s assumption [72,
p. 443], as it reduces the problem of design to symbolic reasoning on fd’s, arguably
neglecting semantic issues. Probabilistic design, however, has roots in statistical
design so that the problem is less amenable to human factors. As we extract the
dependencies from a formal specification, design by synthesis is doing nothing but
translating seamlessly (into fd’s) the reduction made by the user herself in her
tentative model for the studied phenomenon.
The last decade has seen significant research effort to make DB systems really
usable [73]. Our design-by-synthesis framework can also be understood as a technique for user-friendly p-DB design. For instance, in comparison, the CRIUS system supports another kind of user-friendly DB design approach that provides users
with a spreadsheet-like direct manipulation interface to increasingly add structure
to their data [74]. Our dependency extraction and processing, instead, completely
alleviates the user from the burden of data organization.
Also related to probabilistic DB design is the topic of conditioning a p-DB.
It has been firstly addressed by Koch and Olteanu motivated by data cleaning
applications [75]. They have introduced the assert operation to implement, as in
AI, a kind of knowledge compilation, viz., world elimination in face of constraints
(e.g., FDs). For hypothesis management, nonetheless, we need to apply Bayes’
conditioning by asserting observed data, not constraints. In §2.5 we have presented
an example that settles the kind of conditioning problem that is relevant to the
Υ-DB vision. In Chapter 6 we present realistic use cases. We have addressed the
problem at application level only in order to complete the realization of the vision
in a real prototype system. The formulation of Bayes’ conditioning as an extension
of, say, the U-relational data model is open to future work (cf. §7.3).
5.7. SUMMARY OF RESULTS
78
5.7. Summary of Results
In this chapter we have studied and developed our end-purpose technique for
the synthesis of a probabilistic DB geared for predictive analytics. It completes
the pipeline (Fig. 1.4) so that conditioning can then be performed iteratively.
• Algorithm 8 synthesize gives a general formulation of how to perform uncertainty introduction from causal dependencies given in the form of fd’s.
• By Problem 1, we have given a definition of uncertainty factor learning
from data available in a given relation.
• Remark 7 provides an example of the u-factor and predictive projections
resulting from p-DB synthesis and their corresponding probability distributions stored in the world table;
• By Remark 6 and Theorem 6, we have shown that U-relational schema Y k
synthesized over the fd’s processed by causal reasoning is in BCNF. That
is, it is in fact a claim-centered decomposition as desirable for predictive
analytics.
• Theorem 7 ensures that such (uncertainty) decomposition is correct, as the
original (probability distribution) ‘big’ fact table is fully recoverable by a
lossless join.
Chapter 6
Applicability
In this chapter we show the applicability of Υ-DB in real-world scenarios.
We present use cases in Computational Physiology extracted from the Physiome
project.1
In §6.1 we introduce the Physiome project as providing a testbed for
Υ-DB. Then in §6.2 we go through some Physiome case studies to show the construction of Υ-DB and its application for data-driven hypothesis management and
analytics. In §6.3 we present a prototype of the Υ-DB system and demonstrate it
through the running example introduced in §5.2. In §6.4 we present experiments on
Physiome hypotheses. In §6.5 we provide a general discussion on the applicability
of Υ-DB, its assumptions and scope. In §6.6 we conclude the chapter.
6.1. The Physiome Project as a Testbed
The Physiome project is an initiative to seriously address the problems of reproducibility, model integration and sharing in Computational Physiology [34, 33]. It
essentially comprises:
• a curated repository of 380+ computational physiology models available
online for researchers;2
• the Mathematical Modeling Language (MML) to allow models to be written
in declarative form and then exported into a number of XML-compliant
1
http://physiome.org.
The Physiome model repository is expanded to over 73K+ models by including models
extracted from other sources (such as the EBML-EBI BioModels DB, the CellML Archive, and
the Kegg Pathways DB) and converted to MML automatically.
2
6.2. CASE STUDIES
80
interoperable formats;3
• a problem-solving environment called JSim to allow researchers to code
their MML models straightforwardly, run them under different parameter
and solver settings and build customized data plots to see the results.
From the point of view of Υ-DB, Physiome is an external data source that
provides a very interesting testbed with realistic scenarios. We extract Physiome
models into Υ-DB by means of a wrapper we have implemented to read XMML files
(JSim’s XML encoding of MML models). Simulation trial datasets are rendered by
a parametrized UNIX script we have developed to invoke JSim automatically (in
batch mode). Currently, Υ-DB’s Physiome wrapper is designed to read MAT files
to load both the model input (parameter settings data) and its associated model
output (computed predictive data) for each simulation trial.
Physiome does not keep records of phenomena in a repository, but it does
have observational data attached to some of the entries of the model repository.
Such models appear in the filter ‘models with data,’ meaning that they have one
or more observational datasets and plots showing how the model data fits to observations. We shall make use of model entries containing observational data in
the realistic scenarios presented in this paper.
6.2. Case Studies
In this section we present use cases extracted from the Physiome model repository.4
6.2.1 Case: Hemoglobin Oxygen Saturation
In this case we stress the potential of data-driven hypothesis analytics in comparison to handcrafted curve fitting (visual) analysis. We study three different
hypotheses that perform “closely” visually when compared to their target phenomenon dataset, see Fig. 6.1. All of them have been empirically set as fit as
possible to the observations (‘R1s1’ dataset) in their local view (in separate), and
are now compared together in a global view.
3
4
http://www.physiome.org/jsim/docs/MML_Intro.html.
http://www.physiome.org/Models/modelDB/.
6.2. CASE STUDIES
81
Figure 6.1. Plot of hemoglobin oxygen saturation hypotheses (SHbO2.{Ad, H,
D} curves) and their target observations (‘R1s1’ dataset). (source: Physiome).
Example 8 The resources of this example are shown in Fig. 6.2. We consider
the Physiome model entries described in relation HYPOTHESIS, associated to the
phenomenon described in relation PHENOMENON (cf. explanation relation H0 ).
One single hypothesis trial (its best fit) is considered for each hypothesis. 2
HYPOTHESIS
PHENOMENON
υ
28
Name
HbO.Hill
Description
Hill Equation for O2 binding to hemoglobin.
31
HbO.Adair
Hemoglobin O2 saturation curve using Adair’s 4-site
equation.
32
HbO.Dash
Hemoglobin O2 saturation curve at varied levels of
PCO2 and pH.
φ
1
Description
Hemoglobin oxygen saturation with observational
dataset from Sevenringhaus 1979.
H0
φ
1
1
1
υ
28
31
32
Figure 6.2. Descriptive (textual) data of Example 8, with ids υ from Physiome’s
model repository (http://www.physiome.org/Models/modelDB/).
6.2. CASE STUDIES
82
Encoding. The fd encoding of hypotheses υ ∈ {28, 31, 32} is shown (resp.) in
Fig. 6.3, Fig. 6.4 and Fig. 6.5.
Σ28 = {
KO2 n pO2 υ
→
n p50 υ
φ
→
SHbO2 H,
→
KO2,
n p50 pO2 delta pO2 max pO2 min }.
Figure 6.3. Fd set Σ28 of hypothesis υ = 28.
Σ31 = {
φ
→
A1 A2 A3 A4 pO2 υ
p50 υ
→
→
SHbO2 Ad,
A1 A2 A4,
A3 p50 pO2 delta pO2 max pO2 min }.
Figure 6.4. Fd set Σ31 of hypothesis υ = 31.
Σ32 = {
KO2 pO2 υ
→
SHbO2 D,
→
term1 term2 term3 υ
KO2,
HpRBC K6p alphaO2 υ
→
term1,
HpRBC K3 K5 alphaCO2 pCO2 υ
→
term2,
HpRBC K2 K4 alphaCO2 pCO2 υ
→
term3,
pH υ
→
HpRBC,
φ → K2 K3 K4 K5 K6p alphaCO2 alphaO2 pCO2 pH pO2 delta, pO2 max pO2 min }.
Figure 6.5. Fd set Σ32 of hypothesis υ = 32.
Symbol Mappings. As we have seen, the insertion of hypothesis trial datasets
requires users to specify a target phenomenon and the corresponding mappings
from the hypothesis symbols to the target phenomenon symbols. In this use case,
we have:
• M287→1 = { pO2 7→ pO2, SHbO2 H 7→ SHbO2 };
• M317→1 = { pO2 7→ pO2, SHbO2 Ad 7→ SHbO2 };
• M327→1 = { pO2 7→ pO2, SHbO2 D 7→ SHbO2 };
6.2. CASE STUDIES
83
Hypothesis Management. Query Q1 illustrates the feature of hypothesis management for this case. We consider the user is interested in all SHbO2 predictions
over a subset of the pO2 domain. Its result set is shown in Fig. 6.6.
Q1. (select phi, upsilon, tid, “pO2”, “SHbO2 H” as SHbO2 from Y28 claim1
where phi=1 and “pO2”>=20 and “pO2”<=40) union all
(select phi, upsilon, tid, “pO2”, “SHbO2 Ad” as SHbO2 from Y31 claim1
where phi=1 and “pO2”>=20 and “pO2”<=40) union all
(select phi, upsilon, tid, “pO2”, “SHbO2 D” as SHbO2 from Y32 claim1
where phi=1 and “pO2”>=20 and “pO2”<=40) order by “pO2”, upsilon, tid;
Q1
φ
1
1
1
···
1
1
1
υ
28
31
32
···
28
31
32
tid
1
1
1
···
1
1
1
pO2
20
20
20
···
40
40
40
SHbO2
0.329956122828398
0.294443723056007
0.334165672301096
···
0.761898061367189
0.823463829100424
0.759042287799556
Figure 6.6. Result set of hypothesis management query Q1.
Hypothesis Analytics. Fig. 6.7 shows the results of analytics after conditoning the probability distribution in the presence of observations (‘R1s1’ dataset).
The fact that hypothesis υ = 31 provides the best explanation for the studied
phenomenon is enabled by the application of Bayesian inference as implemented
within the Υ-DB system. The contribution of the Υ-DB methodology is to equip
users with a tool for large-scale, data-driven hypothesis management and analytics.
PH1 CONF
φ
1
1
1
···
1
1
1
υ
28
31
32
···
28
31
32
tid
1
1
1
···
1
1
1
pO2
1
1
1
···
100
100
100
SHbO2
0.000151184162020125
0.003789100566457180
0.000178973375779681
···
0.974346796798538
0.990781330988763
0.972764121981342
Prior
.333
.333
.333
···
.333
.333
.333
Posterior
.326
.349
.325
.326
.349
.325
Figure 6.7. Results of analytical study on the hemoglobin phenomenon.
6.2. CASE STUDIES
84
6.2.2 Case: Baroreflex Dysfunction in Dahl SS Rat
This case is extracted from the Virtual Physiological Rat project,5 Here we show
the potential of data-driven hypothesis management and analytics for model tuning. Fig. 6.8 shows the best fit of a baroreflex model for an observational dataset
acquired by experiment on Dahl SS rat [76]. We in turn use Υ-DB to carry out
such hypothesis management and analytics. We generate by a parameter sweep
script 1K trials and insert them into the database. A best fit is then selected
automatically by Bayesian inference.
Figure 6.8. Plot of baroreflex hypothesis (‘HR’) for Dahl SS Rat and its target
observations (‘data’). (source: [76]).
Example 9 The resources of this example are shown in Fig. 6.9. We consider the
single hypothesis entry described in relation HYPOTHESIS, and the phenomenon
described in relation PHENOMENON. By parameter sweep, 1K trials are inserted
into Υ-DB for management and analytics. 2
HYPOTHESIS
υ
1001
PHENOMENON
φ
2
Name
Baroreflex SB CT
Description
Physiological model of the full baroreflex heart control
system based on experimental measurements.
Description
Baroreflex dysfunction in Dahl SS Rat.
H0
φ
2
υ
1001
Figure 6.9. Descriptive (textual) data of Example 9.
Encoding. The fd encoding of hypothesis υ ∈ {1001} is shown in Fig. 6.12.
5
http://virtualrat.org/computational-models/.
6.2. CASE STUDIES
85
Symbol Mappings. We consider that the user provides symbol mappings:
• M10017→2 = { Time 7→ Time, HR 7→ HR };
Hypothesis Management. In query Q2 we consider that the user is interested in
time instants where the heart rate is higher than a threshold, say, 300 beats/min.
The result set is shown in Fig. 6.10.
Q2. select phi, upsilon, tid, “Time”, “HR” from Y1001 claim1
where phi=2 and “HR”>=300 order by “Time”, tid;
Q2
φ
2
2
···
2
2
···
υ
1001
1001
···
1001
1001
···
tid
1
2
···
96
1
···
Time
0.61
0.61
···
0.61
0.62
···
HR
300.013659905941
300.011268345391
···
300.001934440349
300.607671377207
···
Figure 6.10. Result set of hypothesis management query Q2.
Hypothesis Analytics. Fig. 6.11 shows the results of analytics on phenomenon
φ = 2 after conditioning the probability distribution in the presence of observations
(‘SSBN9 HR’ dataset). Since this case deals with model tuning, viz., 1K slightly
different parameter settings, the trial ranking is decided by small differences in the
posterior probability distribution (cf. Fig 6.11).
PH2 CONF
φ
2
2
2
···
2
2
2
···
υ
1001
1001
1001
···
1001
1001
1001
···
tid
491
591
492
···
491
591
492
···
Time
0.4
0.4
0.4
···
118.9
118.9
118.9
···
HR
286.556506432110
286.525209765226
286.555565558231
···
421.251853783050
421.110425308905
421.297317710657
···
Prior
.001000
.001000
.001000
···
.001000
.001000
.001000
···
Posterior
.00103159
.00103144
.00103023
.00103159
.00103144
.00103023
Figure 6.11. Results of analytical study on the baroreflex phenomenon.
6.2.3 Case: Myogenic Behavior of a Blood Vessel
Computational models of physiology may account for diverse effects that take
place at different levels of biological organization from the organ to the cellular
6.2. CASE STUDIES
86
Σ1001 = {
→
HR υ
Period,
→
Beta HR p HR s HRmin HRo υ
→
delta HR pfast delta HR pslow υ
→
→
→
Gamma delta HR ps υ
HR p,
delta HR p,
HRo delta HR s υ
Time delta HR s Time min delta HR ss tau HR nor υ
HR,
→
HRo delta HR p υ
HR s,
delta HR s,
delta HR pfast,
Gamma Time delta HR ps delta HR pslow Time min tau HR ach υ → delta HR pslow,
K nor c nor delta HR smax υ
→
delta HR ss,
C ach K ach delta HR pmax υ
→
delta HR ps,
→
Time Ts c nor Time min q nor, tau nor υ
c nor,
→
C ach Time min Time Tp q ach tau ach υ
C ach,
Gs Tsmax Tsmin alpha cns alpha s0 υ
→
Ts,
Gp Tpmax Tpmin alpha cns alpha p0 υ
→
Tp,
Gcns n υ
→
alpha cns,
→
S Zeta delta delta th υ
Eps 1 Eps wall υ
→
→
B1 Eps 1 Time min Eps 2 Eps 2 Time min Eps wall K1 Kne Time υ
n,
delta,
Eps 1,
B1 B2 Eps 1 Eps 1 Time min Eps 2 Time min Eps 3 Eps 3 Time min K1 K2 Time υ → Eps 2,
→
B2 B3 Eps 2 Eps 2 Time min Eps 3 Time min K2 K3 Time υ
R R0 υ
→
Eps 3,
Eps wall,
Aυ
→
R,
A Time min Bwall Cwall P R0 Time υ
→
A,
HRmax HRo υ
→
delta HR smax,
HRmin HRo υ
→
delta HR pmax,
Time φ
→
data,
φ → A Time min B1 B2 B3 Beta Bwall C ach Time min Cwall Eps 1 Time min
Eps 2 Time min Eps 3 Time min Gamma Gcns Gp Gs HRmax HRmin HRo K1 K2 K3
K ach K nor Kne P R0 S Time delta Time max Time min Tpmax Tpmin Tsmax Tsmin
Zeta alpha p0 alpha s0 c nor Time min delta HR pslow Time min delta HR s Time min
delta th q ach q nor tau HR ach tau HR nor tau ach tau nor
}.
Figure 6.12. Fd set Σ1001 of hypothesis υ = 1001.
and molecular levels [33]. Typically, a sophisticate model is developed incrementally by, say, adding detail into some previously existing model or extending its
dimensionality (e.g., extending it from a stationary to a dynamic account of phenomena). In this case study (cf. Example 10) we consider alternative models of
the myogenic behavior of a reference human blood vessel.
6.2. CASE STUDIES
87
Figure 6.13. Plot of myogenic behavior hypothesis (‘D’) according to υ = 89,
trial tid=2, and its target observations (‘Diameter’).
Example 10 (See Fig. 6.14). We consider the Physiome model entries displayed
in relation HYPOTHESIS, and two phenomena (see relation PHENOMENON). One trial
is considered for hypothesis υ = 60, and two for hypothesis υ = 89. 2
HYPOTHESIS
PHENOMENON
υ
60
Name
Description
Myogenic Compliant This model simulates the flow through a passive
Vessel
and actively responding vessel driven by a sinusoidal pressure input.
89
Myo Dyn Resp wFit
φ
3
This model describes the dynamic response of a
vessel after a step increase in intraluminal pressure.
Description
Dynamics of vessel diameter in response to pulsatile intraluminal pressure.
H0
φ
3
3
υ
60
89
Figure 6.14. Descriptive (textual) data of Example 10.
Encoding. The fd encoding of hypotheses υ ∈ {60, 89} is shown (resp.) in Fig.
6.15 and Fig. 6.16.
Symbol Mappings. We consider that the user provides symbol mappings:
• M607→1 = { t 7→ Time, D 7→ Diameter };
• M897→1 = { t 7→ Time, D 7→ Diameter };
6.2. CASE STUDIES
88
Σ60 = {
→
Fcomp Fout υ
Fin,
Pin Pout R υ → Fout,
V V t min t υ → Fcomp,
D L mu υ
→
R,
D Lυ
→
V,
→
A,
D Pin T υ
A C1a C1p C2a C2p C3a Dp100 Pin T Ttarget υ
→
Atarget,
Atarget Cglobal Cmyo Pin υ
→
D,
D D t min Dc Tc Ttarget t taud υ
→
T,
A A t min Atarget t taua υ
Ac υ
→
→
Ttarget,
A t min,
→
Cglobal Cmyo Tc υ
Dc υ
φ
→
→
Ac,
D t min,
Dc Pmean υ
→
Tc,
Pamp Pmean t tnorm υ
→
Pin,
C1a C1p C2a C2p C3a Cglobal Cmyo Dp100 Pc υ
→
Dc,
C1a C1p C2a C2p C3a Cglobal Cmyo Dp100 L Pamp Pc Pext
Pmean Pout V t min mu t delta t max t min taua taud tnorm }.
Figure 6.15. Fd set Σ60 of hypothesis υ = 60.
Σ89 = {
→
D P Tυ
→
A C1a C1p C2a C2p C3a Dp100 P T Ttarget υ
A,
Atarget,
Atarget Cglobal Cmyo P υ
→
D,
D D t min Dc Tc Ttarget t taud υ
→
T,
→
A A t min Atarget t taua υ
Ac υ
Ttarget,
→
A t min,
→
Dc Pc υ
Tc,
Cglobal Cmyo Dc Pc υ → Ac,
Dc υ → D t min,
→
DelP Pc υ
C1a C1p C2a C2p C3a Cglobal Cmyo Dp100 Pc υ
tφ
→
→
P,
Dc,
DelP,
φ → C1a C1p C2a C2p C3a Cglobal Cmyo Dp100 Pc t delta t max t min taua taud }.
Figure 6.16. Fd set Σ89 of hypothesis υ = 89.
Hypothesis Management. Query Q3 illustrates the feature of hypothesis management for this case. The user selects all diameter predictions within the time
interval t ∈ [100, 300] (cf. plot in Fig. 6.13). Its result set is shown in Fig. 6.17.
6.3. SYSTEM PROTOTYPE
89
Q3. select phi, upsilon, tid, “t”, “D” from Y60 claim1
where phi=3 and “t”>=100 and “t”<=300 union all
select phi, upsilon, tid, “t”, “D” from Y89 claim1
where phi=3 and “t”>=100 and “t”<=300 order by “t”, upsilon, tid;
Q3
φ
3
3
3
3
3
3
···
υ
89
89
89
89
89
89
···
tid
1
1
2
1
1
2
···
t
100.00
100.00
100.00
100.01
100.01
100.01
···
D
194.622865847211
97.3787340059609
126.167727083098
194.626017703936
98.0174705905828
126.161751822302
···
Figure 6.17. Result set of hypothesis management query Q3.
Hypothesis Analytics. Fig. 6.18 shows the results of analytics on phenomenon
φ = 3 after conditoning the probability distribution in the presence of observations,
viz., ‘Davis Sikes Fig3 Myo DigData’ dataset.
PH3 CONF
φ
3
3
3
···
3
3
3
υ
60
89
89
···
60
89
89
tid
1
1
2
···
1
1
2
Time
14.8
14.8
14.8
···
30.5
30.5
30.5
Diameter
194.996792066637
97.0568250956827
116.327813203282
···
195.684170988267
97.0568250767575
116.327813337087
Prior
.5
.25
.25
···
.5
.25
.25
Posterior
.000
.269
.731
.000
.269
.731
Figure 6.18. Results of analytics on the vessel’s myogenic behavior phenomenon.
In this case study two tentative models have been considered under a uniform prior
probability distribution which has been updated to a posterior distribution. Note
that, even though hypothesis υ = 60 has its probability weight concentrated in a
single trial, the Bayesian inference is able to indicate υ = 89 as the best explanation
for φ = 3 and tid = 2, in particular, its best fit.
6.3. System Prototype
A first prototype of the Υ-DB system has been implemented as a Java web application, with the pipeline component in the server side on top of MayBMS (a backend
6.3. SYSTEM PROTOTYPE
90
extension of PostgreSQL). We have developed a demonstration of this prototype
(cf. [28]), in which we go through the whole design-by-synthesis pipeline (Fig. 1.4)
exploring use case scenarios. In this section we provide a brief demonstration of the
system in the population dynamics scenario previously introduced in this thesis.
The demonstration unfolds in three phases. In the first phase, we show
the ETL process to give a sense of what the user has to do in terms of simple
phenomena description, hypothesis naming and file upload to get her phenomena
and hypotheses available in the system to be managed as data. In the second phase,
we reproduce some typical queries of hypothesis management (like those shown
in the previous section). In the third phase, we enter the hypothesis analytics
module. The user chooses a phenomenon for a hypothesis evaluation study, and
the system lists all the predictions with their probabilities under some selectivity
criteria (e.g., population at year 1920). The predictions are ranked according to
their probabilities, which are conditioned on the observational data available for
the chosen phenomenon.
6.3.1 Demo Screenshots
Fig. 6.21 shows screenshots of the system. Fig. 6.21(a) shows the research
projects currently available for a user. Figs. 6.21(b, c) show the ETL interfaces for
phenomenon and hypothesis data definition (by synthesis), and then the insertion
of hypothesis trial datasets, i.e., explanations of a hypothesis towards a target
phenomenon. Fig. 6.21(d) shows the interface for basic hypothesis management by
listing the predictions of a given simulation trial. Figs. 6.21(e, f) show two tabs of
the hypothesis analytics module, viz., selection of observations and then viewing
the corresponding alternative predictions ranked by their conditioned probabilities.
6.3.2 Demo Case: Population Dynamics
In this case we refer to a well-known problem in Computational Science, viz.,
population dynamics scenarios, to demonstrate the Υ-DB system prototype. Fig.
6.19 shows census data collected from in the US from 1790 to 1990.6
Fig. 6.20
shows observational data collected from Hudson’s Bay from 1900 to 1920 on the
Lynx-Hare population [77].
6
Cf. https://www.census.gov/population/censusdata/table-4.pdf.
6.3. SYSTEM PROTOTYPE
91
Figure 6.19. Census US population from 1790 to 1990.
Figure 6.20. Lynx-Hare population observed in Hudon’s Bay from 1900 to 1920.
6.3. SYSTEM PROTOTYPE
92
(a) Research dashboard after login.
(b) Phenomenon data definition.
(c) Hypothesis data definition.
(d) Hypothesis management.
(e) Analytics: selected observations tab.
(f) Analytics: ranked predictions tab.
Figure 6.21. Screenshots of this first prototype of the Υ-DB system.
6.3. SYSTEM PROTOTYPE
93
Example 11 (See Fig. 6.22). We consider the model entries displayed in relation
HYPOTHESIS,
and two phenomena (see relation PHENOMENON). For φ = 1, three
trials are considered for hypothesis υ = 1 and six for hypothesis υ = 2. For φ = 2,
in turn, two trials are considered for hypothesis υ = 1 and υ = 2, and six trials for
hypothesis υ = 3. Note the data definition interfaces in Figs. 6.21(b, c). 2
HYPOTHESIS υ
1
Name
Malthusian growth
model
Description
Exponential growth model ‘growth in population
is proportional to its size’ is considered the first
principle of population dynamics.
2
Logistic equation
This model introduces growth saturation to the
Malthusian model due to the limitation of resources.
3
Lotka-Volterra
model
This model describes predator-prey interactions.
H0
PHENOMENON
φ
1
2
Description
US population from 1790 to 1990.
Lynx population in Hudson’s Bay, Canada, from
1900 to 1920.
φ
1
1
2
2
2
υ
1
2
1
2
3
Figure 6.22. Descriptive (textual) data of Example 11.
Encoding. The fd encoding of hypotheses υ ∈ {1, 2, 3} is shown (resp.) in Fig.
6.23, Fig. 6.24 and Fig. 6.25. See hypothesis structure processing in Fig. 6.21(c).
Σ1 = {
φ
→
b t x t min υ
→
x,
b t delta t max t min x t min }.
Figure 6.23. Fd set Σ1 of hypothesis υ = 1.
Σ2 = {
φ
→
K b t x t min υ
→
x,
K b t delta t max t min x t min }.
Figure 6.24. Fd set Σ2 of hypothesis υ = 2.
Symbol Mappings. We consider that the user provides the following symbol
mappings for (resp.) phenomena φ = 1 and φ = 2, see the interface for mapping
symbols in Fig. 6.21(c).
6.3. SYSTEM PROTOTYPE
94
Σ3 = {
φ
→
bpt yυ
→
x,
d r t x υ → y,
b d p r t delta t max t min x t min y t min }.
Figure 6.25. Fd set Σ3 of hypothesis υ = 3.
• M17→1 = { t 7→ Year, x 7→ Population };
• M27→1 = { t 7→ Year, x 7→ Population };
• M17→2 = { t 7→ Year, x 7→ Lynx };
• M27→2 = { t 7→ Year, x 7→ Lynx };
• M37→2 = { t 7→ Year, x 7→ Lynx };
Hypothesis Management. Query Q4 illustrates the feature of hypothesis management for this case. The user selects hypothesis υ = 3 (the Lotka-Volterra model),
and filters its available data for trial tid = 6 on phenomenon φ = 2. Both the formbased query set-up and its result set are shown in Fig. 6.21(d).
Q4. select “t”, “y”, “x” from Y3 claim1
where upsilon=3 and phi=2 and tid=6 order by “t”;
Hypothesis Analytics. Fig. 6.26 and Fig. 6.27 show the results of analytics on
(resp.) phenomena φ = 1 and φ = 2 after conditoning the probability distribution
in the presence of (resp.) observational datasets ‘US-census’ and ‘Lynx-Hare.’ In
the first one, the user verifies that hypothesis υ = 1 (the Malthusian model) is
unlikely to be competitive with hypothesis υ = 2 (the Logistic equation) as an
approximation of the US population dynamics from 1790 to 1990. That is, if the
user knows her current trials are reasonable, then more trials on the Malthusian
model hardly could outperform trials on the Logistic equation for the studied
phenomenon.
6.4. EXPERIMENTS
PH1 CONF
95
φ
···
1
1
1
1
1
1
···
υ
···
1
1
2
2
2
2
···
tid
···
1
2
1
2
3
4
···
Year
···
1920
1920
1920
1920
1920
1920
···
Population
···
194.102222534948
140.244165184248
82.3031951115155
108.251924734215
105.918217777077
105.988231944275
···
Prior
···
.250
.250
.125
.125
.125
.125
···
Posterior
.000000
.000000
.133038
.239684
.290026
.337251
Figure 6.26. Results of analytics on the US population phenomenon.
PH2 CONF
φ
···
2
2
2
2
2
2
2
2
2
2
···
υ
···
1
1
2
2
3
3
3
3
3
3
···
tid
···
1
2
1
2
1
2
3
4
5
6
···
Year
···
1904
1904
1904
1904
1904
1904
1904
1904
1904
1904
···
Lynx
···
16.49
18.22
79.81
77.82
89.59
65.06
90.08
77.46
88.32
75.92
···
Prior
···
.167
.167
.167
.167
.055
.055
.055
.055
.055
.055
···
Posterior
.047
.000
.013
.017
.131
.184
.124
.176
.127
.180
Figure 6.27. Results of analytics on the Hudson’s Bay lynx population phenomenon, see interfaces in Figs. 6.21(e, f).
6.4. Experiments
The efficiency and scalability of the U-relational representation system and
its p-WSA query algebra have been extensively demonstrated [27]. Υ-DB’s, as
U-relational hypothesis DB’s, must therefore be as efficient and scalable as any
arbitrary U-relational DB.
In these experiments (see Fig. 6.28) we provide some measures of performance of the method of Υ-DB in the particular context of our real-world Physiome testbed. Our purpose here is to provide a concrete feel on how efficient the
Υ-DB methodology can be. However, most of these tests (the four graphs on the
bottom in Fig. 6.28) involve the data level and then require more of the hardware.
Our current experimental setup (personal computer)7 allows us to reach a scale
7
These experiments were performed on a 2.3GHZ/4GB Intel Core i5 running Mac OS X
10.6.8 and MayBMS (a PostgreSQL 8.3.3 extension).
6.4. EXPERIMENTS
96
1,000
XML extraction
time [ms]
time [ms]
75
750
500
250
50
encoding
25
0
0
500
1,000
1,500
0
2,000
500
80
70
u-learning
time [ms]
time [ms]
1,500
2,000
structure length |S|
structure length |S|
60
1,000
30
20
u-factorization
50
0
200
400
600
800
0
1,000
200
400
600
800
1,000
ntrials
ntrials
100
time [s]
time [s]
800
u-propagation
50
600
conditioning
400
200
0
200
400
600
ntrials
800
1,000
200
400
600
800
1,000
ntrials
Figure 6.28. Performance behavior of Υ-DB on a Physiome testbed scenario.
in which the uncertain data being processed in synthesis ‘4U’ is sized up to 1 GB.
For the two first graphs (XML extraction and encoding), we have collected the
response time on the measure of interest over different structure lengths. Each one
corresponds to a real Physiome hypothesis from the table of Fig. 6.29. The last
hypothesis in that table, υ = 379, is used for the tests of the four last graphs in Fig.
6.28, viz., u-learning, u-factorization, u-propagation and conditioning. We have set
different number of trials (ntrials) over it, each one having 1 M B. The last test
in each of such four graphs, with 1K trials, is processing 1 GB of uncertain data
at once and then fits the machine’s main memory. We interpret the performance
results shown in these graphs as follows for each measure of performance.
• Extraction. Some fluctuation may be due to practicalities of XML DOM
access methods. The point of this performance study is to have practical
6.4. EXPERIMENTS
HYPOTHESIS
97
υ
186
89
60
75
70
120
91
93
153
154
379
name
Regulatory Vessel
Myo Dyn Resp wFit
Myogenic Compliant Vessel
Baroreceptor Lu et al 2001
4-State Sarcomere Energetics
Comp four gen weibel lung
CardiopulmonaryMechanics
CardiopulmonMechGasBloodExch
HighlyIntegHuman
HighlyIntHuman wIntervention
Baroreflex SB CT
|S|
40
73
100
153
298
440
1132
1593
1624
1919
171
|E|
20
28
38
74
91
186
412
525
538
634
74
Figure 6.29. Physiome hypotheses used in the experiments.
measures of the amount of time taken to process representative hypothesis
structures. Note that even for structures of size |S| = 2 K the amount
of time required to extract a hypothesis is kept at subsecond order of
magnitude (interactive response time) in a personal machine.
• Encoding. Some fluctuation is expected due to varying degrees of coupling
between variables in the hypothesis structures. Note that, although |S|
provides a very good measure of their size and complexity, the extent to
which they are intricate should cause impact on the encoding procedure,
p
which is any case kept O( |E||S|). Again, the point here is to provide a
notion of the amount of time required to encode representative hypotheses.
For a scalability test on the encoding procedure, cf. Fig. 3.8.
• U-intro. The U-intro procedure is composed of u-factor learning, u-factori zation and u-propagation. We observed in previous tests that it was dominated by the learning component, viz., the discovery of occasional fd’s in
the ‘big’ fact table. However, this is no longer the case once we implemented the workaround of keeping (in addition to the ‘big’ table) a table
containing only the exogenous (input parameter) variables, as it has negligible size w.r.t. the data of the endogenous (predictive output) variables.
Then u-learning became subsecond again and the U-intro procedure became
dominated by u-propagation. In fact, the procedure of u-factorization, carried out once the fd’s are discovered by u-learning, is also sub-second then
6.5. DISCUSSION
98
has negligible processing time w.r.t. u-propagation. The latter is the most
expensive sub-procedure of the synthesis method.
• Conditioning. The conditioning procedure is run for a selected phenomenon.
It is composed of four main parts. First, by operation conf() it performs
a probabilistic inference sub-query on the proper predictive projection of
the ‘big’ fact table of each hypothesis associated with the phenomenon.
Second, it combines the results of each such sub-query through a union
all query whose result set is a multi-hypothesis predictive table. Third,
it loads the phenomenon observation sample data and the predictive data
from the multi-hypothesis table into memory to apply Bayesian inference.
Finally, the prior probability distribution of the predictive table is updated
with the posterior and all the corresponding marginal probabilities are updated in their original U-relational tables. In our tests, this procedure is
carried out over varying number of trials (ntrials). The total response times
are shown in the last plot of Fig. 6.28.
This performance behavior is to be interpreted in the context of ETL in
DW’s. Loading and setting up an Υ-DB has an overhead that shall be, though,
much lower in high-performance machines. Such overhead is nonetheless justified
for the use case of hypothesis management and analytics as opposed to simulation
data management and exploratory analytics (cf. §2.6.2).
6.5. Discussion
We have verified that the hypothesis ratings/rankings shown in §6.2 coincide
with the results (e.g., of model tuning) described in the Physiome model entries and
their related publications. That validates the applicability of the Υ-DB methodology as a tool for data-driven analysis in such realistic scenarios.
The current practice in Computational Science for model evaluation and
comparison in the presence of observational data is somewhat handcrafted: model
agreement is assessed either qualitatively by referring to curve shapes in data plots
or quantitatively by means of ad-hoc scripts. The Υ-DB methodology offers a
tool to perform data-driven hypothesis analytics semi-automatically directly in the
6.5. DISCUSSION
99
database under the support of its querying capabilities. It has, therefore, potential
to be a step towards higher standards of reproducibility and scalability.
Realistic assumptions. The core assumption of our framework is that the
hypotheses are given in a formal specification which is encodable into a SEM that
is complete (satisfies Defs. 1, 2). Also, as a semantic assumption which is standard
in scientific modeling, we consider a one-to-one correspondence between real-world
entities and variable/attribute symbols within a structure, and that all of them
must appear in some of its equations/fd’s. For most science use cases involving
deterministic models (if not all), such assumptions are quite reasonable. It can be
a topic of future work (cf. §7.3) to explore business use cases as well.
Hypothesis learning. The (user) method for hypothesis formation is irrelevant to our framework, as long as the resulting hypothesis is encodable into a
SEM. So, a promising use case is to incorporate machine learning methods into our
framework to scale up the formation/extraction of hypotheses and evaluate them
under the querying capabilities of a p-DB. Consider, e.g., learning the equations,
say, from Eureqa [2].8
Qualitative hypotheses. The Υ-DB methodology is primarely motivated
by computational science (usually involving differential equations). It is, however,
applicable to qualitative deterministic models as well. Boolean Networks, e.g.,
consist in sets of functions f (x1 , x2 ,.., xn ), where f is a Boolean expression. For
instance, Fig. 6.30 presents the system of Boolean equations of a tentative Boolean
Network model for a plant hormone (Fig. 6.31) published in [78].9 The notation,
e.g., SphK*, is read (just like an ordinary differential equation), ‘the next state
value of variable SphK is given by the state value of variable ABA. The parameters
in this kind of model are the variable initial conditions.
8
9
http://creativemachines.cornell.edu/Eureqa.
Cf. http://atlas.bx.psu.edu/booleannet/booleannet.html.
6.5. DISCUSSION
H={
100
SphK*
=
ABA,
S1P*
=
SphK,
GPA1*
=
(S1P or not GCR1) and AGB1,
PLD*
=
GPA1,
PA*
=
PLD,
pHc*
=
ABA,
OST1*
=
ABA,
ROP2*
=
PA,
Atrboh*
=
pHc and OST1 and ROP2 and not ABI1,
ROS*
=
Atrboh
H+ATPase*
=
not ROS and not pHc and not Ca2+ c
ABI1*
=
pHc and not PA and not ROS,
RCN1*
=
ABA,
NIA12*
=
RCN1,
NOS*
=
Ca2+ c,
NO*
=
NIA12 and NOS,
GC*
=
NO,
ADPRc*
=
NO,
cADPR*
=
ADPRc,
cGMP*
=
GC,
PLC*
=
ABA and Ca2+ c,
InsP3*
=
PLC,
InsPK*
=
ABA,
InsP6*
=
InsPK,
CIS*
=
(cGMP and cADPR) or (InsP3 and InsP6),
Ca2 ATPase*
=
Ca2+ c,
Ca2+ c *
=
(CaIM or CIS) and (not Ca2+ ATPase),
AnionEM*
=
((Ca2+ c or pHc) and not ABI1 ) or (Ca2+ c and pHc),
Depolar*
=
KEV or AnionEM or (not H+ATPase) or (not KOUT) or Ca2+ c,
CaIM*
=
(ROS or not ERA1 or not ABH1) and (not Depolar),
KOUT*
=
(pHc or not ROS or not NO) and Depolar,
KAP*
=
(not pHc or not Ca2+ c) and Depolar,
KEV*
=
Ca2+ c,
PEPC*
=
not ABA,
Malate*
=
PEPC and not ABA and not AnionEM,
RAC1*
=
not ABA and not ABI1,
Actin*
=
Ca2+ c or not RAC1,
Closure*
=
(KOUT or KAP ) and AnionEM and Actin and not Malate
+
Figure 6.30. Example of Boolean Network hypothesis.
}.
6.6. CONCLUSIONS
101
Figure 6.31. Example of Boolean Network model (source: [78]).
Several kinds of dynamical system can be modeled in this formalism. Applications have grown out of gene regulatory network to social network and stock
market predictive analytics. Even if richer semantics is considered (e.g., fuzzy logics), our encoding method is applicable likewise, as long as the equations are still
deterministic.
6.6. Conclusions
In this chapter we have demonstrated and discussed the applicability of the Υ-DB
methodology. We have referred to real-world use case scenarios derived from the
Physiome research project. We have shown in some detail the process of building
an Υ-DB with representative models from Physiome’s model repository. That
qualitative assessment is followed by experiments that provide some concrete feel
on the performance behavior of Υ-DB for models with up to 600+ mathematical
variables.
Chapter 7
Conclusions
In this chapter we (§7.1) revisit the research questions addressed by this
thesis, (§7.2) point out its significance and limitations, (§7.3) list open problems
and topics for future work, and (§7.4) conclude with final considerations.
7.1. Revisiting the Research Questions
Let us now revisit the conceptual (RQ1-4) and technical (RQ5-9) research
questions.
RQ1. How to define and encode hypotheses ‘as data’? What are the sources of
uncertainty that may be present and should be considered?
In Chapter 2 we have provided core abstractions that compose the vision
of hypotheses ‘as data,’ or the Υ-DB vision. The problem of hypothesis
encoding has been defined and addressed further in Chapter 3. We have
distinguished two main sources of uncertainty in our model of uncertainty
for hypothesis management, viz., (i) theoretical uncertainty, as arising from
competing hypotheses; and (ii) empirical uncertainty, as arising from the
alternative trial datasets available for each hypothesis.
RQ2. How does hypotheses ‘as data’ relate with observational data or, likewise,
phenomena ‘as data’ from a database perspective?
Also in Chapter 2, we have presented a conceptual framework in which
we have defined hypotheses ‘as data’ and shown how it can be compared
against phenomena ‘as data.’ In fact, hypothesis management is really
7.1. REVISITING THE RESEARCH QUESTIONS
103
significant when it is possible to rate/rank hypotheses in the presence of
(some partial piece of) evidence.
RQ3. Does every piece of simulated data qualify as a scientific hypothesis? What
is the difference between managing ‘simulation’ data from managing ‘hypotheses’ as data?
Early in Table 1.1, we provided a comparison between simulation data
management and hypothesis data management. Furthermore, the scientific
research process is abstracted in Chapter 2 as a well-defined problem of
data cleaning. Hypotheses are seen from an applied science point of view,
and then are reduced into data such that a piece of simulation data is
considered a hypothesis whenever it is assigned to explain some specific
phenomenon.
RQ4. Is there available a proper (machine-readable) data format we can use to
automatically extract mathematically-expressed hypotheses from?
We anticipated in Chapter 2 the adoption of the XML data model as
the general data format for extracting hypothesis specifications from. In
particular, since we deal here with mathematical hypotheses, we refer to
MathML as a standard for hypothesis specification. Concretely, in Chapter
6 we present use case demonstration scenarios for which we have developed
a specific wrapper, viz., for the extraction of hypotheses specified in MML
(Mathematical Modeling Language).
RQ5. Is there an algorithm to, given a SEM, efficiently extract its causal ordering? What are the computational properties of this problem?
As shown in Chapter 3, Simon’s treatment of the problem of causal ordering given a SEM S(E, V) is NP-Hard. In the same chapter, we have
discussed this problem in detail and presented an effective, efficient algorithmic approach to the problem. The computational cost for the whole
p
process of hypothesis encoding is bounded by O( |S||E|). Experiments
show that the approach performs well in practice for large hypotheses.
7.1. REVISITING THE RESEARCH QUESTIONS
104
RQ6. What is the connection between SEM’s and fd’s? Can we devise an encoding scheme to ‘orient equations’ and then effectively transform one into
the other with guarantees? Once we do it, what design-theoretic properties
have such a set of fd’s?
Also in Chapter 3, we have presented an algorithmic encoding scheme to
transform a SEM into a set of fd’s with guarantees in terms of preserving
the hypothesis causal structure. Our study of this problem has revealed
some interesting properties of the resulting fd sets, in particular, that they
are always ‘non-redundant’ and, in comparison with arbitrary information
systems, more precise and economical in the sense that, for any given
attribute, there is exacly one fd with it in its rhs.
RQ7. Is such fd set ready to be used for p-DB schema synthesis as an encoding
of the hypothesis causal structure? If not, what kind of further processing
we have to do? Can we perform it efficiently by reasoning directly on the
fd’s? How does it relate to the SEM’s causal ordering?
As we discuss in Chapter 4, the encoded fd set must be further processed
to find the ‘first causes’ for each of its predictive variable. For addressing
that, in Chapter 4 we have presented the concept of the folding of an fd
set and an efficient algorithm to compute it. Also, we have shown the
equivalence of such fd reasoning with causal ordering processing.
RQ8. Is the uncertainty decomposition required for predictive analytics reducible
to the structure level (fd processing), or do we need to process the simulated
data to identify additional uncertainty factors? Finally, what properties are
desirable for a p-DB schema targeted at hypothesis management? Are they
ensured by this synthesis method?
In Chapter 5 we have presented a conceptual framework to address synthesis for uncertainty ‘4U.’ In particular, we have introduced the need
to process, for each hypothesis, its trial datasets available, and presented
an efficient algorithm to factorize and propagate the overall uncertainty
present in a given hypothesis (as a competing explanation for a target
7.2. SIGNIFICANCE AND LIMITATIONS
105
phenomenon). Then we have motivated BCNF as a notion of “good” design w.r.t. the factorized fd set based on the folding concept, and the
lossless join property as required for the correctness of uncertainty decomposition. We have shown that the synthesized p-DB schema bears both
properties.
RQ9. Given all such a design-theoretic machinery to process hypotheses into
(U-)relational DB’s, what properties can we detect on the hypotheses back
at the conceptual level? Do we have now technical means to speak of hypotheses that are “good” in terms of principles of the philosophy of science?
Equipped with the design-theoretic machinery proposed in this thesis, we
are able to, given a SEM, to automatically (1) extract its causal ordering,
(2) detect its strongly coupled components and decide, for a given predictive projection, what are its associated u-factor projections (if any), and
shall be able as well to (3) query the hypothesis ranking for a phenomenon
of interest. All these are technical means to [15]: (10 ) extract the hypothesis ‘empirical content’ and ‘predictive power;’ (20 ) unravel its cohesiveness
and how parsimonious it is in terms of the number of different claims or
epistemological units carried within it, as well as its empirical grounding
(‘first causes’); and finally, we shall be able to (30 ) appraise it in face of
competing or alternative explanations.
7.2. Significance and Limitations
This thesis addresses the pressing call for large-scale, data-driven hypothesis management and analytics [35, 3, 10]. Some reasons that contribute for its
significance are listed (cf. [10, 29, 28]).
• Structured deterministic hypotheses are now shown to be encodable as
uncertain and probabilistic (U-relational) data based on p-DB principles;
∗ Study of the connection between SEM’s and fd’s, with contribution
both to computational properties of the causal ordering problem, and
to causal reasoning over fd’s;
7.3. OPEN PROBLEMS AND FUTURE WORK
106
∗ First synthesis method for the construction of p-DB’s from some previous existing formal specification.
• Definition of a concrete use case of data-driven hypothesis management
and analytics;
∗ New class of applications introduced for p-DB’s;
∗ Settled the problem of Bayes’ conditioning in p-DB’s.
Now some limitations of the thesis are listed.
• The Bayesian inference is implemented at application level, yet not formulated as a principled technical solution within research in p-DB’s.
• The encoding scheme to transform the mathematical system of a hypothesis into a set of fd’s enabling the synthesis of the p-DB is applicable to
structured deterministic models only, not stochastic ones.
7.3. Open Problems and Future Work
Open problems and topics of future work are listed (no particular order).
(1) The design of a dedicated algebraic operation for Bayes’ conditioning in
p-WSA.
(2) Investigation of other data dependency formalisms (e.g., multi-valued dependencies [20]), approximate fd’s [68], conditional fd’s [79]) to extend the
scope of Υ-DB towards structured stochastic models.
(3) Development of techniques for systematic hypothesis extraction as a welldefined problem of (web) information extraction;
(4) Investigation of business use case scenarios for data-driven decision making
on top of Υ-DB;
(5) Definition of a machine learning use case scenario to industrialize hypothesis formation and assess Υ-DB’s performance feasibility in such a scenario;
7.4. FINAL CONSIDERATIONS
107
(6) Development of automatic data sampling techniques to leverage the data
definition of both hypotheses and phenomena in Υ-DB from a statistical
point of view.
7.4. Final Considerations
In this thesis we have developed the vision of Υ-DB, which is essentially
an abstraction of hypotheses as uncertain and probabilistic data. It comprises
a design-theoretic methodology for the systematic construction and management
of U-relational hypothesis DB’s. It is meant to provide a principled approach
to enable scientists and engineers to manage and evaluate (rate/rank) large-scale
scientific hypotheses as theoretical data. We have addressed some core technical
challenges over the Υ-DB vision in order to properly encode deterministic hypotheses as uncertain and probabilistic data.
As envisioned by Jim Gray [1], the scientific method has been shifting towards
being operated as a data-driven discipline which is rapidly gaining ground [3]. In
this thesis we have strived for proposing some core principles and techniques for
enabling data-driven hypothesis management and analytics, opening a promising
line of research in both probabilistic databases and simulation data management.
Bibliography
[1] HEY, T.; TANSLEY, S.; TOLLE, K.
The fourth paradigm: Data-
intensive scientific discovery. Microsoft Research, 2009.
[2] SCHMIDT, M.; LIPSON, H. Distilling free-form natural laws from experimental data. Science, Washington, v. 324, n. 5923, p. 81–5, 2009.
[3] DHAR, V. Data science and prediction. Communications of the ACM,
v. 56, n. 12, p. 64–73, 2013.
[4] JAGADISH, H. V.; GEHRKE, J.; LABRINIDIS, A.; PAPAKONSTANTINOU, Y.; PATEL, J. M.; RAMAKRISHNAN, R.; SHAHABI, C. Big data
and its technical challenges. Communications of the ACM, v. 57, n. 7, p.
86–94, 2014.
[5] BENIOFF, M. R.; LAZOWSKA, E. D. (Eds.). Computational Science:
Ensuring America’s competitiveness. PITAC (US President’s Information Technology Advisory Committee), 2005.
[6] PERLMAN, E.; BURNS, R.; LI, Y.; MENEVEAU, C. Data exploration of
turbulence simulations using a database cluster. In: . c2007.
[7] MARKRAM, H. The Blue Brain Project. Nature Reviews Neuroscience,
v. 7, p. 153–60, 2006.
[8] AILAMAKI, A.; KANTERE, V.; DASH, D.
Managing scientific data.
Comm. ACM, v. 53, n. 6, p. 68–78, 2010.
[9] AHMAD, Y.; BURNS, R.; KAZHDAN, M.; MENEVEAU, C.; SZALAY, A.;
TERZIS, A. Scientific data management at the Johns Hopkins Institute for
BIBLIOGRAPHY
109
Data Intensive Engineering and Science. SIGMOD Record, v. 39, n. 3, p.
18–23, 2010.
[10] GONCALVES, B.; PORTO, F. Υ-DB: Managing scientific hypotheses as
uncertain data. PVLDB, v. 7, n. 11, p. 959–62, 2014.
[11] CUSHING, J. B. Beyond big data? Computing in Science & Engineering, v. 15, n. 5, p. 4–5, 2013.
[12] WEINBERG, R. Point: Hypotheses first. Nature, London, v. 464, p. 678,
2010.
[13] GOLUB, T. Counterpoint: Data first. Nature, London, v. 464, p. 679, 2010.
[14] FRÉGNAC, Y.; LAURENT, G. Where is the brain in the Human Brain
Project? Nature, London, v. 513, p. 27–9, 2014.
[15] LOSEE, J. A historical introduction to the philosophy of science.
Oxford Univ. Press, 4th ed., 2001.
[16] BOLSTAD, W. M. Introduction to bayesian statistics. 2nd. ed. WileyInterscience, 2007.
[17] SUCIU, D.; OLTEANU, D.; RÉ, C.; KOCH, C. Probabilistic databases.
Morgan & Claypool Publishers, 2011.
[18] KOCH, C. MayBMS: A system for managing large uncertain and
probabilistic databases. In C. Aggarwal (ed.), Managing and Mining Uncertain Data, Chapter 6. Springer-Verlag, 2009.
[19] DARWICHE, A. Bayesian networks. Comm. ACM, v. 53, n. 12, p. 80–90,
2010.
[20] ULLMAN, J. Principles of Databases and Knowledge-Base Systems.
Computer Science Press, 1988.
[21] ABITEBOUL, S.; HULL, R.; VIANU, V. Foundations of databases.
Addison-Wesley, 1995.
BIBLIOGRAPHY
110
[22] MAIER, D. Theory of relational databases. Computer Science Press,
1983.
[23] BADIA, A.; LEMIRE, D. A call to arms: Revisiting database design. SIGMOD Record, v. 40, n. 3, p. 61–9, 2011.
[24] SIMON, H. Causal ordering and identifiability. In Hood & Koopmans
(eds.), Studies in Econometric Methods, Chapter 3, John Wiley & Sons, 1953.
[25] PEARL, J. Causality: Models, reasoning, and inference. Cambridge
Univ. Press, 2000.
[26] CHANG, C.-H.; KAYED, M.; GIRGIS, M. R.; SHAALAN, K. A survey of
web information extraction systems. IEEE Transactions on Knowledge
and Data Engineering, v. 18, n. 10, p. 1411–28, 2006.
[27] ANTOVA, L.; JANSEN, T.; KOCH, C.; OLTEANU, D. Fast and simple
relational processing of uncertain data. In: Proc. of IEEE ICDE. 2008. p.
983–92.
[28] GONCALVES, B.; SILVA, F. C.; PORTO, F. Υ-DB: A system for datadriven hypothesis management and analytics. Technical report, LNCC, 2015.
(available at CoRR abs/1411.7419).
[29] GONCALVES, B.; PORTO, F. Design-theoretic encoding of deterministic
hypotheses as constraints and correlations in U-relational databases. Technical
report, LNCC, 2015.
[30] SIMON, H.; RESCHER, N. Cause and counterfactual. Philosophy of Science, v. 33, n. 4, p. 323–40, 1966.
[31] HINES, M.; MORSE, T.; MIGLIORE, M.; CARNEVALE, N.; SHEPHERD,
G. ModelDB: A database to support computational neuroscience. J. Comput. Neurosci., v. 17, n. 1, p. 7–11, 2004.
[32] CHELLIAH, V.; LAIBE, C.; Le Novère, N. BioModels Database: A repository
of mathematical models of biological processes. Method. Mol. Biol., , n.
1021, p. 189–99, 2013.
BIBLIOGRAPHY
111
[33] HUNTER, P. J.; BORG, T. K. Integration from proteins to organs: the
Physiome Project. Nat. Rev. Mol. Cell. Biol., v. 4, n. 3, p. 237–43, 2003.
[34] BASSINGTHWAIGHTE, J. B. Strategies for the Physiome Project. Ann.
Biomed. Eng., v. 28, p. 1043–58, 2000.
[35] HAAS, P.; MAGLIO, P.; SELINGER, P.; TAN, W. Data is dead... without
what-if models. PVLDB, v. 4, n. 12, p. 1486–9, 2011.
[36] RAHM, E.; Hai Do, H. Data cleaning: Problems and current approaches.
IEEE Data Engineering Bulletin, v. 23, n. 4, p. 3–13, 2001.
[37] BESKALES, G.; SOLIMAN, M. A.; ILYAS, I. F.; BEN-DAVID, S. Modeling
and querying possible repairs in duplicate detection. PVLDB, v. 2, n. 1, p.
598–609, 2009.
[38] PORTO, F.; SPACAPPIETRA, S. Data model for scientific models and
hypotheses. In R. Kaschek, L. Delcambre (ed.), The evolution of Conceptual
Modeling, p. 285-305, Springer-Verlag LNCS vol. 6520, January 2011.
[39] STOUGIANNIS, A.; TAUHEED, F.; PAVLOVIC, M.; HEINIS, T.; AILAMAKI, A. Data-driven Neuroscience: Enabling breakthroughs via innovative
data management. In: . c2013.
[40] KANOV, K.; PERLMAN, E. A.; BURNS, R. C.; AHMAD, Y.; SZALAY, A. S.
I/O streaming evaluation of batch queries for data-intensive computational
turbulence. In: . c2011. p. 1–10.
[41] IDREOS, S.; ALAGIANNIS, I.; JOHNSON, R.; AILAMAKI, A. Here are my
data files. here are my queries. where are my results? In: In Proc. of the
biennial Conf. on Innovative Data Systems Research (CIDR). 2011.
[42] ALAGIANNIS, I.; BOROVICA, R.; BRANCO, M.; IDREOS, S.; AILAMAKI,
A. NoDB: Efficient query execution on raw data files. In: Proc. of ACM
SIGMOD. 2012. p. 241–52.
[43] SOLDATOVA, L.; RZHETSKY, A. Representation of research hypotheses.
J. Biomed. Sem., v. 2, n. S2, 2011.
BIBLIOGRAPHY
112
[44] KING, R. D. et al. The automation of science. Science, Washington, v. 324,
n. 5923, p. 85–9, 2009.
[45] RACUNAS, S. et al. HyBrow: a prototype system for computer-aided hypothesis evaluation. Bioinformatics, v. 20, n. 1, p. 257–64, 2004.
[46] GAO, Y. et al. SWAN: A distributed knowledge infrastructure for alzheimer
disease research. J. Web Semantics, v. 4, n. 3, p. 222–8, 2006.
[47] de WAARD, A. et al. Hypotheses, evidence and relationships: The HypER
approach for representing scientific knowledge claims. In: ISWC Proc. of
Workshop on Semantic Web Applications in Scientific Discourse.
2009.
[48] DASH, D.; DRUZDZEL, M. J. A note on the correctness of the causal ordering
algorithm. Artif. Intell., v. 172, n. 15, p. 1800–8, 2008.
[49] NAYAK, P. P. Causal approximations. Artificial Intelligence, Amsterdam,
v. 70, n. 1-2, p. 277–334, 1994.
[50] DRUZDZEL, M. J.; SIMON, H. A. Causality in Bayesian belief networks. In:
Proc. of Int. Conf. on Uncertainty in Artificial Intelligence. 1993. p.
3–11.
[51] MADEIRA, S. C.; OLIVEIRA, A. L. Biclustering algorithms for biological
data analysis: A survey. IEEE Transactions on Computational Biology
and Bioinformatics, v. 1, n. 1, p. 24–45, 2004.
[52] DHILLON, I. S.; MALLELA, S.; MODHA, D. S. Information-theoretic coclustering. In: Proc. of ACM SIGKDD. 2003. p. 89–98.
[53] EVEN, S. Graph algorithms. 2nd. ed. Cambridge Univ. Press, 2011.
[54] UNO, T. An efficient algorithm for solving pseudo clique enumeration problem. Algorithmica, v. 56, n. 1, p. 3–16, 2010.
[55] GAREY, M. R.; JOHNSON, D. S. Computers and intractability: A
guide to the theory of NP-completeness. 1st. ed., Series of Books in the
Mathematical Sciences. W. H. Freeman, 1979.
BIBLIOGRAPHY
113
[56] JOHNSON, D. S. The NP-completeness column: an ongoing guide. J. Algorithms, v. 8, n. 3, p. 438–88, 1987.
[57] HOPCROFT, J. E.; KARP, R. M. An n5/2 algorithm for maximum matchings
in bipartite graphs. SIAM Journal on Computing, v. 2, n. 4, p. 225–31,
1973.
[58] CORMEN, T. H.; LEISERSON, C. E.; RIVEST, R. L.; STEIN, C. Introduction to algorithms. 3rd. ed. The MIT Press, 2009.
[59] SIMON, H. On the definition of the causal relation. The Journal of Philosophy, v. 49, p. 517–28, 1952.
[60] NAYAK, P. P. Automated modelling of physical systems. SpringerVerlag, 1996.
[61] SERRANO, D.; GOSSARD, D. C. Constraint management in conceptual design. In: Knowledge Based Expert Systems in Engineering: Planning
and Design. Computational Mechanics Publications, 1987. p. 211–24.
[62] BERNSTEIN, P. Synthesizing third normal form relations from functional
dependencies. ACM Trans. on Database Systems, v. 1, n. 4, p. 277–98,
1976.
[63] BEERI, C.; BERNSTEIN, P. Computational problems related to the design
of normal form relational schemas. ACM Trans. on Database Systems,
v. 4, n. 1, p. 30–59, 1979.
[64] MELIOU, A. et al. Causality in databases. IEEE Data Eng. Bull., v. 33,
n. 3, p. 59–67, 2010.
[65] MELIOU, A.; GATTERBAUER, W.; MOORE, K. F.; SUCIU, D. The complexity of causality and responsibility for query answers and non-answers.
PVLDB, v. 4, n. 1, p. 34–45, 2010b.
[66] KANAGAL, B.; LI, J.; DESHPANDE, A. Sensitivity analysis and explanations for robust query evaluation in probabilistic databases. In: Proc. of
ACM SIGMOD. 2011. p. 841–52.
BIBLIOGRAPHY
114
[67] MELIOU, A.; GATTERBAUER, W.; SUCIU, D. Reverse data management.
PVLDB, v. 4, n. 12, p. 1490–3, 2011.
[68] HUHTALA, Y. et al. TANE: An efficient algorithm for discovering functional
and approximate dependencies. Computer Journal, v. 42, n. 2, p. 100–11,
1999.
[69] DARWICHE, A. Modeling and reasoning with Bayesian Networks.
Cambridge University Press, 2009.
[70] TOM S. VERMA, J. P. Causal Networks: Semantics and expressiveness.
In: Proc. of the 4th Conf. on Uncertainty in Artificial Intelligence
(UAI’88). North-Holland Publishing Co., 1988. p. 69–78.
[71] Das Sarma, A.; ULLMAN, J.; WIDOM, J. Schema design for uncertain
databases. In: Proc. of AMW. 2007.
[72] FAGIN, R.
The decomposition versus synthetic approach to relational
database design. In: Proc. of VLDB. 1977. v. 3, p. 441–6.
[73] JAGADISH, H. V. et al. Making database systems usable. In: SIGMOD.
2007. p. 13–24.
[74] QIAN, L.; LEFEVRE, K.; JAGADISH, H. V. CRIUS: User-friendly database
design. PVLDB, v. 4, n. 2, p. 81–92, 2010.
[75] KOCH, C.; OLTEANU, D. Conditioning probabilistic databases. PVLDB,
v. 1, n. 1, p. 313–25, 2008.
[76] BUGENHAGEN, S. M.; COWLEY, A. W. J.; BEARD, D. A. Identifying
physiological origins of baroreflex dysfunction in salt-sensitive hypertension
in the Dahl SS rat. Physiological Genomics, v. 42, p. 23–41, 2010.
[77] ELTON, C.; NICHOLSON, M. The ten-year cycle in numbers of the lynx in
Canada. Journal of Animal Ecology, v. 11, n. 2, p. 215–44, 1942.
[78] LI, S.; ASSMANN, S. M.; ALBERT, R. Predicting essential components of
signal transduction networks: A dynamic model of guard cell abscisic acid
signaling. PLOS Biology, v. 4, n. 10, p. e312, 2006.
BIBLIOGRAPHY
115
[79] FAN, W.; GEERTS, F.; LI, J.; XIONG, M. Discovering conditional functional
dependencies. IEEE Transactions on Knowledge and Data Engineering, v. 23, n. 5, p. 683–98, 2011.
Appendix A
Detailed Proofs
A.1. Proofs of Hypothesis Encoding
A.1.1 Proof of Theorem 1
“Let S(E, V) be a complete structure. Then the extraction of its causal ordering by
Simon’s COA(S) is intractable (NP-Hard).”
Proof 17 We show that, at each recursive step of COA, to find all non-trivial
minimal subsets (i.e., |E 0 | ≥ 2) translates into an optimization problem associated
with the decision problem BPBP, which we know by Lemma 1 to be NP-Complete.
First, recall (Def. 2) that a structure S(E, V) is complete if |E| = |V|; e.g., for
the structure given in Fig. 3.5 (left), note (Def. 9) the minimal structure S 0 (E 0 , V 0 ),
where E 0 = { f1 , f2 , f3 }. For non-trivial minimal structures, i.e., when |E 0 | = K ≥
2, it is easy to see that its corresponding bipartite graph G = (V10 ∪ V20 , E 0 ), where
V1 7→ E, V2 7→ V and E 7→ S must have number of edges |E 0 | ≥ 2K and, for all
its vertices u ∈ V10 ∪ V20 , u must have deg(u) ≥ 2, i.e., G is a pseudo-biclique in
accordance with Def. 6. That intuition is elaborated as follows.
The point is that, no matter how big is such structure S 0 , its equations f ∈ E 0
are such that |V ars(f )| ≥ 2 (as S 0 is non-trivial) and its variables can be grouped
in local patterns from the sparsest kind to the densest. To construct an instance of
the sparsest case, let S 0 be built by setting a first equation where its entry in the
structure matrix AS has form (1, 1, 0+ ) and then, for the next |E 0 | − 2 equations,
shift such pair of 1’s one position right w.r.t. the previous one. Then complete it
with a last equation whose form is form (1, 0+ , 1). That is, the structure is built
A.1. PROOFS OF HYPOTHESIS ENCODING
117
with unique pairs of 1’s spread all over the structure. Then, deciding whether there
is a minimal structure of size |E 0 | = K corresponds exactly to BPBP. It is a special
case (BBP), when such minimal structure is the densest possible, i.e., when AS
has only 1’s and then its corresponding bipartite graph is a K-balanced biclique
with |E 0 | = K 2 , and deg(u) = K for all vertices u ∈ V10 ∪ V20 . For instance, see the
minimal structure with E 0 = { f4 , f5 } found at the second recursive step of COA
in Fig. 3.2. 2
A.1.2 Proof of Proposition 1
“Let S(E, V) be a structure, and ϕ1 : E → V and ϕ2 : E → V be any two total
causal mappings over S. Then C1+ = C2+ .”
Proof 18 The proof is based on an argument from Nayak [49], which we present
here arguably much clearer. Intuitively, it shows that if ϕ1 and ϕ2 differ on the
variable an equation f is mapped to, then such variables, viz., ϕ1 (f ) and ϕ2 (f ),
must be causally dependent on each other (strongly coupled).
To show C1+ = C2+ reduces to C1+ ⊆ C2+ and C2+ ⊆ C1+ . We show the first
containment, with the second being understood as following by symmetry. Closure
operators are extensive, X ⊆ cl(X), and idempotent, cl(cl(X)) = cl(X). That is, if
we have C1 ⊆ C2+ , then we shall have C1+ ⊆ (C2+ )+ and, by idempotence, C1+ ⊆ C2+ .
Then it suffices to show that C1 ⊆ C2+ , i.e., for any (x0 , x) ∈ C1 , we must show
that (x0 , x) ∈ C2+ as well. Observe by Def. 5 that both ϕ1 and ϕ2 are bijections,
−1
0
then, invertible functions. If ϕ−1
1 (x) = ϕ2 (x), then we have (x , x) ∈ C2 and thus,
trivially, (x0 , x) ∈ C2+ . Else, ϕ1 and ϕ2 disagree in which equations they map onto
x. But we show next, in any case, that we shall have (x0 , x) ∈ C2+ .
Take all equations g ∈ E 0 ⊆ E such that ϕ1 (g) 6= ϕ2 (g), and let n ≤ |E| be
the number of such ‘disagreed’ equations. Now, let f ∈ E 0 be such that its mapped
variable is x = ϕ1 (f ). Construct a sequence of length 2n such that, s0 = ϕ1 (f ) = x
and, for 1 ≤ i ≤ 2n, element si is defined si = ϕ2 (ϕ−1
1 (si−1 )). That is, we are
defining the sequence such that, for each equation g ∈ E 0 , its disagreed mappings
ϕ1 (g) = xa and ϕ2 (g) = xb are such that ϕ1 (g) is immediately followed by ϕ2 (g).
As xa , xb ∈ V ars(g), we have (xa , xb ) ∈ C2 and, symmetrically, (xb , xa ) ∈ C1 .
A.1. PROOFS OF HYPOTHESIS ENCODING
118
The sequence is of form s = hx, xf , . . . , xa , xb , . . . , x2n−1 , x2n i.
| {z }
| {z }
| {z }
f
g
h
Since x must be in the codomain of ϕ2 , we must have a repetition of x at
some point 2 ≤ k ≤ 2n in the sequence index, with sk = x and sk−1 = x00 such that
(x00 , x) ∈ C2 . If x00 = x0 , then (x0 , x) ∈ C2 and obviously (x0 , x) ∈ C2+ . Else, note
that xf must also be in the codomain of ϕ1 , while x00 in the codomain of ϕ2 . Let `
be the point in the sequence, 3 ≤ ` ≤ 2n−1, at which s` = xf = xa and s`+1 = xb
for some xb such that (xf , xb ) ∈ C2 . It is easy to see that, either we have xb = x00
or xb 6= x00 but (xb , x00 ) ∈ C2+ . Thus, by transitivity on such a causal chain, we
must have (xf , x00 ) ∈ C2+ and eventually (xf , x) ∈ C2+ . Finally, since x0 ∈ V ars(f )
and ϕ2 (f ) = xf , we have (x0 , xf ) ∈ C2 and, by transitivity, (x0 , x) ∈ C2+ . 2
A.1.3 Proof of Theorem 2.
“Let Σ be an fd set defined Σ , h-encode(S) for some complete structure S. Then
Σ is non-redundant and singleton-rhs but may not be left-reduced (then may not be
canonical).”
Proof 19 We will show that properties (a-b) of Def. 9 hold for Σ produced by
(Alg. 3) h-encode, but property (c) may not hold.
At initialization, the algorithm sets Σ = ∅ and then inserts an fd hX, Ai ∈ Σ
for each hf, xi ∈ ϕt scanned, where x 7→ A and X ∩ A = ∅. At termination, for
all fd’s in Σ we obviously have |A| = 1 then property (a) holds. Also, note that
ϕ : S → V ars(S) is, by Def. 5, a bijection.
Now, for property (b) not to hold there must be some fd hX, Ai ∈ Σ that
is redundant and then can be found in the closure of Γ = Σ \ hX, Ai. By Lemma
4 (below), that can be the case only if A ⊆ X or there is hY, Ai ∈ Γ for some Y .
But from X ∩ A = ∅, we have A * X; and from ϕ being a bijection it follows that
there can be no such fd in Γ. Thus it must be the case that Σ is non-redundant,
i.e., property (b) holds.
Finally, property (c) does not hold if there can be some fd hX, Ai ∈ Σ with
Y ⊂ X such that Γ = Σ \ hX, Ai ∪ hY, Ai has the same closure as Σ. That is, if
we may find hY, Ai ∈ Σ+ . Now, pick structure S whose (3 × 3) matrix As has rows
(1, 0, 0), (1, 1, 0), (1, 1, 1) as an instance. Alg. 3 encodes it into Σ = {φ → x1 , x1 υ →
A.2. PROOFS OF CAUSAL REASONING
119
x2 , x1 x2 υ → x3 }. Let Y = {x1 , υ}, and B = {x2 }. Note that x1 υ → x2 ∈ Σ can be
written as hY, Bi ∈ Σ, and x1 x2 υ → x3 ∈ Σ as hY B, Ai ∈ Σ. Now observe that
hY, Ai ∈ Σ+ can be derived by R5 over hY, Bi, hY B, Ai ∈ Σ, which is sufficient to
show that property (c) may not hold. That is, B is “extraneous” in hY B, Ai ∈ Σ
and can be removed from its lhs without loss of information to Σ. 2
Lemma 4 Let Σ be a (Def. 9-a) singleton-rhs fd set on attributes U . Then X → A
can only be in Σ+ , where XA ⊆ U , if A ⊆ X or there is non-trivial hY, Ai ∈ Σ for
some Y ⊂ U .
Proof 20 By Lemma 5 (below), we know that X → A ∈ Σ+ iff A ⊆ X + . We need
to prove that if A * X and there is no Y → A in singleton-rhs Σ, then A * X + .
But this is equivalent to show that (Alg. 4) XClosure gives only correct answers for
X + w.r.t. Σ, which is known (cf. theorem from Ullman [20, p. 389]). Note that
XClosure(Σ, X) inserts A in X + only if A ⊆ X or there is some fd hY, Ai ∈ Σ. 2
Lemma 5 Let Σ be an fd set. An fd X → Y is in Σ+ iff Y ⊆ X + , where X + is
the attribute closure of X w.r.t. Σ.
Proof 21 This is from Ullman [20, p. 386]. Let Y = A1 ... An and suppose Y ⊆
X + . Then for each Ai , we have Ai ∈ X + and, by the definition of X + , we must
have hX, Ai i ∈ Σ+ . Then it follows by (R4) union that X → Y is in Σ+ as
well. Conversely, suppose hX, Y i ∈ Σ+ . Then, by (R3) decomposition we have
hX, Ai i ∈ Σ+ for each Ai ∈ Y . 2
A.2. Proofs of Causal Reasoning
A.2.1 Proof of Lemma 2
“Let S(E, V) be a complete structure, ϕ a total causal mapping over S and Σ an
fd set encoded through ϕ given S. If hX, Ai ∈ Σ, then A#, the attribute folding of
A (w.r.t. Σ) exists and is unique.”
Proof 22 The existance of A# is ensured by the degenerate case where X = A#,
as X → A is itself in ΣB by an empty application of R5. If X → A is in fact folded
w.r.t. Σ, then the folding of A exists. Else, it is not folded yet X → A is non-trivial
A.2. PROOFS OF CAUSAL REASONING
120
because by Theorem 2 Σ is non-redundant. Then, by Def. 11 there must be some
Y ⊆ U with Y + X such that Y → X is in Σ+ and X 6→ Y . By Def. 10, there
.
is a finite application of R5 over fd’s in Σ to derive Y →
− X. Then by R2 ∼ R5
over X → A, we have Y → A. Although there may be many such (intermediate)
attribute sets Y ⊂ U along the transitive chaining satisfying the conditions above,
we claim there is at least one that is a folding of A. Suppose not. Then, for all
such Y ⊂ U , there is some Y 0 ⊂ U with Y 0 + Y such that Y 0 → Y and Y 6→ Y 0 ,
leading to an infinite regress. Nonetheless, in so far as cycles are ruled out by force
of Def. 11, then Σ+ must have an infinite number of fd’s. But Σ+ is finite, viz.,
bounded by 22|U | (cf. [21, p. 165]).
. Therefore the folding of A must exist.
Moreover, observe that Σ is encoded through ϕ, which is by Def. 5 a bijection.
Then we have hX, Ai ∈ Σ for exactly one attribute set X. Then, as a straightfoward
follow-up of the rationale that led us to infer the folding existance, note that there
.
.
.
.
.
must be a single chaining Y n →
− ... →
− Y1→
− Y0→
− X→
− A. Again, as cycles are
ruled out by force of Def. 11 and Σ+ is finite, then the folding of A is unique. 2
A.2.2 Proof of Theorem 3
“Let S(E, V) be a complete structure, and Σ an fd set encoded given S. Now, let
hX, Ai ∈ Σ. Then AFolding(Σ, A) correctly computes A#, the attribute folding of
A (w.r.t. Σ) in time O(|S|2 ).”
Proof 23 For the proof roadmap, note that AFolding is monotone (size of A? can
only increase) and terminates precisely when A(i+1) = A(i) , where A(i) denotes the
attributes in A? at step i of the outer loop. The folding A# of A at step i is
A(i) \ Λ(i) . We shall prove by induction, given attribute A from fd X → A in Σ,
that A? \ Λ returned by AFolding(Σ, A) is the unique attribute folding A# of A.
(Base case). By Theorem 2, Σ is non-redundant with (then) non-trivial
hX, Ai ∈ Σ for exactly one attribute set X, the algorithm always reaches step
i = 1, which is our base case. Then X is placed in A(1) and A in Λ(1) , and we
have A(1) = XA and Λ(1) = A. Therefore, A(1) \ Λ(1) = X, and in fact we have
hX, Ai ∈ ΣB by an empty application of R5. For it to be specifically in Σ#⊂ ΣB,
it must be folded w.r.t. set ∆ of consumed fd’s at this step, viz., ∆(1) = {X → A}.
In fact, as the only fd in ∆(1) , by Def. 11 it must be folded w.r.t. ∆(1) , and we
A.2. PROOFS OF CAUSAL REASONING
121
have A# = X at step i = 1.
(Induction). Now, let i = k, for k > 1, and assume that hA(k) \ Λ(k) , Ai ∈
Σ# ⊂ ΣB with A(k) 6= Λ(k) . By Lemma 2 we know that A# = A(k) \ Λ(k) is the
unique folding of A at step i = k. For the inductive step, suppose Y is placed in
A(k+1) and B in Λ(k+1) because hY, Bi ∈ Σ \ ∆(k) and B ∈ A(k) .
Since B ∈ A(k) and B ∈
/ Λ(k) (it is yet just be consumed into Λ(k+1) ), we can
write (A(k) \Λ(k) ) = ZB for some Z 6= B, where (A(k) \ Λ(k) ) → A is assumed in Σ#.
|
{z
}
ZB
Now, with the application of R5 consuming Y → B we have (A(k) Y B \ Λ(k) B) → A,
|
{z
}
ZS
where S = Y \ Λ(k) . We claim that ZS → A is folded w.r.t. ∆(k+1) .
Suppose not. Then by Def. 11 there must be some W + ZS such that
W → ZS is in (∆(k+1) )+ but ZS 6→ W . Since ZS 6= ∅, there must be some
C ∈ ZS, i.e., C ∈
/ Λ(k+1) . Note that, as W → ZS is in (∆(k+1) )+ , then by (R3)
decomposition we have W → C in (∆(k+1) )+ as well. But by Lemma 4 that can
only be the case if there is some hT, Ci ∈ ∆(k+1) , which means C has been already
consumed into Λ(k+1) , though C ∈
/ Λ(k+1) .
.
Finally, as for the time bound, note that in worst case, exactly one fd Y → B
is consumed from Σ into ∆ for each step of the outer loop, where |Σ| = |E|. That
is, let n = |E|, then n is decreased stepwise in arithmetic progression such that
n + (n−1) + . . . + 1 = n (n−1)/2 scans are required overall, i.e., O(n2 ). Note also,
however, that B may be the only symbol read at each such fd scan but in worst
case at most |U | = |V| symbols are read. Thus our measure n should be actually
overestimated n = |E| |V| = |S|. Therefore Alg. 6 is bounded by O(|S|2 ). 2
A.2.3 Proof of Corollary 2
“Let S(E, V) be a complete structure, and Σ an fd set encoded given S. Then algorithm folding(Σ) correctly computes Σ#, the folding of Σ in time that is f (|S|) Θ(|E|),
where f (|S|) is the time complexity of (Alg. 6) AFolding.”
Proof 24 By Theorem 3, we know that sub-procedure (Alg. 6) AFolding is correct
and terminates. Then (Alg. 5) folding necessarily inserts in Σ# (initialized empty)
#
exactly one fd Z −
→ A for each fd X → A in Σ scanned. Thus, at termination we
have |Σ#| = |Σ|. Again, as AFolding is correct, we know Z is the unique folding of
A.2. PROOFS OF CAUSAL REASONING
122
A. Therefore it must be the case that Alg. 5 is correct. Finally, for the time bound,
the algorithm iterates over each fd in Σ without having to read its symbols, and at
each such step AFolding takes time that is f (n). Thus folding takes f (|S|) Θ(|E|).
But we know from Theorem 3 and Remark 4 that f (|S|) ∈ O(|S|), then it takes
O(|S| |E|). 2
A.2.4 Proof of Proposition 3
“Let S(E, V) be a complete structure, ϕ a total causal mapping over S and Σ
an fd set encoded through ϕ given S. Let Σ# be the folding of Σ, then Σ# is
parsimonious.”
Proof 25 By Lemma 2 we know that, for each fd hX, Ai ∈ Σ, the attribute folding
#
→ A exists and is unique. That is, for no Y 6= Z we have
Z of A such that Z −
#
Y −
→ A. Thus Σ# , folding(Σ) automatically satisfies Def. 13, as long as we show
it is canonical (cf. Def. 9).
Moreover, by Theorem 2 we know that Σ is both non-redundant and singletonrhs. Now, consider by Lemma 2 that AFolding builds a bijection mapping each
#
hX, Ai ∈ Σ to exactly one hZ, Ai ∈ Σ# such that Z −
→ A. Since Σ is singleton-rhs,
it is obvious that Σ# is as well and covers all attributes in the rhs of fd’s in Σ.
Also, the bijection implies |Σ#| = |Σ|. Since Σ is non-redundant and has exactly
one fd with each attribute A in its rhs, then by Lemma 4 so is Σ#.
Finally we will show that unlike Σ, its folding Σ# must be left-reduced.
Suppose not. Then for some fd Z → A in Σ# there is S ⊂ Z such that non-trivial
S → A is in (Σ#)+ . Since Z → A is the only fd in Σ# with A in its rhs and S → A
M
M
is non-trivial, we must have S −
→Z−
→ A.
Now, suppose S → A is not folded. Then there is W + S such that W → S
is in (Σ#)+ but S 6→ W . Note that W 6= Z, as W + S. Also, W → S and
S → Zimplies W → Z by (R5) transitivity. Note also that S 6→ W and S → Z
implies Z 6→ W . But Z → A is assumed folded.
. That is, S → A must be
folded. Then we have both S → A and Z → A folded, though S 6= Z. That is, the
attribute folding of A is not unique, even though we know by Lemma 2 that it must
be unique.
2
. Thus Σ# must be left-reduced, altogether, therefore, parsimonious.
A.2. PROOFS OF CAUSAL REASONING
123
A.2.5 Proof of Theorem 4
“Let S(E, V) be a complete structure, ϕ a total causal mapping over S and Σ an
fd set encoded through ϕ given S. Then xa , xb ∈ V are such that xb is causally
dependent on xa , i.e., (xa , xb ) ∈ Cϕ+ iff there is some non-trivial fd hX, Bi ∈ ΣB
with A ∈ X, where B 7→ xb and A 7→ xa .”
Proof 26 We prove the statement by induction. We consider first the ‘if’ direction, and then its ‘only if’ converse.
(Base case). Let hX, Bi ∈ Σ be some fd with A ∈ X, where B 7→ xb
and A 7→ xa . By Theorem 2, it is non-trivial and then by default (i.e., an empty
application of R5) it is in ΣB . But as X → B is in Σ, (Alg. 3) h-encode ensures
that there is exactly one equation f ∈ E such that ϕ(f ) = xb and xa ∈ V ars(f )
where B 7→ xb and A 7→ xa . Then by force of Eq. 3.1 we must have (xa , xb ) ∈ Cϕ .
Thus, we obviously have (xa , xb ) ∈ Cϕ+ as well.
(Induction). Now, recall Armstrong’s (R5) pseudo-transitivity rule adapted
here for the particular case of singleton-rhs fd sets, viz., if Y → C and CZ →
B, then Y Z → B. By the inductive hypothesis, take any two non-trivial fd’s
hY, Ci, hCZ, Bi ∈ ΣB with B ∈
/ Y and assume that the causal dependency property
holds for their attributes that encode variables. That is, let D ∈ Y and E ∈ Z,
where D 7→ xd and E 7→ xe for xd , xe ∈ V such that (xd , xc ), (xe , xb ), (xc , xb ) ∈ Cϕ+ .
Note that both Y → C and CZ → B are non-trivial, then C ∈
/ Y, B ∈
/ Z and
B 6= C. Moreover, B ∈
/ Y has been assumed such that the fd hY Z, Bi ∈ ΣB to be
derived by R5 over Y → C and CZ → B is also non-trivial to satisfy the condition
of the theorem. Now, it is easy to see that the property holds likewise for nontrivial fd hY Z, Bi ∈ ΣB . In fact, (xd , xc ), (xc , xb ) ∈ Cϕ+ implies (xd , xb ) ∈ Cϕ+ and
also by the inductive hypothesis we have (xe , xb ) ∈ Cϕ+ . That is, for either some
D ∈ Y or some E ∈ Z, we must have (xd , xb ), (xe , xb ) ∈ Cϕ+ .
The converse ‘only if’ direction can be shown by a symmetrical inductive
argument. That is, for the base case suppose (xa , xb ) ∈ Cϕ . Then, by Eq. 3.1 we
know there is some f ∈ E such that ϕ(f ) = xb and xa ∈ V ars(f ). Moreover, in
that case (Alg. 3) h-encode ensures there must be some non-trivial fd hX, Bi ∈ Σ
with A ∈ X where B 7→ xb and A 7→ xa . Thus by an empty application of R5 we
A.2. PROOFS OF CAUSAL REASONING
124
have hX, Bi ∈ ΣB . The inductive step shows the property still holds for arbitrary
causal dependencies in Cϕ+ . 2
A.2.6 Proof of Proposition 4
“Let S(E, V) be a structure with variable x ∈ V. Then x can only be a first cause
of some y ∈ V if x is exogenous. Accordingly, any variable y ∈ V can only have
some first cause x ∈ V if it is endogenous.”
Proof 27 The proof is straightforward from definitons. For the first statement,
suppose by contradiction that x ∈ V is not exogenous but is a first cause of some
y ∈ V. By Def. 5, ϕ is bijective then there is some f ∈ E such that ϕ(f ) = x.
Moreover, as x is not exogenous then by Def. 8 it must be endogenous. In other
words, there must be some xa ∈ V such that xa 6= x ∈ V ars(f ) and then by Eq.
3.1 we have (xa , x) ∈ Cϕ hence (xa , x) ∈ Cϕ+ . However, as x is a first cause, by
Def. 14 there can be no y ∈ V such that (y, x) ∈ Cϕ+ .
.
Now, a symmetrical argument proves the second statement. Also by contradiction take a variable y ∈ V that is not endogenous and suppose it has some first
cause x ∈ V. As variable y is not endogenous then by Def. 8 it must be exogenous.
In other words, there must be f ∈ E such that V ars(f ) = {y}. Thus for any total
causal mapping ϕ over S, we must have ϕ(f ) = y and, for no x ∈ V, we have
(x, y) ∈ Cϕ . Therefore it is not possible to derive (x, y) ∈ Cϕ+ for some x ∈ V. But
as y has some first cause x ∈ V by assumption, we must have (x, y) ∈ Cϕ+ .
. 2
A.2.7 Proof of Lemma 3
“Let S(E, V) be a complete structure, ϕ a total causal mapping over S and Σ an
fd set encoded through ϕ given S. Then a variable xa ∈ V can only be a first cause
of some variable xb ∈ V, where hX, Bi ∈ Σ, and B 7→ xb , A 7→ xa , if either (i)
A ∈ X or (ii) A ∈
/ X but there is hZ, Ci ∈ ΣB with A ∈ Z and C ∈ X.”
Proof 28 We prove the statement by construction out of Theorem 4.
By Def. 14, one of the conditions for xa to be a first cause of xb is that
(xa , xb ) ∈ Cϕ+ . Moreover, by Theorem 4 we know that (xa , xb ) ∈ Cϕ+ can only
hold if there is some non-trivial fd hZ, Bi ∈ ΣB with A ∈ Z, where B 7→ xb and
A 7→ xa . Now, by Def. 5 ϕ is bijective then there is hX, Bi ∈ Σ. Moreover, since
A.2. PROOFS OF CAUSAL REASONING
125
Σ is parsimonious, X → B is the only fd in Σ with B in its rhs. So let A ∈
/ X.
Then we know X 6= Z hence X → B cannot be the fd required by Theorem 4.
.
Then such fd Z →
− B with A ∈ Z can only exist if derived by some finite
application of R5. That is, there must be some hZ, Ci ∈ Σ with A ∈ Z such that
X = CW for some W and then R5 can be applied over hZ, Ci, hCW, Bi ∈ Σ to
get non-trivial hZW, Bi ∈ ΣB where A ∈ Z.
Now, it is easy to see that when such fd Z → C with A ∈ Z does not exists
in ΣB (the second condition of the lemma), then obviously Z → C cannot exist in
Σ to satisfy the requirement imposed by Theorem 4. That is, (xa , xb ) ∈
/ Cϕ+ . 2
A.2.8 Proof of Theorem 5
“Let S(E, V) be a complete structure, ϕ a total causal mapping over S and Σ an
fd set encoded through ϕ given S. Now, let B be an attribute that encodes some
variable xb ∈ V. If hX, Bi ∈ Υ(Σ)#,1 then every first cause xa of xb (if any) is
encoded by some attribute A ∈ X.”
Proof 29 We show that the existance of a missing first cause xc of xb for folded
#
X−
→ B, where B 7→ xb and C 7→ xc but C ∈
/ X leads to a contradiction.
Suppose, by contradiction, that there is some missing first cause xc ∈ V of
xb , where C 7→ xc and C ∈
/ X. Then, by Lemma 3, since variable xc is a first cause
of variable xb , it must be exogenous and, for hY, Bi ∈ Υ(Σ) either (i) C ∈ Y or
(ii) C ∈
/ Y but there is hZ, Di ∈ Υ(Σ)B with C ∈ Z and some D ∈ Y .
In the first case (i), since xc is exogenous and Σ is parsimonious, we have
hφ, Ci ∈ Σ but by Def. 15 there can be no W → C in the υ-projection Υ(Σ) of Σ.
That is, C cannot be ‘consumed’ by R5 and then hY, Bi ∈ Υ(Σ) with C ∈ Y implies
that, for any hW, Bi ∈ Υ(Σ)B , we must have C ∈ W . However, by assumption we
have hX, Bi ∈ Υ(Σ)# then, by Def. 12, hX, Bi ∈ Υ(Σ)B , yet C ∈
/ X.
.
In the second case (ii), observe that C ∈ Z and D ∈ Y , and let Y = DS.
Then by R5 over Z → D and DS → B we get hZS, Bi ∈ Υ(Σ)B , where C ∈ Z.
Well, either ZS → B is folded or it is not, rendering two cases for analysis. If
ZS → B is folded, then both ZS → B and X → B are folded. But as Υ(Σ) is
1
Note that the folding is taken w.r.t. the υ-projection of Σ, then xb where B 7→ xb is an
endogenous variable.
A.3. PROOFS OF PROBABILISTIC DB SYNTHESIS
126
parsimonious, then by Lemma 2 the folding of B must be unique. Therefore we
must have ZS = X, with C ∈ Z but C ∈
/ X.
.
Else, assume ZS → B is not folded. Then by Def. 11 there is some W with
W + ZS such that non-trivial W → ZS is in Υ(Σ)+ and ZS 6→ W .
However, as C ∈ Z and W → ZS, by (R3) decomposition we must have
W → C in Υ(Σ)+ , either with C ∈ W or with C ∈
/ W and then W → C is nontrivial. But we know the latter cannot be the case by the same argument used in
the first case (i), viz., xc is exogenous with C 7→ xc and Σ is parsimonious. That
is, we must have C ∈ W . Furthermore, as W → ZS in Υ(Σ)+ , then by R5 over
ZS → B we get hW, Bi ∈ Υ(Σ)B with C ∈ W . Now it is easy to see that the same
situation recurs to W → B. If it is not folded, eventually for some T we will have
.
.
− W →
− B with C ∈ T , where T → B will be folded just like X → B. That is,
T →
by (Lemma 2) the uniqueness of the folding of B, we will have T = X with C ∈ T
and C ∈
/ X.
. 2
A.3. Proofs of Probabilistic DB Synthesis
A.3.1 Proof of Theorem 6
“Let Sk and Hk be (resp.) the complete structure and ‘big’ fact table of hypothesis
k, and let Γ0k be the repaired factorization of Sk over Hk , and Y0 the ‘explanation’
table where hypothesis k is recorded. Now, let Y k be a U-relational schema defined
Y k , synthesize4u(Sk , Hk , Y0 ). Then Y k is in BCNF w.r.t. Γ0k and is minimalcardinality.”
Proof 30 Let Yki [ Vi Di | φ Ai Gi ] and Ykj [ Vj Dj | S T ] be (resp.) any u-factor
projection and predictive projection of Hk . Note that all fd’s in Γ0k are either in
Φ(Γ0k ) of form φ Ai → B or φ → B, or in Υ(Γ0k ) of form A1 A2 ... A` S → T with
υ ∈ S. We must show that no fd in (Γ0k )+ can violate Yki or Ykj . It is easy to see
that the projection (cf. Def. 20) of non-trivial fd’s in Φ(Γ0k )+ onto Ykj is empty,
just like the projection of non-trivial fd’s in Υ(Γ0k )+ onto Yki .
For the u-factor projections, note by Def. 21 that for any fd X → C in
(Γ0k )+ to violate BCNF in Yki [ Vi Di | φ Ai Gi ], it must be non-trivial (C 6∈ X) with
XC ⊆ φ Ai Gi but X 6→ φ Ai Gi (that is, X is not a superkey for Yki ). Note that we
A.3. PROOFS OF PROBABILISTIC DB SYNTHESIS
127
have both φ Ai → B and φ → Ai B in (Γ0k )+ for any B ∈ Gi , but both φ Ai and φ
are superkeys for Yki . Also, that there can be no non-trivial fd’s hX, Ci ∈ Φ(Γ0k )+
with φ 6∈ X, and by the definition of Problem 1 we know that Ai Gi is a maximal
group. So, for any non-trivial hX, Ci ∈ Φ(Γ0k )+ , X must be a superkey for Yki .
Thus no u-factor projection can be subject of BCNF violation w.r.t. Γ0k .
Now, for predictive projection Ykj [ Vj Dj | S T ] let us reconstruct the process
towards deriving hS, T i ∈ Υ(Γ0k )B . Note that it is derived by synthesize4u simulating ` applications of R5 over hφ, Ai i, hA1 A2 ...A` S, T i ∈ Γ0k for 0 ≤ i ≤ `. Note
also that (i) no cyclic fd’s can be involved in such R5 applications, as they must always be over an fd in Φ(Γ0k ); and (ii) as a result of (Alg. 7) merge, A1 A2 ...A` S → T
was the only non-trivial fd in the projection πA1 A2 ...A` ST (Γ0k ). Thus, the only nontrivial fd’s in the projection πST ( (Γ0k ) )+ in addition to S → T itself must be of
form S → C rendered out of (R3) decomposition from it for all C ∈ T . In any
such fd’s, we have S as a superkey for Ykj . Therefore no predictive projection can
be subject of BCNF violation w.r.t. Γ0k .
For the minimality note, as a consequence of (Alg.
7) merge, any two
schemes Ykp [XZ], Ykq [V W ] are rendered by synthesize4u into Y k iff we have fd’s
hX, Zi, hV, W i ∈ (Γ0k )+ and X 6↔ V , i.e., it is not the case that both X → V and
V → X hold in (Γ0k )+ . Now, to prove that Y k is minimal-cardinality, we have to
find that merging any such pair of arbitrary schemes shall hinder BCNF in Y k .
In fact, take Y 0k := Y k \ (Ykp [XZ] ∪ Ykq [V W ]) ∪ Yk` [XZV W ]. As X 6↔ V , then
neither X nor V can be a superkey for Yk` , which therefore cannot be in BCNF. 2
A.3.2 Proof of Theorem 7
“Let Sk be the complete structure of hypothesis k, and Hk [U ] its ‘big’ fact table
such that Γ0k is the repaired factorization of Sk over Hk and Y0 is the ‘explanation’
table where hypothesis k is recorded. Now, let Y k be a U-relational schema defined
Y k , synthesize4u(Sk , Hk , Y0 ). Then,
i
(a) the join ./ m
i=1 Yk [ Vi Di | φ Ai Gi ] of any subset of the u-factor projections
of Hk is lossless w.r.t. Γ0k .
(b) any predictive projection Ykj [ Vj Dj | S T ], result of a join of the theoretical
u-factor Y0 [ V0 D0 | φ υ ] with the ‘big’ fact table Hk [U ] and in turn with
A.3. PROOFS OF PROBABILISTIC DB SYNTHESIS
128
u-factor projections Yki [ Vi Di | φ Ai Gi ], is lossless w.r.t. Γ0k .”
Proof 31 For item (a), by Lemma 6, we know that any pair Yki [ Vi Di | φ Ai Gi ],
Ykj [ Vj Dj | φ Aj Gj ] of u-factor projections of Hk will have a lossless join w.r.t. Γ0k
iff (φ Ai Gi ∩ φ Aj Gj ) → (φ Ai Gi \ φ Aj Gj ) or (φ Ai Gi ∩ φ Aj Gj ) → (φ Aj Gj \
φ Ai Gi ) hold in (Γ0k )+ . By Def. 17, we know that (φ Ai Gi ∩ φ Aj Gj ) = {φ}, and
φ Ai Gi \ φ Aj Gj = Ai Gi . In fact φ → Ai Gi is a repaired fd in Γ0k , therefore Yki
and Ykj have a lossless join. Now, since the join is an associative operation [20, p.
62], and as we have chosen Yki and Ykj arbitrarily, then clearly any subset of the
u-factor projections must have a lossless join.
For item (b), for any predictive projection Ykj [ Vj Dj | S T ] take the join ./
Yki [ Vi Di | φ Ai Gi ] of u-factor projections such that, for all Ai ∈ Ai Gi , we have
Ai ∈ W ⊂ Z where S = Z \ W and hZ, T i ∈ Γ0k . That is, Ai is a pivot attribute
representing a first cause of some C ∈ T . By item (a), we know that such join is
lossless.
We must show that the join ./ Yki with ‘big’ fact table Hk [U ] is also lossless.
By Lemma 6, that is the case iff (φ Ai Gi ∩ U ) → (φ Ai Gi \ U ) or (φ Ai Gi ∩
U ) → (U \ φ Aj Gj ) hold in (Γ0k )+ . In fact, we have (φ Ai Gi ∩ U ) = φ Ai Gi and
(φ Ai Gi \ U ) = ∅ such that φ Ai Gi → ∅ is trivially in (Γ0k )+ .
Finally, the join of theoretical u-factor Y0 [ V0 D0 | φ υ ] with big fact table
Hk [U ] must be lossless likewise. In fact, note that ( φ υ ∩ U ) = φ υ, and (φ υ \U ) =
∅. Then also trivially we have φ υ → ∅, which is in (Γ0k )+ as well. Since the join
is commutative [20, p. 62], the order of application is irrelevant therefore the join
of all joins examined above taken together must be lossless. 2
Lemma 6 Let Σ be a set of fd’s on attributes U , and Ri [S], Rj [T ] ∈ R[U ] be
relation schemes with ST ⊆ U ; and let πST (Σ) be the projection of Σ onto ST .
Then Ri [S] and Rj [T ] have a lossless join w.r.t. πST (Σ) iff (S ∩ T ) → (S \ T ) or
(S ∩ T ) → (T \ S) hold in πST (Σ)+ .
Proof 32 See Ullman [20, p. 397]. 2
| 5cs.CE
|
arXiv:1704.07798v3 [quant-ph] 13 Aug 2017
Limitations on Transversal Computation through Quantum
Homomorphic Encryption
Michael Newman1 and Yaoyun Shi2
1
2
Department of Mathematics
Department of Electrical Engineering and Computer Science
University of Michigan, Ann Arbor, MI 48109, USA
[email protected], [email protected]
Abstract
Transversality is a simple and effective method for implementing quantum computation faulttolerantly. However, no quantum error-correcting code (QECC) can transversally implement a
quantum universal gate set (Eastin and Knill, Phys. Rev. Lett., 102, 110502). Since reversible
classical computation is often a dominating part of useful quantum computation, whether or not
it can be implemented transversally is an important open problem. We show that, other than
a small set of non-additive codes that we cannot rule out, no binary QECC can transversally
implement a classical reversible universal gate set. In particular, no such QECC can implement
the Toffoli gate transversally.
We prove our result by constructing an information theoretically secure (but inefficient) quantum homomorphic encryption (ITS-QHE) scheme inspired by Ouyang et al. (arXiv:1508.00938).
Homomorphic encryption allows the implementation of certain functions directly on encrypted
data, i.e. homomorphically. Our scheme builds on almost any QECC, and implements that
code’s transversal gate set homomorphically. We observe a restriction imposed by Nayak’s
bound (FOCS 1999) on ITS-QHE, implying that any ITS quantum fully homomorphic scheme
(ITS-QFHE) implementing the full set of classical reversible functions must be highly inefficient. While our scheme incurs exponential overhead, any such QECC implementing Toffoli
transversally would still violate this lower bound through our scheme.
1
1.1
Introduction
Restrictions on transversal gates
Transversal gates are surprisingly ubiquitous objects, finding applications in quantum cryptography
[29], [25], quantum complexity theory [11], and of course quantum fault-tolerance. The instability
of quantum information is well-documented, and quantum error-correcting codes [24] allow the
encoding of single qubits into multiple qubit systems so that errors on small subsets of physical
qubits can be corrected [21]. Performing computations on these codes carries the risk of propagating
errors between different subsystems, unless the code can implement the computation in a way that
preserves the subsystem structure. Informally, these types of logical operators that decompose as
a product across the subsystems are called transversal, and the oft-cited Eastin-Knill theorem [15],
[39] limits the ability of quantum codes to prevent this error propagation.
Theorem 1 (Eastin-Knill). No quantum error-correcting code can implement a quantum universal
transversal gate set.
These transversal gate sets are valuable as most models of fault-tolerant quantum computation implement associated transversal gate sets fault-tolerantly “for free”. Incurring comparatively
significant overhead, often in the form of magic state distillation [18], [22], gauge fixing [5], [26],
or more recently deconstructions of non-transversal gates into fault-tolerant pieces [37], one can
fault-tolerantly implement some remaining gate set making the computation space universal. Improving the efficiency of this overhead and designing new fault tolerant architectures to supplement
transversal gates is central to quantum fault tolerance.
Implementing fault-tolerant classical reversible computation efficiently would be extremely desirable as many quantum algorithms are primarily classical subroutines with a relatively small
number of quantum gates, and there have been several proposals for doing so [31], [13], [23].
For example, factoring a cryptographically large RSA key using Shor’s algorithm requires around
3 × 1011 Toffoli gates to perform modular exponentiation alone, and is the dominating portion of
the circuit [19]. As Toffoli is universal for classical reversible computation, one might ask if there
are any quantum error-correcting codes that can naturally implement Toffoli, and thus classical
computations, transversally? We give restrictions on the ability of QECCs to do this.
Theorem 2 (Informal). Almost no quantum error-correcting code can implement a classical universal transversal gate set. In particular, almost no quantum error-correcting code can implement
the Toffoli gate transversally.
The only exceptions to our theorem are non-additive distance d codes that decompose as d-fold
product states in their logical computational basis, where each “subcode” itself fails to be erasurecorrecting. Essentially, one can think of these as maximally redundant quantum codes: they are the
concatenation of a repetition code with some distance 1 inner code, similar to Shor’s stabilizer code
written as a 3-fold product of GHZ states. We do not expect that any such code can implement
Toffoli transversally, but it remains a case our proof technique cannot rule out. In particular, our
proof does apply to all binary additive codes. The result is perhaps slightly surprising since there
exist QECCs (e.g. triorthogonal codes) that can implement the CCZ gate transversally [31], and in
fact transversal Toffoli gates can map between different quantum Reed-Solomon codes by increasing
the degree of the underlying polynomial [13].
1
1.2
Quantum homomorphic encryption
The main ingredient in our proof is an information-theoretically secure homomorphic encryption
scheme. Generally, homomorphic encryption [20] is a means of delegating computation on sensitive
data securely. It allows for the encryption of data in such a way that another party can perform
meaningful computation on the ciphertext without decoding, preserving the security of the underlying plaintext. A scheme is termed fully homomorphic encryption (FHE) if it can implement a
universal class of functions in some computation space.
Recently, extensions of homomorphic encryption to the quantum setting have been considered.
Instead of encrypting classical data and implementing addition and multiplication gates homomorphically, quantum homomorphic encryption aims to encrypt quantum data and implement unitary
gates homomorphically. Progress was made in [10], and recently [14] extended this work to a leveled
scheme that could homomorphically implement all polynomial-sized quantum circuits.
The aforementioned schemes are only computationally secure, since they use classical FHE as
a subroutine. This is no great indictment: FHE is built on the difficulty of certain hard lattice
problems that are leading candidates for quantum-secure encryption [12], [33]. However, quantum
information often promises information-theoretic security (ITS) guarantees that are impossible
classically. Intermediate advances have also been made in this more restrictive setting. One such
scheme allows for the implementation of a large class of unitaries homomorphically, but with less
stringent ITS guarantees [36].
More recently, [29] proposed a compact ITS-QHE scheme in which the size of the encoding scales
polynomially with size of the input for the limited Clifford circuit class. This scheme achieves
the strongest notion of imperfect ITS, with the probability of distinguishing between any two
ciphertexts exponentially suppressed in the size of the encoding.
The scheme in [29] is based on a “noisy” quantum encoding of the data. They take an encoding
circuit for a particular quantum code and replace the ancilla bits of the encoding with uniformly
random noise. Their encryption is then choosing a random embedding of this code into yet more
uniformly random noise. This scheme links ITS-QHE to transversal gates: the transversal gates
for their code are exactly those gates that can be implemented homomorphically.
1.3
Limitations on ITS-QHE
There are fundamental limitations on what ITS homomorphic encryption can do. It is known
that for a purely classical scheme, efficient ITS-FHE is impossible, violating lower bounds in the
setting of single server private information retrieval [16]. It was further shown that in the best case
scenario, when the mutual information between the plaintext and the ciphertext is precisely zero,
efficient quantum FHE is impossible [38].
This no-go result actually applies to the more restrictive setting of classical data being encrypted
into quantum data, while allowing only classical reversible functions to be evaluated homomorphically. Both [38] and [29] ask whether relaxing to imperfect ITS-security might allow for efficient
ITS-QFHE. Unfortunately, this is not the case.
Proposition 3. [Informal] Efficient ITS-QFHE is impossible.
Concurrent to this work, this proposition was observed in [25]. We provide a precise statement
and proof of this restriction in Appendix A. This result can be seen by combining the proof technique
in [16] with similar single server private information retrieval bounds in the quantum setting [4].
In essence, the inefficiency of ITS-QFHE follows from viewing ITS-QHE (on classical data using a
quantum encoding) as a certain quantum random access encoding (QRAC) (see [1]) of the function
2
class we wish to implement homomorphically. Well-known bounds on QRACs [27] place lower
bounds on the encoding size of such a scheme, precluding efficiency. Using a variant of the codebased ITS-QHE scheme proposed in [29], we can then argue that (almost) any QECC implementing
the Toffoli gate transversally would yield a scheme violating this lower bound.
It is worth noting that similar tasks such as blind quantum computation [9] and computing on
encrypted data [17], [8] allow ITS solutions, but they do so at the cost of interactivity between the
client and server. We do not allow this interactivity in our definition of homomorphic encryption.
1.4
Comparison to related works
The five works the most closely resemble our results are [15],[39] and [6], which place restrictions
on transversal gate sets for QECCs, and [29] and [25], which use similar ITS-QHE constructions.
We very roughly summarize these results and compare them to our own.
In [39], Zeng et al. were some of the first to place restrictions on quantum universal transversal
gate sets for additive quantum codes by elucidating the stabilizer group structure. Further work in
[2] classified the set of diagonal gates that can implement one and two qubit logical operations in
stabilizer codes. Shortly thereafter, [15] showed that for any QECC, the transversal gate set must
be finite, and so cannot approximate with arbitrary precision the full unitary group. Intuitively,
they make a Lie type argument by showing that infinitesimal transversal operations are themselves
linear combinations of local error operators. Since these unitaries must act identically on the
codespace, it follows that the group of transversal operations must be finite.
More recently, [6] placed restrictions on the more general class of topologically protected logical
gates in topological stabilizer codes, which include transversal gates as an optimal subset. They
showed that for a topological stabilizer code defined on a d-dimensional lattice, any such gate must
lie in the dth level of the Clifford hierarchy. These results were extended in [32] to more general
stabilizer subsystem codes, and in Appendix D, we detail how these arguments can be used to rule
out classical reversible transversal computation for the subclass of stabilizer codes.
In the direction of ITS-QHE, [29] gave a compact and efficient ITS-QHE scheme for the restricted
class of Clifford circuits. Using magic state injection, they complete a universal gate set by adding
the T -gate. However, because the client and server cannot communicate during the protocol, they
must limit themselves to circuits using a constant number of T -gates. Again, their encryption is a
“noisy” encoding of the data into some code followed by a secret embedding into random noise. With
this encryption, they are able to generate indistinguishable outputs using only polynomial overhead
in the input size. In the more recent work [25], the authors independently observe Proposition 3.
They take the more positive approach of arguing what can be done in spite of this limitation,
extending the ITS-QHE schematic in [29] to other particular error-correcting codes and using code
concatenation to achieve security with only polynomial overhead. This achieves ITS-QHE on the
larger circuit class IQP + , which is probably not classically simulable [7].
Because of the stringent lower bounds placed by Nayak, we actually forgo the noisy encoding
circuit and embed QECCs directly into random noise after removing a correctable set of qubits.
This has the effect of increasing the overhead by an exponential factor in order to achieve security,
but thanks to the roomy lower bound, this factor is still too small to allow an ITS-QFHE scheme.
We can argue directly about the security of this scheme using the nonlocality of the quantum
information being encoded in almost any QECC. The idea is conceptually simple: in order to obtain
encryptions of the data that are both secure and (sufficiently) short, we must inject randomness
into the encodings themselves by withholding qubits from the code. While ordinarily this would
negatively affect the correctness of homomorphic evaluation, the error-correcting property allows us
to inject this randomness while still maintaining perfect recoverability. Then intuitively, spreading
3
the information across the subsystems limits the complexity of the class of logical operators that
don’t couple the subsystems, i.e. the transversal operators. This differs fundamentally from the
approaches in [39] and [15] in that it is a quantitative information-type bound.
It is not without its drawbacks however, as these maximally redundant codes fail to “spread out”
the information sufficiently. The prototypical example is Shor’s code, which is the concatenation of
a bit-flip and phase-flip code. However, we can argue directly using the stabilizer group structure
that no such additive code can implement Toffoli transversally.
2
Preliminaries
2.1
Quantum Information
We quickly review some standard notation, followed by some less standard tools we will need from
quantum information theory. For a more complete view, see [28].
Throughout, we will be working with 2-level qubit quantum systems. We denote by |H| =
log(dim(H)), the number of qubits constituting state space H. We define a general quantum state to
be a positive semi-definite operator ρ ∈ L(H) of trace one. We call such a state pure if rank(ρ) = 1,
otherwise we call it mixed, and note that such an operator is mixed if and only if Tr(ρ2 ) < 1. For
any operator U ∈ HA , we use the notation U A to indicate the operator UA ⊗ IB ∈ L(HA ⊗ HB ).
When it is unclear which space a state lives in, we will denote its state space as a superscript (e.g.
|ψiA ). We also sometimes adopt the notation that for ρ ∈ L(A ⊗ B), ρA = TrB (ρ). By slight abuse
of notation, we also adopt the convention that for any permutation π ∈ Sn , π can also indicate
the unitary permutation operator corresponding to the physical permutation of qubits. We also
sometimes omit the dimension of an identity operator I, but usually the dimension is implicitly its
trace normalization factor, e.g. I/D acts on a space of dimension D.
The norm k · kp refers the usual Schatten p-norm, so that for any A ∈ L(H) with singular values
(a1 , . . . , an ),
kAkp =
n
X
i=0
Further recall that
api
!1/p
for p > 1, and kAk1 =
n
X
|ai |.
i=0
1
kρ − σk1 = max Tr(P (ρ − σ)),
P ≤I
2
and so we can think of the 1-norm as a means of bounding the ability to distinguish two
quantum states, where ≤ refers to the positive semidefinite partial ordering. For a collection of
quantum states {ρS } indexed by S ∈ S, we sometimes
write ES [ρS ] to denote the expectation over
1 P
a uniformly random choice of S, ES [ρS ] = |S|
ρ
. We will regularly be referring to several
S
S∈S
particular gates, and so list them here.
0 1
0 −i
1 0
X=
Y =
Z=
1 0
i 0
0 −1
1 0 0 0
1 0 0 0
0 1 0 0
0 1 0 0
CX =
CZ
=
0 0 0 1
0 0 1 0
0 0 1 0
0 0 0 −1
4
Toff =
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
CCZ =
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0 0
0 0
0 0
0 0
0 0
0 0
1 0
0 −1
Definition 4. An (n, m, p)-quantum random access code (QRAC) is a mapping from an n-bit string
x to an m-qubit quantum state ρx along with a family of measurements {Mi0 , Mi1 }ni=1 satisfying,
for all x ∈ {0, 1}n , i ∈ [n],
Tr(Mij ρx ) ≥ p if xi = j.
More generally, we can consider some protocol Mi for retrieving xi and we call this protocol the
ith query of the QRAC, satisfying Pr[Mi (ρx ) = xi ] ≥ p.
Definition 5. An n-qubit quantum code is simply a subspace C of an n-body Hilbert space H
along with a fiducial orthonormal logical basis. Let PC denote the projection onto C. The code is
further called an [[n, k, d]] quantum error-correcting code (QECC) if it is a subspace C of dimension
2k satisfying, for any ⌊ d−1
2 ⌋-local operators Ea , Eb ,
PC Ea† Eb PC = λab PC .
for some Hermitian (λab ). We call the Ea , Eb correctable errors and we say d is the distance of the
code.
Operationally, this means that there exists a recovery channel R by which any ⌊ d−1
2 ⌋-local error
d−1
(acting nontrivially on at most ⌊ 2 ⌋ qubits) can be corrected. Recall also that any [[n, k, d]] QECC
can correct up to (d − 1) errors in known locations. We call these types of errors erasure errors,
and so call codes satisfying d ≥ 2 erasure-correcting codes.
For simplicity, we will restrict our discussion to QECCs encoding a single qubit, i.e. [[n, 1, d]]
QECCs. A quick review of the proof shows that we can make this assumption without loss of
generality by arguing against a classical universal transversal gate set on any single encoded logical
qubit. By similar reasoning, the proof applies to subsystem codes as well, where we require the
logical operators of such a code to be independent of the state of the gauge qubits.
When we have a collection of p logical qubits, each encoded into an n-qubit code, we can
decompose the np physical qubits into a fixed n-wise partition of p-qubits each so that every
partitioning set contains exactly one qubit from each code block. We refer to these partitioning
sets as the subsystems of the collective code.
Definition 6. An n-qubit stabilizer group S is an abelian subgroup of the n-qubit Pauli group not
containing −I. An n-qubit additive (or stabilizer) code encoding k logical qubits can be described
as the simultaneous (+1)-eigenspace of the Pauli operators comprising an n-qubit stabilizer group S
with n − k generators. The logical Pauli operators of this code correspond to the normalizer cosets
N (S)/S, and it follows that the distance of the code is the minimal weight operator in N (S)/S.
Definition 7. For any quantum error correcting code C, we define its logical states |ψiL to be the
physical encoding of |ψi in C. We define the logical gate UL to be a codespace preserving physical
gate that satisfies, for all |ψiL ∈ C, UL |ψiL = (U |ψi)L .
5
The set of transversal gates TC associated to C are those logical gates that decompose as a
product across the subsystems. That is to say, UL ∈ TC if UL = U1 ⊗ . . . ⊗ Un , where n is the
length of the code, each Ui acts on a single subsystem, and UL is a codespace preserving map on
the code C ⊗r for U an r-qubit gate. We further define a logical gate to be strongly transversal if it
decomposes as UL = U ⊗n . Following the example of [39], we do not allow coordinate permutations
in our definition of transversality.
Definition 8. We say a quantum code C = SpanC (|0̃i, |1̃i) is an r-fold code if it can be written as
|iiL =
r
O
|ψij i.
j=1
where each vector |ψij i does not further decompose as a product state across any bipartition. We
additionally assume that r ≤ d, that |ψ0j i and |ψ1j i occupy the same subsystem, and that |ψ0j i ⊥
|ψ1j i. It then makes sense to refer to Span{|ψ0j i, |ψ1j i} as the jth subcode. These assumptions are
natural, and we justify them in our discussion.
If the code is additionally an [[n, 1, d]] QECC with r = d ≥ 3 and each subcode has distance 1,
we simply call the resulting code a maximally redundant code. Note that any (pure state) code is
at least a 1-fold code.
The guiding example is Shor’s code, which can be seen as the concatenation of a repetition outer
code and a complementary GHZ inner code, neither of which is quantum erasure correcting. In the
case that the subcodes are identical, any maximally redundant code is just the concatenation of a
repetition code with some distance 1 subcode. Intuitively, these are codes for which you can’t erase
enough qubits to mix the state while still remaining perfectly correctable: while redundancy can be
used in classical error-correcting codes to protect information, quantum error-correcting codes must
“spread out” information to protect it. In this sense, these codes are maximally redundant because
they “spread out” the information the least. We show that non-additive maximally redundant codes
(i.e. maximally redundant codes for which the subcodes are comprised of non-stabilizer subspaces)
are the only binary QECCs with the hope of implementing logical Toffoli transversally.
2.2
Homomorphic Encryption
We define an ITS-QHE scheme as three algorithms performed between two parties which we will
call Client and Server. We restrict ourselves to the more limited setting of a quantum scheme implementing Boolean functions on classical data using quantum encodings. Of course, any impossibility
result then extends to the more difficult task of quantum computations on quantum inputs.
The parameters of such a scheme are given by (n, m, m′ , ǫ, ǫ′ ) and some gate set F. Formally, we
define the algorithms of an ITS-QHE scheme as acting on Client’s private workspace C, a message
space M sent from Client to Server after encryption, and a message space M′ sent from Server to
Client after evaluation.
(i) QHE.Enc(x) = ρCM , in which the client chooses an n-bit input x and encrypts with some
private randomness to obtain ρ. We assume that any quantum evaluation key is appended to
the encryption. Client then sends the message portion of the encryption ρM to the Server.
We define m to be the length of this message, the size of the encoding.
(ii) QHE.Evalf : L(M) −→ L(M′ ), in which Server, with description of some circuit f , applies an evaluation map to an encrypted state, possibly consuming an evaluation key in the
6
′
process. Server then sends his portion of the state σ CM := (I C ⊗ QHE.Evalf )(ρCM ) back
to Client, and we define the length of this message to be m′ , the size of the evaluated encoding.
′
(iii) QHE.Dec(σ CM ) = y, in which Client decrypts the returned evaluated encoding using her
side information C and recovers some associated plaintext y.
As an encryption scheme, the above should certainly satisfy
QHE.Dec(QHE.Enc(x)) = x.
and as an ITS-QHE scheme, there are three additional properties the scheme should satisfy as well.
(i) ǫ-Information-theoretic security: for any inputs x, y, letting ρx , ρy ∈ L(M) denote the outputs of QHE.Enc on M (thinking of these states as mixtures of encryptions under uniformly
random choices of secret key),
kρx − ρy k1 ≤ ǫ.
(ii) F-homomorphic: for any circuit f ∈ F and for any input x,
Pr[QHE.Dec(QHE.Evalf (QHE.Enc(x)))1 6= f (x)] ≤ ǫ′
where the probability is over the randomness of the protocol and the subscript 1 denotes
the first bit of the output. This restriction to the first bit is just to argue directly about
Boolean functions. For ease of exposition, we also assume without loss of generality that our
protocols are perfectly correct, and allow ǫ′ = 0. We call the scheme fully homomorphic if it
is homomorphic on the set of all classical Boolean circuits.
(iii) Compactness: a priori, the server could do nothing except append a description of the circuit
f to be run by the decryption function after decrypting. To avoid trivial solutions like this,
we demand that the total time-complexity of Client’s actions in the protocol do not scale with
the complexity of the functions to be evaluated, but only with some fixed function on the
size of the input. Intuitively, this captures the motivation behind homomorphic encryption:
limiting the computational cost to Client. However, we note that the standard definition of
compactness refers to the time-complexity of the decrypt function specifically.
We denote a scheme homomorphic for some class of functions F and satisfying all of these
properties as an F-ITS-QHE scheme. If F is the set of all Boolean circuits, we denote such a
scheme as an ITS-QFHE scheme. We observe that such a scheme must be inefficient. For a precise
statement and proof, see Appendix A.
Proposition 9. The communication cost of ITS-QFHE must be exponential in the size of the input.
3
A coding based ITS-QHE scheme
We now consider a strategy for implementing compact QHE using quantum codes. This will be a
simple “block” embedding encryption scheme homomorphically implementing quantum circuits on
classical input, and is similar to the construction in [29]. We will use the error-correcting property
to withhold a correctable set of qubits from the encoding.
7
Coding QHE Scheme:
Arguments:
C
~i
m
S
=
∈
=
∈
an [[n + r, 1, d]] r-fold QECC with r < d
{0, 1}p
the size of each noise code block
[m]n , the secret key
1. On input ~i ∈ {0, 1}p , encode ~i as the pure state
the logical computational basis defining C.
p
N
|iℓ iL , for {|0iL , |1iL }
ℓ=1
2. Let R be a collection of r subsystems, each of p-qubits, comprised of one
N
~
subsystem from each subcode. Then form γ i = TrR ( ℓ |iℓ iL ). Essentially,
~
γ i is the state of the collection of codewords with each codeword missing
one subsystem from each of its subcodes.
3. Initialize n (p × m) arrays of maximally mixed qubits, and replace the
~
Sj -th column of each array with the j-th subsystem of γ i . This forms the
encrypted state.
4. Publish a constant number of labeled encryptions of 0 and 1, to be used
as ancilla in homomorphic evaluation.
Figure 1: A description of the encryption procedure for the code based QHE scheme.
8
n+1
b
b
b
b
= |x1 iL
b
b
b
b
= |x2 iL
b
b
b
b
= |xp iL
x1
x2
Encryption
bbbb
Encoding
xp
S1
bc
bc
bc
bc
b
bc
bc
b
bc
S2
b
bc
bc
bc
b
bc
bc
b
bc
Sn
b
b
bc
bc
b
bc
bc
b
bc
m
n
Figure 2: A diagram illustrating the code-based QHE scheme for an (n + 1)-length 1-fold quantum
code while withholding a single subsystem. The (n + 1)-th subsystem remains in the hands of
Client. The arrows connecting the subsystems indicate where each subsystem (i.e. column) is
being mapped. The filled dots represent code qubits, while the empty dots represent maximally
mixed qubits.
The scheme is detailed in Figures 1 and 2. Using that notation to summarize, our encryption
~
channel E is defined, for secret key S and input string ~i, as E(S,~i) = γSi . We sometimes use
~
~
the notation γS instead of γSi or γ instead of γ i , omitting ~i when we are unconcerned with the
underlying plaintext. The total size of our encrypted input is mnp qubits. In our preceding
notation, the described scheme has parameters (p, mnp, mnp, ǫ(m, p), 0), implementing the set of
gates TC homomorphically.
Lemma 10. Let E be the encryption scheme detailed in Figure 1. Let TC denote the group of
transversal operators associated to the underlying quantum code C. Then, E is TC -homomorphic.
Proof. Let UL be the logical operator we wish to apply to some codestate |ψiL . By definition,
UL ∈ TC implies UL can be decomposed as a product operator U1 ⊗ . . . ⊗ Un+r where Ui is an
operator that acts only on the i-th subsystem of the code. Then, without knowledge of the secret
key S, a third party can implement UL by applying the operator
m
n O
O
Ui
i=1 j=1
where each Ui is an operator local to some subsystem in Server’s possession (that is to say, on
one of the columns in the corresponding array). Returning the resulting data to a party with the
secret key, that party can decrypt to obtain a state of the form V R UL |ψiL , where V R is supported
on the r subsystems that Client has withheld. Since r < d, viewing V R as an erasure error on r
subsystems, there exists some recovery channel R such that R(V R UL |ψL i) = UL |ψiL . Decoding,
we obtain U |ψi as desired.
9
Note that this scheme is a TC -homomorphic, non-leveled, and compact QHE scheme, since the
recovery and decryption channel do not depend on the complexity of U .
We now aim to compute the security ǫ(m, p) of the proposed scheme, namely the tradeoff
between the size of the input p, the size of the encoding mnp, and the ITS guarantee. To avoid
confusion, we point out here that the code size n is a constant, as we are not concatenating to
achieve security, just amplifying the size of the noise into which we are embedding.
We want to show that while the scheme is inefficient, its parameters still defeat Nayak’s bound.
To simplify the security proof, we use the more stringent requirement that the outputs are indistinguishable from uniformly random noise. Here, we will see that the nonlocality of the information
stored in QECCs is essential in its allowing us to withhold qubits while still delegating computation
to Server. This imposes the requirement of using quantum error-correcting codes, as evidenced by
the following observation.
Lemma 11. Suppose we replace the preceding scheme with one that does not withhold any of the
physical qubits comprising the (pure state) code. Then if m = o(2p ), ǫ must be bounded away from
zero.
Proof. Counting the rank of the encrypted state, note that rank(γS ) = 2np(m−1) . Then,
rank(ES [γS ]) ≤ mn 2np(m−1)
≤ 2n(p(m−1)+log(m)) .
Thus, the fraction of nonzero eigenvalues must be at most (2n )log(m)−p . Since log(m) = o(p), the
fraction of nonzero eigenvalues goes to zero, and so kES [γS ] − I/2mnp k1 must be bounded away
from zero as claimed.
4
Security tradeoff for the QHE coding scheme
Our aim is to give (inefficient, but sufficient) security parameters for the coding QHE scheme. We
will then argue that if there were a QECC implementing a sufficiently large transversal gate set
(such as the set of all classical reversible gates), then it would violate Nayak’s bound with these
parameters.
We will first need a small lemma on the structure of the partial trace operator. The proof can
be found in Appendix B.
Lemma 12. For Hilbert space decomposition H = H∆
¯ 2,
¯ 1 ⊗ H∆ ⊗ H∆
¯
¯
¯
¯
Tr (ρ∆1 ∆ ⊗ I ∆2 )(I ∆1 ⊗ σ ∆∆2 ) = Tr Tr∆
¯ 1 (ρ)Tr∆
¯ 2 (σ) .
With this we are ready to prove the security tradeoff between ǫ, p, and m. We adopt the same
notation used in the proposed scheme for convenience, and note that we are demanding the stronger
condition that outputs are indistinguishable from random noise.
Proposition 13. For the scheme described in Figure 1, letting K = 2p be the dimension of any
subsystem and for some c ∈ (0, 1), we have
k (I/K mn ) − ES [γS ]k1 ≤ ǫ(K, m)
n
1/2
−c 2K n
−
1
+
K
.
for ǫ(K, m) = m−1
m
m
10
Proof. By Cauchy-Schwartz,
k (I/K mn ) − ES [γS ]k21 ≤ K mn k (I/K mn ) − ES [γS ]k22
2
1
mn
2
≤ K Tr(ES [γS ] ) −
Tr(I)
Tr(ES [γS ]) +
K mn
K 2(mn)
≤ K mn Tr(ES [γS ]2 ) − 1.
where the third line follows by noting that, as a quantum state, Tr(ES [γS ]) = 1. We write |S ∩ S ′ |
to denote the size of the intersection of S and S ′ considered as sets. We can then decompose, for
pℓ = PrS,S ′ [|S ∩ S ′ | = ℓ],
mn X
K
mn
2
Tr(γS γS′ )
K Tr(ES [γS ] ) =
m2n
′
S,S
(∗) = K mn
n
X
ℓ=0
pℓ Tr E[(γS γS ′ ) |S ∩ S ′ | = ℓ] .
(n)(m−1)(n−ℓ)
n
≤ nl /mℓ and that p0 = ( m−1
Note that pℓ = ℓ mn
m ) . Furthermore, up to a permutation
mn−2n
on the coordinates, we may write for dim(I) = K
,
2
mn
′
mn
n
n
(mn−2n)
K E[(γS γS ′ ) |S ∩ S | = 0] = K Tr (γ/K ) ⊗ (γ/K ) ⊗ I/K
=1
again by noting that γ is a quantum state of trace one and by multiplicativity of trace over tensor
products. Next consider the general case |S ∩ S ′ | = ℓ. Then up to a permutation on the coordinates
and for some π ∈ Sn , for ∆ the subsystem of the intersection S ∩ S ′ ,
2
mn
n−ℓ
†
n−ℓ
(mn−2n+ℓ)
mn
⊗ γ)(πγπ ⊗ I/K
) ⊗ I/K
K Tr(γS γS ′ ) = K Tr (I/K
= K ℓ Tr (I ⊗ γ)(πγπ † ⊗ I)
†
= K ℓ Tr Tr∆
¯ (πγπ )
¯ (γ)Tr∆
where the final line follows from Lemma 12. Then, because we have withheld a subsystem from
i
each subcode of the underlying
QECC, in any row i we have that Tr∆
¯ (γ ) is mixed. It follows
i
†
i
that Tr Tr∆
¯ (πγ π ) < 1. So by separability across each encoded qubit and again by
¯ (γ )Tr∆
multiplicativity of trace across tensor products,
p
p
Y
O
ij †
ij
ij †
ij
Tr Tr∆
=
Tr∆
Tr
¯ (πγ π ) .
¯ (γ )Tr∆
¯ (πγ π )
¯ (γ )Tr∆
j=1
j=1
It follows that there exists some c ∈ (0, 1) so that
K mn Tr(γS γS ′ ) ≤ K ℓ−c .
11
Putting this all together, we observe that
K
mn
n
X
n
K ℓ
−c
pℓ Tr E[(γS γS ′ ) |S ∩ S | = ℓ] ≤ K
m
ℓ
ℓ=1
ℓ=1
n
K
−c
−1
≤K
1+
m
2K n
≤ K −c
m
n
X
′
Including the first term in the sum, we get,
n
m−1 n
−c 2K
+K
(∗) ≤
m
m
and so,
ǫ(K, m) =
m−1
m
n
−1+K
−c
2K
m
n 1/2
as desired.
5
Limitations on classical transversal computation
We are left with two competing bounds. On the one hand, it follows from Nayak’s bound (Appendix
A) that, for any F-ITS-QHE encryption scheme with security ǫ and communication size s,
s ≥ log(|F|)(1 − H(ǫ)).
If we choose parameters that do not leak some constant fraction of information about our input,
then as ǫ → 0 we see that for s chosen as some fixed function of the input size, it must be that
s = Ω(log(|F|)). Using the notation and parameters from the aforementioned coding scheme, this
means that mnp = Ω(log(|Fp |)) for Fp the restriction of functions in F to p-bit inputs. Note that
we can assume no ancilla overhead since the constant gets absorbed into this asymptotic bound.
Now by construction of the scheme, F is the transversal gate set for the underlying choice of
quantum error-correcting code. Next, we would like to choose m as a function of K so that ǫ → 0.
For this, it suffices to choose m as a function of K so that
n
−c 2K
= 0.
lim K
K→∞
m
Equivalently, we require m = ω(K 1−( n ) ). Then for some c′ < 1, we can select m = K c and still
have ǫ → 0. Plugging this back into Nayak’s bound, we see that asymptotically
c
′
′
K c log(K) = Ω(log(|Fp |))
for |Fp | the size of the function class, seen itself as a function returning the number of unique
members in the class on p-bit inputs. In particular, Fp cannot be the set of all Boolean functions,
for then log(|Fp |) = K. This shows that no code satisfying the hypotheses of our scheme can
implement Toffoli transversally.
12
We now justify our earlier assumptions on the structure of candidate r-fold codes. Suppose
an r-fold [[n, 1, d]] QECC could implement a logical Toffoli gate transversally. First note that the
tensor decomposition between the logical states must align, or else the restriction of logical Toffoli
to one element of the product would unitarily map a pure state to a mixed state. Furthermore, we
can think of the QECC criterion in Definition 5 as a diagonal and off-diagonal condition: for all
|E| < d,
h0L |E|0L i = h1L |E|1L i,
h0L |E|1L i = 0.
Since the Paulis form an operator basis, we can always assume
N that E is an element of the Pauli
group. Then, for r-fold codes with logical basis states |iiL = rj=1 |ψij i, this becomes
r
Y
hψik |Ek |ψjk i = cE δij
k=1
where E = E1 ⊗ . . . ⊗ Er . Note then that if |ψ0j i 6⊥ |ψ1j i, we can trace out the corresponding
subsystem and obtain a code with the same correctable error set on the complement of that system.
Furthermore, if r > d, then we can again trace out any r − d subcode subsystems to obtain a code
with the same correctable error set on the complement. Both of these observations follow from
noticing that these subcodes must themselves satisfy the diagonal condition,
hψ0j |E|ψ0j i = hψ1j |E|ψ1j i.
It follows from the security proof that if r < d, then the code would satisfy the hypotheses
of our scheme and violate the lower bound in Proposition 20. Thus, r = d. Furthermore, logical
transversal Toffoli on the entire code must restrict (up to global phase) to a logical transversal
Toffoli gate on the subcodes, each of which is 1-fold by definition. Thus, each subcode must itself
have distance 1. To summarize,
Theorem 14. If a QECC is not a maximally redundant code, then it does not admit a classicalreversible universal transversal gate set. In particular, no such code can implement the Toffoli gate
transversally.
c
Note also that for the scheme in Figure 1, for any m = ω(K 1−( n ) ), ǫ(K) is negligible in p.
Summarizing the parameters of the coding scheme:
Proposition 15. For any r-fold [[n, 1, d]] quantum error-correcting code C with r < d and with
transversal gate set TC , the described protocol is a compact quantum TC -homomorphic encryption
′
scheme with security ǫ = negl(p) for p the input size and with encoding size m = 2pc for some
c′ < 1.
While this is highly inefficient, we pause to give some intuition for why it suits our purposes.
On the one hand, we can envision trivial “hiding” schemes that have encoding length 2p in each
bit. Nayak’s bound allows for higher efficiency, roughly demanding that encodings implementing
the set of all classical functions on p bits homomorphically must have length at least (2p /p) in each
′
bit. Finally our scheme, with encoding length 2pc for some c′ ∈ (0, 1), is just efficient enough to
defeat this bound and allow us to argue Theorem 14.
Because these maximally redundant codes have a simple design, if we further assume that
they are additive, we can use the additional stabilizer structure to argue directly that they cannot
implement logical Toffoli transversally. From this observation, we directly obtain the following.
13
Corollary 16. No additive QECC can implement transversal Toffoli.
For proof, see Appendix C. Note that this also follows from the arguments in Appendix D. Our
central result follows.
Corollary 17. If a QECC is not a non-additive maximally redundant code, then it cannot implement the Toffoli gate transversally.
Finally, note that by concatenating an [[n, 1, d]] d-fold code with itself, the code remains d-fold
while the distance must increase to at least d2 . Furthermore, if such a code implements Toffoli
strongly transversally, then so does its concatenation with itself. As a result, we observe the
following.
Corollary 18. No QECC can implement strongly transversal Toffoli.
6
Discussion
Do there exist non-additive maximally redundant codes that can then implement Toffoli transversally? One can essentially think of these as QECCs formed by concatenating an outer repetition
code with a distance 1 inner code that is not a stabilizer subspace. Intuitively, since the inner code
is not quantum error-correcting, the code only “spreads out the information in one basis”. More
precisely, the inner code only satisfies the diagonal QECC criterion. While this is a less restrictive
condition, it still must be “complementary” to the outer code, and this allows us to argue impossibility in the additive case. Unfortunately by comparison, the structure of general non-additive
codes is less well-understood – in particular, we know of no examples of such a code. We expect
that no QECC can implement Toffoli transversally, and view this exception as a consequence of
the lack of structure on general non-additive codes. We hope to resolve this exception in upcoming
work.
The QHE scheme we have detailed is non-leveled and compact, but highly inefficient. An
immediate question would be to refine the security proof, which uses too strong a security demand.
It would be most interesting to see if a modified approach can achieve efficient ITS-QHE for
transversal gate sets of general quantum error-correcting codes, where the size of the encoding is
some fixed polynomial of the input length. There are certain quantitative properties of “nonlocality”
in QECCs (see e.g. [3], [34]) that might be helpful in such an endeavor. Following the outline of
[29], we could also expect to extend a scheme built on a code with desirable transversal gates
to accommodate a constant number of non-transversal gates. Just as one might tailor a QECC
for a specific algorithm that makes heavy use of its transversal gate set, one might also tailor an
ITS-QHE scheme to homomorphically implement that algorithm. Furthermore, it would be of
theoretical interest to find a protocol matching the lower bound implicit in Proposition 20.
Another interesting open question is to consider leveled ITS-QHE schemes: allow the client
some preprocessing to scale with the size of the circuit. Can this relaxation allow more efficient or
universal schemes for polynomial sized circuits, mirroring the computational security case? A first
step might be to try to apply the techniques of instantaneous nonlocal computation [35] that proved
invaluable in the computationally secure scheme. Moreover, through gauge-fixing, we have ways
of converting between codes that together form a universal transversal gate set. Its not clear how
to implement such a strategy, since the noisy embedding and non-interactivity present barriers to
measuring syndromes, but these elements taken together might be useful in extending the current
scheme.
14
Finally, one could ask if there is a correspondence between transversal gates for quantum codes
and nontrivial ITS homomorphically-implementable gate sets, based on the “richness” of the function classes they can realize. In particular, [15] asked: what is the maximum size of finite group
that can be implemented logically and transversally? Indeed, since the Clifford group on p-qubits
2
is of size at most 22p +3p [30], one could reasonably expect to efficiently implement the Clifford
gates homomorphically with information theoretic security, as was done in [29]. We hope that our
arguments might extend past classical reversible circuit classes to address this question, although it
is unclear how to generalize Nayak’s bound to apply to these general finite subgroups of the unitary
group.
7
Acknowledgments
The authors would like to thank Cupjin Huang and Fang Zhang for useful discussions, in particular concerning the security proof. They are grateful to Audra McMillan, Anne Broadbent, and
Zhengfeng Ji for their comments on an earlier draft of this paper. In particular, we thank Anne
for bringing [4] to our attention and pointing out that Proposition 3 follows from the result of
that paper and an argument similar to the classical case. They also thank Fernando Pastawksi for
pointing us to Lemma 23. This research was supported in part by NSF Awards 1216729, 1318070,
and 1526928.
15
References
[1] A. Ambainis, D. Leung, L. Mancinska, and M. Ozols. Quantum random access codes with
shared randomness, October 2008. arXiv:0810.2937.
[2] J. T. Anderson and T. Jochym-O’Connor. Classification of transversal gates in qubit stabilizer
codes, 2016. Quantum Information and Computation 16, 0771-0802.
[3] L. Arnaud and N. J. Cerf. Exploring pure quantum states with maximally mixed reductions,
January 2013. Phys. Rev. A 87, 012319 (2013).
[4] A. Baumeler and A. Broadbent. Quantum private information retrieval has linear communication complexity, April 2013. Journal of Cryptology. Volume 28, Issue 1, pp 161-175 (2015).
[5] H. Bombin. Gauge color codes: Optimal transversal gates and gauge fixing in topological
stabilizer codes, August 2015. New J. Phys. 17 (2015) 083002.
[6] S. Bravyi and R. Koenig. Classification of topologically protected gates for local stabilizer
codes, 2013. Phys. Rev. Lett. 110, 170503 (2013).
[7] M. J. Bremner, R. Jozsa, and D. J. Shepherd. Classical simulation of commuting quantum
computations implies collapse of the polynomial hierarchy, August 2010. Proceedings of the
Royal Society A, Volume 467, Issue 2126.
[8] A. Broadbent. Delegating private quantum computations, June 2015. Canadian Journal of
Physics, 2015, 93(9): 941-946.
[9] A. Broadbent, J. Fitzsimons, and E. Kashefi. Universal blind quantum computation, 2009.
Proceedings of the 50th Annual IEEE Symposium on Foundations of Computer Science (FOCS
2009), pp. 517-526.
[10] A. Broadbent and S. Jeffery. Quantum homomorphic encryption for circuits of low T-gate
complexity, 2015. In Proceedings of Advances in Cryptology – CRYPTO 2015, pp 609-629.
[11] A. Broadbent, Z. Ji, F. Song, and J. Watrous. Zero-knowledge proof systems for QMA, April
2016. Proceedings of the 2016 IEEE 57th Annual Symposium on Foundations of Computer
Science (FOCS 2016) pp.31-40.
[12] R. Cramer, L. Ducas, C. Peikert, and O. Regev. Recovering short generators of principal ideals
in cyclotomic rings. Cryptology ePrint Archive, Report 2015/313, 2015.
[13] A. W. Cross. Fault-tolerant quantum computer architectures using hierarchies of quantum
error-correcting codes, 2008. PhD Thesis.
[14] Y. Dulek, C. Schaffner, and F. Speelman. Quantum homomorphic encryption for polynomialsized circuits, August 2016. CRYPTO 2016: Advances in Cryptology - CRYPTO 2016, pp
3-32.
[15] B. Eastin and E. Knill. Restrictions on transversal encoded quantum gate sets, July 2009.
Phys. Rev. Lett. 102, 110502.
[16] M. Fillinger.
Lattice based cryptography and fully homomorphic encryption, 2012.
http://homepages.cwi.nl/~schaffne/courses/reports/MaxFillinger_FHE_2012.pdf.
16
[17] K. Fisher, A. Broadbent, L. Shalm, Z. Yan, J. Lavoie, R. Prevedel, T. Jennewein, and K. Resch.
Quantum computing on encrypted data, January 2014. Nature Communications 5, Article
number: 3074.
[18] A. G. Fowler, S. J. Devitt, and C. Jones. Surface code implementation of block code state
distillation, January 2013. Scientific Reports 3, 1939.
[19] A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland. Surface codes: Towards
practical large-scale quantum computation, August 2012. Phys. Rev. A 86, 032324 (2012).
[20] C. Gentry. A fully homomorphic encryption scheme, 2009. Ph.D. Thesis, Stanford University.
[21] D. Gottesman. Stabilizer codes and quantum error correction, 1997. Caltech Ph.D. Thesis.
[22] C. Jones. Composite toffoli gate with two-round error detection, March 2013. Phys. Rev. A
87, 052334.
[23] C. Jones. Novel constructions for the fault-tolerant toffoli gate, 2013. Phys. Rev. A 87, 022328.
[24] E. Knill and R. Laflamme.
Phys.Rev.Lett.84:2525-2528.
A theory of quantum error-correcting codes, April 1996.
[25] C.-Y. Lai and K.-M. Chung. On statistically-secure quantum homomorphic encryption. In
preparation.
[26] H. P. Nautrup, N. Friis, and H. J. Briegel. Topological code switching in two dimensions,
September 2016. arXiv:1609.08062.
[27] A. Nayak. Optimal lower bounds for quantum automata and random access codes, April 1999.
FOCS 1999.
[28] M. A. Nielsen and I. L. Chuang. Quantum computation and quantum information, 2011.
Cambridge University Press New York, NY.
[29] Y. Ouyang, S.-H. Tan, and J. Fitzsimons. Quantum homomorphic encryption from quantum
codes, August 2015. arXiv:1508.00938.
[30] M.
Ozols.
Notes
on
the
clifford
group,
July
http://home.lu.lv/~sd20008/papers/essays/Clifford%20group%20[paper].pdf.
2008.
[31] A. Paetznick and B. W. Reichardt. Universal fault-tolerant quantum computation with only
transversal gates and error correction, April 2013. Phys. Rev. Lett. 111, 090505 (2013).
[32] F. Pastawksi and B. Yoshida. Fault-tolerant logical gates in quantum error-correcting codes,
2015. Phys. Rev. A 91, 012305.
[33] C. Peikert. A decade of lattice cryptography, March 2016. Foundations and Trends in Theoretical Computer Science 10(4):283-424.
[34] A. J. Scott. Multipartite entanglement, quantum-error-correcting codes, and entangling power
of quantum evolutions, May 2004. Phys. Rev. A 69, 052330.
[35] F. Speelman. Instantaneous non-local computation of low T-depth quantum circuits, November
2015. arXiv:1511.02839.
17
[36] S.-H. Tan, J. A. Kettlewell, Y. Ouyang, L. Chen, and J. F. Fitzsimons. A quantum approach
to homomorphic encryption, 2016. Sci. Rep. 6, 33467.
[37] T. J. Yoder, R. Takagi, and I. L. Chuang. Universal fault-tolerant gates on concatenated
stabilizer codes, March 2016. Phys. Rev. X 6, 031039 (2016).
[38] L. Yu, C. A. Perez-Delgado, and J. F. Fitzsimons. Limitations on information theoretically
secure quantum homomorphic encryption, June 2014. Phys. Rev. A 90, 050303 (2014).
[39] B. Zeng, A. Cross, and I. L. Chuang. Transversality versus universality for additive quantum
codes, September 2011. IEEE Transactions on Information Theory, Volume: 57, Issue: 9, 6272
- 6284.
18
A
A no-go result for ITS-QFHE
We now set out to show that efficient ITS-QFHE is impossible. Define for any x ∈ {0, 1}n the state
|sx iCM = QHE.Enc(x).
Note that for any |sx ihsx |M , we can always purify to a system of size at most 2|M|. So without
loss of generality, we may assume that |C| = m, the size of the message sent from Client to Server.
Next, by information theoretic security, the state of the encryption on subsystem M must be almost
independent of x. Formally,
kTrC (|sx ihsx |) − TrC (|s0n ihs0n |)k1 ≤ ǫ
for ǫ the security of the scheme. Equivalently, there exists a VxC so that, defining s′x = (VxC ⊗
I M )|s0n ihs0n |(VxC ⊗ I M )† and sx = |sx ihsx |,
ks′x − sx k1 ≤ ǫ.
Furthermore, for any f ∈ F, we have by the homomorphic property that, abbreviating QHE.Evalf
as fev (·) and QHE.Dec(·) as D(·),
C
′
M CM CM
=: ηf,x ,
I ⊗ fev
(sx )
′
D CM (ηf,x ) = f (x).
′
by replacing sx with s′x in the definition of ηf,x , by contractivity of trace
But now, defining ηf,x
distance we also have
′
′
Pr[D CM (ηf,x
) 6= f (x)] ≤ ǫ.
To elucidate the underlying QRAC, define the mapping
f 7→ ηf,0n ,
i
h
′
′
′ . So let D CM′ (V C ⊗ I M′ )(·) denote the query
and note that (VxC ⊗ I M )ηf,0n (VxC ⊗ I M )† = ηf,x
x
for index x of f , thinking of f as a 2n length bit string, with the xth bit defined as f (x). Then
we have a (2n , m + m′ , 1 − ǫ)-QRAC for the set of all Boolean functions, where m + m′ is the
communication cost of the protocol. We now recall a well-known bound on the efficiency of QRACs
[27].
Theorem 19 (Nayak’s Bound). If there exists an (n, m, p)-QRAC, then for H(·) the binary entropy
function,
m ≥ n(1 − H(p)).
So it must be that the total communication cost of the protocol (m + m′ ) ≥ 2n (1 − H(ǫ)).
For security, allowing ǫ → 0 and noting that H(ǫ) → 0, we see that the communication cost
(m + m′ ) = Ω(2n ). Thus, either the size of the encoding or the evaluated ciphertext must be
exponentially long in the input, precluding efficiency. In short,
Proposition 20. The communication cost of ITS-QFHE must be exponential in the size of the
input.
19
B
Proof of Lemma 12
Proof. Expanding in terms of outer products,
Tr ((ρ ⊗ I)(I ⊗ σ)) = Tr
XXX
i,i′ j,j ′
k
¯
¯
ai,i′ ,j,j ′ |iihi|∆1 ⊗ |jihj ′ |∆ ⊗ |kihk|∆2 ·
X X X
bm,m′ ,n,n′ |ℓihℓ|
¯1
∆
¯2
′ ∆
⊗ |mihm′ |∆ ⊗ |nihn |
ℓ m,m′ n,n′
= Tr
XXX
i,i′ n,n′ j,m′
=
XXX
n
i
!
!
X
′
′
ai,i′ ,j,j ′ bj ′ ,m′ ,n,n′ |iihi| ⊗ |jihm | ⊗ |nihn |
j′
ai,i,j,j ′ bj ′ ,j,n,n .
j,j ′
On the other hand, we have
X X
XX
Tr Tr∆
bm,m′ ,n,n |mihm′ |
ai,i,j,j ′ |jihj ′ |
¯ 1 (ρ)Tr∆
¯ 2 (σ) = Tr
i
= Tr
XXXX
i
=
n
XXX
i
n m,m′
j,j ′
n
j,j ′
j′
!
ai,i,j,j ′ bj ′ ,m′ ,n,n |jihm′ |
ai,i,j,j ′ bj ′ ,j,n,n
j,j ′
!
as claimed.
C
Proof of Corollary 16
Proof. By Theorem 14, it suffices to consider maximally redundant codes. So suppose, for the sake
of contradiction, that an [[n, 1, d]] additive d-fold code could implement Toffoli transversally. Let
[·, ·] denote the group commutator. We denote by ¯· states and operations acting on the subcodes,
and ˜· those on the full code. We will assume that each subcode is the same, e.g. |ĩi = |īi⊗d , so that
we can speak directly about the inner and outer codes. The general argument follows similarly.
Since the code is additive, the code distance is the minimal weight logical Pauli operator acting
on the code. For any Z̄L , by multiplicativity of the inner product over tensor products,
1
1
|0̃i + |1̃i Z̄L |0̃i − |1̃i =
h0̃|Z̄L |0̃i − h0̃|Z̄L |1̃i + h0̃|Z̄L |0̃i − h1̃|Z̄L |1̃i
2
2
n
n
1
h0̄|0̄i d + h1̄|1̄i d 6= 0.
=
2
Since the outer code has distance d, it follows from the QECC criterion that Z̄L must have weight at
least d. Then X̄L must have weight 1, since the underlying inner code has distance 1 by assumption.
g L on the
Because the outer classical repetition code factors as a tensor product, transversal Toff
outer code must restrict (up to a global phase) to transversal ToffL on the inner code. Since we’re
now working with multiqubit gates, let GL (i) denote the logical gate for G acting on the ith code
block. We can compute directly,
20
[ToffL (1, 2, 3), X̄L (1)] = CX L (2, 3).
Furthermore, because ToffL and X̄L are transversal, it follows that CX L has a representative
that is also transversal and is supported on the subsystems that support X̄L . By a similar argument
[CX L (1, 2), Z̄L (1)] = Z̄L (2)
so that Z̄L must also be contained in the subsystems supporting CX L , and in turn X̄L . As
we have already observed, the minimal weight of any representative of Z̄L must be at least d, a
contradiction as X̄L has a representative of weight 1.
D
An alternate proof for stabilizer codes
Here we offer an alternate proof limiting universal transversal reversible computation for the subclass of stabilizer codes. The arguments here are based off of results from [6] and [32], which we
reproduce for completeness.
Definition 21. The Clifford hierarchy C is a sequence of gate sets {Ck }k≥1 defined recursively by
Ck = {U : U C1 U † ⊆ Ck−1 }, where we define C1 to be the Pauli group.
Note that C2 is the Clifford group, and Ck fails to be a group for k > 2. Further note that
reversible circuits saturate the Clifford hierarchy (and in fact can lie outside it entirely) by the gate
C k X, the k-controlled bit-flip gate, which lies in Ck+1 . Toffoli is simply C 2 X, and so lies in the
third level of the Clifford hierarchy. We next recall the stabilizer cleaning lemma, which can be
found in [6].
Lemma 22. Let S be a stabilizer code, and let R be any subset of physical qubits of the code
such that any logical operator supported on R acts trivially on S. Then, for any logical operator
UL , there exists a representative of UL supported on Rc .
We call such subsets R cleanable. Equipped with the cleaning lemma, we can now summarize
the following lemma from [32].
Lemma 23. Let S be a stabilizer code and let {R0 , . . . , Rk } be a set of cleanable subsets of the
physical qubits comprising S. Let U be a logical operator supported on ∪ki=0 Ri such that U is
transversal with respect to the Ri . Then, UL ∈ Ck .
Proof. We proceed by induction on k. In the base case, we have a logical operator U supported
on cleanable subsets R0 ∪ R1 . Let P be any logical Pauli operator cleaned off of R1 , and let [·, ·]
denote the group commutator. Since in a stabilizer code the logical Pauli operators are transversal,
we have Supp([U, P ]) ⊆ R0 , which by cleanability implies that [UL , PL ] = cIL . Since this is true
for any PL , it must be that UL ∈ C1 .
Similarly, suppose U is supported on ∪ki=0 Ri . Then, cleaning any logical Pauli P off of Rk ,
k−1
we see that Supp([U, P ]) ⊆ ∪i=0
Ri . By our inductive hypothesis, [UL , PL ] ⊆ Ck−1 , which implies
†
UL PL UL ∈ Ck−1 for any logical Pauli PL . Thus UL ∈ Ck , completing the proof.
This argument generalizes to subsystem codes, and we refer the reader [32] for a more complete
description. As a consequence we obtain the following.
21
Corollary 24. No erasure-correcting stabilizer code can implement a classical reversible universal
transversal gate set.
Proof. Partition the code block into single subsystem subsets {R1 , . . . , Rn } where n is the length
of the code. Then, since the code is erasure-correcting, any logical operator supported on a single
subsystem must act trivially on the codespace, and so these subsets are cleanable. By the lemma,
any transversal logical gate must lie in Cn . Since reversible circuits saturate C, they cannot be
logically transversally implementable.
22
| 7cs.IT
|
Optimal Densification for Fast and Accurate Minwise Hashing
Anshumali Shrivastava
Rice University, Houston, TX, 77005, USA
arXiv:1703.04664v1 [cs.DS] 14 Mar 2017
Abstract
Minwise hashing is a fundamental and one of
the most successful hashing algorithm in the literature. Recent advances based on the idea of
densification (Shrivastava & Li, 2014a;c) have
shown that it is possible to compute k minwise hashes, of a vector with d nonzeros, in
mere (d + k) computations, a significant improvement over the classical O(dk). These advances have led to an algorithmic improvement
in the query complexity of traditional indexing
algorithms based on minwise hashing. Unfortunately, the variance of the current densification
techniques is unnecessarily high, which leads to
significantly poor accuracy compared to vanilla
minwise hashing, especially when the data is
sparse. In this paper, we provide a novel densification scheme which relies on carefully tailored
2-universal hashes. We show that the proposed
scheme is variance-optimal, and without losing
the runtime efficiency, it is significantly more accurate than existing densification techniques. As
a result, we obtain a significantly efficient hashing scheme which has the same variance and
collision probability as minwise hashing. Experimental evaluations on real sparse and highdimensional datasets validate our claims. We believe that given the significant advantages, our
method will replace minwise hashing implementations in practice.
ANSHUMALI @ RICE . EDU
speech, and text are quite popular (Broder, 1997; Fetterly
et al., 2003). Enriching features with co-occurrence information leads to blow up in the dimensionality. 5-grams
are common for text representations. With vocabulary
size of 106 , 5-grams representation requires dimensionality of 1030 . Representing genome sequences with features
consisting of 32-contiguous characters (or higher) (Ondov
et al., 2016) leads to around 432 = 264 dimensions.
To deal with the overwhelming dimensionality, there is an
increased emphasis on the use of hashing algorithms, such
as minwise hashing. Minwise hashing provides a convenient way to obtain a compact representation of the data,
without worrying about the actual dimensionality. These
compact representations are directly used in large scale
data processing systems for a variety of tasks.
Minwise hashing is defined for binary vectors. Binary vectors can also be equivalently viewed as sets, over the universe of all the features, containing only attributes corresponding to the non-zero entries. Minwise hashing belongs
to the Locality Sensitive Hashing (LSH) family (Broder
et al., 1998; Charikar, 2002). The method applies a random permutation (or random hash function) π : Ω → Ω,
on the given set S ⊂ Ω, and stores the minimum value after
the permutation mapping. Formally,
hπ (S) = min(π(S)).
Given sets S1 and S2 , it can be shown by elementary probability arguments that
P r(hπ (S1 ) = hπ (S2 )) =
1. Introduction and Motivation
Recent years have witnessed a dramatic increase in the dimensionality of modern datasets. (Weinberger et al., 2009)
show dataset with 16 trillion (1013 ) unique features. Many
studies have shown that the accuracy of models keeps
climbing slowly with exponential increase in dimensionality. Large dictionary based representation for images,
(1)
|S1 ∩ S2 |
= R.
|S1 ∪ S2 |
(2)
The quantity
R=
|S1 ∩ S2 |
a
=
,
|S1 ∪ S2 |
f1 + f2 − a
(3)
is the well known Jaccard Similarity (or resemblance) R
which is the most popular similarity measure in information retrieval applications (Broder, 1997).
The probability of collision (equality of hash values), under minwise hashing, is equal to the similarity of interest
Optimal Densification for Fast and Accurate Minwise Hashing
R. This particular property, also known as the LSH property (Indyk & Motwani, 1998; Charikar, 2002), makes minwise hash functions hπ suitable for creating hash buckets,
which leads to sublinear algorithms for similarity search.
Because of this same LSH property, minwise hashing is
a popular indexing technique for a variety of large-scale
data processing applications, which include duplicate detection (Broder, 1997; Henzinger, 2006), all-pair similarity (Bayardo et al., 2007), temporal correlation (Chien
& Immorlica, 2005), graph algorithms (Buehrer & Chellapilla, 2008; Chierichetti et al., 2009; Najork et al., 2009),
and more. It was recently shown that the LSH property of
minwise hashes can be used to generate kernel features for
large-scale learning (Li et al., 2011).
Minwise hashing is known to be theoretical optimal in
many scenarios (Bavarian et al., 2016). Furthermore, it was
recently shown to be provably superior LSH for angular
similarity (or cosine similarity) compared to widely popular Signed Random Projections (Shrivastava & Li, 2014b).
These unique advantages make minwise hashing arguably
the strongest hashing algorithm both in theory and practice.
Hashing Cost is Bottleneck: The first step of algorithms
relying on minwise hashing is to generate, some large
enough, k minwise hashes (or fingerprints) of the data vectors. In particular, for every data vector x, hi (x) ∀i ∈
{1, 2, ..., k} is repeatedly computed with independent permutations (or hash functions). These k hashes are used
for a variety of data mining tasks such as cheap similarity estimation, indexing for sub-linear search, kernel features for large scale learning, etc. Computing k hashes of a
vector x with traditional minwise hashing requires O(dk)
computation, where d is the number of non-zeros in vector x. This computation of the multiple hashes requires
multiple passes over the data. The number of required
hashes typically ranges from few hundreds to several thousand. For example, the number of hashes required by the
famous LSH algorithm is O(nρ ) which grows with the size
of the data. (Li, 2015) showed the necessity of around
4000 hashes per data vector in large-scale learning. Hashing time is the main computational and resource bottleneck
step in almost all applications using minwise hashing.
Other Related Fast Sketches are not LSH: Two notable techniques for estimating Jaccard Similarity are: 1)
bottom-k sketches and 2) one permutation hashing (Li
et al., 2012). Although these two sketches are cheap
to compute, they do not satisfy the key LSH property
and therefore are unsuitable for replacing minwise hashing (Shrivastava & Li, 2014a;c). There are also substantial
empirical evidence that using these (non-LSH) sketches for
indexing leads to a drastic bias in the expected behavior,
leading to poor accuracy.
The Idea of “Densified” One Permutation Hashing: Re-
cently, (Shrivastava & Li, 2014a) showed a technique for
densifying sparse sketches from one permutation hashing
which provably removes the bias associated with one permutation hashing. Please see Section 3.3 for details. This
was the first success in creating efficient hashing scheme
which satisfies the LSH property analogous to minwise
hashing and at the same time the complete process only
requires O(d + k) computations instead of the traditional
bottleneck of O(dk). Having such an efficient scheme directly translates into algorithmic improvements for a variety of machine learning and data mining tasks.
Current Densification is Inaccurate For Very Sparse
Datasets: The densification process although efficient and
unbiased was shown to have unnecessarily higher variance.
It was shown in (Shrivastava & Li, 2014c) that the traditional “densification” lacks sufficient randomness. It was
further revealed that densification could be provably improved by using k extra random bits. The improved scheme
has reduced variance, and it retained the computational efficiency, see (Shrivastava & Li, 2014c) for more details.
An improved variance was associated with a significant
performance gain in the task of near-neighbor search. In
this work, we show that even the improved densification
scheme is far from optimal. The findings of (Shrivastava
& Li, 2014c) leaves an open curiosity: What is the best
variance that can be achieved with “densification” without
sacrificing the running time? We close this by providing a
variance-optimal scheme.
Our Contributions: We show that the existing densification schemes, for fast minwise hashing, are not only suboptimal but, worse, their variances do not go to zero with
increasing number of hashes. The variance with an increase
in the number of hashes converges to a positive constant.
This behavior implies that increasing the number of hashes
after a point will lead to no improvement, which is against
the popular belief that accuracy of randomized algorithms
keeps improving with an increase in the number of hashes.
To circumvent these issues we present a novel densification
scheme which has provably superior variance compared to
existing schemes. We show that our proposal has the optimal variance that can be achieved by densification. Furthermore, the variance of new methodology converges to zero
with an increase in the number of hashes, a desirable behavior absent in prior works. Our proposal makes novel use of
2-universal hashing which could be of independent interest
in itself. The benefits of improved accuracy come with no
loss in computational requirements, and our scheme retains
the running time efficiency of the densification.
We provide rigorous experimental evaluations of existing
solutions concerning both accuracy and running time efficiency, on real high-dimensional datasets. Our experiments
validate all our theoretical claims and show significant im-
Optimal Densification for Fast and Accurate Minwise Hashing
provement in accuracy, comparable to minwise hashing,
with a significant gain in computational efficiency.
2. Important Notations and Concepts
Equation 2 (i.e. the LSH Property) leads to an estimator of
Jacard Similarity R, using k hashes, defined by:
R̂ =
k
1X
1(hi (S1 ) = hi (S2 )).
k i=1
(4)
Here 1 is the indicator function. In the paper, by variance,
we mean the variance of the above estimator. Notations like
V ar(h+ ), will mean the variance of the above estimator
when the h+ is used as the hash function.
[k] will denote the set of integers {1, 2, ..., k}. n denotes
the number of points (samples) in the dataset. D will be
used for dimensionality. We will use min{S} to denote the
minimum element of the set S. A permutation π : Ω → Ω
applied to a set S is another set π(S), where x ∈ S if and
only if π(x) ∈ π(S). Our hashing will generate k hashes
hi i ∈ {1, 2, ..., k}, generally from different bins. Since
they all have same distribution and properties we will drop
subscripts. We will use h and h+ to denote the hashing
schemes of (Shrivastava & Li, 2014a) and (Shrivastava &
Li, 2014c) respectively.
3. Background: Fast Minwise Hashing via
Densification
3.1. 2-Universal Hashing
Definitions: A randomized function huniv : [l] → [k] is
2-universal if for all, i, j ∈ [l] with i 6= j, we have the
following property for any z1 , z2 ∈ [k]
P r(huniv (i) = z1 and huniv (j) = z2 ) =
1
k2
(5)
(Carter & Wegman, 1977) showed that the simplest way
to create a 2-universal hashing scheme is to pick a prime
number p ≥ k, sample two random numbers a, b and compute
huniv (x) = ((ax + b) mod p) mod k
3.2. One Permutation Hashing and Empty Bins
It was shown in (Li et al., 2012; Dahlgaard et al., 2015)
that instead of computing the global minimum in Equation 2, i.e., h(S) = min(π(S)), an efficient way to generate k sketches, using one permutation, is to first bin the
range space of π, i.e. Ω, into k disjoint and equal partitions
followed by computing minimum in each bin (or partition).
Let Ωi denote the ith partition of the range space of π, i.e.
Ω. Formally, the ith one permutation hashes (OPH) of a set
S is defined as
(
min{π(S) ∩ Ωi }, if {π(S) ∩ Ωi } =
6 φ
H
hOP
(S) =
i
E,
otherwise.
(6)
An obvious computational advantage of this scheme is that
it is likely to generate many hash values, at most k, and only
requires one permutation π and only pass over the sets (or
binary vectors) S. It was shown that for any two sets S1
and S2 we have a conditional collision probability similar
to minwise hashing.
H
H
Let Ei = 1 hOP
(S2 ) = hOP
(S2 ) = E
i
i
(7)
Pr
=
Ei = 0 = R
(8)
H
H
However, P r hOP
(S1 ) = hOP
(S2 ) Ei = 1 6= R
i
i
(9)
H
hOP
(S1 )
i
H
hOP
(S2 )
i
Here Ei is an indicator random variable of the event that
the ith partition corresponding to both S1 and S2 are empty.
See Figure 1
Any bin has a constant chance of being empty. Thus, there
is a positive probability of the event {Ei = 1}, for any
given pair S1 and S2 and hence for large datasets (big n)
a constant fraction of data will consist of simultaneously
empty bins (there are n2 × k trials for the bad event {Ei =
1} to happen). This fraction further increases significantly
with the sparsity of the data and k, as both sparsity and k
increases the probability of the bad event {Ei = 1}. See
Table 4 for statistics of empty bins on real scenarios.
Unfortunately, whenever the outcome of the random permutation leads to simultaneous empty bins, i.e. event
Ei = 1, the LSH Property is not valid. In fact, there is not
sufficient information present in the simultaneous empty
partitions for any meaningful statistics. Hence, one permutation hashing cannot be used as an LSH. Simple heuristics
of handling empty bins as suggested in (Shrivastava & Li,
2014a) leads to a significant bias and it was shown both theoretically and empirically that this bias leads to significant
deviation from the expected behavior of one permutation
hashing when compared with minwise hashing. Thus, one
permutation hashing although computationally lucrative is
not a suitable replacement for minwise hashing.
3.3. The Idea of Densification
In (Shrivastava & Li, 2014a), the authors proposed “densification” or reassignment of values to empty bins by
reusing the information in the non-empty bins to fix the
bias of one permutation hashing. The overall procedure is
quite simple. Any empty bin borrows the values of the clos-
Optimal Densification for Fast and Accurate Minwise Hashing
𝑺𝟏 = {𝟏𝟎, 𝟑, 𝟏𝟖, 𝟏, 𝟐𝟏, 𝟐, 𝟏𝟐, 𝟐𝟐} , 𝑺𝟐 = {𝟏𝟎, 𝟏𝟓, 𝟑, 𝟔, 𝟏𝟖, 𝟐𝟏, 𝟕}
𝑺𝟏 , 𝑺𝟐 ∈ 𝛀 = {𝟏, 𝟐, 𝟑, … , 𝟐𝟑}. 𝑳𝒆𝒕 𝑹𝒂𝒏𝒅𝒐𝒎 𝝅: 𝛀 → 𝛀 leads to
𝛑(𝟏𝟎) = 𝟓, 𝝅(𝟏𝟓) = 6, 𝝅(𝟑) = 𝟕, 𝝅(𝟔) = 𝟏𝟐, 𝝅(𝟏𝟖) = 𝟏𝟒,
𝝅(𝟏) = 𝟏𝟓, 𝝅(𝟐𝟏) = 𝟏𝟔, 𝝅(𝟕) =17, 𝝅(𝟐) = 𝟏𝟖, 𝝅(𝟏𝟐) = 𝟐𝟏,
𝝅(𝟐𝟐) = 𝟐𝟐
𝝅(𝑺𝟏 ) = {𝟓, 𝟕, 𝟏𝟒, 𝟏𝟓, 𝟏𝟔, 𝟏𝟖, 𝟐𝟏, 𝟐𝟐}
𝒎𝒊𝒏(𝝅(𝑺𝟏 )) = 𝟓
𝝅(𝑺𝟐 ) = {𝟓, 𝟔, 𝟕, 𝟏𝟐, 𝟏𝟒, 𝟏𝟔, 𝟏𝟕}
𝒎𝒊𝒏(𝝅(𝑺𝟐 )) = 𝟓
Partition Range 𝛀 into 6 Bins: [0,3],[4,7], [8,11],[12,15],[16,19],[20,23]
One Permutation Hashing: GET MIN IN BIN (E if EMPTY BIN)
𝒉𝑶𝑷𝑯 (𝑺𝟏 ) =
𝒉
𝑶𝑷𝑯 (𝑺
𝟐)
=
𝑬,
𝟓,
𝑬,
𝟏𝟒,
𝟏𝟔,
𝟐𝟏
𝑬,
𝟓,
𝑬,
𝟏𝟐,
𝟏𝟔,
𝑬
Densification: REASSIGN FROM RIGHT (CIRCULAR)
𝒉(𝑺𝟏 ) =
𝟓,
𝟓,
𝟏𝟒,
𝟏𝟒,
𝟏𝟔,
𝟐𝟏
𝒉(𝑺𝟐 ) =
𝟓,
𝟓,
𝟏𝟐,
𝟏𝟐,
𝟏𝟔,
𝟓
Improved Densification: REASSIGN FROM LEFT OR RIGHT
(CIRCULAR) DEPENDING ON RANDOM BITS 0/1.
randbits =
0
1
1
0
1
0
=
𝟐𝟏,
𝟓,
𝟏𝟒,
𝟏𝟒,
𝟏𝟔,
𝟐𝟏
𝒉+ (𝑺𝟐 ) =
𝟏𝟔,
𝟓,
𝟏𝟐,
𝟏𝟐,
𝟏𝟔,
𝟏𝟔
𝒉
+ (𝑺
𝟏)
Figure 1. Illustration of One Permutaion Hashing (OPH) and
the two existing Densification Schemes of (Shrivastava & Li,
2014a;c). Densification simply borrows the value from nearby empty bins. Different sets Si with have different pattern of
empty/nonempty bins. Since the pattern is random, the process is
similar to random re-use of unbiased values, which satisfies LSH
property in each bin (Shrivastava & Li, 2014c).
est non-empty bins towards the circular right (or left)1 . See
Figure 1 for an illustration. Since the positions of empty
and non-empty bins were random, it was shown that densification (or reassignment) was equivalent to a stochastic
reselection of one hash from a set of existing informative
(coming from non-empty bins) hashes which have the LSH
property. This kind of reassignment restores the LSH property and collision probability for any two hashes, after reassignment, is exactly same as that of minwise hashing.
The densification generates k hashes with the required LSH
Property and only requires two passes over the one permutation sketches making the total cost of one permutation
1
In (Shrivastava & Li, 2014a) they also needed an offset because the value of a hash in any bin was always reset between
[0,k]. We do not need the offset if we use the actual values of
π(S).
hashing plus densification O(d + k). This was a significant
improvement over O(dk) with classical minwise hashing.
O(d + k) led to an algorithmic improvement over randomized algorithms relying on minwise hashing, as hash computation cost is bottleneck step in all of them.
3.4. Lack of Randomness in Densification
It was pointed out in (Shrivastava & Li, 2014c) that the
densification scheme of (Shrivastava & Li, 2014a) has unnecessarily high variance. In particular, the probability of
two empty bins borrowing the information of the same nonempty bin was significantly higher. This probability was
due to poor randomization (load balancing) which hurts the
variance. (Shrivastava & Li, 2014c) showed that infusing
more randomness in the reassignment process by utilizing
k extra random bits provably improves the variance. See
Figure 1 for an example illustration of the method. The running time of the improved scheme was again O(d + k) for
computing k hashes. This improvement retains the required
LSH property, however this time with improved variance.
An improved variance led to significant savings in the task
of near-neighbor search on real sparse datasets.
4. Issues with Current Densification
Our careful analysis reveals that the variance, even with the
improved scheme, is still significantly higher. Worse, even
in the extreme case when we take k → ∞ the variance
converges to a positive constant rather than zero, which implies that even with infinite samples, the variance will not
be zero. This positive limit further increases with the sparsity of the dataset. In particular, we have the following theorem about the limiting variances of existing techniques:
Theorem 1 Give any two finite sets S1 , S2 ∈ Ω, with A =
|S1 ∪ S2 | > a = |S1 ∩ S2 | > 0 and |Ω| = D → ∞. The
limiting variance of the estimators from densification and
improved densification when k = D → ∞ is given by:
A−a
a
> 0 (10)
lim V ar(h) =
k→∞
A A(A + 1)
a 3(A − 1) + (2A − 1)(a − 1)
a
>0
lim V ar(h+ ) =
−
k→∞
A
2(A + 1)(A − 1)
A
(11)
This convergence of variance to a constant value, despite
infinite samples, of the existing densification is also evident in our experimental findings (see Figure 3) where we
observe that the MSE (Mean Square Error) curves go flat
with increasing k. Similar phenomenon was also reported
in (Shrivastava & Li, 2014c). It should be noted that for
classical minwise hashing the variance is R(1−R)
→ 0 for
k
any pair S1 and S2 . Thus, current densification, although
fast, loses significantly in terms of accuracy. We remove
this issue with densification. In particular, we show in The-
Optimal Densification for Fast and Accurate Minwise Hashing
Example of Poor Load Balancing
𝒉𝑶𝑷𝑯 (𝑺𝟏 ) =
𝟓,
𝑬,
𝑬,
𝑬,
𝟏𝟔,
𝟐𝟏,
𝟐𝟓
𝒉𝑶𝑷𝑯 (𝑺𝟐 ) =
𝟓,
𝑬,
𝑬,
𝑬,
𝟏𝟕,
𝟐𝟏,
𝟐𝟓
Densification: ONLY 1 INFORMATIVE VALUE USED
𝒉(𝑺𝟏 ) =
𝟓, 𝟏𝟔, 𝟏𝟔, 𝟏𝟔,
𝟏𝟔,
𝟐𝟏,
𝟐𝟓
𝒉(𝑺𝟐 ) =
𝟓, 𝟏𝟕, 𝟏𝟕, 𝟏𝟕,
𝟏𝟕,
𝟐𝟏,
𝟐𝟓
Improved Densification: ONLY 2 INFORMATIVE VALUES USED
randbits =
𝒉(𝑺𝟏 ) =
0
0 1
0
𝟓, 𝟏𝟔, 𝟓, 𝟏𝟔,
1
𝟏𝟔,
0
𝟐𝟏,
1
𝟐𝟓
𝒉(𝑺𝟐 ) =
𝟓, 𝟏𝟕, 𝟓, 𝟏𝟕,
𝟏𝟕,
𝟐𝟏,
𝟐𝟓
Figure 2. Illustration of Poor Load Balancing in the existing Densification strategies. The re-assignment of bins is very local and
its not uniformly distributed. Thus, we do not use the information
in far off non-empty bins while densification.
orem 6 that the limiting variance of the proposed optimal
densification goes to 0. Our experimental findings suggests
that the new variance is very close to the classical minwise hashing. In addition, the new densification retains the
speed of existing densified hashing thereby achieving the
best of the both worlds.
5. Optimal Densification
We argue that even with the improved densification there
is not enough randomness (or load balancing) in the reassignment process which leads to reduced variance.
For given set S, the densification process reassigns every
empty bin with a value from one of the existing non-empty
bins. Note, the identity of empty and non-empty bins are
different for different sets. To ensure the LSH property,
the re-assignment should be consistent for any given pair
of sets S1 and S2 . In particular, as noted in (Shrivastava
& Li, 2014a), given any arbitrary pair S1 and S2 , whenever any given bin i is simultaneously empty, i.e. Ei = 1,
the reassignment of this bin i should mimic the collision
probability of one of the simultaneously non-empty bin j
with Ej = 0. An arbitrary reassignment (or borrow) of
values will not ensure this consistency across all pairs. We
would like to point out that the reassignment of S1 has no
idea about S2 or any other object in the dataset. Thus, ensuring the consistency is non-trivial. Although the current
densification schemes achieve this consistency by selecting
the nearest non-empty bin (as shown in (Shrivastava & Li,
2014c)), they lack sufficient randomness.
5.1. Intuition: Load Balancing
In Figure 2, observe that if there are many contiguous
non-empty bins (Bins 2, 3 and 4), then with densification
schemes h, all of them are forced to borrow values from the
same non-empty bin (Bin 5 in the example). Even though
there are other informative bins (Bins 1, 6 and 7), their information is never used. This local bias increases the probability (p) that two empty bins get tied to the same information, even if there are many other informative non-empty
bins. Adding k random bits improves this to some extent by
allowing load sharing between the two ends instead of one
(Bins 1 and 5 instead of just 5). However, the load balancing is far from optimal. The locality of information sharing is the main culprit with current densification schemes.
Note, the poor load balancing does not change the expectation but affects the variance significantly.
For any given pairs of vectors S1 and S2 , let m be the
number of simultaneous non-empty bins (out of k), i.e.
Pk
i=1 Ei = k − m. Note, m is a random variable whose
value is different for every pair and depends on the outcome
of random π. Formally, the variance analysis of (Shrivastava & Li, 2014c) reveals that the probability that any two
simultaneous empty bin p and q (Ep = Eq = 1) reuses the
2
with the densification scheme h.
same information is m+1
1.5
This probability was reduced down to m+1
with h+ by utilizing k extra random bits to promote load balancing.
1.5
is not quite perfect load balancing. In a perp = m+1
fect load balancing with m simultaneous non-empty bins,
the probability of two empty bins hitting the same non1
empty bins is at best p = m
. Can we design a densification
scheme which achieves this p while maintaining the consistency of densification and at the same time does not hurt the
running time? It is not clear if such a scheme even exists.
We answer this question positively by constructing a densification method with precisely all the above requirements.
1
Furthermore we show that achieving p = m
is sufficient
for having the limiting variance of zero.
5.2. Simple 2-Universal Hashing Doesn’t Help
To break the locality of the information reuse and allow a
non-empty bin to borrow information from any other far
off bin consistently, it seems natural to use universal hashing. The hope is to have a 2-universal hash function (Section 3.1) huniv : [k] → [k]. Whenever a bin i is empty,
instead of borrowing information from neighbors, borrow
information from bin huniv (i). The hash function allows
consistency across any two S1 and S2 hence preserves LSH
property. The value of huniv (i) is uniformly distributed, so
any bin is equally likely. Thus, it seems to break the locality on the first thought. If huniv (i) is also empty then we
continue using huniv (huniv (i)) until we reach a non-empty
bins whose value we re-use.
One issue is that of cycles. If i = huniv (i) (which
has k1 chance), then this creates a cycle and the assign-
Optimal Densification for Fast and Accurate Minwise Hashing
ment will go into infinite loop. A cycle can even occur if
huniv (huniv (i)) = i with both i and huniv (i) being empty.
Note, the process runs until it finds a non-empty bin. However, cycles are not just our concern. Even if we manage
to get away with cycles, this scheme does not provide the
required load balancing.
i which is independent of the huniv (j, attemptj ) due to
2-universality of the hash function huniv . Thus, the probability that any two empty bins reuse the information of the
1
same non-empty bin is m
A careful inspection reveals that there is a very significant
chance that both i and huniv (i) to be empty for any given
set S. Observe that if i and huniv (i) are both empty, then
we are bound to reuse the information of the same nonempty bin for both empty bins i and huniv (i). We should
note that we have no control over the positions of empty
and non-empty bins. In fact, if no cycles happen then it
is not difficult to show that the simple assignment using
universal hashing is equivalent to the original densification
h with the order of bins reshuffled using huniv (.). It has
worse variance than h+ .
We denote the final k hashes generated by the proposed
densification scheme of Algorithm 1 using h∗ (* for optimality). Formally, with the optimal densification h∗ , we
have the following:
5.3. The Fix: Carefully Tailored 2-Universal Hashing
Algorithm 1 Optimal Densification
input k One Permutation Hashes hOP H [ ] of S.
input huniv (., .)
Initialize h∗ [ ] = 0
for i = 1 to k do
if OP H[i] 6= E then
h∗ [i] = hOP H [i]
else
attempt = 1
next = huniv (i, attempt)
while hOP H [next] 6= E do
attempt + +
next = huniv (i, attempt)
end while
h∗ [i] = hOP H [next]
end if
end for
RETURN h∗ [ ]
It turns out that there is a way to use universal hashing that
ensures no cycles as well as optimal load balancing. We
describe the complete process in Algorithm 1. The key is
to use a 2-universal hashing huniv : [k] × N → [k] which
takes two arguments: 1) The current bin id that needs to be
reassigned and 2) the number of failed attempt made so far
to reach a non-empty bin. This second argument ensures
no infinite loops as it changes with every attempt. So even
if we reach the same non-empty bin back (cycle), the next
time we will visit a new set of bins. Also, even if both i and
j = huniv (i, attempti ) are empty, i and j are not bound
to end to the same non-empty bin. This is because in the
next attempt we seek bin value huniv (i, attempti + 1) for
5.4. Analysis and Optimality
Theorem 2
|S1 ∩ S2 |
P r h∗ (S1 ) = h∗ (S2 ) =
=R
|S1 ∩ S2 |
RR̄
R
R
+ A 2 + B 2 − R2
V ar(h∗ ) =
k
k
k
lim V ar(h∗ ) = 0
k→∞
(12)
(13)
(14)
where Nemp is the number of simultaneous empty bins between S1 and S2 and the quantities A and B are given by
Nemp (Nemp − 1)
A = E 2Nemp +
k − Nemp
B = E (k − Nemp )(k − Nemp − 1) + 2Nemp (k − Nemp − 1)
Nemp (Nemp − 1)(k − Nemp − 1)
+
k − Nemp
Using the formula for P r(Nemp = i) from (Li et al., 2012),
we can precisely compute the theoretical variance. The interesting part is that we can formally show that the variance
of the proposed scheme is strictly superior compared to the
densification scheme with random bits improvements.
Theorem 3
V ar(h∗ ) ≤ V ar(h+ ) ≤ V ar(h)
(15)
Finally, due to optimal pairwise load balancing the variance
is the best possible we can hope with independent reassignments. Formally, we have
Theorem 4 Among all densification schemes, where the
reassignment process for bin i is independent of the reassignment process of any other bin j, Algorithm 1 achieves
the best possible variance.
Note: The variance can be reduced if we allow correlations
in the assignment process, for example if we force bin i and
bin j to not pick the same bin during reassignments, this
will reduce p beyond the perfectly random load balancing
1
. However, such tied reassignment will require
value of m
more memory and computations for generating structured
hash functions.
Optimal Densification for Fast and Accurate Minwise Hashing
Pair 1
Pair 2
Pair 3
Pair 4
Pair 5
Pair 6
Pair 7
Pair 8
Sim R
0.94
0.87
0.74
0.61
0.54
0.47
0.20
0.02
|S1 |
208
47
91
15
419
76
264
655
|S2 |
196
41
67
14
232
36
182
423
Table 1. Statistics of Pairs vectors (binary) with their Similarity
and Sparsity (non-zeros).
5.5. Running Time
We show that the expected running time of our proposal,
including all constants, is very similar to the running time
of the existing densification schemes.
Given set S with |S| = d, we are interested in computing
k hash values. The first step involves computing k one permutation hashes (or sketches) which only requires a single
pass over the elements of S. This takes max{d, k} ≤ d + k
time. Now the densification Algorithm 1 requires a for loop
of size k and within each for loop, if the bin is empty, it
requires an additional while loop. Let Nemp be the numk−Nemp
ber of empty bins, and therefore
is the probability
k
that the while loop will terminate in one iteration (next is
not empty). Therefore, the expected number of iteration
that each while loop will run is a binomial random variable
with expectation k−Nkemp . Thus, the expected running time
of the algorithm is given by
k
k − Nemp
Nemp
≤ d + (r + 2)k where r =
Nnot−emp
E[Running Time] = d + 2k − Nemp + Nemp
The quantity r, which is the ratio of number of empty bins
to the number of non-empty bins, is generally very small.
It is rarely more than 2 to 3 in practice. Observe that randomly throwing d items into k bins, the expected number
d
of empty bins is E[Nemp ] ≈ k(1 − k1 )d ≈ ke− k . Which
−d/k
e
makes r ≈ 1−e
−d/k . The number of sketches is usually of
the order of non-zeros. Even for very good concentration,
the size of the sketches k is rarely much larger than the size
of the set d. Even when k is 4 times d the value of r is
approximately 3.5. Thus, the quantity r is negligible.
It should be further noted that the implementation cost of
densification scheme h+ is d + 4k which in not very different from the cost of our proposal.
6. Evaluations
Our aim is to verify the theoretical claims of these papers
empirically. We show that our proposal can replace minwise hashing for all practical purposes. To establish that,
we focus on experiments with the following objectives:
1. Verify that our proposed scheme has a significantly
better accuracy (variance) than the existing densification schemes. Validate our variance formulas.
2. Empirically quantify the impact of optimal variance
in practice. How does this quantification change with
similarity and sparsity? Verify that the proposal has
accuracy close to vanilla minwise hashing.
3. Verify that there is no impact on running time of the
proposed scheme over existing densification schemes,
and our proposal is significantly faster than vanilla
minwise hashing. Understand how the running time
changes with change in sparsity and k?
6.1. Accuracy
For objectives 1 and 2, we selected 9 different word pairs
embedding, generated from new20 corpus, with varying
level of similarity and sparsity. We use the popular termdocument vector representation for each word. The statistics of these word vector pairs are summarized in Table 1
For each word pairs, we generated k hashes using three
different schemes: 1) Densification h, 2) Improved Densification h+ and the proposed densification h∗ (Algorithm 1). Using these hashes, we estimate the Jaccard similarity (Equation 4). We plot the mean square error (MSE)
with varying the number of hashes. Since the process is
randomized, we repeat the process 5000 times, for every k,
and report the average over independent runs. We report all
integer values of k in the interval [1, 214 ].
It should be noted that since all three schemes have the
LSH Property, the bias is zero and hence the MSE is the
theoretical variance. To validate our variance formula, we
also compute and plot the theoretical value of the variance
(Equation 19) of the optimal scheme. Also, to understand
how all these fast methodologies compare with the accuracy of vanilla minwise hashing we also plot the theoretical
.
variance of minwise hashing which is R(1−R)
k
From the results in Figure 3, we can conclude.
Conclusion 1: The proposed densification is significantly
more accurate, irrespective of the choice of sparsity and
similarity, than the existing densification schemes especially for large k. Note the y-axis of plots is on log scale,
so the accuracy gains are drastic.
Conclusion 2: The gains with optimal densification is
more for sparse data.
Conclusion 3: The accuracy of optimal densification is
very close the accuracy of costly minwise hashing.
Optimal Densification for Fast and Accurate Minwise Hashing
10
-4
10
-5
Pair 1: R=0.94
10 -6
10 0 10 1 10 2
10
3
10
4
10
10
10 -2
10
-3
10
-4
10
-5
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
10 -6
10 0
5
Num of Hashes (k)
10 0
10
10 -2
MSE
MSE
10
10 -3
10 -4
10 -5
10
-6
10 0
10 2
10
10
-4
10
-5
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
10 -6
10 0
5
10 4
10 5
10 -2
10 -3
10
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
Num of Hashes (k)
10 0
10 1
10 2
10
10
3
10
4
10
-5
10 -6
10 0
5
10 5
10 -3
10
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
Num of Hashes (k)
10 1
10 2
10 1
10 2
10 3
10 4
10 5
10 0
Pair 8: R=0.08
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
10 -1
10 -2
10 -3
10 -4
10 -5
-6
10 0
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
Num of Hashes (k)
10 -2
10 -5
10 4
10 -3
10 -4
Pair 7: R=0.2
-1
10 -4
10 3
10
2
10 0
10
-6
10
1
10 -2
Num of Hashes (k)
Pair 6: R=0.47
10 -5
10 3
10
4
-1
10 -4
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
10 1
10
3
10
-3
Pair 4: R=0.61
10 -1
10 -2
Num of Hashes (k)
Pair 5: R=0.54
-1
10
2
MSE
10 0
10
1
10 0
Pair 3: R=0.74
-1
MSE
10
-3
10 0
Pair 2: R=0.87
-1
MSE
10 -2
10
MSE
MSE
10
10 0
Vanilla
Improved
Proposed
Prop-Theory
MinHash-Theory
-1
MSE
10 0
10 3
10 4
10 5
10 -6
10 0
Num of Hashes (k)
10 1
10 2
10 3
10 4
10 5
Num of Hashes (k)
Figure 3. Average MSE in Jaccard Similarity Estimation with the Number of Hash Values (k). Estimates are averaged over 5000
repetitions. With Optimal densification the variance is very close to costly minwise hashing, which is significantly superior to existing
densification schemes. Note the log scale of y-axis.
RCV1
URL
NEWS20
Existing h+
100 200 300
62
86
104
373 584 624
154 179 205
This Paper h∗
100 200 300
56
90
116
287 459 631
139 158 174
Vanilla MinHash
100
200
300
376
733
1098
2760 6247 8158
1201 2404 3650
Table 2. Time (in milliseconds) requires to compute 100, 200 and 300 hashes of the full data using the existing densification, proposed
densification and vanilla minwise hashing. Minhash can be 10-18x slower for computing 300 hashes on these datasets.
RCV1
URL
News20
Avg. non-zeros (d)
73
115
402
Dim (D)
47,236
3,231,961
1,355,191
Samples
20,242
100,000
19,996
Table 3. Basic Statistics of Datasets.
Conclusion 4: The theoretical variance of our proposal
overlaps with the empirical estimates, and it seems to go
to zero validating Theorem 19.
Conclusion 5: The variances (or the MSE) of existing densification seems to converge to constant and do not go to
zero confirming Theorem 5.
6.2. Speed
RCV1
URL
News20
100 Bins
53
33
14
200 Bins
143
112
58
300 Bins
238
202
120
Table 4. Avg. Number of Empty Bins per vector (rounded) generated with One Permutation Hashing (Li et al., 2012). For sparse
datasets, a significant (80% with 300 hashes) of the bins can be
empty and have no information. They cannot be used for indexing
and kernel learning. Fortunately, we can efficiently densify them
with optimal densification, such that, all the generated k hashes
have variance similar to minwise hashing. The overall computational cost is significantly less compared to minwise hashing
To compute the runtime, we use three publicly available
text datasets: 1) RCV1, 2) URL and 3) News20. The dimensionality and sparsity of these datasets are an excellent representative of the scale and the size frequently encountered in large-scale data processing systems, such as
Google’s SIBYL (Chandra et al., 2010). The statistics of
these datasets are summarized in Table 3.
We implemented three methodologies for computing
hashes: 1) Densification Scheme h+ , 2) The Proposed h∗
(Algorithm 1 and 3) Vanilla Minwise Hashing. The methods were implemented in C++. Cheap hash function replaced costly permutations. Clever alternatives to avoid
mod operations were employed. These tricks ensured that
our implementations2 are as efficient as the possible. We
compute the wall clock time required to calculate 100,
200 and 300 hashes of all the three datasets. The time
include the end-to-end hash computation of the complete
data. Data loading time is not included. The results are
presented in Table 2. All the experiments were done on
Intel i7-6500U processor laptop with 16GB RAM.
Also, to get an estimate of the importance of densification,
we also show the average number of empty bins generated
2
Codes are available at http://rush.rice.edu/
fastest-minwise.html
Optimal Densification for Fast and Accurate Minwise Hashing
by only using one permutation hashing and report the numbers in Table 4. We can clearly see that the number of
empty bins is significantly larger and the hashes are unusable without densification.
Chierichetti, Flavio, Kumar, Ravi, Lattanzi, Silvio, Mitzenmacher, Michael, Panconesi, Alessandro, and Raghavan,
Prabhakar. On compressing social networks. In KDD,
pp. 219–228, Paris, France, 2009.
From the running time numbers Table 2, we conclude:
Dahlgaard, Søren, Knudsen, Mathias Bæk Tejs, Rotenberg,
Eva, and Thorup, Mikkel. Hashing for statistics over kpartitions. In Foundations of Computer Science (FOCS),
2015 IEEE 56th Annual Symposium on, pp. 1292–1310.
IEEE, 2015.
Conclusion 1: Optimal densification is as fast as traditional densification irrespective of k and the sparsity. However, optimal densification is significantly more accurate.
Conclusion 2: Both the densification scheme is significantly faster than minwise hashing. They are 10-18x faster
for computing 300 hashes on the selected datasets.
Given the simplicity, we hope our work gets adopted.
References
Bavarian, Mohammad, Ghazi, Badih, Haramaty, Elad,
Kamath, Pritish, Rivest, Ronald L., and Sudan,
Madhu. The optimality of correlated sampling. CoRR,
abs/1612.01041, 2016. URL http://arxiv.org/
abs/1612.01041.
Bayardo, Roberto J., Ma, Yiming, and Srikant, Ramakrishnan. Scaling up all pairs similarity search. In WWW, pp.
131–140, 2007.
Broder, Andrei Z. On the resemblance and containment
of documents. In the Compression and Complexity of
Sequences, pp. 21–29, Positano, Italy, 1997.
Broder, Andrei Z., Charikar, Moses, Frieze, Alan M., and
Mitzenmacher, Michael. Min-wise independent permutations. In STOC, pp. 327–336, Dallas, TX, 1998.
Buehrer, Gregory and Chellapilla, Kumar. A scalable pattern mining approach to web graph compression with
communities. In WSDM, pp. 95–106, Stanford, CA,
2008.
Carter, J. Lawrence and Wegman, Mark N. Universal
classes of hash functions. In STOC, pp. 106–112, 1977.
Fetterly, Dennis, Manasse, Mark, Najork, Marc, and
Wiener, Janet L. A large-scale study of the evolution of
web pages. In WWW, pp. 669–678, Budapest, Hungary,
2003.
Henzinger, Monika Rauch. Finding near-duplicate web
pages: a large-scale evaluation of algorithms. In SIGIR,
pp. 284–291, 2006.
Indyk, Piotr and Motwani, Rajeev. Approximate nearest
neighbors: Towards removing the curse of dimensionality. In STOC, pp. 604–613, Dallas, TX, 1998.
Li, Ping. 0-bit consistent weighted sampling. In KDD,
2015.
Li, Ping, Shrivastava, Anshumali, Moore, Joshua, and
König, Arnd Christian. Hashing algorithms for largescale learning. In NIPS, Granada, Spain, 2011.
Li, Ping, Owen, Art B, and Zhang, Cun-Hui. One permutation hashing. In NIPS, Lake Tahoe, NV, 2012.
Najork, Marc, Gollapudi, Sreenivas, and Panigrahy, Rina.
Less is more: sampling the neighborhood graph makes
salsa better and faster.
In WSDM, pp. 242–251,
Barcelona, Spain, 2009.
Ondov, Brian D, Treangen, Todd J, Melsted, Páll,
Mallonee, Adam B, Bergman, Nicholas H, Koren,
Sergey, and Phillippy, Adam M. Mash: fast genome
and metagenome distance estimation using minhash.
Genome Biology, 17(1):132, 2016.
Chandra, Tushar, Ie, Eugene, Goldman, Kenneth, Llinares,
Tomas Lloret, McFadden, Jim, Pereira, Fernando, Redstone, Joshua, Shaked, Tal, and Singer, Yoram. Sibyl: a
system for large scale machine learning. Technical report, 2010.
Shrivastava, Anshumali and Li, Ping. Densifying one
permutation hashing via rotation for fast near neighbor
search. In ICML, Beijing, China, 2014a.
Charikar, Moses S. Similarity estimation techniques from
rounding algorithms. In STOC, pp. 380–388, Montreal,
Quebec, Canada, 2002.
Shrivastava, Anshumali and Li, Ping. In defense of minhash over simhash. In Proceedings of the Seventeenth
International Conference on Artificial Intelligence and
Statistics, pp. 886–894, 2014b.
Chien, Steve and Immorlica, Nicole. Semantic similarity
between search engine queries using temporal correlation. In WWW, pp. 2–11, 2005.
Shrivastava, Anshumali and Li, Ping. Improved densification of one permutation hashing. In UAI, Quebec, CA,
2014c.
Optimal Densification for Fast and Accurate Minwise Hashing
Weinberger, Kilian, Dasgupta, Anirban, Langford, John,
Smola, Alex, and Attenberg, Josh. Feature hashing for
large scale multitask learning. In ICML, pp. 1113–1120,
2009.
The variance is little involved. From the collision probability, we have the following is unbiased estimator.
k−1
1X
1{h∗j (S1 ) = h∗j (S2 )}.
R̂ =
k j=0
(21)
For variance, define the number of simultaneously empty
bins by
A. Proofs
Theorem 5 Give any two finite sets S1 , S2 ∈ Ω, with A =
|S1 ∪ S2 | > a = |S1 ∩ S2 | > 0 and |Ω| = D → ∞. The
limiting variance of the estimators from densification and
improved densification when k = D → ∞ is given by:
a
A−a
lim V ar(h) =
> 0 (16)
k→∞
A A(A + 1)
a
a 3(A − 1) + (2A − 1)(a − 1)
−
>0
lim V ar(h+ ) =
k→∞
A
2(A + 1)(A − 1)
A
(17)
Nemp =
k−1
X
1{Ej = 1},
(22)
j=0
where 1 is the indicator
function. We partition the event
h∗j (S1 ) = h∗j (S2 ) into two cases depending on Ej . Let
MjN (Non-empty Match at j) and MjE (Empty Match at
j) be the events defined as:
MjN = 1{Ej = 0 and h∗j (S1 ) = h∗j (S2 )}
MjE
= 1{Ej = 1 and
h∗j (S1 )
=
h∗j (S2 )}
(23)
(24)
Proof: When k = D, then Nemp = D − A. Substituting
this value in the variance formulas from (Shrivastava & Li,
2014c) and taking the limit as D = k → ∞, we get the
a
above expression after manipulation. When 0 < R = A
<
1, they both are strictly positive.
Note that, MjN = 1 =⇒ MjE = 0 and MjE = 1 =⇒
MjN = 0. From the LSH property of estimator we have
Theorem 6
It is not difficult to show that,
|S1 ∩ S2 |
=R
P r h (S1 ) = h (S2 ) =
|S1 ∩ S2 |
R
R
RR̄
V ar(h∗ ) =
+ A 2 + B 2 − R2
k
k
k
lim V ar(h∗ ) = 0
∗
k→∞
∗
(18)
E(MjN |Ej = 0) = E(MjE |Ej = 1)
= E(MjE + MjN ) = R ∀j
E MjN MiN i 6= j, Ej = 0 and Ei = 0 = RR̃,
where R̃ =
a−1
f 1+f 2−a−1 .
(19)
R̂ =
(20)
where Nemp is the number of simultaneous empty bins between S1 and S2 and the quantities A and B are given by
Nemp (Nemp − 1)
A = E 2Nemp +
k − Nemp
B = E (k − Nemp )(k − Nemp − 1) + 2Nemp (k − Nemp − 1)
Nemp (Nemp − 1)(k − Nemp − 1)
+
k − Nemp
Proof:
The collision probability is easy using a simple observation that values coming from different bin numbers can
never match across S1 and S2 , i.e. h∗i (Si ) 6= h∗j (S2 ) if
i 6= j, as they have disjoint different range. So whenever, for a simultaneous empty bin i, i.e. Ei = 1, we get
h∗i (S1 ) = h∗i (S2 ) after reassignment, the value must be
coming from same non-empty bin, say numbers k which is
not not empty. Thus,
P r(h∗i (S1 ) = h∗i (S2 )) = P r(h∗k (S1 ) = h∗k (S2 )|Ek = 0) = R
(25)
Using these new events, we have
k−1
1 X E
Mj + MjN
k j=0
(26)
We are interested in computing
2
k−1
X
1
V ar(R̂) = E
MjE + MjN − R2 (27)
k j=0
For notational convenience we will use m to denote the
event k − Nemp = m, i.e., the expression E(.|m) means
E(.|k − Nemp = m). To simplify the analysis, we will first
compute the conditional expectation
2
k−1
1 X E
Mj + MjN m
(28)
f (m) = E
k j=0
By expansion and linearity of expectation, we obtain
X
X
k 2 f (m) = E
MiN MjN m + E
MiN MjE m
i6=j
i6=j
#
" k
X
X
E
N
2
E
2
E
+E
Mi Mj m + E
(Mj ) + (Mj ) m
i6=j
i=1
Optimal Densification for Fast and Accurate Minwise Hashing
MjN = (MjN )2 and MjE = (MjE )2 as they are indicator
functions and can only take values 0 and 1. Hence,
k−1
X
E
(MjN )2 + (MjE )2 m = kR
(29)
Under any independent re-assigment, the probability that
two empty bins chooses the same non-empty bin out of m
1
non-empty bins is lower bounded by m
which is achieved
by optimal densification.
j=0
The values of the first three terms are given by the following 3 expression using simple binomial enpension and using the fact that we are dealing with indicator random variable which can only take values 0 or 1.
X
(30)
E
MiN MjN m = m(m − 1)RR̃
i6=j
"
#
X
R
(m
−
1)R
R̃
E
MiN MjE m = 2m(k − m)
+
m
m
i6=j
(31)
Let p be the probability that two simultaneously empty bins
i and j finally picks the same non-empty bin for reassignment. Then we have
h
i
X
E
MiE MjE m = (k − m)(k − m − 1) pR + (1 − p)RR̃
i6=j
(32)
because with probability (1 − p), it uses estimators from
different simultaneous non-empty bin and in that case the
MiE MjE = 1 with probability RR̃. We know that Algo1
rithm 1 which uses 2-universal hashing the value of p = m
.
This is because any pairwise assignment is perfectly random with 2-universal hashing.
Substituting for all terms with value of p and rearranging
terms gives the required expression.
When k = D, then Nemp = D − A. Substituting this value
in the variance formulas and taking the limit as D = k →
∞, we get 0 for all R.
Theorem 7
V ar(h∗ ) ≤ V ar(h+ ) ≤ V ar(h)
(33)
1
1.5
2
Proof: We have p∗ = m
≤ p+ = m+1
≤ p = m+1
. The
+
value of p and p comes from analysis in (Shrivastava &
Li, 2014c)
Theorem 8 Among all densification schemes, where the
reassignment process for bin i is independent of the reassignment process of any other bin j, Algorithm 1 achieves
the best possible variance.
| 8cs.DS
|
arXiv:1804.00102v2 [stat.ME] 5 Apr 2018
Collaborative targeted inference from continuously
indexed nuisance parameter estimators
Cheng Ju
Division of Biostatistics, UC Berkeley
and
Antoine Chambaz
MAP5 (UMR 8145), Université Paris Descartes
and
Mark J. van der Laan
Division of Biostatistics, UC Berkeley
April 9, 2018
Abstract
Suppose that we wish to infer the value of a statistical parameter at a law from which we
sample independent observations. Suppose that this parameter is smooth and that we can
define two variation-independent, infinite-dimensional features of the law, its so called Q- and
G-components (comp.), such that if we estimate them consistently at a fast enough product of
rates, then we can build a confidence interval (CI) with a given asymptotic level based on a
plain targeted minimum loss estimator (TMLE). The estimators of the Q- and G-comp. would
typically be by products of machine learning algorithms. We focus on the case that the machine
learning algorithm for the G-comp. is fine-tuned by a real-valued parameter h. Then, a plain
TMLE with an h chosen by cross-validation would typically not lend itself to the construction
of a CI, because the selection of h would trade-off its empirical bias with something akin to
the empirical variance of the estimator of the G-comp. as opposed to that of the TMLE. A
collaborative TMLE (C-TMLE) might, however, succeed in achieving the relevant trade-off. We
prove that this is the case indeed.
We construct a C-TMLE and show that, under high-level empirical processes conditions,
and if there exists an oracle h that makes a bulky remainder term asymptotically Gaussian,
then the C-TMLE is asymptotically Gaussian hence amenable to building a CI provided that
its asymptotic variance can be estimated too. The construction hinges on guaranteeing that
an additional, well chosen estimating equation is solved on top of the estimating equation that
a plain TMLE solves. The optimal h is chosen by cross-validating an empirical criterion that
guarantees the wished trade-off between empirical bias and variance.
We illustrate the construction and main result with the inference of the so called average
treatment effect, where the Q-comp. consists in a marginal law and a conditional expectation,
and the G-comp. is a propensity score (a conditional probability). We also conduct a multifaceted simulation study to investigate the empirical properties of the collaborative TMLE when
the G-comp. is estimated by the LASSO. Here, h is the bound on the `1 -norm of the candidate
coefficients. The variety of scenarios shed light on small and moderate sample properties, in the
face of low-, moderate- or high-dimensional baseline covariates, and possibly positivity violation.
Keywords: cross-validation, empirical process theory, semiparametric models
1
1
Introduction
We wish to infer the value of a statistical parameter at a law from which we sample independent
observations. The parameter is a smooth function of the data distribution. We assume that we
can define two variation-independent, infinite-dimensional features of the law, its so called Qand G-components, such that if we estimate them consistently at a fast enough joint rate, then
we can build a confidence interval (CI) with a given asymptotic level based on a plain targeted
minimum loss estimator (TMLE) [30, 29]. Typically, the parameter depends on the law only
through its Q-component, whereas its canonical gradient depends on the law through both its Qand G-components. The estimators of the Q- and G-components would typically be by products
of machine learning algorithms. We focus on the case that the machine learning algorithm for the
G-component is fine-tuned by a real-valued parameter h. Is it possible to construct an estimator
that will lend itself to the construction of a CI, by fine-tuning data-adaptively and in a targeted
fashion both the algorithm for the estimation of the G-component and the resulting estimator of
the parameter of interest?
Literature overview. The general problem that we address is often encountered in observational
studies of the effect of an exposure, for instance when one wishes to infer the average effect of a
two-level exposure. It is then necessary to account for the fact that the level of exposure is not
fully randomized in the observed population. A pivotal object of interest in such studies, the so
called exposure mechanism (that is, the conditional law of exposure given baseline covariates) is an
example of what we generally call a G-component of the law of the experiment.
A wide range of estimators of the average effect of a two-level exposure require the estimation
of the propensity score: Horvitz-Thompson estimators [9]; estimators based on propensity score
matching [23, 8, 7] or stratification [1, 24]; any estimator relying on the efficient influence curve,
among which double-robust inverse probability of exposure weighted estimators [20, 22, 18] or
estimators built based on the targeted minimum loss estimation (TMLE) methodology [30, 29].
Common methods for the estimation of the propensity score are multivariate logistic regression [14], high-dimensional propensity score adjustment [25, 2], and a variety of machine learning
algorithms [15, 6, 10]. Except in the so called collaborative variant of TMLE that we will discuss
shortly, the estimators of the propensity score can be derived at a preliminary step, regardless essentially of why they are needed and how they are used at the subsequent step. This is problematic
because optimality at the preliminary step has little if any relation to optimality at the subsequent
step. For instance, the optimal estimator of the propensity score at the preliminary step might take
values very close to zero, therefore disqualifying it as a viable estimator at the subsequent step, not
to mention an optimal one. In a less dramatic scenario, using an instrumental variable (which only
influences exposure but not the outcome) to estimate the propensity score could concomitantly
yield a better estimator thereof and only increase the variance of the resulting estimator of the
effect of exposure [31, 29].
This prompted the development of the so called collaborative version of the targeted minimum
loss estimation methodology [31, 29], where the estimation of the G-component is not separated
from that of the parameter of main interest anymore. More concretely, collaborative TMLE (CTMLE) consists in building a sequence of estimators of the G-component and in selecting one
of them by optimizing a criterion that targets the parameter of main interest. For instance, in
the above less dramatic scenario, covariates that are strongly predictive of exposure but not of
the outcome would be removed, resulting in less bias for the estimator of the parameter of main
interest.
The C-TMLE methodology has been adapted to a wide range of fields, including genomics [4, 34],
2
survival analysis [27], and clinical studies[11]. Because the derivation of C-TMLE estimators is often
computationally demanding, scalable versions have also been developed [11].
In [26], the authors propose a C-TMLE algorithm that uses regression shrinkage of the exposure
model for the estimation of the propensity score. It sequentially reduces the parameter that determines the amount of penalty placed on the size of the coefficient values, and selects the appropriate
parameter by cross-validation. The methodology for continuously fine-tuned, collaborative targeted
learning that we develop in this article encompasses the algorithm of [26]. Its statistical analysis
sheds light on why, and under which assumptions, it would provide valid statistical inference.
The present study builds upon [28]. The methodology is also studied in [12, 13], the latter an
example of real-life application.
At this point in the introduction, we wish to formalize what is the problem at stake. What
follows recasts the introductory paragraph in the theoretical framework that we adopt in the article.
Setting the scene. Let O1 , . . . , On be n independent draws from a law P0 on a set O. We view
P0 as an element of the statistical model M, a collection of plausible laws for O1 , . . . , On . The
more we know about P0 , the smaller is M. Our primary goal is to infer the value of parameter
Ψ : M → R at P0 , namely, ψ0 ≡ Ψ(P0 ). Our statistical analysis is asymptotic in the number of
observations.
We consider the case that Ψ is pathwise differentiable at every P ∈ M with respect to (w.r.t.)
a tangent set SP ⊂ L20 (P ): there exists D∗ (P ) ∈ L20 (P ) such that, for every s ∈ SP , there exists a
submodel {Pt : t ∈ R, |t| < c} ⊂ M satisfying (i) Pt |t=0 = P , (ii) Pt P for all t ∈] − c, c[, (iii)
d
dPt
log
(O)
dt
dP
= s(O)
t=0
(the submodel’s score function equals s), and (iv) the real valued mapping t 7→ Ψ(Pt ) is differentiable at t = 0 with a derivative equal to P D∗ (P )s, where P f is a shorthand notation for EP (f (O))
(any measurable f ). It is assumed moreover that every P ∈ M is associated with two possibly
infinite-dimensional features Q ∈ Q and G ∈ G such that (i) Q and G are unrelated (i.e., variation
independent: knowing anything about Q tells nothing about G and vice versa), (ii) Ψ(P ) depends
on P only through Q, (iii) D∗ (P ) depends on P only through Q and G, and (iv) G is a mapping
from O to R. At this early stage, we can introduce the pivotal
Rem20 (Q, G) ≡ Ψ(P ) − Ψ(P0 ) + P0 D∗ (P )
for every P ∈ M. The notation is justified (i) because we wish to think of the right-hand-side
expression as a remainder term, and (ii) by the fact that Ψ(P ) and D∗ (P ) depend on P only
through Q and G. We consider the case that parameter Ψ is such that, for some pseudo-distances
dQ and dG on Q and G,
|Rem20 (Q, G)| . dQ (Q, Q0 ) × dG (G, G0 ),
(1)
where a . b stand for “there exists a universal positive constant c > 0 such that a ≤ bc”. A
remainder term satisfying (1) is said double-robust.
Let Q̂ be an algorithm for the estimation of Q0 , the Q-component of the true law P0 . Likewise,
let Ĝh (h ∈ H, an open interval of R∗+ of which the closure contains 0) be an h-specific algorithm
for the estimation of G0 , the G-component of P0 . Formally, we view Q̂ and each Ĝh as mappings
from
(
)
N
[
X
N −1
Dirac(oi ) : o1 , . . . , oN ∈ O
N ≥1
i=1
3
to Q and G, respectively, that can “learn” from the empirical measure Pn some estimators Q̂(Pn )
and Ĝh (Pn ) of Q0 and G0 . Set Q0n ≡ Q̂(Pn ) (the superscript 0 stands for “initial”), Gn,h ≡ Ĝh (Pn ),
0 ∈ M be any element of the model of which the Q- and G-components equal Q0 and
and let Pn,h
n
0 ) is a natural estimator of
Gn,h . Derived by the mere substitution of Pn0 for P0 in Ψ(P0 ), Ψ(Pn,h
Ψ(P0 ). It is not targeted toward the inference of Ψ(P0 ) in the sense that none of the known features
0 was derived specifically for the sake of ultimately estimating Ψ(P ).
of Pn,h
0
0 ) toward Ψ(P ) is
It is well documented in the TMLE literature that one way to target Ψ(Pn,h
0
∗ ∈ M from P 0 in such a way that
to build Pn,h
n,h
√
∗
Pn D∗ (Pn,h
) = oP (1/ n)
∗ ). This can be achieved, in such a way that G
and to infer Ψ(P0 ) with Ψ(Pn,h
n,h is not modified,
0
by “fluctuating” Pn,h , a procedure that we will develop in details in the specific example studied
in the article. Then, by (1), the estimator satisfies the asymptotic expansion:
√
∗
∗
Ψ(Pn,h
) − Ψ(P0 ) = (Pn − P0 )D∗ (Pn,h
) + Rem20 (Q∗n,h , Gn,h ) + oP (1/ n).
(2)
By convention, we agree that small values of h correspond with less bias for Gn,h as an estimator
of G0 . Moreover, we assume that there exists hn ∈ H, hn = o(1), such that dG (Gn,hn , G0 ) = oP (ρ1,n )
for some ρ1,n = o(1), i.e., that Gn,hn consistently estimates G0 at rate ρ1,n . If Q∗n,hn is also such
√
that dQ (Q∗n,hn , Q0 ) = oP (ρ2,n ) for some ρ2,n = o(1), and if ρ1,n ρ2,n = o(1/ n), then (1) and (2)
yield
√
∗
∗
) + oP (1/ n)
) − Ψ(P0 ) = (Pn − P0 )D∗ (Pn,h
Ψ(Pn,h
n
n
which may in turn imply the asymptotic linear expansion
√
∗
Ψ(Pn,h
) − Ψ(P0 ) = (Pn − P0 )IF + oP (1/ n),
n
(3)
with influence function IF ≡ D∗ (P0 ), depending in particular on how data-adaptive are algorithms
√
∗
Q̂ and Ĝh (h ∈ H). By the central limit theorem, (3) guarantees that n(Ψ(Pn,h
) − Ψ(P0 )) is
n
asymptotically Gaussian.
√
We focus on a more challenging situation, where ρ1,n ρ2,n is not necessarily o(1/ n). We
anticipate that our analysis is also very relevant at small and moderate sample sizes when ρ1,n ρ2,n =
√
o(1/ n).
In order to derive an asymptotic linear expansion similar to (3) from (2) in this situation, we
would have to derive an asymptotic expansion of Rem20 (Q∗n,hn , Gn,hn ). Unfortunately, we have
reasons to believe that this is not possible without targeting (their presentation in an example is
deferred to Section 3.3).
∗ ) (h ∈ H) do not cooperate in the sense that, although
Now, observe that the estimators Ψ(Pn,h
Q∗n,h and Q∗n,h0 (for any two h, h0 ∈ H, h 6= h0 ) share the same initial estimator Q0n , the construction
of the latter does not capitalize on that of the former. In contrast, we propose to build collabo∗ ) (h ∈ H) and to select data-adaptively one
ratively a continuum of estimators of the form Ψ(Pn,h
among them that will be asymptotically Gaussian, under conditions often encountered in empirical
process theory.
Organization of the article. In Section 2, we lay out a high-level presentation of collaborative
TMLE, and state a high-level result. In Sections 3, 4, 5 and 6, we consider a specific example. In
Section 3, we particularize the theoretical construction and analysis. In Section 4, we describe two
practical instantiations of the estimator developed in Section 3. In Sections 5 and 6, we carry out
a mutli-faceted simulation study of their performances and comment upon its results. In Section 7,
we summarize the content of the article. All the proofs are gathered in the appendix.
4
2
High-level presentation and result
We now state and prove a general result about continuously fine-tuned, collaborative targeted
minimum loss estimation, a version of [Theorem 10.1 in 28]. Its high-level assumptions are clarified
in the particular example that we study in the next sections.
From now on, we slightly abuse notation and denote D∗ (Q, G) instead of D∗ (P ), where Q and
G are the Q- and G-components of P . Let G· ≡ {Gt : t ∈ T } ⊂ G be a (one-dimensional) subset
of G (indexed by a real parameter ranging in an open subset T of H) such that t 7→ D∗ (Q, Gt )(O)
is twice differentiable over T for all Q ∈ Q (P0 -almost surely). We characterize ∂D∗ and ∂ 2 D∗ by
setting, for every h ∈ T and Q ∈ Q,
∂h D∗ (Q, G· )(O) ≡
∂h2 D∗ (Q, G· )(O) ≡
d ∗
D (Q, Gt )(O)|t=h ,
dt
d2 ∗
D (Q, Gt )(O)|t=h .
dt2
(4)
Consider the following inter-dependent assumptions. The first one is indexed by (Q, h, c) ∈
Q × H × R∗+ .
A1(Q, h, c) There exists an open neighborhood T ⊂ H of h ∈ H for which the set Gn,· ≡ {Ĝh (Pn ) ≡
Gn,h : h ∈ T } ⊂ G is such that t 7→ D∗ (Q, Gn,· )(O) is twice differentiable over T (P0 -almost
surely). Moreover, P0 -almost surely,
sup |∂h2 D∗ (Q, Gn,· )(O)| ≤ c.
h∈T
∗ ∈ M, with Q- and G-components denoted by Q∗
A2 For all h ∈ H, we know how to build Pn,h
n,h
√
and Gn,h , in such a way that Pn D∗ (Q∗n,h , Gn,h ) = oP (1/ n). Moreover, we know how to
choose hn ∈ H such that
√
Pn D∗ (Q∗n,hn , Gn,hn ) = oP (1/ n).
(5)
and, for some deterministic c2 > 0, A1(Q∗n,hn , hn , c2 ) is met and
Pn ∂hn D∗ (Q∗n,hn , Gn,· ) = oP (1/n1/4 ).
(6)
A3 It holds that dG (Gn,hn , G0 ) = oP (1), and there exists Q1 ∈ Q such that dQ (Q∗n,hn , Q1 ) = oP (1).
In addition,
√
(Pn − P0 ) D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q1 , G0 ) = oP (1/ n),
(7)
√
Rem20 (Q∗n,hn , Gn,hn ) − Rem20 (Q1 , Gn,hn ) = oP (1/ n).
(8)
A4 Let Φ0 : G → R be given by Φ0 (G) ≡ P0 D∗ (Q1 , G). There exist h̃n ∈ H and ∆(P1 ) ∈ L20 (P0 )
such that
√
Φ0 (Gn,h̃n ) − Φ0 (G0 ) = (Pn − P0 )∆(P1 ) + oP (1/ n).
(9)
√
A5 It holds that (hn − h̃n )2 = oP (1/ n). Moreover, there exists a deterministic c5 > 0 such that
A1(Q1 , hn , c5 ) is met, and
√
(Pn − P0 ) D∗ (Q1 , Gn,hn ) − D∗ (Q1 , Gn,h̃n )
= oP (1/ n),
(10)
√
(hn − h̃n ) × P0 ∂hn D∗ (Q∗n,hn , Gn,· ) − ∂hn D∗ (Q1 , Gn,· ) = oP (1/ n),
(11)
√
∗
∗
∗
(Pn − P0 ) ∂hn D (Qn,hn , Gn,· ) − ∂hn D (Q1 , Gn,· ) = oP (1/ n).
(12)
5
Now that we have introduced our high-level assumptions, we can state the corresponding highlevel result that they entail. The proof is relegated to the appendix.
Theorem 1 (Asymptotics of the collaborative TMLE – a high-level result). Under assumptions
A2 to A5, it holds that
√
∗
) − Ψ(P0 ) = (Pn − P0 ) (D∗ (Q1 , G0 ) + ∆(P1 )) + oP (1/ n).
Ψ(Pn,h
(13)
n
Commenting on the high-level assumptions. Assumption A1(Q, h, c) concerns both D∗
(specifically, how D∗ (Q, G)(O) depends on G(O)) and algorithms Ĝt , t ∈ H (specifically, how
smooth is t 7→ Ĝt (Pn )(O) around h). In the particular example studied in the following sections,
the counterpart C1 of A1(Q, h, c) concerns only algorithms Ĝt , t ∈ H.
∗
In the example, we show how Pn,h
can be built collaboratively in such a way that A2 is
n
met, under a series of nested assumptions about the smoothness of data-dependent, real-valued
functions over H, the construction of which notably involve algorithms Ĝt , t ∈ H. To understand
why achieving (6) is relevant, observe that the following oracle version of Pn ∂hn D∗ (Q∗n,hn , Gn,· ),
1
lim P0 D∗ (Q∗n,hn , Gn,hn +t ) − D∗ (Q∗n,hn , Gn,hn ) ,
t→0 t
t6=0
can be rewritten as
1
Rem20 (Q∗n,hn , Gn,hn +t ) − Rem20 (Q∗n,hn , Gn,hn )
t→0 t
lim
t6=0
in view of (1). Thus, achieving (5) relates to finding critical points of h 7→ Rem20 (Q∗n,hn , Gn,h ).
Assumption A3 formalizes the convergence of Gn,hn to its target G0 w.r.t. dG , and that of Q∗n,hn
to some limit Q1 ∈ Q w.r.t. dQ . It does not require that Q1 be equal to the target Q0 of Q∗n,hn ,
but A4 may be impossible to meet when Q1 6= Q0 (see below). Condition (7) in A3 is met for
instance if the L2 (P0 )-norm of D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q1 , G0 ) goes to zero in probability and if the
difference falls in a P0 -Donsker class with probability tending to one. As for (8), it typically holds
√
whenever the product of the rates of convergence of Q∗n,hn and Gn,hn to their limits is oP (1/ n).
The counterpart of A3 in the example studied in the following sections is C2.
With A4, we assume the existence of an oracle h̃n that undersmoothes Gn,h enough so that
Φ0 (Gn,h̃n ) is an asymptotically linear estimator of Φ0 (G0 ), where we note that Φ0 is pathwise
differentiable in a similar way as Ψ. We say that h̃n is an oracle because the definition of Φ0
involves P0 and Q1 . It happens that
√
Lemma 2. Under A2 and A3, if Q1 = Q0 , dG (Gn,hn , G0 ) × dQ (Q∗n,hn , Q0 ) = oP (1/ n), and if
(3) is met with IF = D∗ (P0 ), then A4 holds with hn = h̃n and ∆(P1 ) = 0.
It is difficult to assess whether or not A4 is a tall order when dG (Gn,hn , G0 ) × dQ (Q∗n,hn , Q0 ) is
√
not necessarily oP (1/ n), or if Q1 6= Q0 .
Finally, A5 states that the distance between h̃n and hn , introduced in A2, is of order oP (1/n1/4 )
at most. Its conditions (10) and (12) are of similar nature as (7). As for (11), the Cauchy-Schwarz
inequality reveals that it is met if the L2 (P0 )-norm of ∂hn D∗ (Q∗n,hn , Gn,· ) − ∂hn D∗ (Q1 , Gn,· ) is
oP (1/n1/4 ).
6
3
Collaborative TMLE for continuous tuning when inferring the
average treatment effect: presentation and analysis
In this section, we specialize the discussion to the inference of a specific statistical parameter,
the so called average treatment effect. Section 3.1 introduces the parameter and recalls what
are the corresponding D∗ and Rem20 from Section 1. Section 3.2 describes the uncooperative
construction of a continuum of uncooperative TMLEs. Section 3.3 argues why the selection of one
of the uncooperative TMLEs is unlikely to yield a well behaved (i.e., asymptotically Gaussian)
estimator when the product of the rates of convergence of the estimators of Q0 and G0 to their
√
limits is not fast enough (i.e., o(1/ n)). Then, Sections 3.4 and 3.5 present the collaborative
construction of collaborative TMLEs and how to select one among them that is well behaved,
under assumptions that are spelled out in Section 3.6, where the high-level Theorem 1 and its
assumptions are specialized.
3.1
Preliminary
We observe n independent draws O1 ≡ (W1 , A1 , Y1 ), . . ., On ≡ (Wn , An , Yn ) from P0 , the true law
of O ≡ (W, A, Y ). It is known that Y takes its values in [0, 1]. We consider the statistical model
M that leaves unspecified the law QW,0 of W and the conditional law of Y given (A, W ), while we
might know that the conditional expectation G0 of A given W belongs to a set G.
Introduce
Q̄0 (A, W ) ≡ EP0 (Y |A, W ), G0 (W ) ≡ P0 (A = 1|W ).
The parameter of interest is the average treatment effect,
ψ0 ≡ EQW,0 Q̄0 (1, W ) − Q̄0 (0, W ) .
We choose it because its study provides a wealth of information and paves the way for the analysis
of a variety of other parameters often encountered in the statistical literature.
More generally, every P ∈ M gives rise to QW , Q̄(A, W ), G(W ) and Q ≡ (QW , Q̄), which are
respectively the marginal law of W under P , the conditional expectation of Y given (A, W ) under
P , the conditional probability that A = 1 given W under P , and the couple consisting of QW and
Q̄. For each of them, the average treatment effect is Ψ(P ), where Ψ : M → [0, 1] is given by
Ψ(P ) ≡ EQW Q̄(1, W ) − Q̄(0, W ) .
For notational conciseness, we let `G be given by
`G(A, W ) ≡ AG(W ) + (1 − A)(1 − G(W ))
(14)
for every G ∈ G. Note that `G(A, W ) is the conditional likelihood of A given W when A given W
is drawn from the Bernoulli law with parameter G(W ), hence the “`” in the notation. Parameter
Ψ viewed as a real-valued mapping over M is pathwise differentiable at every P ∈ M w.r.t. the
maximal tangent set SP = L20 (P ). The efficient influence curve D∗ (P ) of Ψ at P ∈ M is given by
D∗ (P )(O) ≡ D2∗ (Q̄, G)(O) + Q̄(1, W ) − Q̄(0, W ) − Ψ(P )
where
(15)
2A − 1
D2∗ (Q̄, G)(O) ≡
(Y − Q̄(A, W )).
`G(A, W )
Recall definition (1). It is easy to check that, for every P ∈ M,
`G0 (A, W )
Rem20 (Q̄, G) = EP0 (2A − 1) 1 −
Q̄(A, W ) − Q̄0 (A, W ) .
`G(A, W )
7
(16)
Writing Rem20 (Q̄, G) instead of Rem20 (Q, G) slightly abuses notation, but is justified because
integrating out A in the RHS of (16) reveals that it only depends on P0 , Q̄ and G. Furthermore,
by the Cauchy-Schwartz inequality, it holds that
2
2
Rem20 (Q̄, G) ≤ P0 (Q̄ − Q̄0 ) × P0
3.2
G − G0
`G
2
.
(17)
Uncooperative construction of a continuum of uncooperative TMLEs
ˆ (P ) be an initial estimator of Q̄ and {G
Prerequisites. Let Q̄0n ≡ Q̄
n
0
n,h ≡ Ĝh (Pn ) : h ∈ H} be
a continuum of candidate estimators of G0 indexed by a real-valued tuning parameter h ∈ H, an
open interval of R∗+ . By convention, we agree that small values of h correspond with less bias for
Gn,h as an estimator of G0 . Specifically, denoting L1 the valid loss function for the estimation of
G0 given by
L1 (G)(A, W ) ≡ − log `G(A, W ) = −A log G(W ) − (1 − A) log(1 − G(W )),
(18)
for every G ∈ G, where `G was defined in (14), we assume from now on that the empirical risk
h 7→ Pn L1 (Gn,h ) increases.
For example, Ĝh could correspond to fitting a logistic linear regression maximizing the loglikelihood under the constraint that the sum of the absolute values of the coefficients is smaller
than or equal to 1/h with h ∈ H ≡ R∗+ . We will refer to this algorithm as the LASSO logistic
regression algorithm.
Uncooperative TMLEs. Let QW,n be the empirical law of {W1 , . . . , Wn }. Set arbitrarily h ∈ H
0 ∈ M denote any element of M such that the marginal law of W under P 0 equals Q
and let Pn,h
W,n
n,h
0
0
0
and the conditional expectation of Y given (A, W ) under Pn is equal to Q̄n , hence Qn = (QW,n , Q̄0n )
0 coincide with G
on the one hand; and the conditional expectation of A given W under Pn,h
n,h on
0
the other hand. Evaluating Ψ at Pn,h yields an estimator of Ψ(P0 ),
0
Ψ(Pn,h
)=
n
1X 0
Q̄n (1, Wi ) − Q̄0n (0, Wi ) ,
n
i=1
which is not targeted toward the inference of Ψ(P0 ) in the sense that none of the known features
of Pn0 was derived specifically for the sake of ultimately estimating Ψ(P0 ).
0 ) toward Ψ(P ) is to build P ∗ ∈ M from P 0 in such a way that
One way to target Ψ(Pn,h
0
n,h
n,h
√
∗
Pn D∗ (Pn,h
) = oP (1/ n)
∗ ). This can be achieved by “fluctuating” P 0 in the following sense.
and to infer Ψ(P0 ) with Ψ(Pn,h
n,h
For every G ∈ G, introduce the so called “clever covariate” C(G) given by
C(G)(A, W ) ≡
2A − 1
.
`G(A, W )
(19)
Now, for every ε ∈ R, let Q̄0n,h,ε be characterized by
logit Q̄0n,h,ε (A, W ) ≡ logit Q̄0n (A, W ) + εC(Gn,h )(A, W )
8
(20)
0
0 except that the conditional expectation of Y given (A, W )
and Pn,h,ε
∈ M be defined like Pn,h
0
0
0 when ε = 0. Moreover, denoting L
under Pn,h,ε
equals Q̄0n,h,ε (and not Q̄0n ). Clearly, Pn,h,ε
= Pn,h
2
the loss function given by
L2 (Q̄)(O) ≡ −Y log Q̄(A, W ) − (1 − Y ) log 1 − Q̄(A, W )
for every Q̄ induced by a P ∈ M, it holds that
d
L2 (Q̄0n,h,ε )(O) = −D2 (Q̄0n,h,ε , Gn,h )(O),
dε
0
a property that prompts us to say that the one-dimensional submodel {Pn,h,ε
: ε ∈ R} ⊂ M
0
0
“fluctuates” Pn,h “in the direction of” D2 (Q̄n , Gn,h ).
0 along the above submodel is indexed by the minimizer of the
The optimal fluctuation of Pn,h
empirical risk
εn,h ≡ arg min Pn L2 (Q̄0n,h,ε ),
(21)
ε∈R
of which the existence is assumed (note that ε 7→ Pn L2 (Q̄0n,h,ε ) is twice differentiable and strictly
∗ ≡ P0
convex). We call Pn,h
n,h,εn,h the TMLE of P0 , and the resulting estimator
n
∗
∗
ψn,h
≡ Ψ(Pn,h
)=
1 X 0
Q̄n,h,εn,h (1, Wi ) − Q̄0n,h,εn,h (0, Wi )
n
(22)
i=1
the TMLE of Ψ(P0 ). It is readily seen that (22) is equivalent to
∗
Pn D∗ (Pn,h
) − D2∗ (Q̄∗n,h , Gn,h ) = 0
where Q̄∗n,h ≡ Q̄0n,h,εn,h . Since εn,h minimizes the differentiable mapping ε 7→ Pn L2 (Q0n,h,ε ), it holds
moreover that
Pn D2∗ (Q̄∗n,h , Gn,h ) = 0
(23)
which, combined with the previous display, yields
∗
) = 0;
Pn D∗ (Pn,h
(24)
∗ is targeted toward Ψ(P ) indeed. Furthermore, in view of (16) and (24), ψ ∗ satisfies
in words, ψn,h
0
n,h
∗
∗
ψn,h
− Ψ(P0 ) = (Pn − P0 )D∗ (Pn,h
) + Rem20 (Q̄∗n,h , Gn,h ).
(25)
∗
Finally, the TMLEs ψn,h
(h ∈ H) are said uncooperative because, although they share the same
0
∗ does not capitalize on
initial estimator Q̄n , for any two h, h0 ∈ H, h 6= h0 , the construction of ψn,h
∗ .
that of ψn,h
0
3.3
Selecting one of the uncooperative TMLEs
At this stage of the procedure, a crucial question is to select one TMLE in the collection of uncooperative TMLEs, one that lends itself to the construction of a CI for Ψ(P0 ) with a given asymptotic
∗
level. Such a TMLE necessarily writes as ψn,h
for some well chosen hn ∈ H. This could possibly
n
be a deterministic (fixed in n) or a data-driven (random and n-dependent) element of H.
The risk R1 generated by L1 (18) is given by
R1 (G, G0 ) ≡ EQ0,W [KL(G0 (W ), G(W ))] ,
9
where KL(p, q) is the Kullback-Leibler divergence between the Bernoulli laws with parameters
p, q ∈ [0, 1]. By Pinsker’s inequality, it holds that
0 ≤ 2P0 (G − G0 )2 ≤ R1 (G, G0 )
for all G ∈ G. Therefore, if G is bounded away from zero and one, then (17) implies
Rem20 (Q̄, G)2 . P0 (Q̄ − Q̄0 )2 × R1 (G, G0 ).
(26)
If the deterministic hn ∈ H is such that (i) there exist two rates ρ1,n = o(1) and ρ2,n = o(1) such
2
∗
∗ (P )
=
that R1 (Gn,hn , G0 ) = oP (ρ21,n ) and P0 (Q̄∗n,hn − Q̄0 )2 = oP (ρ22,n ), (ii) P0 D∗ (Pn,h
)
−
D
0
n
∗
∗
oP (1), (iii) D (Pn,hn ) falls in a P0 -Donsker class with P0 -probability tending to one, (iv) oP (ρ1,n ρ2,n ) =
√
oP (1/ n), then [32, Lemma 19.24], (25) and (26) guarantee that (3) is met (with IF = D∗ (P0 )).
[This argument will be used repeatedly throughout the article.] Thus, by the central limit theorem,
√
∗
n(ψn,h
− Ψ(P0 )) converges in law to the centered Gaussian law with variance VarP0 (D∗ (P0 )(O)).
n
So, if the synergy between the convergences of Q̄∗n,hn and Gn,hn to their respective limits Q̄0 and
∗
G0 is sufficient, then the TMLE ψn,h
can be used to build CIs.
n
√
The argument falls apart if oP (ρ1,n ρ2,n ) is not oP (1/ n) (or, worse, if the L2 (P0 )-limit Q̄1 of
Q̄∗n,hn is not Q̄0 , because we do not expect that R1 (Gn,hn , G0 ) = oP (1/n)). In that case, whether or
∗
similar to (3) will
not it is possible to derive a useful asymptotic linear expansion of a TMLE ψn,h
n
√
depend on whether or not we can derive an asymptotic linear expansion for n Rem20 (Q̄∗n,hn , Gn,hn ).
If Gn,hn was derived by maximizing the likelihood over a correctly specified, finite-dimensional
√
and fine-tune-parameter-free parametric model, then n Rem20 (Q̄∗n,hn , Gn,hn ) would be asymptotically linear. Because of how we estimate G0 , we now argue that there is little chance that we can
√
select hn ∈ H such that the remainder term n Rem20 (Q̄∗n,hn , Gn,hn ) is asymptotically linear. A
natural choice would be to use the likelihood-based cross-validation selector hn,CV . Let us recall
how it is derived and explain why we do not believe it will solve our problem.
Let Bn ∈ {0, 1}n be a cross-validation scheme. For instance, Bn could be a V -fold crossvalidation scheme, i.e., a random vector taking VPdifferent values b1 , . . . , bV ∈ {0, 1}n , each with
probability 1/V , such that (i) the proportion n−1 ni=1 bv (i) of ones among the coordinates of each
P
0
be the empirical probability
bv is close to 1/V , and (ii) Vv=1 bv (i) = 1 for all 1 ≤ i ≤ n. Let Pn,B
n
1
law of the training subsample {Oi : Bn (i) = 0, 1 ≤ i ≤ n} and Pn,Bn be the empirical probability
law of the validation subsample {Oi : Bn (i) = 1, 1 ≤ i ≤ n}. The likelihood-based cross-validation
selector hn,CV of h ∈ H is given by
h
i
1
0
hn,CV ≡ arg min EBn Pn,B
L
(
Ĝ
(P
))
.
(27)
1
h
n,Bn
n
h∈H
√
Unfortunately, we do not expect that n Rem20 (Q̄∗n,h , Gn,h ) is asymptotically linear. Heuristically, hn,CV trades off the bias and variance of Gn,h as an estimator of G0 , whereas we wish to
∗ . Clearly, the variance of the estimator ψ ∗ = Ψ(P ∗ ),
trade off this bias with the variance of ψn,h
n,h
n,h
where Ψ is a smooth functional, is significantly smaller than that of the infinite-dimensional object
Gn,h .
3.4
Collaborative construction of finitely many collaborative TMLEs
The take-home message of Sections 3.2 and 3.3 is that the uncooperative construction of a continuum
of standard TMLEs will typically fail to produce one asymptotically linear TMLE if the product
of the rates of convergence of the estimators of Q̄0 and G0 to their limits is not fast enough (i.e.,
10
√
o(1/ n)). In Sections 3.4 and 3.5, we demonstrate how a collaborative construction of a continuum
of standard TMLEs can produce one asymptotically linear TMLE in this challenging situation,
under appropriate assumptions.
Recursive construction. We now present the collaborative construction of finitely many TMLEs.
In the forthcoming theoretical presentation, we make on the fly a series of assumptions. The most
important ones will be emphasized.
We argued that the cross-validated selector hn,CV (27) does not sufficiently undersmooth Gn,h
√
to make of n Rem20 (Q̄∗n,h , Gn,h ) an asymptotically linear term. Since we have assumed that
h 7→ Pn L1 (Gn,h ) increases, we can focus on those tuning parameters h in H∩]0, hn,CV ], a set
assumed non-empty from now (an assumption that we call B1(Pn , 1)).
The construction is recursive. It unfolds as follows.
(∗)
(∗)
Initialization. We begin as in Section 3.2: for every h ∈ H∩]0, hn,CV ], we build Q̄n,h and Pn,h
using Q̄0n as an initial estimator of Q̄0 and Gn,h as the estimator of G0 . Note that placing
(∗)
(∗)
the star symbol between parentheses suggests that Q̄n,h and Pn,h are the tentative h-specific
estimator of Q̄0 and TMLE. Specifically, for every h ∈ H∩]0, hn,CV ], we define Q̄0n,h,ε as in
(20), εn,h,1 as in (21), assuming that it exists (an assumption that we call B2(Pn , 1)), then
(∗)
(∗)
(∗)
set Q̄n,h ≡ Q̄0n,h,εn,h,1 and find Pn,h ∈ M such that the marginal law of W under Pn,h is the
empirical law QW,n of {W1 , . . . , Wn } and the conditional expectation of Y given (A, W ) under
(∗)
(∗)
(∗)
(∗)
Pn,h equals Q̄n,h , hence Qn,h = (QW,n , Q̄n,h ) on the one hand; and the conditional expectation
(∗)
of A given W under Pn,h coincides with Gn,h on the other hand.
(∗)
We assume that h 7→ Pn L2 (Qn,h ) is minimized globally at hn,1 in the interior of H∩]0, hn,CV ]
(an assumption that we call B3(Pn , 1)). If there are several minimizers, then hn,1 is the
largest of them by choice. Observe that, for every h ∈ H∩]0, hn,CV ],
(∗)
(∗)
Pn L2 (Q̄n,hn,1 ) ≤ Pn L2 (Q̄n,h ) ≤ Pn L2 (Q̄0n,h )
and, in particular,
(∗)
Pn L2 (Q̄n,hn,1 ) < Pn L2 (Q̄∗n,hn,CV ) ≤ Pn L2 (Q̄0n,hn,CV ).
Let us assume now that, in addition, h 7→ εn,h,1 , h 7→ 1/Gn,h (Wi ) and h 7→ 1/(1 − Gn,h (Wi ))
(all 1 ≤ i ≤ n) are differentiable in an open neighborhood of hn,1 (an assumption that we call
(∗)
B4(Pn , 1)). Consequently, (i) ∂hn,1 D∗ (Q̄n,hn,1 , Gn,· )(Oi ) is well defined for each 1 ≤ i ≤ n
(∗)
(see (4)), and (ii) h 7→ Pn L2 (Q̄n,h ) is differentiable in that neighborhood. Moreover, since
hn,1 minimizes the previous mapping, we have
0 = −
=
d
(∗)
Pn L2 (Q̄n,t )
dt
t=hn,1
!
d
(∗)
(∗)
εn,t,1
× Pn D2∗ (Q̄n,hn,1 , Gn,hn,1 ) + εn,h,1 × Pn ∂hn,1 D∗ (Q̄n,hn,1 , Gn,· )
dt
t=hn,1
(∗)
= εn,h,1 × Pn ∂hn,1 D∗ (Q̄n,hn,1 , Gn,· ),
where the third equality holds because
(∗)
Pn D2∗ (Q̄n,hn,1 , Gn,hn,1 ) = Pn D2∗ (Q̄0n,h,εn,h,1 , Gn,hn,1 ) = 0
11
in light of (23). If εn,h,1 6= 0 (an assumption that we call B5(Pn , 1)), then we thus have
proven that the following equation is solved
(∗)
Pn ∂hn,1 D∗ (Q̄n,hn,1 , Gn,· ) = 0.
(∗)
(∗)
To complete the initialization, we define hn,0 ≡ hn,CV , Q̄∗n,hn,1 ≡ Q̄n,hn,1 , Q∗n,hn,1 ≡ Qn,hn,1 ,
(∗)
∗
∗
∗
Pn,h
≡ Pn,hn,1 , ψn,h
≡ Ψ(Pn,h
), and note that they satisfy
n,1
n,1
n,1
∗
)=0
Pn ∂hn,1 D∗ (Q̄∗n,hn,1 , Gn,· ) = Pn D∗ (Pn,h
n,1
Pn L2 (Q̄∗n,hn,1 )
<
and
Pn L2 (Q̄∗n,hn,0 )
(recall how (23) implied (24) earlier).
Recursion. Let k ≥ 2 be arbitrarily chosen. Suppose that, for all 1 ≤ ` < k, we have already built
∗
∗
the 5-tuples (hn,` , Q̄∗n,hn,` , Q∗n,hn,` , Pn,h
, ψn,h
) under assumptions B1(Pn , `) to B5(Pn , `),
n,`
n,`
and also that H∩]0, hn,k−1 ] 6= ∅ (an assumption that we call B1(Pn , k)). Let us now present
∗
the construction of (hn,k , Q̄∗n,hn,k , Q∗n,hn,k , Pn,h
) under assumptions B1(Pn , k) to B5(Pn , k).
n,k
Because the presentation is very similar to that of the initialization, it is more laid out more
directly.
(∗)
(∗)
For every h ∈ H∩]0, hn,k−1 ], we build again Q̄n,h and Pn,h but using Q̄∗n,hn,k−1 as an initial
estimator of Q̄0 and Gn,h as the estimator of G0 . Specifically, for every h ∈ H∩]0, hn,k−1 ],
k−1
we define Q̄n,h,ε
as in (20) with Q̄∗n,hn,k−1 substituted for Q̄0n , εn,h,k as in (21) with Q̄k−1
n,h,ε
(∗)
substituted for Q̄0n,h,ε (B2(Pn , k) assumes the existence of εn,h,k ), then set Q̄n,h ≡ Q̄kn,h,εn,h,k
(∗)
(∗)
and find Pn,h ∈ M such that the marginal law of W under Pn,h is the empirical law QW,n
(∗)
(∗)
of {W1 , . . . , Wn } and the conditional expectation of Y given (A, W ) under Pn,h equals Q̄n,h ,
(∗)
(∗)
hence Qn,h = (QW,n , Q̄n,h ) on the one hand; and the conditional expectation of A given W
(∗)
under Pn,h coincides with Gn,h on the other hand.
(∗)
We assume that h 7→ Pn L2 (Qn,h ) is minimized globally at hn,k in the interior of H∩]0, hn,k−1 ]
(an assumption that we call B3(Pn , k)). If there are several minimizers, then hn,k is the
largest of them by choice. Moreover, we also assume that h 7→ εn,h,k , h 7→ 1/Gn,h (Wi ) and
h 7→ 1/(1 − Gn,h (Wi )) (all 1 ≤ i ≤ n) are differentiable in an open neighborhood of hn,k (an
(∗)
assumption that we call B4(Pn , k)). Consequently, ∂hn,k D∗ (Q̄n,hn,k , Gn,· )(Oi ) is well defined
(∗)
for each 1 ≤ i ≤ n (see (4)), h 7→ Pn L2 (Q̄n,h ) is differentiable in that neighborhood and, since
hn,k minimizes the previous mapping,
(∗)
εn,h,k × Pn ∂hn,k D∗ (Q̄n,hn,k , Gn,· ) = 0.
If εn,h,k 6= 0 (an assumption that we call B5(Pn , k)), then it holds that
(∗)
Pn ∂hn,k D∗ (Q̄n,hn,k , Gn,· ) = 0.
(∗)
(∗)
To complete the presentation and the recursion, we define Q̄∗n,hn,k ≡ Q̄n,hn,k , Q∗n,hn,k ≡ Qn,hn,k ,
(∗)
∗
∗
∗
Pn,h
≡ Pn,hn,k , ψn,h
≡ Ψ(Pn,h
), and note that they satisfy
n,k
n,k
n,k
∗
Pn ∂hn,k D∗ (Q̄∗n,hn,k , Gn,· ) = Pn D∗ (Pn,h
)=0
n,k
12
and
(28)
Pn L2 (Q̄∗n,hn,` ) < Pn L2 (Q̄∗n,hn,`−1 )
for all 1 ≤ ` ≤ k.
∗
We discuss when to stop the loop in the next paragraph. The collection {Pn,h
: 0 ≤ k ≤ Kn }
n,k
∗
of TMLEs is arguably built collaboratively, as the derivation of every Pn,hn,` heavily depends on
∗
Pn,h
.
n,`−1
The loop is iterated until a stopping criterion is met. The instantiations of the collaborative
TMLE laid out in Section 4 rely on the LASSO logistic regression algorithm. It is thus possible to
pre-specify an upper bound on Kn . In general, we may decide to stop the recursive construction
whenever a maximal number K of iterations has been reached, or hn,k ≤ ~, or M successive
∗
TMLEs ψn,h
(0 ≤ m < M ) all belong to an interval of length smaller than ηn,k , for some
n,k+m
user-supplied integers Kmax , M and small positive numbers hmin and ηn,k , the former chosen such
that H∩]0, hmin [ is non-empty and the latter possibly sample-size- and data-driven. The choice of
Kmax would typically be driven by considerations about the computational time. The choice of
hmin would typically depend on the collection {Ĝh : h ∈ H} of h-specific algorithms, h ≤ hmin
meaning that too much undersmoothing is certainly at play when using Ĝh . We would suggest
∗
2 ≡ Υ (P ∗
choosing M ≡ 3 and characterizing ηn,k by ηn,k
Pn
n,hn,k )/10n with ΥPn : M → R+ given by
n
∗
ΥPn (P ) ≡ EPn D (P )(O)
2
1X ∗
=
D (P )(Oi )2 .
n
i=1
∗
The definition of ΥPn is justified by the fact that ΥPn (D∗ (Pn,h
)) estimates the asymptotic variance
n
∗
)
in
the
context
where
we
prove
(3)
(with
IF = D∗ (P0 )) in Section 3.3.
of the TMLE Ψ(Pn,h
n
3.5
Selecting one of the finitely many collaborative TMLEs
It remains to determine which TMLE to select among the collection of collaborative TMLEs that
we constructed in Section 3.4. Again, the selection hinges on the cross-validation principle.
The recursive construction described in Section 3.4 can be applied to the empirical measure Pn of any subset of the complete data set. Starting from hn,CV (as defined in (27) even
when Pn differs from Pn ), let the 5-tuple (Hn,1 , Q̄∗n,Hn,1 , Q∗n,Hn,1 , P∗n,Hn,1 , Ψ(P∗n,Hn,1 )) be defined like
∗
∗
the 5-tuple (hn,1 , Q̄∗n,hn,1 , Q∗n,hn,1 , Pn,h
, ψn,h
) with Pn substituted for Pn , under assumptions
n,1
n,1
B1(Pn , 1) to B5(Pn , 1). Then, recursively, let (Hn,k , Q̄∗n,Hn,k , Q∗n,Hn,k , P∗n,Hn,k , Ψ(P∗n,Hn,k )) be defined
∗
∗
like (hn,k , Q̄∗n,hn,k , Q∗n,hn,k , Pn,h
, ψn,h
) with Pn substituted for Pn , under assumptions B1(Pn , k)
n,k
n,k
to B5(Pn , k). The recursive construction is stopped when Kn 5-tuples have been derived, where
Kn is defined like Kn with Pn substituted for Pn .
The collection
n
o
(Hn,k , Q̄∗n,Hn,k , Q∗n,Hn,k , P∗n,Hn,k , Ψ(P∗n,Hn,k )) : 1 ≤ k ≤ Kn
(29)
of Kn collaborative TMLEs is used to define a continuum of collaborative TMLEs in the following
straightforward way. The challenge is to associate a 4-tuple (Q̄∗n,h , Q∗n,h , P∗n,h , Ψ(P∗n,h )) to any h ∈
H∩]0, hn,CV ]. To do so, we simply let Hn (h) be the element of {Hn,k : 1 ≤ k ≤ Kn } that is closest
to h (with a preference for the larger of the two closer ones when h is right in the middle), that is,
formally, we set
n
o
Hn (h) ≡ max Hn,k : |h − Hn,k | = min{|h − Hn,` | : 1 ≤ ` ≤ Kn }
(30)
13
and associate to h the corresponding 4-tuple
(Q̄∗n,Hn (h) , Q∗n,Hn (h) , P∗n,Hn (h) , Ψ(P∗n,Hn (h) )).
Let Bn be the cross-validation scheme introduced in Section 3.3. By convention, let the max of
the empty set be 0. The collaborative TMLE that we select is
∗
∗
)
, ψn,h
(Q̄∗n,hn,κn , Q∗n,hn,κn , Pn,h
n,κn
n,κn
where κn is given by
(
κn ≡ 1 ∨ max 1 ≤ k ≤ Kn : hn,k ≥
arg min
h∈H∩]0,hn,CV ]
1
L (Q̄∗n,Hn (h)
EBn Pn,B
n 2
(31)
0
Pn =Pn,B
n
)
)) .
(32)
In words, κn is the unique element of {1, . . . , Kn } such that hn,κn is the smallest element of
{hn,1 , . . . , hn,Kn } that is larger than the minimizer of the cross-validated L2 -risk of the collaborative TMLE, if there exists such an element, and 1 otherwise. In (32), Q̄∗n,Hn (h) Pn =P 0
equals
n,Bn
0
Q̄∗n,Hn (h) when Pn = Pn,B
.
n
The contrast between hn,κn and hn,CV is stark. At first glance, the main difference is that the
role play by cross-validated L1 -risks of algorithms to estimate G0 in (27) is played by cross-validated
L2 -risks of algorithms to estimate Q̄0 in (32). A closer examination reveals that the difference is
∗
0∗
0
deeper. Replacing L1 (Ĝh (Pn,B
)) by L2 (Q̄0∗
n,Bn h ) (with Q̄n,Bn ,h defined like Q̄n,h in Section 3.2 but
n
0
instead of Pn ) would not make of the resulting alternative cross-validated selector of
based on Pn,B
n
h a good candidate: because of the inherent lack of cooperation between the uncooperative TMLEs
∗ , the resulting estimator of G would not even be consistent. This fact motivates the general
ψn,h
0
C-TMLE methodology, of which the present instantiation includes a twist consisting in solving two
critical equations, see (28).
3.6
Asymptotics
∗
hinges on Theorem 1.
The study of the asymptotic properties of the collaborative TMLE ψn,h
n,κn
We first specify two pseudo-distances dG and dQ on G and Q in light of requirement (1). On the
one hand, because we will eventually assume that G0 is bounded away from zero and one, (16)
yields that we can choose dG such that, for each G1 , G2 ∈ G,
dG (G1 , G2 )2 ≡ P0 (G1 − G2 )2 .
On the other hand, note that any data-dependent Qn ≡ (QW,n , Q̄n ) ∈ Q naturally gives rise to a
substitution estimator ψn of ψ0 :
ψn ≡ EQW,n
n
1X
Q̄n (1, W ) − Q̄n (0, W ) =
Q̄n (1, Wi ) − Q̄n (0, Wi ) .
n
i=1
It is easy to check (see the appendix) that the following result holds.
Lemma 3. Assume that G0 is bounded away from zero and one and that Q̄n (1, ·) − Q̄n (0, ·) falls
in a P0 -Donsker class with P0 -probability tending to one. If P0 (Q̄n − Q̄1 )2 = oP (1) for some Q̄1 ,
then ψn = P0 (Q̄1 (1, ·) − Q̄1 (0, ·)) + oP (1).
14
Since we always estimate the marginal law of W under P0 , QW,0 , with its empirical counterpart
QW,n , we can thus define the pseudo-distance dQ by setting, for each Q1 , Q2 ∈ Q,
dQ (Q1 , Q2 )2 ≡ P0 (Q̄1 − Q̄2 )2 ,
an expression that does not depend on the first components of Q1 and Q2 .
Consider the following inter-dependent assumptions. The first one is related to A1(Q, h, c) and
completes B5(Pn , hn,κn ).
C1 There exists a universal constant C1 ∈]0, 1/2[ such that G0 and any by-product Gn,h of algorithm Ĝh (any h ∈ H) trained on the empirical measure Pn take their values in [C1 , 1 − C1 ].
Moreover, there exists an open neighborhood T ⊂ H of hn,κn and a universal constant C2 > 0
such that t 7→ Gn,t (W ) is twice differentiable over T (P0 -almost surely) and, P0 -almost surely,
sup
h∈T
d2
d
Gn,t (W )|t=h ∨ sup
G (W )|t=h ≤ C2 .
2 n,t
dt
h∈T dt
When C1 is met, we denote G0n,h (W ) the first derivative of t 7→ Gn,t (W ) at h ∈ H.
C2 Both Gn,hn,κn and Q̄∗n,hn,κn converge in L2 (P0 ), to G0 and Q̄1 respectively. Moreover, it holds
√
that P0 (Gn,hn,κn − G0 )2 = oP (1/ n) and P0 (Gn,hn,κn − G0 )2 × P0 (Q̄∗n,hn,κn − Q̄1 )2 = oP (1/n).
√
C3 Assumption A4 is met, (hn,κn − h̃n )2 = oP (1/ n) and (hn,κn − h̃n )2 × P0 (Q̄∗n,hn,κn − Q̄1 )2 =
oP (1/n).
C4 With P0 -probability tending to one, Q̄∗n,hn,κn , Gn,hn,κn , Gn,h̃n and G0n,hn,κn fall in P0 -Donsker
classes.
We are now in a position to state the corollary of Theorem 1 that describes the asymptotic
properties of the collaborative TMLE targeting the average treatment effect.
Corollary 4 (Asymptotics of the collaborative TMLE – targeting the average treatment effect).
Suppose that assumptions B1(·, ·) to B5(·, ·) that we made in Sections 3.4 and 3.5 when constructing
the collaborative TMLE given in (31) are met. In addition, suppose that C1 to C4 are satisfied.
Then
√
∗
ψn,h
− Ψ(P0 ) = (Pn − P0 ) (D∗ (Q1 , G0 ) + ∆(P1 )) + oP (1/ n).
n,κn
√
∗
− Ψ(P0 )) converges in law
By the central limit theorem, the corrolary implies that n(ψn,h
n,κn
2
∗
to the centered Gaussian law with a variance σ ≡ P0 (D (Q1 , G0 ) + ∆(P1 ))2 . Therefore, provided
that we can estimate σ 2 consistently (or conservatively), we can build CIs for Ψ(P0 ) with a given
asymptotic level. Sections 4, 5 and 6 investigate the practical implementation of the collaborative
TMLE and its performances in a simulation study.
4
Collaborative TMLE for continuous tuning when inferring the
average treatment effect: practical implementation
In this section, we describe the practical implementation of the two instantiations of the collaborative TMLE algorithm presented and studied in Section 3. In both of them, the collection
ˆ is left
{Ĝh : h ∈ H} is embodied in R by the glmnet algorithm [3]. The nature of algorithm Q̄
ˆ on P at the beginining of the
unspecified. As for Q̄0n , it is obtained once and for all by training Q̄
n
ˆ
procedure. More specifically, we never evaluate Q̄(Pn ) for Pn 6= Pn .
15
4.1
LASSO-C-TMLE
We now describe our LASSO-C-TMLE algorithm. Recall that Pn denotes the empirical measure
of a generic subset of the complete data set. The following algorithm implements the theoretical
procedure laid out in Sections 3.4 and 3.5.
1. Build a sequence {Gn,h ≡ Ĝh (Pn ) : h ∈ H100 } by computing a discretized version of the path
of the LASSO logistic regression of A on W with a regularization parameter h ranging in the
set H100 provided by cv.glmnet with options nlambda=100 (hence card(H100 ) = 100) and
nfolds=10. Set hmin ≡ min H100 and let hn,CV be equal to lambda.min.
2. Build a sequence {Gn,h ≡ Ĝh (Pn ) : h ∈ H100 ∩ [hmin , hn,CV )} by computing a discretized
version of the path of the LASSO logistic regression of A on W with a regularization parameter
h ranging in H100 ∩ [hmin , hn,CV ) using glmnet with a lambda set to H100 ∩ [hmin , hn,CV ) from
step 1.
Set k ≡ 1 and Hn,k−1 ≡ hn,CV .
based on Gn,h (and
3. For every h ∈ H100 ∩ [hmin , Hn,k−1 ), determine Q̄kn,h by fluctuating Q̄k−1
n
Pn ) as Section 3.2.
4. Identify the minimizer Hn,k of h 7→ Pn L2 (Q̄kn,h ) over H100 ∩ [hmin , Hn,k−1 ), define and store
Q̄∗n,h ≡ Q̄kn,h for every h ∈ H100 ∩ [Hn,k , Hn,k−1 ), and finally define Q̄kn ≡ Q̄kn,hn,k .
5. As long as Hn,k > hmin , set k ← k + 1 and repeat steps 3 and 4 recursively.
The algorithm necessarily converges in a finite number of repetitions of step 4. Let Kn be the number
of repetitions. For every 1 ≤ k ≤ Kn , set Q∗n,Hn,k ≡ (QW,n , Q̄∗n,Hn,k ) ∈ Q (its first component is
the empirical law of W under Pn ) and let P∗n,Hn,k ∈ M be any element of model M of which
the Q-component equals Q∗n,Hn,k . The collection (29) of Kn collaborative TMLEs and mapping
h 7→ Hn (h) over H100 [hmin , hn,CV ) as in (30) are thus now well defined.
Recall the definition of the cross-validation scheme Bn introduced in Section 3.3. Set
1
~(Pn ) ≡
arg min
EBn Pn,Bn L2 ( Q̄n,Hn (h) P =P 0 )
n
h∈H100 ∩[hmin ,hn,CV )
n,Bn
and run once steps 1 to 5 with Pn ≡ Pn , hence the collection
n
o
∗
∗
(hn,k , Q̄∗n,hn,k , Q∗n,hn,k , Pn,h
,
Ψ(P
))
:
1
≤
k
≤
K
n
n,h
n,k
n,k
of collaborative TMLEs. Finally, set
κn ≡ 1 ∨ max {1 ≤ k ≤ Kn : hn,k ≥ ~(Pn )} .
∗
The collaborative TMLE that we select, our LASSO-C-TMLE estimator, is Ψ(Pn,h
), as in (31).
n,κn
4.2
LASSO-PSEUDO-C-TMLE
The LASSO-C-TMLE procedure described in Section 4.1 is quite demanding computationally. It
is thus tempting to try and develop an alternative algorithm that would mimick LASSO-C-TMLE
but be simpler.
16
In Section 3.5, we emphasized (see comment before statement of theorem) that one of the keys
of LASSO-C-TMLE is to ensure the existence of hn ∈ H and Q̄∗n,hn such that
Pn ∂hn D∗ (Q̄∗n,hn , Gn,· ) = 0.
If we knew how to compute the derivative G0n,h (W ) of t 7→ Gn,t (W ) at t = h, then this could be
ˆ (P ). Specifically, in light of (19)
easily achieved by enriching the fluctuation of the initial Q̄0 ≡ Q̄
n
n
and (20), given any h ∈ H, we would define
Ch+ (Gn,· )(A, W ) ≡ C(Gn,h )(A, W ) 1, G0n,h (W ) ,
(33)
introduce Q̄0n,h,ε+ characterized for any ε+ ∈ R2 by
logit Q̄0n,h,ε+ (A, W ) ≡ logit Q̄0n (A, W ) + Ch+ (Gn,· )(A, W )ε+
0
0
and Pn,h,ε
+ ∈ M defined like Pn,h,ε except that the conditional expectation of Y given (A, W )
0
0
0
under Pn,h,ε
+ equals Q̄n,h,ε+ (and not Q̄n,h,ε ). Then, the optimal fluctuation would be indexed by
the minimizer of the empirical risk
0
ε+
n,h ≡ arg min Pn L2 (Q̄n,h,ε+ ).
ε+ ∈R2
∗+
0
0
It would result in Q̄∗+
n,h ≡ Q̄n,h,ε+ , Pn,h ≡ Pn,h,ε+ and the TMLE
n,h
n
∗+
∗+
)=
≡ Ψ(Pn,h
ψn,h
1 X ∗+
(0,
W
)
Q̄n,h (1, Wi ) − Q̄∗+
i
n,h
n
i=1
where, by construction, we would have
∗+
∗
Pn D∗ (Q̄∗+
n,h , Gn,h ) = Pn ∂h D (Q̄n,h , Gn,h ) = 0.
The LASSO-PSEUDO-C-TMLE algorithm that we now describe adapts the above procedure.
It unfolds as follows.
1. Build a sequence {Gn,h ≡ Ĝh (Pn ) : h ∈ H100 } by computing a discretized version of the path
of the LASSO logistic regression of A on W with a regularization parameter h ranging in the
set H100 provided by cv.glmnet with option nlambda=100 (hence card(H100 ) = 100). Let hn
be equal to lambda.min. Our estimator of G0 is Gn,hn .
2. Choose arbitrarily h+
n ∈ arg min{|h − hn | : h ∈ H100 , h 6= hn } and, for every 1 ≤ i ≤ n, define
G0+
n,hn (Wi ) ≡
Gn,h+
(Wi ) − Gn,hn (Wi )
n
h+
n − hn
,
a rudimentary numerical approximation of the derivative G0n,hn (Wi ) of t 7→ Gn,t (Wi ) at t = hn .
∗+
0+
0
3. Determine Q̄∗+
n,hn and Pn,hn as described above, with h = hn and Gn,hn substituted for Gn,hn
in (33).
∗+
.
The LASSO-PSEUDO-C-TMLE estimator is ψn,h
n
17
5
Main simulation study
In this section, we present the results of a multi-faceted simulation study of the behaviors and
performances of the two instantiations of the collaborative TMLE described in Section 4. Section 5.1
specifies the synthetic data-generating distribution P0 that we use, Section 5.2 introduces the
competing estimators, Section 5.3 outlines the structure of the simulation study, and Section 5.4
gathers its results. Written in R [16], our code makes extensive use of the C-TMLE package [5].
5.1
Synthetic data-generating distribution
Our synthetic data-generating distribution P0 = Π0,p,δ depends on two fine-tune parameters: the
dimension p of the baseline covariate W and a nonnegative constant δ ≥ 0. Sampling O ≡ (W, A, Y )
under Π0,p,δ unfolds sequentially along the following steps.
1. Sample W̃ from the centered Gaussian law on RM , M = dp/10e, of which the covariance
matrix Σ is the block-diagonal matrix (Akl )1≤k,l≤M where: A11 is the 10 × 10 identity matrix;
each Akk for 1 < k ≤ M is the block-diagonal matrix (Bk,st )1≤s,t≤4 with
Bk,11
1
0 .25
1 .25 ,
= 0
.25 .25 1
Bk,22 = Bk,33 =
1 .5
.5 1
,
Bk,44
1 .5 0
= .5 1 0
0 0 1
and Bk,st is a zero matrix for 1 ≤ s 6= t ≤ 4; each Akl for 1 ≤ k 6= l ≤ M is a zero matrix. If
p = 10, then Σ = A11 and we set W ≡ W̃ . If M > 10p, then we set W ≡ (W̃1 , . . . , W̃p )> .
2. Sample A conditionally on W from the Bernoulli law with paramater
!
p
X
βk Wk ,
G0 (W ) ≡ expit δ +
k=1
where (β1 , · · · , βp ) = (1, 1, 3/(p − 2), . . . , 3/(p − 2)).
3. Sample Ỹ conditionally on (A, W ) from the Gaussian law with (conditional) variance 1/25
and expectation
2
f0 (A, W ) ≡ (1 + W1 + W2 + W5 + W6 + W8 + A),
5
then define Y ≡ expit(Ỹ ).
The covariance matrix Σ induces a loose dependence structure. The components of W̃ can be
gathered in 1 + 4 × (M − 1) independent groups, one group consisting of 10 + (M − 1) independent
random variables, and the other groups consisting of either two or three mildly dependent random
variables (with correlations equal to either 0.25 or 0.5). Neither
Z
Q̄0 (A, W ) ≡
[0,1]
2 /50
e−[logit(u)−f0 (A,W )]
√
10 π(1 − u)
du
nor Ψ(Π0,p,δ ) has a closed form expression. Independently of p and δ, Ψ(Π0,p,δ ) ≈ 0.0799.
18
5.2
Competing estimators
Let O1 , . . . , On be independent draws from P0 . Recall the characterization of Ĝh (h ∈ H100 ) and
ˆ for the estimation of Q̄ consist
definition of hn,CV given in Section 4.1, step 1. Let the algorithm Q̄
0
in fitting the working model {Q̄θ : θ = (θ0 , θ1 ), θ0 , θ1 ∈ R8 } where Q̄θ is given by
Q̄θ (A, W ) ≡ Φ (Aθ1> + (1 − A)θ0> ))(W3 , . . . , W10 )>
with Φ the distribution function of the standard normal law. Note that the working model is
necessarily mis-specified, notably because of the absence of W1 and W2 in the above definition.
ˆ
ˆ on the whole data set once and for all. To emphasize, Q̄
Recall that Q̄0n is obtained by training Q̄
is never re-trained during the cross-validation procedure. This is consistent with implementation
the original instantiation of the C-TMLE algorithm and of its scalable instantiations.
We compare the LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators of Ψ(Π0,p,δ ) from
Section 4 with the following commonly used competitors:
• the unadjusted estimator:
ψnunadj
Pn
Pn
Ai Yi
(1 − Ai )Yi
i=1
≡ Pn
;
− Pi=1
n
i=1 Ai
i=1 (1 − Ai )
• the so called G-comp estimator [17]:
ψnG-comp
n
1X 0
≡
Q̄n (1, Wi ) − Q̄0n (0, Wi ) ;
n
i=1
• the so called IPTW estimator [9, 21]:
n
ψnIPTW ≡
1 X (2Ai − 1)Yi
;
n
Gn,hCV (Ai , Wi )
i=1
• the so-called A-IPTW estimator [19]:
ψnA-IPTW ≡
n
n
i=1
i=1
1X 0
1X
(2Ai − 1)
Yi − Q̄0n (Wi , Ai ) +
Qn (1, Wi ) − Q0n (0, Wi ) ;
n
Gn,hCV (Ai , Wi )
n
∗
• and the plain TMLE estimator ψn,h
, see (22).
n,CV
5.3
Outline of the structure of the simulation study
We consider six different scenarios. In each of them, we repeat independently B = 200 times the
following steps: for each (n, p, δ) in a collection of scenario-specific triplets,
1. simulate a data set of n independent observations drawn from Π0,p,δ ;
2. derive the LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators of Sections 4.1 and
4.2 as well as the competing estimators presented in Section 5.2;
∗
3. for the double-robust estimators only, i.e., ψnA-IPTW , ψn,h
and our two collaborative TMn,CV
LEs, construct 95% CIs and check whether or not each of them contains Ψ(Π0,p,δ ).
19
Building confidence intervals based on the collaborative TMLEs. By Corollary 4, the
asymptotic variances of our collaborative TMLEs both write as
VarP0 [(D∗ (Q1 , G0 ) + ∆(P1 )) (O)] .
(34)
∗
Because ∆(P1 ) is difficult to estimate, we estimate (34) with the empirical variance of D∗ (Pn,h
),
n,κn
i.e., with
n
1X ∗ ∗
2
∗
)
=
Pn D∗ (Pn,h
D (Pn,hn,κn )(Oi )2
n,κn
n
i=1
∗
Pn D∗ (Pn,h
)
n,κn
(recall that
TMLEs take the form
= 0 by construction). Therefore, the 95% CIs based on our collaborative
q
∗
∗
±
1.96
ψn,h
)2 /n.
Pn D∗ (Pn,h
n,κn
n,κ
n
We anticipate that the asymptotic variances are over-estimated, resulting in CIs that are too wide.
However, we also anticipate that the omitted correction term is of second order relative to main
term, or, put in other words, that the difference between (34) and VarP0 (D∗ (Q1 , G0 )(O)) is small.
Six scenarios. The three first scenarios investigate what happens when δ = 0 and the number of
covariates p increases as a function of sample size n. In scenario 1, p = 0.2 × n and we increase n.
√
In scenario 2, p = b2.83 × nc and we increase n. In scenario 3, p = b7.6 × log nc and we increase n.
The values of the pairs (p, n) used in these scenarios are presented in Table 0. The constants 0.2,
2.83 and 7.6 are chosen so that p = 40 at sample size n = 200 in the three scenarios.
Table 0: Values of p and n in scenarios 1, 2 and 3.
n
p in scenario 1
p in scenario 2
p in scenario 3
200
40
40
40
400
80
56
45
600
120
69
48
800
160
80
50
1000
200
89
52
1200
240
98
53
1400
280
105
55
1600
320
113
56
1800
360
120
56
2000
400
126
57
In scenarios 4 and 5, we still set δ = 0 and either keep p fixed and increase n (scenario 4) or
keep n fixed and increase p (scenario 5). Finally, in scenario 6, we keep n and p fixed and challenge
the positivity assumptions that G0 is bounded away from 0 and 1 by progressively increasing δ.
In each scenario and for all estimators, we report in a table the average bias (bias, multiplied by
10), standard error (SE, multiplied by 10) and mean squared error (MSE, multiplied by 100) across
(b)
the B = 200 repetitions. Specifically, if {φn : 1 ≤ b ≤ B} are the B realizations of an estimator of
P
1:B
(b)
ψ0 = Ψ(Π0,p,δ ) based on n independent draws from Π0,p,δ , then we call φn ≡ B −1 B
b=1 (φn − ψ0 )
P
P
1:B
1:B
(b)
(b)
B
2 1/2 the standard error, and (φ
2
−1
the average bias, (B −1 B
n ) +B
b=1 (φn − φn ) )
b=1 (φn −
1:B
φn )2 the mean squared error.
We also represent in a series of figures how MSE, the empirical coverage of the 95% CIs and
their widths evolve as the sample size (scenarios 1 to 4) or number of covariates (scenario 5) or
parameter δ (scenario 6) increase. To ease comparisons, all similar figures share the same x- and
y-axes.
5.4
Results
√
Scenarios 1, 2, and 3: increasing n and setting p = 0.2n, b2.83 nc, b7.6 log nc. The
results of the three simulation studies under scenarios 1, 2 and 3 are best presented and commented
20
upon altogether. Figure 1 and Table 1 summarize the numerical findings under scenario 1; Figure 2
and Table 2 summarize the numerical findings under scenario 2; Figure 3 and Table 3 summarize
the numerical findings under scenario 3.
Figures 1a, 2a and 3a reveal a general trend: MSE decreases as sample size n increases, despite
the fact that the number of covariates p also increases (at different n-rates in each scenario). Overall, LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE perform similarly and better than TMLE;
TMLE outperforms IPTW, and IPTW outperforms A-IPTW. Moreover, the gap between LASSOC-TMLE, LASSO-PSEUDO-C-TMLE on the one hand and TMLE on the other hand (i) reduces
as sample size n increases (in each scenario), and (ii) reduces as p decreases (for each sample size
n, across scenarios).
Judging by Tables 1, 2 and 3, the unadjusted, G-comp, IPTW and A-IPTW estimators are
strongly biased. The TMLE estimator is strongly biased too, even for large sample size n, when
the number of covariates p is not sufficiently small. Note however that the bias of TMLE vanishes
at sample size n = 2000 in scenario 2 (then, p = 126) and at sample size n ∈ {1000, 2000} (then,
p ∈ {52, 57}). Double-robustness is in action. In contrast, the LASSO-C-TMLE and LASSOPSEUDO-C-TMLE estimators are both essentially unbiased in all configurations.
Tables 1, 2 and 3 also reveal that the variance of the TMLE estimator tends to be smaller than
those of the LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators, those last two variances
being very similar. Moreover, the gap between them tends to diminish as sample size n increases,
in all scenarios.
We now turn to Figures 1b, 2b and 3b. The LASSO-C-TMLE estimator performs best in terms
of empirical coverage, followed by the LASSO-PSEUDO-C-TMLE, TMLE and A-IPTW estimators,
in that order. At moderate sample size, the superiority of LASSO-C-TMLE-based CIs over the
others is striking. However, even they fail to provide the wished coverage except when sample size
n is large (say, larger than 1500).
As a side note, we recall that if S is drawn from the Binomial law with parameter (B, q) =
(200, q), then S ≤ 185 with probability approximately 8% for q = 95%, 22% for q = 94% and
43% for q = 93%. In this light, an empirical coverage of 7.5% is not that abnormal for B = 200
independent CIs of exact coverage q = 93%, and even q = 94%. Moreover, we anticipated to
get conservative CIs because of how we estimate the asymptotic variance of the LASSO-C-TMLE
estimator, see Section 5.3. The fact that the “ratio” entries of Table 3, scenario 3, are that close to
one for the LASSO-C-TMLE estimator at sample size n ∈ {1000, 2000} (not to mention at sample
size n = 2000 in Table 1, scenario 1) reveals that little over-estimation of the asymptotic variance
is at play. Finally, we see in Figures 1c, 2c and 3c that the CIs based on the LASSO-C-TMLE
and LASSO-PSEUDO-C-TMLE estimators are systematically slightly wider and slightly narrower
than those based on the TMLE estimator, all much narrower than those based on the A-IPTW
estimator.
Scenario 4: keeping p fixed and increasing n. Figure 4 and Table 4 summarize the numerical
findings under scenario 4, where the number of covariates p is set to 40 and sample size n goes from
200 to 2000 by steps of 200.
We observe the same trend in Figure 4a as in Figures 1a, 2a and 3a: MSE decreases as sample
size n increases; overall, LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE perform similarly and
better than TMLE; TMLE outperforms IPTW, and IPTW outperforms A-IPTW. Moreover, the
gap between LASSO-C-TMLE, LASSO-PSEUDO-C-TMLE on the one hand and TMLE on the
other hand vanishes completely as sample size increases, whereas it only got smaller in scenarios 1,
2, 3.
21
estimator
0.100
●
●
MSE
0.075
L−C−TMLE
●
0.050
TMLE
●
LP−C−TMLE
●
●
0.025
●
●
●
A−IPTW
●
IPTW
0.000
500
1000
1500
2000
sample size
●
●
●
●
●
●
●
●
●
●
2.0
estimator
0.50
0.40
0.30
0.20
●
ratio of CI width
1 − CI coverage
(a) MSE for five of the seven estimators. The
MSE of the A-IPTW estimator is so large that it
does not fit in the picture. MSE is multiplied by
100.
TMLE
A−IPTW
500
1000
1500
LP−C−TMLE
●
●
●
●
●
●
●
●
●
●
A−IPTW
0.8
2000
500
sample size
TMLE
L−C−TMLE
1.2
LP−C−TMLE
0.05
0.04
0.03
0.02
●
1.6
L−C−TMLE
0.10
estimator
1000
1500
2000
sample size
(b) Coverage of 95% CIs based on the doublerobust estimators.
(c) Relative width of 95% CIs based on the
double-robust estimators w.r.t. that of the plain
∗
TMLE, ψn,h
.
n,CV
Figure 1: Scenario 1. We fix the ratio p/n = 0.2, and increase the sample size n from 200 to 2000.
n
200
1000
2000
bias
SE
MSE
ratio
bias
SE
MSE
ratio
bias
SE
MSE
ratio
ψnunadj
1.259
0.236
1.641
ψnG-comp
1.212
0.152
1.491
ψnIPTW
0.435
0.320
0.291
1.171
0.110
1.382
1.206
0.066
1.459
0.279
0.157
0.103
1.175
0.076
1.386
1.217
0.047
1.483
0.232
0.120
0.068
ψnA-IPTW
0.632
0.137
0.418
1.414
0.407
0.061
0.169
1.765
0.339
0.050
0.118
1.666
∗
ψn,h
n,CV
0.327
0.151
0.130
0.882
0.189
0.064
0.040
0.969
0.134
0.045
0.020
0.959
L-C-TMLE
LP-C-TMLE
-0.007
0.242
0.059
0.577
0.032
0.083
0.008
0.882
0.014
0.050
0.003
1.062
0.039
0.260
0.069
0.466
0.026
0.091
0.009
0.632
0.020
0.069
0.005
0.626
Table 1: Scenario 1. The performance of each estimator at sample size n ∈ {200, 1000, 2000},
with ratio p/n = 0.2. The columns named L-C-TMLE and LP-C-TMLE correspond to the LASSOC-TMLE and LASSO-PSEUDO-C-TMLE estimators, respectively. Rows ratio report the ratios of
the average of the SE estimates across the B repetitions to the empirical SE. Bias and SE are
multiplied by 10, and MSE is multiplied by 100.
22
estimator
0.100
●
TMLE
0.075
MSE
L−C−TMLE
0.050
●
LP−C−TMLE
●
●
0.025
●
●
A−IPTW
●
●
●
●
IPTW
0.000
500
1000
1500
2000
sample size
●
0.50
0.40
0.30
0.20
●
●
●
●
●
●
●
●
2.0
estimator
●
●
ratio of CI width
1 − CI coverage
(a) MSE for five of the seven estimators.
TMLE
A−IPTW
500
1000
1500
LP−C−TMLE
●
●
●
●
●
●
●
●
●
●
A−IPTW
0.8
2000
500
sample size
TMLE
L−C−TMLE
1.2
LP−C−TMLE
0.05
0.04
0.03
0.02
●
1.6
L−C−TMLE
0.10
estimator
1000
1500
2000
sample size
(b) Coverage of 95% CIs based on the doublerobust estimators. MSE is multiplied by 100.
(c) Relative width of 95% CIs based on the
double-robust estimators w.r.t. that of the plain
∗
TMLE, ψn,h
.
n,CV
√
Figure 2: Scenario 2. We increase the sample size n from 200 to 2000 and set p = b2.83 nc.
n
200
1000
2000
bias
SE
MSE
ratio
bias
SE
MSE
ratio
bias
SE
MSE
ratio
ψnunadj
1.242
0.221
1.592
ψnG-comp
1.173
0.157
1.401
ψnIPTW
0.469
0.278
0.297
1.216
0.104
1.489
1.214
0.068
1.479
0.271
0.184
0.107
1.192
0.075
1.426
1.214
0.051
1.477
0.201
0.140
0.060
ψnA-IPTW
0.614
0.135
0.395
1.377
0.361
0.077
0.136
1.505
0.274
0.061
0.079
1.488
∗
ψn,h
n,CV
0.322
0.156
0.128
0.857
0.126
0.064
0.020
0.965
0.080
0.051
0.009
0.870
L-C-TMLE
LP-C-TMLE
0.014
0.226
0.051
0.630
0.003
0.076
0.006
0.862
0.007
0.053
0.003
0.898
0.018
0.217
0.048
0.553
0.019
0.074
0.006
0.790
0.018
0.049
0.003
0.904
Table 2: Scenario 2. The performance of each estimator at sample size n ∈ {200, 1000, 2000},
√
with ratio p = b2.83 nc. The columns named L-C-TMLE and LP-C-TMLE correspond to the
LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators, respectively. Rows ratio report the
ratios of the average of the SE estimates across the B repetitions to the empirical SE. Bias and SE
are multiplied by 10, and MSE is multiplied by 100.
23
estimator
0.100
●
TMLE
0.075
MSE
L−C−TMLE
0.050
LP−C−TMLE
●
0.025
A−IPTW
●
●
0.000
500
●
●
1000
●
●
1500
●
IPTW
●
2000
sample size
●
0.50
0.40
0.30
0.20
2.0
estimator
●
●
●
●
0.10
ratio of CI width
1 − CI coverage
(a) MSE for five of the seven estimators. MSE is
multiplied by 100.
TMLE
●
●
●
0.05
0.04
0.03
0.02
●
A−IPTW
500
1000
1500
LP−C−TMLE
●
●
●
●
●
●
●
●
●
●
A−IPTW
0.8
2000
500
sample size
TMLE
L−C−TMLE
1.2
LP−C−TMLE
●
●
1.6
L−C−TMLE
●
estimator
1000
1500
2000
sample size
(b) Coverage of 95% CIs based on the doublerobust estimators.
(c) Relative width of 95% CIs based on the
double-robust estimators w.r.t. that of the plain
∗
TMLE, ψn,h
.
n,CV
Figure 3: Scenario 3. We increase the sample size n from 200 to 2000 and keep p = b7.6 log(n)c.
n
200
1000
2000
bias
SE
MSE
ratio
bias
SE
MSE
ratio
bias
SE
MSE
ratio
ψnunadj
1.244
0.228
1.600
ψnG-comp
1.190
0.148
1.439
ψnIPTW
0.427
0.316
0.282
1.242
0.114
1.555
1.208
0.077
1.465
0.224
0.219
0.098
1.227
0.074
1.511
1.205
0.050
1.453
0.159
0.157
0.050
ψnA-IPTW
0.634
0.136
0.420
1.363
0.286
0.086
0.089
1.544
0.185
0.066
0.039
1.552
∗
ψn,h
n,CV
0.314
0.170
0.128
0.777
0.056
0.065
0.007
0.995
0.019
0.046
0.003
1.023
L-C-TMLE
LP-C-TMLE
-0.009
0.239
0.057
0.598
-0.008
0.071
0.005
0.992
-0.006
0.053
0.003
0.978
0.014
0.241
0.058
0.502
-0.006
0.069
0.005
0.894
-0.003
0.050
0.002
0.942
Table 3: Scenario 3. The performance of each estimator at sample size n ∈ {200, 1000, 2000}, with
p = b7.6 log(n)c. The columns named L-C-TMLE and LP-C-TMLE correspond to the LASSO-CTMLE and LASSO-PSEUDO-C-TMLE estimators, respectively. Rows ratio report the ratios of
the average of the SE estimates across the B repetitions to the empirical SE. Bias and SE are
multiplied by 10, and MSE is multiplied by 100.
24
Judging by Table 4, the unadjusted, G-comp, IPTW and A-IPTW estimators are strongly biased whereas the LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators are both essentially
unbiased even at small sample size n = 200. The TMLE estimator is strongly biased too at sample
size n = 200, but much less so as n increases, with no bias at all at n = 2000. Again, doublerobustness is in action. Moreover, there is little if any difference between the LASSO-C-TMLE,
LASSO-PSEUDO-C-TMLE and TMLE estimators in terms of bias, SE and MSE for sample size
n ∈ {1000, 2000}.
Figure 4b reveals that, at sample sizes n ≥ 1000, the empirical coverage of the CIs based on the
LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators is satisfactory, and that CIs based
on the TMLE estimator may provide more coverage than wished. By Table 4 (ratio rows), the
estimation of the actual variance of the LASSO-C-TMLE and TMLE estimator is quite good at
sample size n ∈ {1000, 2000}. Apparently, the variance of the LASSO-PSEUDO-C-TMLE estimator
is under-estimated at sample size n = 1000, but much better estimated at sample size n = 2000.
Scenario 5: keeping n fixed and increasing p. Figure 5 and Table 5 summarize the numerical
findings under scenario 5, where the sample size n is set to 1000 and the number of covariates p
ranges over {50, 75, 100, 150, 200}.
The take home message of Figure 5a is that, in terms of MSE, the LASSO-C-TMLE and LASSOPSEUDO-C-TMLE estimators outperform the TMLE estimator, which outperforms the IPTW and
A-IPTW estimators. Figure 5b further shows that the above message is also valid when considering
the empirical coverage of the CIs based on the different estimators. As the number of covariates
p increases, all the empirical coverage degrade. However, the CIs based on the LASSO-C-TMLE
estimator behave remarkably better than those based on the LASSO-PSEUDO-C-TMLE estimator,
which are themselves superior to those based on the TMLE estimator.
Examining Table 5 helps to better understand the general pattern. The unadjusted, G-comp,
IPTW and A-IPTW estimators are too strongly biased to compete. The TMLE estimator performs
rather well when the number of covariates p equals 50, like the LASSO-C-TMLE and LASSOPSEUDO-C-TMLE estimators. However, when p ∈ {100, 200}, then the TMLE estimator is too
biased to compete too – even double-robustness does not help yet at the moderate sample size
of n = 1000. In contrast, the LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators are
essentially unbiased, and exhibit relatively small variances (compared to all the variances reported
in Tables 1, 2, 3, 4. Finally, let us note that the estimation of the variance of the LASSO-C-TMLE
estimator is rather good (see the ratio rows of Table 5), as opposed to that of the variance of the
LASSO-PSEUDO-C-TMLE estimator.
Scenario 6: keeping n and p fixed and challenging the positivity assumption. In this
scenario, we study how the level of posivitity violation influences the performance of the estimators,
at small sample size n = 100 and with p = 50 covariates, by progressively increasing δ ∈ {0.5+k/10 :
0 ≤ k ≤ 15}. Figure 6a illustrates how the positivity violation is challenged. We recover the fact
that δ 7→ Π0,50,δ (A = 1|W ) is increasing. When δ = 2, the law is highly skewed to 1, and
the positivity assumption is practically violated. Figures 6b, 6c, 6d and Table 6 summarize the
numerical findings under scenario 6.
We see in Figure 6 that, overall, the TMLE estimator is much more affected than the LASSOC-TMLE and LASSO-PSEUDO-C-TMLE estimators by the near violation of the positivity assumption at sample size n = 500, and that the LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE
estimators behave similarly in terms of MSE and empirical coverage. Judging by Table 6, The unadjusted, G-comp, IPTW, A-IPTW and TMLE estimators are too strongly biased to compete with
25
estimator
0.100
●
TMLE
0.075
MSE
L−C−TMLE
0.050
LP−C−TMLE
●
0.025
A−IPTW
●
●
0.000
500
●
●
1000
●
●
1500
●
IPTW
●
2000
sample size
●
0.50
0.40
0.30
0.20
2.0
estimator
●
●
●
TMLE
●
●
●
●
500
1000
1500
●
●
A−IPTW
LP−C−TMLE
●
●
●
●
●
●
●
●
●
●
0.8
●
2000
500
sample size
TMLE
L−C−TMLE
1.2
LP−C−TMLE
0.05
0.04
0.03
0.02
estimator
1.6
L−C−TMLE
●
0.10
ratio of CI width
1 − CI coverage
(a) MSE for five of the seven estimators. MSE is
multiplied by 100.
1000
1500
A−IPTW
2000
sample size
(b) Coverage of 95% CIs based on the doublerobust estimators.
(c) Relative width of 95% CIs based on the
double-robust estimators w.r.t. that of the plain
∗
TMLE, ψn,h
.
n,CV
Figure 4: Scenario 4. We fix the number of covariates p = 40, and increase the sample size n
from 200 to 2000.
n
200
1000
2000
bias
SE
MSE
ratio
bias
SE
MSE
ratio
bias
SE
MSE
ratio
ψnunadj
1.286
0.226
1.705
ψnG-comp
1.215
0.159
1.501
ψnIPTW
0.485
0.345
0.354
1.259
0.109
1.597
1.191
0.073
1.425
0.213
0.211
0.090
1.260
0.080
1.595
1.202
0.049
1.448
0.117
0.165
0.041
ψnA-IPTW
0.649
0.147
0.443
1.283
0.251
0.090
0.071
1.490
0.140
0.066
0.024
1.673
∗
ψn,h
n,CV
L-C-TMLE
LP-C-TMLE
0.317
0.172
0.130
0.761
0.028
0.062
0.005
1.062
-0.002
0.046
0.002
1.062
0.020
0.231
0.054
0.594
-0.020
0.076
0.006
0.947
-0.001
0.049
0.002
1.014
0.020
0.259
0.068
0.469
-0.021
0.074
0.006
0.866
-0.002
0.048
0.002
0.974
Table 4: Scenario 4. The performance of each estimator at sample size n ∈ {200, 1000, 2000}, with
p = 40. The columns named L-C-TMLE and LP-C-TMLE correspond to the LASSO-C-TMLE and
LASSO-PSEUDO-C-TMLE estimators, respectively. Rows ratio report the ratios of the average of
the SE estimates across the B repetitions to the empirical SE. Bias and SE are multiplied by 10,
and MSE is multiplied by 100.
26
estimator
0.100
TMLE
●
0.075
MSE
L−C−TMLE
0.050
●
0.025
A−IPTW
●
●
0.000
LP−C−TMLE
●
●
IPTW
50
100
150
200
number of covariates
●
0.10
●
2.0
estimator
●
0.50
0.40
0.30
0.20
●
●
ratio of CI width
1 − CI coverage
(a) MSE for five of the seven estimators. MSE is
multiplied by 100.
TMLE
A−IPTW
TMLE
L−C−TMLE
1.2
LP−C−TMLE
0.05
0.04
0.03
0.02
●
1.6
L−C−TMLE
●
estimator
LP−C−TMLE
●
●
●
●
●
100
150
200
A−IPTW
0.8
50
100
150
200
50
number of covariates
number of covariates
(b) Coverage of 95% CIs based on the doublerobust estimators.
(c) Relative width of 95% CIs based on the
double-robust estimators w.r.t. that of the plain
∗
TMLE, ψn,h
.
n,CV
Figure 5: Scenario 5. We fix the sample size n = 1000, and increase the number of covariates p
from 50 to 200.
p
50
100
200
bias
SE
MSE
ratio
bias
SE
MSE
ratio
bias
SE
MSE
ratio
ψnunadj
1.237
0.107
1.542
ψnG-comp
1.205
0.072
1.457
ψnIPTW
0.228
0.211
0.096
1.179
0.102
1.402
1.199
0.069
1.443
0.252
0.167
0.091
1.190
0.107
1.428
1.221
0.067
1.494
0.297
0.159
0.114
ψnA-IPTW
0.285
0.090
0.089
1.472
0.357
0.073
0.133
1.561
0.417
0.067
0.178
1.591
∗
ψn,h
n,CV
0.053
0.065
0.007
0.983
0.130
0.071
0.022
0.867
0.179
0.060
0.036
1.009
L-C-TMLE
LP-C-TMLE
-0.017
0.071
0.005
0.998
0.005
0.072
0.005
0.896
0.024
0.077
0.006
0.933
-0.008
0.072
0.005
0.864
0.013
0.075
0.006
0.791
0.024
0.089
0.009
0.645
Table 5: Scenario 5. The performance of each estimator at sample size n = 1000, with p ∈
{50, 100, 200}. The columns named L-C-TMLE and LP-C-TMLE correspond to the LASSO-CTMLE and LASSO-PSEUDO-C-TMLE estimators, respectively. Rows ratio report the ratios of
the average of the SE estimates across the B repetitions to the empirical SE. Bias and SE are
multiplied by 10, and MSE is multiplied by 100.
27
the nearly unbiased LASSO-C-TMLE and LASSO-PSEUDO-C-TMLE estimators. The rather poor
performance in terms of empirical coverage of the CIs based on the LASSO-C-TMLE and LASSOPSEUDO-C-TMLE estimators may be due to the apparent failure in estimating well their variance
(see the ratio rows of Table 6).
6
Secondary simulation study: LASSO-C-TMLE as a fine-tuning
procedure
In this shorter section, we describe a second, less ambitious simulation study. Its aim is to evaluate
the interest in using the LASSO-C-TMLE procedure as a fine-tuning procedure. Specifically we wish
to investigate, in the same context as in Section 5, how the rivals of the LASSO-C-TMLE estimator
that also rely on the estimation of G0 (i.e., the IPTW, A-IPTW, TMLE and LASSO-PSEUDOC-TMLE estimators) perform when they are provided with the estimator Gn,hn,κn indexed by the
data-adaptive, targeted, fine-tune parameter hn,κn .
We thus choose to repeat independently B = 200 times the following steps: for each number of
covariates p ∈ {100, 200},
1. simulate a data set of n = 1000 independent observations drawn from Π0,p,0 ;
2. derive the LASSO-C-TMLE estimator of Sections 4.1
3. derive the LASSO-PSEUDO-C-TMLE estimator of Section 4.2 as well as the competing
IPTW, A-IPTW and TMLE estimators exactly as presented in Section 5.2, and also using
Gn,hn,κn in place of Gn,hn (LASSO-PSEUDO-C-TMLE) and Gn,hn,CV (the others).
The results are reported in Table 7. A clear pattern emerges from Table 7: the bias is systematically reduced when using Gn,hn,κn in place of Gn,hn,CV or Gn,hn . Nevertheless, the MSE of
the IPTW estimator increases, with a two-fold increase when the number of covariates p = 200.
In contrast, the A-IPTW estimator benefits more from the substitution, with a stark decrease of
the MSE on top of that of the bias, the latter being still far too large. This makes only more
remarkable the fact that the TMLE estimator greatly benefits from the substitution on all fronts,
bias and MSE. On the contrary, the benefit for the LASSO-PSEUDO-C-TMLE estimator is not
convincing. In summary, Gn,hn,κn is targeted even “out of context”, i.e., even when it is used to
build a plain TMLE estimator as opposed to the full-fledged C-TMLE estimator.
7
Discussion
We study the inference of the value of a smooth statistical parameter at a law P0 from which
we sample n independent observations, in situations where (i) we rely on a machine learning
algorithm fine-tuned by a real-valued parameter h to estimate the G-component G0 of P0 , possibly
consistently, and (ii) the product of the rates of convergence of the estimators of the Q- and G√
components of P0 to their targets may be slower than the convenient o(1/ n). A plain TMLE with
an h chosen by cross-validation would typically not lend itself to the construction of a CI, because
the selection of h would trade-off its empirical bias with something akin to the empirical variance of
the estimator of G0 as opposed to that of the TMLE. We develop a collaborative TMLE procedure
that succeeds in achieving the relevant trade-off: under high-level empirical processes conditions,
and if there exists an oracle h that makes a bulky remainder term asymptotically Gaussian, then the
C-TMLE is asymptotically Gaussian hence amenable to building a CI provided that its asymptotic
variance can be estimated too.
28
●
1.00
0.05
0.75
0.04
MSE
delta
ECDF
0.5
0.50
●
0.03
1.2
2.0
●●
●
●
●
●
●
●
●● ●
●
●●●
●
●●
●
estimator
●
L−C−TMLE
0.02
0.25
TMLE
LP−C−TMLE
0.01
0.00
0.00
0.25
0.50
0.75
0.0
1.00
0.5
● ● ●● ● ● ● ● ●● ● ● ●● ● ●
●
●
●
●
●
TMLE
●
A−IPTW
TMLE
L−C−TMLE
1.2
LP−C−TMLE
0.05
0.04
0.03
0.02
estimator
1.6
L−C−TMLE
0.10
2.0
2.0
estimator
●
1.5
(b) MSE for three of the seven estimators. MSE
is multiplied by 100.
ratio of CI width
1 − CI coverage
(a) For every δ ∈ {0.5, 1.2, 2}, we simulate n =
1000 observations (W1 , A1 , Y1 ), . . . , (Wn , An , Yn )
from Π0,50,δ , compute {Π0,50,δ (A = 1|W = Wi ) :
1 ≤ i ≤ n}, and finally plot the corresponding
empirical cumulative distribution.
0.50
0.40
0.30
0.20
1.0
delta
propensity score
LP−C−TMLE
●●●●●●●●●●●●●●●●●●●●●
A−IPTW
0.8
0.0
0.5
1.0
1.5
2.0
0.0
0.5
delta
1.0
1.5
2.0
delta
(c) Coverage of 95% CIs based on the doublerobust estimators.
(d) Relative width of 95% CIs based on the
double-robust estimators w.r.t. that of the plain
∗
TMLE, ψn,h
.
n,CV
Figure 6: Scenario 6. We fix n = 500, p = 50, and vary δ ∈ {0.5 + k/10 : 0 ≤ k ≤ 15}. As the
∗
MSEs for IPTW and A-IPTW are too large, we only plot the MSEs of the plain TMLE ψn,h
n,CV
and the two collaborative TMLEs to ease comparisons.
δ
1.0
2.0
bias
SE
MSE
ratio
bias
SE
MSE
ratio
ψnunadj
1.283
0.170
1.675
ψnG-comp
1.252
0.105
1.579
ψnIPTW
0.995
0.319
1.092
1.391
0.223
1.984
1.340
0.142
1.817
1.620
0.409
2.791
ψnA-IPTW
0.484
0.117
0.248
1.385
0.625
0.154
0.414
1.283
∗
ψn,h
n,CV
0.145
0.108
0.033
0.816
0.185
0.143
0.055
0.650
L-C-TMLE
LP-C-TMLE
0.006
0.120
0.014
0.774
0.053
0.168
0.031
0.597
0.009
0.128
0.017
0.650
0.063
0.186
0.039
0.474
Table 6: Scenario 6. The performance of each estimator at sample size n = 500, with p = 50
and δ ∈ {1, 2}. The columns named L-C-TMLE and LP-C-TMLE correspond to the LASSO-CTMLE and LASSO-PSEUDO-C-TMLE estimators, respectively. Rows ratio report the ratios of
the average of the SE estimates across the B repetitions to the empirical SE. Bias and SE are
multiplied by 10, and MSE is multiplied by 100.
29
The construction of the C-TMLE and the main result about its empirical behavior are illustrated
with the inference of the average treatment effect, both theoretically and numerically. In the
simulation study, the G-component is estimated by the LASSO, and h is the bound on the `1 norm of the candidate coefficients. Overall, the resulting LASSO-C-TMLE estimator is superior
to all its competitors, including a plain TMLE estimator. Evaluated in terms of empirical bias,
standard error, mean squared error and coverage of CIs, the superiority is striking in small and
moderate sample sizes. It is also strong when the number of covariates increases, or when the
positivity assumption is increasingly challenged, thus making the inference task progressively even
more delicate.
The simulation study suggests that the CIs based on the C-TMLE do not provide the wished
coverage, especially in small sample sizes. Obviously, this may be explained by the need for the
C-TMLE estimator to reach its asymptotic regime. More subtly, this may also be related to highlevel assumption A4, that states the existence of an oracle h making a bulky remainder term
asymptotically Gaussian. The assumption may fail to hold in practice. We will devote future
research to understanding better A4 and finding strategies to avoir relying on it.
In conclusion, we believe that the present study further demonstrates the high versatility and
potential of the collaborative targeted minimum loss estimation methodology. For (relative) simplicity, we focused on the inference of a smooth, real-valued statistical parameter from independent
and identically distributed observations, assuming that the machine learning algorithm is finetuned by a real-valued parameter. Our instantiation of the collaborative targeted minimum loss
estimation methodology can be extended to other statistical parameters, sampling schemes, and
fine-tuning of machine learning algorithms.
A
Proofs
Sections A.1, A.2, A.3 and A.4 respectively prove Lemma 2, Theorem 1, Lemma 3 and Corollary 4.
A.1
Proof of Lemma 2
Proof. By (1),
∗
Ψ(Pn,h
) − Ψ(P0 ) + P0 D∗ (Q∗n,hn , Gn,hn ) = Rem20 (Q∗n,hn , Gn,hn ),
n
Ψ(P0 ) − Ψ(P0 ) + P0 D∗ (Q1 , Gn,hn ) = Rem20 (Q1 , Gn,hn ),
hence, by (3) and (8),
P0 D∗ (Q∗n,hn , Gn,hn )−D∗ (Q1 , Gn,hn )
∗
= Rem20 (Q∗n,hn , Gn,hn ) − Rem20 (Q1 , Gn,hn ) − Ψ(Pn,h
) − Ψ(P0 )
n
√
= −(Pn − P0 )D∗ (P0 ) + oP (1/ n).
But (5) and Φ0 (G0 ) = P0 D∗ (Q1 , G0 ) = 0 also imply that
−Φ0 (Gn,hn ) + Φ0 (G0 ) = P0 D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q1 , Gn,hn )
+(Pn − P0 ) D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q1 , G0 )
√
+(Pn − P0 )D∗ (Q1 , G0 ) + oP (1/ n)
which, combined with the previous display and (7), yield
√
√
Φ0 (Gn,hn ) − Φ0 (G0 ) = (Pn − P0 ) (D∗ (P0 ) − D∗ (Q1 , G0 )) + oP (1/ n) = oP (1/ n)
showing that (9) is satisfied with h̃n = hn and ∆(P1 ) ≡ 0.
30
A.2
Proof of Theorem 1
Proof. The proof unfolds in two parts.
Step one: extracting the would-be first order term. Equality (5) in A2 rewrites as
√
oP (1/ n) = Pn D∗ (Q∗n,hn , Gn,hn )
= (Pn − P0 )D∗ (Q∗n,hn , Gn,hn ) + P0 D∗ (Q∗n,hn , Gn,hn )
= (Pn − P0 )D∗ (Q1 , G0 ) + (Pn − P0 ) D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q1 , G0 )
+ P0 D∗ (Q∗n,hn , G0 ) + P0 D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q∗n,hn , G0 ) .
√
The second term of the sum between the first pair of brackets is oP (1/ n) by (7) in A3. As for
the first term between the second pair of brackets, (1) and (1) entail that it satisfies
∗
∗
),
) + Rem20 (Q∗n,hn , G0 ) = Ψ(P0 ) − Ψ(Pn,h
P0 D∗ (Q∗n,hn , G0 ) = Ψ(P0 ) − Ψ(Pn,h
n
n
∗
∗
where we also use the fact that Ψ(Pn,h
) depends on Pn,h
only through Q∗n,hn . Thus, it holds that
n
n
√
∗
Ψ(Pn,h
) − Ψ(P0 ) − (Pn − P0 )D∗ (Q1 , G0 ) + oP (1/ n)
n
= P0 D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q∗n,hn , G0 ) ≡ T1,n . (35)
Let us now study T1,n , the right-hand side expression in (35). It rewrites as
T1,n = P0 (D∗ (Q1 , Gn,hn ) − D∗ (Q1 , G0 ))
+ P0 D∗ (Q∗n,hn , Gn,hn ) − D∗ (Q∗n,hn , G0 ) − P0 (D∗ (Q1 , Gn,hn ) − D∗ (Q1 , G0 )) .
Consider the three terms in the right-hand side of the above equation. Combining (1), (1) and the
fact that Ψ(P ) = Ψ(P 0 ) whenever P and P 0 have the same Q-component reveals that the first and
third terms equal both Φ0 (Gn,hn ) − Φ0 (G0 ) and Rem20 (Q1 , Gn,hn ). For similar reasons, the second
term equals Rem20 (Q∗n,hn , Gn,hn ). Therefore, by using successively (8) in A3 then (9) from A4, we
obtain that
T1,n = Φ0 (Gn,hn ) − Φ0 (G0 ) + Rem20 (Q∗n,hn , Gn,hn ) − Rem20 (Q1 , Gn,hn )
√
= Φ0 (Gn,h̃n ) − Φ0 (G0 ) + Φ0 (Gn,hn ) − Φ0 (Gn,h̃n ) + oP (1/ n)
√
= (Pn − P0 )∆(P1 ) + Φ0 (Gn,hn ) − Φ0 (Gn,h̃n ) + oP (1/ n).
(36)
Step two: showing that the would-be first order term is complete. The rest of the proof consists
√
in showing that the term between brackets in (36), say T2,n , is oP (1/ n). The inequality below
follows from the definition of Φ0 and the triangle inequality, and the equality from (10) in A5:
|T2,n | ≤ Pn D∗ (Q1 , Gn,h̃n ) − D∗ (Q1 , Gn,hn )
+ (Pn − P0 ) D∗ (Q1 , Gn,hn ) − D∗ (Q1 , Gn,h̃n )
√
= Pn D∗ (Q1 , Gn,h̃n ) − D∗ (Q1 , Gn,hn ) + oP (1/ n).
Therefore, it suffices to prove that the absolute value in the above right-hand side expression is
√
oP (1/ n). Under A1(Q1 , hn , c5 ) (guaranteed by A5), for every 1 ≤ i ≤ n, the Taylor-Lagrange
inequality yields
D∗ (Q1 , Gn,h̃n ) − D∗ (Q1 , Gn,hn ) − (h̃n − hn ) × ∂hn D∗ (Q1 , Gn,· ) (Oi ) . (h̃n − hn )2
31
hence, by convexity,
Pn D∗ (Q1 , Gn,h̃n ) − D∗ (Q1 , Gn,hn ) − (h̃n − hn ) × ∂hn D∗ (Q1 , Gn,· ) . (h̃n − hn )2 .
Since Pn ∂hn D∗ (Q∗n,hn , Gn,· ) and (h̃n − hn ) are both oP (1/n1/4 ) by (6) in A2 and A5, we get
Pn D∗ (Q1 , Gn,h̃n ) − D∗ (Q1 , Gn,hn )
√
= (h̃n − hn ) × Pn ∂hn D∗ (Q∗n,hn , Gn,· ) − ∂hn D∗ (Q1 , Gn,· ) + oP (1/ n)
= (h̃n − hn ) × (Pn − P0 ) ∂hn D∗ (Q∗n,hn , Gn,· ) − ∂hn D∗ (Q1 , Gn,· )
√
+(h̃n − hn ) × P0 ∂hn D∗ (Q∗n,hn , Gn,· ) − ∂hn D∗ (Q1 , Gn,· ) + oP (1/ n).
√
In light of (11) and (12) in A5, the righ-hand side expression is oP (1/ n). This completes the
√
proof: T2,n = oP (1/ n), hence (36) rewrites as
√
T1,n = (Pn − P0 )∆(P1 ) + oP (1/ n),
and (13) finally follows from (35).
A.3
Proof of Lemma 3
Proof. Set q̄n ≡ Q̄n (1, ·) − Q̄n (0, ·), q̄1 ≡ Q̄1 (1, ·) − Q̄1 (0, ·), and ψ1 ≡ P0 q̄1 . Using inequality
(a + b)2 ≤ 2(a2 + b2 ) (valid for all real numbers a, b), we first remark that
P0 (q̄n − q̄1 )2 ≤ 2P0 [Q̄n (1, ·) − Q̄1 (1, ·)]2 + 2P0 [Q̄n (0, ·) − Q̄1 (0, ·)]2
= 2P0 (Q̄n − Q̄0 )2 /`G0 . P0 (Q̄n − Q̄0 )2 .
Therefore, it also holds that P0 (q̄n − q̄1 )2 = oP (1). Second, we decompose the difference ψn − ψ1 as
ψn − ψ1 = Pn q̄n − P0 q̄1 = (Pn − P0 )(q̄n − q̄1 ) + (Pn − P0 )q̄1 + P0 (q̄n − q̄1 ).
√
Lemma 19.24 in [32] guarantees that the first term in the above RHS expression is oP (1/ n).
Because q̄1 is uniformly bounded, the standard central limit theorem (for sequences of independent
and identically distributed, real-valued random variables with finite variance) implies that the
√
second term is OP (1/ n). Finally, the third term is oP (1) by the Cauchy-Schwarz inequality and
the remark we previously made. In summary, ψn − ψ1 = oP (1), as stated.
A.4
Proof of Corollary 4
The proof of Corollary 4 uses repeatedly the fact that some specific random functions fall in P0 Donsker classes with P0 -probability tending to one. Specifically, the proof will refer several times
to the following lemma (its proof is deferred to the end of this section).
Lemma 5. Suppose that the assumptions of Corollary 4 are met. Then, with P0 -probability
∗
tending to one, Q̄∗n,hn,κn (1, ·) − Q̄∗n,hn,κn (0, ·), D∗ (Pn,h
), D∗ (Q1 , Gn,hn,κn ) − D∗ (Q1 , Gn,h̃n ) and
n,κn
∂hn,κn D∗ (Q∗n,hn,κn , Gn,· ) − ∂hn,κn D∗ (Q∗n,hn,κn , Gn,· ) also fall in P0 -Donsker classes.
We can now present the proof of Corollary 4.
32
Proof of Corollary 4. There is no obvious counterpart in C1, C2 and C4 to (7) and (8) appearing
within A3. Yet, under C2 and C4, Gn,hn,κn consistently estimates G0 and Q̄∗n,hn,κn converges to
a limit Q̄1 that may differ from Q̄0 . Moreover, since Gn,hn,κn and G0 are bounded away from zero
by C1, we have
∗
) − D∗ (P1 ))2 = oP (1),
(37)
P0 (D∗ (Pn,h
n,κn
where P1 ∈ M is any element of model M of which the Q- and G-components equal Q1 ≡ (QW,0 , Q̄1 )
∗
and G0 (see proof below). By Lemma 5, D∗ (Pn,h
) falls in a P0 -Donsker class with P0 -probability
n,κn
tending to one. It thus holds that
√
∗
) − D∗ (P1 )) = oP (1/ n),
(Pn − P0 )(D∗ (Pn,h
n,κn
as requested in (7) of A3. In addition, the following convergence also occurs,
√
Rem20 (Q̄∗n,hn,κn , Gn,hn,κn ) − Rem20 (Q̄1 , Gn,hn,κn ) = oP (1/ n),
(38)
as requested in (8) of A3 (see proof below). Consequently, C1, C2 and C4 imply A3.
Let us now turn to assumption A5. To alleviate notation, let G00n,h (W ) be the second order
derivative of t 7→ Gn,t (W ) at h ∈ T under C1. Given the definition of D∗ (Q1 , Gn,t )(O), see (15),
t 7→ D∗ (Q1 , Gn,t )(O) is twice differentiable on T and, for each h ∈ T ,
!
G0n,h (W )2
G00n,h (W )
2 ∗
∂h D (Q1 , Gn,· )(O) = (Y − Q̄1 (A, W )) ×
− 2(2A − 1)
.
`Gn,h (A, W )2
`Gn,h (A, W )3
Obviously, under C1, there exists a universal constant C3 > 0 such that the supremum in h ∈ T of
∂h2 D∗ (Q1 , Gn,· )(O) is P0 -almost surely smaller than C3 . Consequently, assumption A1(Q̄1 , hn,κn , C3 )
is met. In addition, we show below that (10), (11) and (12) hold true whenever C1 to C4 are met.
∗
, see (28); A4 is assumed to hold true;
In summary, A2 is satisfied by construction of ψn,h
n,κn
A3 and A5 are met. Thus, Theorem 1 applies and implies the result stated in Corollary 4. This
completes the proof.
Proof of (37). Suppose that the assumptions of Corollary 4 are met and recall decomposition (15).
To alleviate notation, introduce Q̄n ≡ Q̄∗n,hn,κn , q̄n ≡ Q̄∗n,hn,κn (1, ·) − Q̄∗n,hn,κn (0, ·) and Gn ≡
Gn,hn,κn . By Lemma 5, q̄n falls in a P0 -Donsker class with P0 -probability tending to one. Using
repeatedly inequality (a + b)2 ≤ 2(a2 + b2 ), we obtain
∗
P0 (D∗ (Pn,h
) − D∗ (P1 ))2 . P0 (`Gn − `G0 )2 + P0 (Q̄n `G0 − Q̄1 `Gn )2
n,κn
∗
+P0 (q̄n − q̄1 )2 + (ψn,h
− ψ0 )2
n,κn
. P0 (Gn − G0 )2 + P0 (Q̄n − Q̄1 )2
∗
+P0 (q̄n − q̄1 )2 + (ψn,h
− ψ0 )2
n,κn
∗
= P0 (q̄n − q̄1 )2 + (ψn,h
− ψ0 )2 + oP (1).
n,κn
The assumptions of Lemma 3 are met too. Therefore, we can retrieve the bound
P0 (q̄n − q̄1 )2 . P0 (Q̄n − Q̄1 )2 = oP (1)
∗
from its proof and assert that its conclusion holds: (ψn,h
− ψ0 ) = oP (1). This completes the
n,κn
proof of (37).
33
Proof of (38). Suppose that the assumptions of Corollary 4 are met. In view of (16), we have
Tn ≡ |Rem20 (Q̄∗n,hn,κn , Gn,hn,κn ) − Rem20 (Q̄1 , Gn,hn,κn )|
`G0 (A, W )
∗
= EP0 (2A − 1) 1 −
(Q̄n,hn,κn (A, W ) − Q̄1 (A, W )) .
`Gn,hn,κn (A, W )
Therefore, the Cauchy-Schwarz inequality and equality (`Gn,hn,κn − `G0 )2 = (Gn,hn,κn − G0 )2 yield
Tn2 . P0 (Gn,hn,κn − G0 )2 × P0 (Q̄∗n,hn,κn − Q̄1 )2 = oP (1/n),
which completes the proof of (38).
Proof of (10) in the context of Section 3. Suppose that the assumptions of Corollary 4 are met.
Using C1 and the Taylor-Lagrange inequality yields
|Gn,h̃n (W ) − Gn,hn,κn (W )| . |h̃n − hn,κn |
hence P0 (Gn,h̃n − Gn,hn,κn )2 . (h̃n − hn,κn )2 = oP (1) (with much to spare). Now, observe that
D∗ (Q1 , Gn,hn,κn ) − D∗ (Q1 , Gn,h̃n ) (O) = (Y − Q̄1 (A, W ))(2A − 1)
×
which evidently implies the upper bound
D∗ (Q1 , Gn,hn,κn ) − D∗ (Q1 , Gn,h̃n ) (O)
1
1
−
`Gn,hn,κn (A, W ) `Gn,h̃n (A, W )
!
,
. |`Gn,hn,κn (A, W ) − `Gn,h̃n (A, W )|
= |Gn,hn,κn (W ) − Gn,h̃n (W )|.
Therefore, P0 (D∗ (Q1 , Gn,hn,κn )−D∗ (Q1 , Gn,h̃n ))2 = oP (1). Furthermore, Lemma 5 guarantees that
D∗ (Q1 , Gn,hn,κn ) − D∗ (Q1 , Gn,h̃n ) falls in a P0 -Donsker class with P0 -probability tending to one.
The same argument as the one that lead to (3) in Section 3.3 thus completes the proof of (10).
Proof of (11) in the context of Section 3. Suppose that the assumptions of Corollary 4 are met. In
view of (15), we have
∂hn,κn D∗ (Q∗n,hn,κn , Gn,· )(O) =
∂hn,κn D∗ (Q1 , Gn,· )(O) =
2A − 1
G0
(W )(Y − Q̄∗n,hn,κn (A, W )),
`Gn,hn,κn (A, W ) n,hn,κn
2A − 1
G0
(W )(Y − Q̄1 (A, W )),
`Gn,hn,κn (A, W ) n,hn,κn
(39)
(40)
hence
∂hn,κn D∗ (Qn,hn,κn , Gn,· ) − ∂hn,κn D∗ (Q1 , Gn,· ) . |Q̄1 − Q̄∗n,hn,κn |.
(41)
Therefore, the Cauchy-Schwarz inequality implies the bound
(hn,κn − h̃n ) × P0 ∂hn,κn D∗ (Qn,hn,κn , Gn,· ) − ∂hn,κn D∗ (Q1 , Gn,· )
2
. (hn,κn − h̃n )2 × P0 (Q̄∗n,hn,κn − Q̄1 )2 = oP (1/n),
thus completing the proof of (11).
34
Proof of (12) in the context of Section 3. Suppose that the assumptions of Corollary 4 are met.
By Lemma 5, ∂hn,κn D∗ (Q∗n,hn,κn , Gn,· ) − ∂hn,κn D∗ (Q1 , Gn,· ) falls in a P0 -Donsker class with P0 probability tending to one. In view of (41), it holds that
2
P0 ∂hn,κn D∗ (Qn,hn,κn , Gn,· ) − ∂hn,κn D∗ (Q1 , Gn,· ) = oP (1).
The same argument as the one that lead to (3) in Section 3.3 thus completes the proof of (12).
Proof of Lemma 5. We proceed by order of appearance in the statement of the lemma. First, note
that
Q̄∗n,hn,κn (1, W ) − Q̄∗n,hn,κn (0, W ) = (2A − 1)Q̄∗n,hn,κn (A, W ).
∗
Second, derive from (15) the explicit forms of D∗ (Pn,h
), D∗ (Q1 , Gn,hn,κn ), D∗ (Q1 , Gn,h̃n ), and
n,κn
of the difference of the two last ones. Third, recall the explicit forms of ∂hn,κn D∗ (Q∗n,hn,κn , Gn,· )
and D∗ (Q1 , Gn,hn,κn ) given in (39) and (40), and derive from them that of their difference. Thanks
to C4 and the above explicit forms, straightforward applications of [33, Theorem 2.10.6] yield the
result.
We conclude this article on a final remark about B4(Pn , k). Suppose that B2(Pn , k) is met. If,
in light of C1, we also assume that t 7→ Gn,t (W ) is twice differentiable in a neighborhood of hn,k ,
then the assumptions of the implicit function theorem are satisfied and h 7→ εn,h,k is differentiable
around hn,k .
References
[1] W. G. Cochran. The effectiveness of adjustment by subclassification in removing bias in
observational studies. Biometrics, pages 295–313, 1968.
[2] J. M. Franklin, W. Eddings, R. J. Glynn, and S. Schneeweiss. Regularized regression versus the
high-dimensional propensity score for confounding adjustment in secondary database analyses.
American journal of epidemiology, 187(7):651–659, 2015.
[3] J. Friedman, T. Hastie, and R. Tibshirani. Regularization paths for generalized linear models
via coordinate descent. Journal of Statistical Software, 33(1):1–22, 2010. URL http://www.
jstatsoft.org/v33/i01/.
[4] S. Gruber and M. J. van der Laan. An application of collaborative targeted maximum likelihood
estimation in causal inference and genomics. The International Journal of Biostatistics, 6(1):
Article 18, 2010.
[5] S. Gruber and M. J. van der Laan. ctmle: an r package for collaborative targeted maximum
likelihood estimation. Technical report, University of California, Berkeley, 2010.
[6] S. Gruber, R. W. Logan, I. Jarrı́n, S. Monge, and M. A. Hernán. Ensemble learning of inverse
probability weights for marginal structural modeling in large observational datasets. Statistics
in medicine, 34(1):106–117, 2015.
[7] D. Ho, K. Imai, G. King, and E. Stuart. Matchit: Matchit: Nonparametric preprocessing for
parametric casual inference. R package version, pages 2–2, 2006.
35
[8] D. E. Ho, K. Imai, G. King, and E. A. Stuart. Matching as nonparametric preprocessing for
reducing model dependence in parametric causal inference. Political analysis, 15(3):199–236,
2007.
[9] D. G. Horvitz and D. J. Thompson. A generalization of sampling without replacement from a
finite universe. Journal of the American statistical Association, 47(260):663–685, 1952.
[10] C. Ju, M. Combs, S. D. Lendle, J. M. Franklin, R. Wyss, S. Schneeweiss, and M. J. van der
Laan. Propensity score prediction for electronic healthcare databases using super learner and
high-dimensional propensity score methods. arXiv preprint arXiv:1703.02236, 2017.
[11] C. Ju, S. Gruber, S. D. Lendle, A. Chambaz, J. M. Franklin, R. Wyss, S. Schneeweiss, and
M. J. van der Laan. Scalable collaborative targeted learning for high-dimensional data. arXiv
preprint arXiv:1703.02237, 2017.
[12] C. Ju, J. Schwab, and M. J. van der Laan. On adaptive propensity score truncation in causal
inference. arXiv preprint arXiv:1707.05861, 2017.
[13] C. Ju, R. Wyss, J. M. Franklin, S. Schneeweiss, J. Häggström, and M. J. van der Laan.
Collaborative-controlled LASSO for constructing propensity score-based estimators in highdimensional data. Statistical methods in medical research, 2017.
[14] T. Kurth, A. M. Walker, R. J. Glynn, K. A. Chan, J. M. Gaziano, K. Berger, and J. M. Robins.
Results of multivariable logistic regression, propensity matching, propensity adjustment, and
propensity-based weighting under conditions of nonuniform effect. American Journal of Epidemiology, 163(3):262–270, 2006.
[15] B. K. Lee, J. Lessler, and E. A. Stuart. Improving propensity score weighting using machine
learning. Statistics in medicine, 29(3):337–346, 2010.
[16] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for
Statistical Computing, Vienna, Austria, 2017. URL https://www.R-project.org/.
[17] J. Robins. A new approach to causal inference in mortality studies with a sustained exposure
periodapplication to control of the healthy worker survivor effect. Mathematical Modelling, 7
(9-12):1393–1512, 1986.
[18] J. Robins. Robust estimation in sequentially ignorable missing data and causal inference
models. In Proceedings of the American Statistical Association: Section on Bayesian Statistical
Science, pages 6–10, 2000.
[19] J. M. Robins and A. Rotnitzky. Semiparametric efficiency in multivariate regression models
with missing data. Journal of the American Statistical Association, 90(429):122–129, 1995.
[20] J. M. Robins and A. Rotnitzky. Comment on the Bickel and Kwon article, ‘Inference for
semiparametric models: Some questions and an answer’. Statistica Sinica, 11(4):920–936,
2001.
[21] J. M. Robins, M. A. Hernan, and B. Brumback. Marginal structural models and causal
inference in epidemiology. Epidemiology, 11(5):550–560, 2000.
[22] J. M. Robins, A. Rotnitzky, and M. van der Laan. Comment on “On Profile Likelihood”
by S.A. Murphy and A.W. van der Vaart. Journal of the American Statistical Association –
Theory and Methods, 450:431–435, 2000.
36
[23] P. R. Rosenbaum and D. B. Rubin. The central role of the propensity score in observational
studies for causal effects. Biometrika, pages 41–55, 1983.
[24] P. R. Rosenbaum and D. B. Rubin. Reducing bias in observational studies using subclassification on the propensity score. Journal of the American statistical Association, 79(387):516–524,
1984.
[25] S. Schneeweiss, J. A. Rassen, R. J. Glynn, J. Avorn, H. Mogun, and M. A. Brookhart. Highdimensional propensity score adjustment in studies of treatment effects using health care claims
data. Epidemiology (Cambridge, Mass.), 20(4):512, 2009.
[26] M. E. Schnitzer and M. Cefalu. Collaborative targeted learning using regression shrinkage.
Statistics in Medicine, 2017.
[27] O. M. Stitelman and M. J. van der Laan. Collaborative targeted maximum likelihood for time
to event data. The International Journal of Biostatistics, 6(1):Article 21, 2010.
[28] M. van der Laan, A. Chambaz, and C. Ju. Targeted Learning in Data Science. Causal Inference
for Complex Longitudinal Studies, chapter CTMLE for continuous tuning. Springer Series in
Statistics. Springer, 2018.
[29] M. J. van der Laan and S. Rose. Targeted Learning: Causal Inference for Observational and
Experimental Data. Springer Science & Business Media, 2011.
[30] M. J. van der Laan and D. Rubin. Targeted maximum likelihood learning. The International
Journal of Biostatistics, 2(1), 2006.
[31] M. J. van der Laan, S. Gruber, et al. Collaborative double robust targeted maximum likelihood
estimation. The International Journal of Biostatistics, 6(1):Article 17, 2010.
[32] A. W. van der Vaart. Asymptotic statistics, volume 3 of Cambridge Series in Statistical and
Probabilistic Mathematics. Cambridge University Press, Cambridge, 1998.
[33] A. W. van der Vaart and J. A. Wellner. Weak convergence and empirical processes. Springer
Series in Statistics. Springer-Verlag, New York, 1996. With applications to statistics.
[34] H. Wang, S. Rose, and M. J. van der Laan. Finding quantitative trait loci genes with collaborative targeted maximum likelihood learning. Statistics & Probability Letters, 81(7):792–796,
2011.
37
p
100
200
p
100
200
bias
SE
MSE
bias
SE
MSE
bias
SE
MSE
bias
SE
MSE
ψnIPTW
0.252
0.167
0.091
0.297
0.159
0.114
∗
ψn,hn,CV
0.130
0.071
0.022
0.179
0.060
0.036
(ψnIPTW )0
0.078
0.325
0.112
0.106
0.480
0.242
∗
ψn,hn,κn
0.017
0.101
0.011
-0.037
0.140
0.021
ψnA-IPTW
0.357
0.073
0.133
0.417
0.067
0.178
(ψnA-IPTW )0
0.151
0.167
0.050
0.152
0.150
0.045
L-C-TMLE
LP-C-TMLE
LP-C-TMLE0
0.005
0.072
0.005
0.024
0.077
0.006
0.013
0.075
0.006
0.024
0.089
0.009
-0.000
0.094
0.009
-0.072
0.148
0.027
Table 7: Using LASSO-C-TMLE as a fine-tuning procedure. The performance of each
estimator at sample size n = 1000 with p ∈ {100, 200}. The prime symbol indicates the use of
Gn,hn,κn as an estimator of G0 in place of Gn,hn,CV or Gn,hn . Bias and SE are multiplied by 10, and
MSE is multiplied by 100.
38
| 10math.ST
|
Clustering implies geometry in networks
Dmitri Krioukov1
arXiv:1604.01575v2 [cond-mat.stat-mech] 19 May 2016
1
Northeastern University, Departments of Physics, Mathematics,
and Electrical&Computer Engineering, Boston, MA, USA
Network models with latent geometry have been used successfully in many applications in network
science and other disciplines, yet it is usually impossible to tell if a given real network is geometric,
meaning if it is a typical element in an ensemble of random geometric graphs. Here we identify
structural properties of networks that guarantee that random graphs having these properties are
geometric. Specifically we show that random graphs in which expected degree and clustering of
every node are fixed to some constants are equivalent to random geometric graphs on the real line,
if clustering is sufficiently strong. Large numbers of triangles, homogeneously distributed across all
nodes as in real networks, are thus a consequence of network geometricity. The methods we use to
prove this are quite general and applicable to other network ensembles, geometric or not, and to
certain problems in quantum gravity.
In equilibrium statistical mechanics it is often possible
to tell if a given system state is a typical state in a given
ensemble. In network science, where statistical mechanics methods have been used successfully in a variety of
applications [1–3], the same question is often intractable.
Stochastic network models define ensembles of random
graphs with usually intractable distributions. Therefore
it is usually unknown if a given real network is a typical
element in the ensemble of random graphs defined by a
given model, i.e., if the model is appropriate for the real
data, so that it can yield reliable predictions. Progress
has been made in addressing this problem in some classes
of models, such as the configuration [4–10] and stochastic
block models [11–13].
Here we are interested in latent-space network models [14]. In these models, nodes are assumed to populate
some latent geometric space, while the probability of connections between nodes is usually a decreasing function
of their distance in this space. Latent-space models were
first introduced in sociology in the 70ies [15] to model homophily in social networks—the more similar two people
are, the closer they are in a latent space, the more likely
they are connected [16]. Since then, latent-space models
have been used extensively in many applications, ranging from predicting social behavior and missing or future
links [17–19], to designing efficient information routing
algorithms in the Internet [20] and identifying connections in the brain critical for its function [21], to inferring community structure in networks [14]—see [22, 23]
for surveys.
The simplest network model with a latent space is the
model with the simplest latent space, which is the real
line R1 . Nodes are points sprinkled randomly on R1 , and
two nodes are connected if the distance between them on
R1 is below a certain threshold µ. This random graph
ensemble is known as the Gilbert model of random geometric graphs [24, 25]. Even in this simplest model,
the ensemble distribution is intractable and unknown.
Therefore it is impossible to tell if a given (real) network
is “geometric”—that is, if it is a typical element in the
ensemble. One can always check (in simulations) a subset of necessary conditions: if the network is geometric,
then all its structural properties must match the corresponding ensemble averages. By “network property” one
usually means a function of the adjacency matrix. The
simplest examples of such functions are the numbers of
edges, triangles, or subgraphs of different sizes in the network [26]. The distributions of betweenness or shortestpath lengths correspond to much less trivial functions of
adjacency matrices. Since the number of such propertyfunctions is infinite, and since their inter-dependencies
are in general intractable and unknown [26], it is impossible to check if all properties match and all conditions
necessary for network geometricity are satisfied. Do any
sufficient conditions exist? That is, are there any structural network properties such that random networks that
have these properties are typical elements in the ensemble of random geometric graphs?
Here we answer this question positively for random geometric graphs on R1 . We show that the set of sufficientcondition properties is surprisingly simple. These properties are only the expected numbers of edges k̄ and triangles t̄, or equivalently, expected degree k̄ and clustering c̄ = 2t̄/k̄ 2 of every node. Specifically, we consider a maximum-entropy ensemble of random graphs
in which the expected degree of every node is fixed to
the same value k̄, while the expected number of triangles to which every node belongs is also fixed to some
other value t̄. There is seemingly nothing geometric
about this ensemble since it is defined in purely networkstructural terms—edges and triangles, in combination
with the maximum-entropy principle [27, 28]. Yet we
show that if clustering is sufficiently strong, then this ensemble is equivalent to the ensemble of random geometric
graphs on R1 . In general, the ensemble is not sharp but
soft [29, 30]—the probability of connections is not 0 or
1 depending on if the distance between nodes is larger
or smaller than µ, but the grand canonical Fermi-Dirac
probability function in which energies of edges are distances they span on R1 . Strong clustering, a fundamen-
2
tally important property of real networks [31, 32], thus
appears as a consequence of their latent geometry.
The simplest model of networks with strong clustering
is the Strauss model [33] of random graphs with given
expected numbers of edges and triangles. The Strauss
model is well studied, but many of its problematic features, including degeneracy and phase transitions with
hysteresis caused by statistical dependency of edges and
non-convexity of the constraints, are not observed in real
networks [27, 34, 35]. In particular, in the Strauss model
all the triangles coalesce into a maximal clique, so that
a portion of nodes have a large degree and clustering
close to 1, while the rest of the nodes have a low degree and zero clustering [35, 36]. This clustering organization differs drastically from the one in real networks,
where triangles are homogeneously distributed across all
nodes, modulo Poisson fluctuations and structural constraints [37, 38]. If we want to fix the expected number
of edges and triangles of every node to the same values k̄ and t̄, then the Strauss model cannot be “fixed”
to accomplish this. Therefore instead we begin with the
canonical ensemble of random graphs in which every edge
{i, j} occurs, independently from other edges, with given
probability pij , which in general is different for different
edges. This ensemble is well-behaved and void of any
Strauss-like pathologies [2]. The expected degree hki i
and number of triangles
hti i at node i inPthe ensemble
P
are simply hki i = j pij and hti i = (1/2) j,k pij pjk pki .
Any connection probability matrix {pij } satisfying constraints hki i = k̄ and hti i = t̄ for some k̄, t̄ will yield a
canonical ensemble in which all nodes will have the same
expected degree k̄ and number of triangles t̄. However
we cannot claim that such an ensemble will be an unbiased ensemble with these constraints, because a particular matrix {pij } satisfying them may enforce additional
constraints on the expected values of some other network
properties. In other words, we first have to find a way
to sample matrices {pij } from some maximum-entropy
distribution subject only to the desired constraints.
This seemingly intractable problem finds a solution using the theory of graph limits known as graphons [39],
with basic formalism introduced in network models with
latent variables [40, 41]. Graphon p(x, y) is a symmetric
integrable function p : [0, 1]2 → [0, 1], which is essentially the thermodynamic n → ∞ limit of matrix {pij }.
For a fixed graph size n, graphon p defines graph ensemble Gn (p) by sprinkling n nodes uniformly at random on
interval [0, 1], and then connecting nodes i and j with
probability pij = p(xi , xj ), where xi , xj are sprinkled positions of i, j on [0, 1]. In the n → ∞ limit, the discrete node index i becomes continuous x ∈ [0, 1]. Graphs
in ensemble Gn (p) are dense, because the expected deR1
gree of a node at x ∈ [0, 1] is hk(x)i = n 0 p(x, y) dy.
Here we are interested in sparse ensembles, since most
real networks are sparse. Their average degrees are either constant or growing at most logarithmically with the
network size n [42]. To model sparse networks, one can
replace p(x, y) by a rescaled graphon pn (x, y) = p(x, y)/n
which depends on n [41, 43]. The expected degrees do not
then depend on n, but the
RR 1 number of triangles vanishes
as 1/n, ht(x)i = (1/2n) 0 p(x, y)p(y, z)p(z, x) dy dz, as
opposed to clustering in real networks, where it does not
depend on the size of growing networks either [42].
The solution to this impasse is a linearly growing support of graphon p. That is, let p : R2 → [0, 1] be a
graphon on the whole infinite plane R2 . For any finite
n we simply consider its restriction to a finite square
of size n × n, e.g., In2 , where In = [−n/2, n/2], so that
pn : In2 → [0, 1] and pn (x, y) = p(x, y). Graphon p(x, y)
is then the connection probability in the thermodynamic
limit. In this case, both the expected degree and number of triangles at any node in the thermodynamic
limit
R
can be finite and
positive:
hk(x)i
=
p(x,
y)
dy
and
R
RR
ht(x)i = (1/2) R2 p(x, y)p(y, z)p(z, x) dy dz. For a finite graph size n, the graph ensemble Gn (p) is defined
by sprinkling n points xi uniformly at random on interval In , and then connecting nodes i and j with probability pij = p(xi , xj ). The only difference between Gn (p)
and the infinite graph ensemble G∞ (p) in the thermodynamic limit is that in the latter case this sprinkling
is a realization Π = {xi } of the unit-rate Poisson point
process on the whole infinite real line R.
The main utility of using graphons here is that they
allow us to formalize our entropy-maximization task as
a variational problem which we will now formulate. We
first observe that for a fixed sprinkling Π, the connection probability matrix {pij } is also fixed. Since with
fixed {pij }, all edges are independent Bernoulli random
variables albeit with different success probabilities, the
entropy of a graph ensemble S[Gn (p|Π)] with fixed sprinkling Π
Pis the sum of entropies of all edges, S[Gn (p|Π)] =
(1/2) i,j h(pi,j ), where h(p) = −p log p − (1 − p) log(1 −
p) is the entropy of a Bernoulli random variable with the
success probability p. Unfixing Π now, the distribution
of entropy S[Gn (p|Π)] as a function of random sprinkling
Π in ensemble Gn (p) is known [44] to converge in the
thermodynamic limit to the delta function centered at
the graphon entropy s[p] defined below:
ZZ
1
h[p(x, y)] dx dy,
S[Gn (p|Π)] → S[Gn (p)] → s[p] =
2
R2
(1)
where S[Gn (p)] P
is the Gibbs entropy of ensemble Gn (p),
S[Gn (p)] = − G∈Gn (p) P (G) log P (G). Bernoulli entropy S[Gn (p|Π)] is thus self-averaging, and for large n,
any graph sampled from Gn (p) is a typical representative of the ensemble. The proof in [44] is for dense
graphons, but we show in the appendix that S[Gn (p|Π)]
is self-averaging in our sparse settings as well. Therefore,
our sparse ensemble Gn (p) is unbiased if it is defined by
graphon p∗ (x, y) that maximizes graphon entropy s[p]
above, subject to the constraints that the expected num-
3
bers of edges and triangles at every node are fixed to the
same values k̄, t̄,
Z
hk(x)i =
p(x, y) dy = k̄,
(2)
R
ZZ
1
ht(x)i =
p(x, y)p(y, z)p(z, x) dy dz = t̄.
(3)
2
R2
∗
To find graphon p (x, y) that maximizes entropy (1)
and satisfies constraints (2,3), we observe that constraint
(2) implies thatRp∗ (x, y) cannot be integrable since
RR
p(x, y) dx dy = k̄ R dx. Therefore we first have to
R2
solve the problem for finite n and then consider the thermodynamic limit. Using the method
RR of Lagrange multipliers, we define Lagrangian L = I 2 dx dy { 21 h[p(x, y)]+
n
R
λk p(x, y)+ 12 λt p(x, y) In p(y, z)p(z, x) dz} with Lagrange
multipliers λk , λt coupled to the degree and triangle constraints. Equation δL/δp = 0 leads to the following integral equation
Z
1
− 1 + 2λk + 3λt
p(x, z)p(z, y) dz = 0,
log
p(x, y)
In
(4)
which appears intractable. However, inspired by the
grand canonical formulation of edge-independent graph
ensembles [45], we next show that for sufficiently
large n, k̄, t̄, its approximate solution is the following
Fermi-Dirac graphon
(
1
1
if 0 ≤ r ≤ 1,
β(ε−µ) = 1+e2α(r−1/2)
∗
p (x, y) = 1+e1
(5)
1
∗
= 1+eα ≡ pα
if r > 1,
1+eβµ
where energy ε = |x − y| ≥ 0 of edge-fermion (x, y) is
the distance between nodes x and y on R1 , the chemical potential µ ≥ 0 and inverse temperature β ≥ 0 are
functions of k̄ and t̄, while α = βµ and r = ε/2µ are the
rescaled inverse temperature—the logarithm of thermodynamic activity—and energy-distance.
To show this, we first notice that if p∗ (x, y) is a solution, then the degree constraint (2) becomes
Z
k̄ =
p∗ (x, y) dy = 2µ + p∗α (n − 4µ) ≈ 2µ + p∗α n. (6)
In
Therefore if the average degree k̄ is fixed and does not
depend on n, then p∗α ∼ 1/n and α ∼ log n. If p∗α is
small, then the last integral term in (4)—the expected
number of common neighbors between nodes x and y—is
negligible for r > 1 (|x − y| > 2µ), and Eq. (4) simplifies to the equation for Erdős-Rényi graphs in which
only the expected degree is fixed. Its solution is constant
p∗ (x, y) = 1/(1 + e−2λk ), so that λk = −α/2, cf. (5).
If r < 1, then the common-neighbor integral in (4) is
no longer negligible, but we can evaluate it exactly for
p∗ (x, y). The exact expression for
Z
1
Cn (r, α) =
p∗ (x, z)p∗ (z, y) dz,
(7)
2µ In
1.0
1−r
Cn (r, 5)
Cn (r, 10)
Cn (r, 15)
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
r
0.6
0.8
1.0
FIG. 1. Rescaled number of common neighbors Cn (r, α) (7)
versus 1 − r for different values of α = βµ (µ = 1, n = 103 ).
where r = |x−y|/2µ, is terse and non-informative, so that
we omit it for brevity. Its important property is that for
large α it is closely approximated by Cn (r, α) ≈ 1 − r,
Fig. 1. In the α → ∞ limit this approximation becomes
exact since p∗ (x, y) → Θ(µ − |x − y|) = Θ(1/2 − r),
where Θ() is the Heaviside step function—x and y are
connected if |x − y| < µ. Approximating the commonneighbor integral in (4) by 2µ(1 − r), and noticing that
log(1/p∗ (x, y) − 1) = β(ε − µ) = 2α(r − 1/2), we transform (4) into
α(r − 1/2) + λk + 3µλt (1 − r) = 0.
(8)
This equation has a solution with λk = −α/2 and
λt = β/3. This solution is consistent with the solution in the r > 1 regime. First, the value of λk is the
same in both regimes r < 1 and r > 1. Second, one can
check
that the expected number of common neighbors
R
∗
p
(x,
z)p∗ (z, y) dz decays exponentially with α for any
In
r > 1. Therefore the common neighbor term in (4) is indeed negligible in the r > 1 regime, even though the
prefactor 3λt = α/µ is large for fixed µ and large α.
Figure 2 illustrates that if α is large, then the expected
average degree k̄ (6) and clustering
ZZ
1
2t̄
p∗ (x, y)p∗ (y, z)p∗ (z, x) dy dz
(9)
c̄ = 2 = 2
2
k̄
k̄
In
in ensemble Gn (p∗ ) are functions of only µ and α, respectively. Given values of the two constraints k̄ and t̄ (or c̄)
define the two ensemble parameters µ and β (or α) as the
solution of Eqs. (6,9). We note that for large α (α > 10
in Fig. 2), clustering is close to its maximum c̄max = 3/4
(t̄max = 3µ2 /2), which can be computed analytically.
Since our approximations are valid only for large α, they
apply only to graphs with strong clustering. In the sparse
thermodynamic limit n → ∞ with a finite average degree
k̄, the chemical potential µ must be finite and α must diverge (temperature T = 1/β must go to zero) because
of (6), so that only graphs with the strongest clustering
are the exact solution to our entropy-maximization problem. For finite n however, higher-temperature graphs
with weaker clustering are an approximate solution.
We emphasize that the fact that graphon (5), in which
the dependency on x and y is only via distance ε = |x−y|,
4
still homogeneously distributed across all nodes, albeit
subject to non-trivial structural constraints imposed by
the power-law degree distribution [26, 37, 38]. As shown
in [46, 47], random geometric graphs on R1 can be generalized to satisfy an additional constraint enforcing a
power-law degree distribution. This generalization still
uses the grand canonical Fermi-Dirac connection probability, albeit in hyperbolic geometry, and reproduces the
clustering organization in real networks. These observations lead to the conjecture that real scale-free networks
are typical elements in ensembles of soft random geometric graphs with non-trivial degree distribution constraints. If so, then non-trivial community structure, another common feature of real networks, is a reflection of
non-uniform node density in latent geometry [14, 48].
FIG. 2. Average degree (a,c) and clustering (b,d) in soft
random geometric graphs with connection probability (5) as
functions of µ and β. The dashed curves in (a,b) show the
simulation results averaged over 100 random graphs of size
n = 103 on the interval [−500, 500] with periodic boundary
conditions. The solid curves in (a,b) and color in (c,d) are
the (corresponding) analytic results using (6) in (a,c), and
numeric evaluation of (9) in (b,d) with n = 103 . The color
axes in (c,d) are in the logarithmic scale, with color ticks
evenly spaced in log k̄ in (c) and log(3/4 − c̄) in (d).
is an approximate entropy maximizer, means that the ensemble of random graphs in which the expected degree
and clustering of every node are fixed to given constants,
is approximately equivalent to the ensemble of soft random geometric graphs with the specific form of the connection probability, i.e., the grand canonical Fermi-Dirac
distribution function that maximizes ensemble entropy
constrained by fixed average energy and number of particles. In our ensemble, Fermi particles are graph edges
(0 or 1 edge between a pair of nodes), and their energy
is the distance they span on R1 . The average number
of particles m̄ = k̄n/2 is fixed by chemical potential µ.
Fixing average energy ε̄ and fixing the average number
of triangles t̄ are equivalent because the smaller the ε̄,
the more likely the lower-energy/smaller-distance states,
the larger the t̄ thanks to the triangle inequality in R1 .
This equivalence explains why the Fermi-Dirac distribution (5) appears as an approximate solution to our entropy maximization problem constrained by fixed k̄ and
t̄. In the zero-temperature limit β → ∞, graphon (5)
becomes the step function p∗ (x, y) = Θ(µ − ε), meaning that these soft random geometric graphs become the
traditional sharp random geometric graphs in which any
pair of nodes is connected if their distance-energy is at
most µ. All the approximations become exact in this
limit.
The degree distribution in (soft) random geometric
graphs is the Poisson distribution [25], while in many real
networks it is a power law. Triangles in real networks are
As a final remark we note that the graphon-based
methodology we developed here is quite general and can
be applied to other network models with latent variables,
geometric or not, to tell if a given model is adequate for
a given network. We also note that a very similar class of
problems underlies approaches to quantum gravity with
emerging geometry [23, 49] where one expects continuous spacetime to emerge in the classical limit from fundamentally discrete physics at the Planck scale. Perhaps
the most directly related example is the Hauptvermutung
problem in causal sets [50, 51]. Given a Lorentzian spacetime, causal sets are random geometric graphs in it with
edges connecting timelike-separated pairs of events sprinkled randomly onto the spacetime at the Planck density.
If no continuous spacetime is given to begin with, then
what discrete physics can lead to an ensemble of random
graphs equivalent to the ensemble of causal sets sprinkled onto the spacetime that we observe? To answer this
question, one has to solve the same ensemble equivalence
problem as we solved here, except not for R1 , but for the
spacetime of our Universe.
Appendix
Here we show that entropy of the considered sparse
graph ensemble is self-averaging. For completeness, we
first show that average entropy density converges to
graphon entropy density in the thermodynamic limit, and
then show that the relative variance (coefficient of variation) of the entropy distribution goes to zero in this limit.
We begin with notations and definitions.
Notations and definitions. Let In = [−n/2, n/2]
be the interval of length n, and Π = {xi }, i = 1, 2, . . . , n
be n real numbers sampled uniformly at random from
In . For large n, binomial sampling Π approximates the
Poisson point process of unit rate on In . Since every xi
is uniformly distributed on In , and since all xi s are independent, the probability density function of sprinklings
5
Π is
P (Π) =
1
.
nn
(10)
We impose the periodic boundary conditions on In making it a circle, so that the distance between point i and
j is
xij =
n
n
−
− |xi − xj | .
2
2
(11)
Distances {xij } are uniformly distributed on [0, n/2].
Given Π, ensemble Gn (p|Π) is the ensemble of graphs
whose edges, or elements of adjacency matrix {aij }, are
independent Bernoulli random variables: abusing notation for p, aij = 1 with probability pij = p(xi , xj ) =
p(xij ), and aij = 0 with probability 1 − pij . There are
no self-edges, so that pii = 0. The entropy of random
variable aij is h(pij ), where
h(x) = −x log x − (1 − x) log(1 − x)
(12)
is the entropy of the Bernoulli random variable with success probability x. Since all aij s are independent in ensemble Gn (p|Π), its entropy is
Sn ≡ Sn (Π) ≡ S [Gn (p|Π)] =
n
1 X
h(pij ),
2 i,j=1
(13)
which is fixed for a given sprinkling Π. Ensemble Gn (p) is
the ensemble of graphs sampled by first sampling random
sprinkling Π, and then sampling a random graph from
Gn (p|Π).
We consider entropy Sn as a random variable defined
by Π. This random variable is self-averaging if its relative
variance vanishes in the thermodynamic limit,
p
hSn − hSn ii2
cv =
−−−−→ 0,
(14)
n→∞
hSn i
where h·i stands for averaging across random sprinklings Π. We first show that average entropy density—
that is, average entropy per node hSn i/n—converges to
graphon entropy density σ,
hSn i
σ
−−−−→ ,
n→∞
n
2
(15)
Z
σ = lim σn =
h [p(xij )] dxij ,
n→∞
R
Z
sn
σn =
=
h [p(xij )] dxij ,
n
In
ZZ
sn =
h [p(xi , xj )] dxi dxj ,
(16)
Average ensemble entropy density converges to
graphon entropy density. Using the definitions and
observations above, we get
Z
Y
hSn i
1
=
Sn (Π)P (Π)
dxk
(19)
n
n Inn
k
Z X
Y
1
hSn i
=
h [p(xi , xj )] P (Π)
dxk
(20)
n
2n Inn i,j
k
Z X
Y
1
= n+1
h [p(xi , xj )]
dxk .
(21)
2n
n
In
i,j
k
The integration over n − 2 variables xk with indices k not
equal to either i or j yields the factor of nn−2 :
ZZ
1 X
hSn i
h [p(xi , xj )] dxi dxj ,
(22)
= 3
n
2n i,j
2
In
where we have also swapped the summation and integration. Changing variables from xi and xj to xij and xj ,
and integrating over xj yields another factor of n:
Z
hSn i
1 X
= 2
h [p(xij )] dxij .
(23)
n
2n i,j In
Since xij are uniformly distributed on [0, n/2], all terms
in the sum contribute equally, bringing another factor of
n(n − 1) ≈ n2 , the total number of terms in the sum:
Z
hSn i
1
σn
=
h [p(xij )] dxij =
.
(24)
n
2 In
2
We thus have that
1
σ
hSn i
−−−−→
lim σn = .
(25)
n n→∞ 2 n→∞
2
Ensemble entropy is self-averaging. To compute
p
hSn2 i − hSn i2
cv =
(26)
hSn i
we must calculate
2 +
*
X
1
hSn2 i =
h [p(xi , xj )]
2 i,j
(27)
= I1 + I2 , where
(28)
Z X
Y
1
I1 =
h2 [p(xi , xj )] P (Π)
dxm ,
(29)
4 Inn i,j
m
Z X
Y
1
I2 =
h [p(xi , xj )] h [p(xk , xl )] P (Π)
dxm .
4 Inn
m
i,j;k,l
(30)
(17)
(18)
2
In
and then prove (14). For notational convenience in the
equations above, we have extended the support of p(xij )
from R+ to R by p(−xij ) = p(xij ).
The first integral I1 is different from (20) only in that
instead of h/2 we now have (h/2)2 . Therefore we immediately conclude that
n
I1 = γn , where
(31)
Z4
h2 [p(x)] dx.
γn =
In
(32)
6
R
If σn = In h [p(x)] dx converges to finite σ in the n → ∞
limit, as it does for the Fermi-Dirac p∗ , then so does γn ,
γn −−−−→ γ < ∞, because h(p) ∈ [0, 1].
[5] G. Bianconi, P. Pin, and M. Marsili, Proc Natl Acad Sci
USA 106, 11433 (2009).
[6] K. Anand and G. Bianconi, Phys Rev E 80, 045102(R)
n→∞
(2009).
To calculate the second integral I2 , we use P (Π) =
[7] K. Anand, D. Krioukov, and G. Bianconi, Phys Rev E
1/nn and integrate over n − 4 variables xm with indices
89, 062807 (2014).
[8] D. Garlaschelli and M. Loffredo, Phys Rev E 78, 015101
m not equal to any i, j, k, l, bringing the factor of nn−4 :
(2008).
ZZZZ
1 X
[9] D. Garlaschelli and M. Loffredo, Phys Rev Lett 102,
h [p(xi , xj )] h [p(xk , xl )] dxi dxj dxk dxl . 38701 (2009).
I2 = 4
4n
4
In
i,j;k,l
[10] T. Squartini, R. Mastrandrea, and D. Garlaschelli, New
(33)
J Phys 17, 023052 (2015).
Changing variables from xi , xj , xk , and xl , to xij , xj ,
[11] T. P. Peixoto, Phys Rev E 85, 056122 (2012).
xkl , and xl , and integrating over xj and xl , thus bringing
[12] T. P. Peixoto, Phys Rev Lett 110, 148701 (2013).
[13] T. P. Peixoto, Phys Rev X 4, 011047 (2014).
another factor of n2 , we get:
[14] M. E. J. Newman and T. P. Peixoto, Phys Rev Lett 115,
Z
Z
1 X
088701 (2015).
I2 = 2
h [p(xij )] h [p(xkl )] dxij dxkl . (34)
[15] D. D. McFarland and D. J. Brown, in Bonds of Pluralism:
4n
2
In
i,j;k,l
The Form and Substance of Urban Social Networks (John
Wiley, New York, 1973) pp. 213–252.
Since xij and xkl are independent and uniformly dis[16] M. McPherson, L. Smith-Lovin, and J. M. Cook, Annu
tributed on [0, n/2], every term in the double sum conRev Sociol 27, 415 (2001).
tributes equally, while the total number of terms is
[17] P. D. Hoff, A. E. Raftery, and M. S. Handcock, J Am
Stat Assoc 97, 1090 (2002).
[n(n − 1)]2 ≈ n4 , yielding
[18] P. Sarkar, D. Chakrabarti, and A. W. Moore, in IJCAI
ZZ
(2011) pp. 2722–2727.
n2
I2 =
h [p(xij )] h [p(xkl )] dxij dxkl
(35)
[19]
G.
Tita, P. J. Brantingham, A. Galstyan, and Y.-S. Cho,
4
2
In
Discret
Contin Dyn Syst - Ser B 19, 1335 (2014).
Z
2
[20] M. Boguñá, F. Papadopoulos, and D. Krioukov, Nat
n
=
h [p(x)] dx
(36)
Commun 1, 62 (2010).
2 In
[21] A. Gulyás, J. J. Bı́ró, A. Kőrösi, G. Rétvári, and D. Kri n 2
oukov, Nat Commun 6, 7651 (2015).
σn = hSn i2 .
=
(37)
[22] M. Barthélemy, Phys Rep 499, 1 (2011).
2
[23] G. Bianconi, EPL 111, 56001 (2015).
Collecting the calculations of hSn i and hSn2 i, we finally
[24] E. N. Gilbert, J Soc Ind Appl Math 9, 533 (1961).
obtain
[25] M. Penrose, Random Geometric Graphs (Oxford University Press, Oxford, 2003).
p
p
√
γn
I1 + I2 − hSn i2
(n/4)γn
1
[26] C. Orsini, M. M. Dankulov, P. Colomer-de Simón, A. Ja=
cv =
=√
. (38)
makovic, P. Mahadevan, A. Vahdat, K. E. Bassler,
hSn i
(n/2)σn
n σn
Z. Toroczkai, M. Boguñá, G. Caldarelli, S. Fortunato,
and D. Krioukov, Nat Commun 6, 8627 (2015).
If σn −−−−→ σ < ∞, then γn −−−−→ γ < ∞, and
n→∞
n→∞
[27] S. Horvát, É. Czabarka, and Z. Toroczkai, Phys Rev Lett
√
114, 158701 (2015).
γn
1
cv = √
−−−−→ 0.
(39)
[28] T. Squartini, J. de Mol, F. den Hollander, and D. Garn σn n→∞
laschelli, Phys Rev Lett 115, 268701 (2015).
[29] C. P. Dettmann and O. Georgiou, Phys Rev E 93, 032313
(2016).
[30] M. Penrose, Ann Appl Probab 26, 986 (2016).
[31] F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and
We thank G. Lippner, P. Topalov, M. Piskunov,
D. Parisi, Proc Natl Acad Sci 101, 2658 (2004).
M. Kitsak, M. Boguñá, S. Horvát, Z. Toroczkai, and
[32] F. Radicchi and C. Castellano, Phys Rev E 93, 030302
Y. Baryshnikov for useful discussions and suggestions.
(2016).
This work was supported by NSF CNS-1442999.
[33] D. Strauss, SIAM Rev 28, 513 (1986).
[34] D. Foster, J. Foster, M. Paczuski, and P. Grassberger,
Phys Rev E 81, 046115 (2010).
[35] J. Park and M. E. J. Newman, Phys Rev E 72, 026136
(2005).
[1] R. Albert and A.-L. Barabási, Rev Mod Phys 74, 47
[36] C. Radin, K. Ren, and L. Sadun, J Phys A Math Theor
(2002).
47, 175001 (2014).
[2] J. Park and M. E. J. Newman, Phys Rev E 70, 66117
[37] P. Colomer-de Simón, M. Á. Serrano, M. G. Beiró, J. I.
(2004).
Alvarez-Hamelin, and M. Boguñá, Sci Rep 3, 2517
[3] J. Gao, B. Barzel, and A.-L. Barabási, Nature 530, 307
(2013).
(2016).
[38] V. Zlatić, D. Garlaschelli, and G. Caldarelli, EPL 97,
[4] G. Bianconi, Eur Lett 81, 28005 (2008).
7
28005 (2012).
[39] L. Lovász, Large Networks and Graph Limits (American
Mathematical Society, Providence, RI, 2012).
[40] G. Caldarelli, A. Capocci, P. D. L. Rios, and M. A.
Muñoz, Phys Rev Lett 89, 258702 (2002).
[41] M. Boguñá and R. Pastor-Satorras, Phys Rev E 68,
36112 (2003).
[42] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and
D.-U. Hwanga, Phys Rep 424, 175 (2006).
[43] C. Borgs, J. T. Chayes, H. Cohn, and Y. Zhao,
arXiv:1401.2906.
[44] S. Janson, NYJM Monogr 4 (2013), arXiv:1009.2376.
[45] D. Garlaschelli, S. E. Ahnert, T. M. A. Fink, and G. Cal-
darelli, Entropy 15, 3148 (2013).
[46] M. Á. Serrano, D. Krioukov, and M. Boguñá, Phys Rev
Lett 100, 78701 (2008).
[47] D. Krioukov, F. Papadopoulos, M. Kitsak, A. Vahdat,
and M. Boguñá, Phys Rev E 82, 36106 (2010).
[48] K. Zuev, M. Boguñá, G. Bianconi, and D. Krioukov, Sci
Rep 5, 9421 (2015).
[49] Z. Wu, G. Menichetti, C. Rahmede, and G. Bianconi,
Sci Rep 5, 10073 (2015).
[50] R. Sorkin, in Lectures on Quantum Gravity, edited by
A. Gomberoff and D. Marol (Springer, New York, 2005)
pp. 305–328.
[51] L. Bombelli, J. Lee, D. Meyer, and R. Sorkin, Phys Rev
Lett 59, 521 (1987).
| 10math.ST
|
Consensus Attention-based Neural Networks for Chinese Reading
Comprehension
Yiming Cui†∗ , Ting Liu‡ , Zhipeng Chen† , Shijin Wang† and Guoping Hu†
†
iFLYTEK Research, Beijing, China
‡
Research Center for Social Computing and Information Retrieval,
Harbin Institute of Technology, Harbin, China
†
{ymcui,zpchen,sjwang3,gphu}@iflytek.com
‡
[email protected]
arXiv:1607.02250v3 [cs.CL] 15 Mar 2018
Abstract
Reading comprehension has embraced a booming in recent NLP research. Several institutes
have released the Cloze-style reading comprehension data, and these have greatly accelerated the
research of machine comprehension. In this work, we firstly present Chinese reading comprehension datasets, which consist of People Daily news dataset and Children’s Fairy Tale (CFT)
dataset. Also, we propose a consensus attention-based neural network architecture to tackle the
Cloze-style reading comprehension problem, which aims to induce a consensus attention over
every words in the query. Experimental results show that the proposed neural network significantly outperforms the state-of-the-art baselines in several public datasets. Furthermore, we
setup a baseline for Chinese reading comprehension task, and hopefully this would speed up the
process for future research.
1
Introduction
The ultimate goal of machine intelligence is to read and comprehend human languages. Among various machine comprehension tasks, in recent research, the Cloze-style reading comprehension task has
attracted lots of researchers. The Cloze-style reading comprehension problem (Taylor, 1953) aims to
comprehend the given context or document, and then answer the questions based on the nature of the
document, while the answer is a single word in the document. Thus, the Cloze-style reading comprehension can be described as a triple:
hD, Q, Ai
where D is the document, Q is the query and A is the answer to the query.
By adopting attention-based neural network approaches (Bahdanau et al., 2014), the machine is able to
learn the relationships between document, query and answer. But, as is known to all, the neural network
based approaches need large-scale training data to train a reliable model for predictions. Hermann et al.
(2015) published the CNN/Daily Mail news corpus for Cloze-style reading comprehensions, where the
content is formed by the news articles and its summarization. Also, Hill et al. (2015) released the Children’s Book Test (CBT) corpus for further research, where the training samples are generated through
automatic approaches. As we can see that, automatically generating large-scale training data for neural
network training is essential for reading comprehension. Furthermore, more difficult problems, such as
reasoning or summarization of context, need much more data to learn the higher-level interactions.
Though we have seen many improvements on these public datasets, some researchers suggested that
these dataset requires less high-level inference than expected (Chen et al., 2016). Furthermore, the public
datasets are all automatically generated, which indicate that the pattern in training and testing phase are
nearly the same, and this will be easier for the machine to learn these patterns.
In this paper, we will release Chinese reading comprehension datasets, including People Daily news
datasets and Children’s Fairy Tale datasets. As a highlight in our datasets, there is a human evaluated
∗
This work was done by the Joint Laboratory of HIT and iFLYTEK (HFL).
This work is licensed under a Creative Commons Attribution 4.0 International Licence.
http://creativecommons.org/licenses/by/4.0/
Licence details:
dataset for testing purpose. And this will be harder for the machine to answer these questions than the
automatically generated questions, because the human evaluated dataset is further processed, and may
not be accordance with the pattern of automatic questions. More detailed analysis will be given in the
following sections. The main contributions of this paper are as follows:
• To our knowledge, this is the first released Chinese reading comprehension datasets and human
evaluated test sets, which will benefit the research communities in reading comprehension.
• Also, we propose a refined neural network that aims to utilize full representations of query to deal
with the Cloze-style reading comprehension task, and our model outperform various state-of-the-art
baseline systems in public datasets.
The rest of the paper will be organized as follows. In Section 2, we will briefly introduce the existing
Cloze-style datasets, and describe our Chinese reading comprehension datasets in detail. In Section
3, we will show our refined neural network architecture for Cloze-style reading comprehension. The
experimental results on public datasets as well as our Chinese reading comprehension datasets will be
given in Section 4. Related work will be described in Section 5, and we make a brief conclusion of our
work at the end of this paper.
2
Chinese Reading Comprehension Datasets
We first begin with a brief introduction of the existing Cloze-style reading comprehension datasets, and
then introduce our Chinese reading comprehension datasets: People Daily and Children’s Fairy Tale.
2.1
Existing Cloze-style Datasets
Typically, there are two main genres of the Cloze-style datasets publicly available, which all stem from
the English reading materials.
CNN/Daily Mail.1 The news articles often come with a short summary of the whole report. In the spirit
of this, Hermann et al. (2015) constructed a large dataset with web-crawled CNN and Daily Mail news
data. Firstly, they regard the main body of the news article as the Document, and the Query is formed
through the summary of the article, where one entity word is replaced by a placeholder to indicate the
missing word. And finally, the replaced entity word will be the Answer of the Query. Also, they have
proposed the anonymize the named entity tokens in the data, and re-shuffle the entity tokens for every
sample in order to exploit general relationships between anonymized named entities, rather than the
common knowledge. But as Chen et al. (2016)’s studies on these datasets showed that the anonymization
is less useful than expected.
Children’s Book Test. 2 There was also a dataset called the Children’s Book Test (CBT) released by
Hill et al. (2015), which is built from the children’s book story. Different from the previously published
CNN/Daily Mail datasets, they formed the Document with 20 consecutive sentences in the book, and
regard the 21st sentence as the Query, where one word is blanked with a placeholder. The missing word
are chosen from named entities (NE), common nouns (CN), verbs and prepositions. As the verbs and
prepositions are less dependent with the document, most of the studies are focusing on the NE and CN
datasets.
2.2
People Daily and Children’s Fairy Tale Datasets
In this part, we will introduce our Chinese reading comprehension datasets in detail3 . Though many solid
works on previously described public datasets, there is no studies on Chinese reading comprehension
datasets. What makes our datasets different from previous works are listed as below.
• As far as we know, the proposed dataset is the first Chinese Cloze-style reading comprehension
datasets, which will add language diversity in the community.
1
The pre-processed CNN and Daily Mail datasets are available at http://cs.nyu.edu/˜kcho/DMQA/
The CBT datasets are available at http://www.thespermwhale.com/jaseweston/babi/CBTest.tgz
3
Our datasets are available at http://hfl.iflytek.com/chinese-rc/.
2
Document
Query
Answer
1 ||| People Daily (Jan 1). According to report of “New York
Times”, the Wall Street stock market continued to rise as the global
stock market in the last day of 2013, ending with the highest
1 ||| 人民网 1月 1日 讯 据 《 纽约 时报 》 报道 ， 美国 华尔街 股市
record or near record of this year.
在 2013年 的 最后 一 天 继续 上涨 ， 和 全球 股市 一样 ， 都 以 最
2 ||| “New York times” reported that the S&P 500 index rose
高 纪录 或 接近 最高 纪录 结束 本年 的 交易 。
29.6% this year, which is the largest increase since 1997.
2 ||| 《 纽约 时报 》 报道 说 ， 标普 500 指数 今年 上升 29.6% ，
3 ||| Dow Jones industrial average index rose 26.5%, which is the
为 1997年 以来 的 最 大 涨幅 ；
largest increase since 1996.
3 ||| 道琼斯 工业 平均 指数 上升 26.5% ， 为 1996年 以来 的 最 大
4 ||| NASDAQ rose 38.3%.
涨幅 ；
5 ||| In terms of December 31, due to the prospects in employment
4 ||| 纳斯达克 上涨 38.3% 。
and possible acceleration of economy next year, there is a rising
5 ||| 就 12月 31日 来说 ， 由于 就业 前景 看好 和 经济 增长 明年 可 confidence in consumers.
能 加速 ， 消费者 信心 上升 。
6 ||| As reported by Business Association report, consumer
6 ||| 工商 协进会 报告 ， 12月 消费者 信心 上升 到 78.1 ， 明显 高
confidence rose to 78.1 in December, significantly higher than 72
于 11月 的 72 。
in November.
7 ||| 另 据 《 华尔街 日报 》 报道 ， 2013年 是 1995年 以来 美国
7 ||| Also as “Wall Street journal” reported that 2013 is the best U.S.
股市 表现 最 好 的 一 年 。
stock market since 1995.
8 ||| 这 一 年 里 ， 投资 美国 股市 的 明智 做法 是 追 着 “ 傻钱 ” 跑 。
8 ||| In this year, to chase the “silly money” is the most wise way to
9 ||| 所谓 的 “ 傻钱 ” XXXXX， 其实 就 是 买 入 并 持有 美国 股票
invest in U.S. stock.
这样 的 普通 组合 。
9 ||| The so-called “silly money” is that, to buy and hold the
10 ||| 这个 策略 要 比 对冲 基金 和 其它 专业 投资者 使用 的 更为 复
common combination of U.S. stock.
杂 的 投资 方法 效果 好 得 多 。
10 ||| This strategy is better than other complex investment
methods, such as hedge funds and the methods adopted by other
professional investors.
所谓 的 “ 傻钱 ” XXXXX， 其实 就 是 买 入 并 持有 美国 股票 这样
的 普通 组合 。
The so-called “silly money” XXXXX is that, to buy and hold the
common combination of U.S. stock.
策略
strategy
Figure 1: Example training sample in People Daily datasets (the English translation is given in the
right box). The ”XXXXX“ represents the missing word. In this example, the document consists of 10
sentences, and the 9th sentence is chosen as the query.
• We provide a large-scale Chinese reading comprehension data in news domain, as well as its validation and test data as the in-domain test.
• Further, we release two out-of-domain test sets, and it deserves to highlight that one of the test sets
is made by the humans, which makes it harder to answer than the automatically generated test set.
People Daily. We roughly collected 60K news articles from the People Daily website4 . Following Liu
et al. (2016), we process the news articles into the triple form hD, Q, Ai. The detailed procedures are as
follows.
• Given a certain document D, which is composed by a set of sentences D = {s1 , s2 , ..., sn }, we
randomly choose an answer word A in the document. Note that, we restrict the answer word A
to be a noun, as well as the answer word should appear at least twice in the document. The partof-speech and sentence segmentation is identified using LTP Toolkit (Che et al., 2010). We do not
distinguish the named entities and common nouns as Hill et al. (2015) did.
• Second, after the answer word A is chosen, the sentence that contains A is defined as the query Q,
in which the answer word A is replaced by a specific placeholder hXi.
• Third, given the query Q and document D, the target of the prediction is to recover the answer A.
In this way, we can generate tremendous triples of hD, Q, Ai for training the proposed neural network,
without any assumptions on the nature of the original corpus. Note that, unlike the previous work, using
the method mentioned above, the document can be re-used for different queries, which makes it more
general to generate large-scale training data for neural network training. Figure 1 shows an example of
People Daily datasets.
Children’s Fairy Tale. Except for the validation and test set of People Daily news data, we also present
two out-of-domain test sets as well. The two out-of-domain test sets are made from the Children’s Fairy
Tale (CFT), which is fairly different from the news genre. The reason why we set out-of-domain test sets
is that, the children’s fairy tale mainly consists of the stories of animals or virtualized characters, and
4
http://www.people.com.cn
People Daily
Train Valid
Test
# Query
Max # tokens in docs
Max # tokens in query
Avg # tokens in docs
Avg # tokens in query
Vocabulary
870,710
618
502
379
38
3,000
536
153
425
38
248,160
Children’s Fairy Tale
Test-auto Test-human
3,000
634
265
410
41
1,646
318
83
122
20
1,953
414
92
153
20
N/A
Table 1: Statistics of People Daily datasets and Children’s Fairy Tale datasets.
this prevents us from utilizing the gender information and world knowledge in the training data, which
is important when solving several types of questions, such as coreference resolutions etc.
In CFT dataset, one test set is automatically generated using the algorithms described above, and the
other one is made by the human, which suggest that the latter is harder than the former one. Because
the automatically generated test sets are aware of the co-occurence or fixed collocation of words, and
thus when the pattern around the query blank exactly appeared in the document, it is much easier for the
machine to identify the correct answer. While in building human evaluation test set, we have eliminated
these types of samples, which makes it harder for the machine to comprehend. Intuitively, the human
evaluation test set is harder than any other previously published Cloze-style test sets.
The statistics of People Daily news datasets as well as Children’s Fairy Tale datasets are listed in the
Table 1.
3
Consensus Attention Sum Reader
In this section, we will introduce our attention-based neural network model for Cloze-style reading comprehension task, namely Consensus Attention Sum Reader (CAS Reader). Our model is primarily motivated by Kadlec et al. (2016), which aims to directly estimate the answer from the document, instead of
making a prediction over the full vocabularies. But we have noticed that by just concatenating the final
representations of the query RNN states are not enough for representing the whole information of query.
So we propose to utilize every time slices of query, and make a consensus attention among different
steps.
Formally, when given a set of training triple hD, Q, Ai, we will construct our network in the following
way. We first convert one-hot representation of the document D and query Q into continuous representations with a shared embedding matrix We . As the query is typically shorter than the document, by
sharing the embedding weights, the query representation can be benefited from the embedding learning
in the document side, which is better than separating embedding matrices individually.
Then we use two different bi-directional RNNs to get the contextual representations of document and
query, which can capture the contextual information both in history and future. In our implementation,
we use the bi-directional Gated Recurrent Unit (GRU) for modeling. (Cho et al., 2014)
e(x) = We ∗ x, where x ∈ D, Q
(1)
→
−
−−−→
hs = GRU (e(x))
(2)
←
− ←−−−
hs = GRU (e(x))
(3)
→
− ←
−
hs = [hs ; hs ]
(4)
We take hdoc and hquery to represent the contextual representations of document and query, both of
which are in 3-dimension tensor shape. After that, we directly make a dot product of hdoc and hquery (t)
to get the “importance” of each document word, in respect to the query word at time t. And then,
!(“$%&'”|*, ,) =
Sum Attention
Layer
/
01 = 02 + 04
1∈7(“89:;”,<)
Merging Function
Individual
Attention Layer
bi-GRU Layer
Embedding
Layer
Mary
sits
beside
...
says
he
love
Mary
Document
he
loves
<BLANK>
Query
Figure 2: Architecture of the proposed Consensus Attention Sum Reader (CAS Reader).
we use the softmax function to get a probability distribution α over the document hdoc , also known as
“attention”.
α(t) = sof tmax(hdoc
hquery (t))
(5)
In this way, for every time step t in the query, we can get a probability distribution over the document,
denoted as α(t), where α(t) = [α(t)1 , α(t)1 , ..., α(t)n ], α(t)i means the attention value of ith word in
the document at time t, and n is the length of the document. To get a consensus attention over these
individual attentions, we explicitly define a merging function f over α(1)...α(m). We denote this as
s = f (α(1), ..., α(m))
(6)
where s is the final attention over the document, m is the length of the query. In this paper, we define
the merging function f as one of three heuristics, shown in equations below.
m
P
sof
tmax(
α(t)),
if mode = sum;
t=1
m
P
s ∝ sof tmax( 1
(7)
α(t)), if mode = avg;
m
t=1
sof tmax( max α(t)), if mode = max.
t=1...m
Finally, we map the attention result s to the vocabulary space V , and sum the attention value which
occurs in different place of the document but shares the same word, as Kadlec et al. (2016) do.
P (w|D, Q) =
X
si , w ∈ V
(8)
i∈I(w,D)
where I(w, D) indicate the position that word w appear in the document D. Figure 2 shows the
proposed neural network architecture.
4
4.1
Experiments
Experimental Setups
Training details of neural network models are illustrated as follows.
Embed. # units
Hidden # units
Dropout
384
384
384
256
256
384
384
256
None
None
None
0.1
CNN News
CBTest NE
CBTest CN
People Daily & CFT
Table 2: Other neural network setups for each task. Note that, the dropout is only applied to the output
of the GRUs.
CNN News
Train Valid
# Query
Max # candidates
Avg # candidates
Avg # tokens
Vocabulary
380,298
527
26
762
3,924
187
26
763
118,497
Test
3,198
396
25
716
CBT NE
Train Valid
108,719
10
10
433
2,000
10
10
412
53,063
Test
2,500
10
10
424
CBT CN
Train Valid
120,769
10
10
470
2,000
10
10
448
53,185
Test
2,500
10
10
461
Table 3: Statistics of public Cloze-style reading comprehension datasets: CNN news data and CBTest
NE(Named Entites) / CN(Common Nouns).
• Embedding Layer: We use randomly initialized embedding matrix with uniformed distribution in
the interval [−0.1, 0.1]. Note that, no pre-trained word embeddings are used in our experiments.
• Hidden Layer: We initialized the GRU units with random orthogonal matrices (Saxe et al., 2013).
As GRU still suffers from the gradient exploding problem, we set gradient clipping threshold to 10
in our experiments (Pascanu et al., 2013) .
• Vocabulary: For training efficiency and generalization, in People Daily and CFT datasets, we truncate the full vocabulary (about 200K) and set a shortlist of 100K. All unknown words are mapped to
10 different specific symbols using the method proposed by Liu et al. (2016). There is no vocabulary
truncation in CNN and CBTest dataset.
• Optimization: We used the ADAM update rule (Kingma and Ba, 2014) with an initial learning rate
lr = 0.0005, and used negative log-likelihood as the training objective function. The batch size is
set to 32.
Other neural network setups, such as dimensions of embedding layer and hidden layer, and dropout
(Srivastava et al., 2014) for each task, are listed in Table 2. We trained model for several epochs and
choose the best model according to the performance of validation set. All models are trained on Tesla
K40 GPU. Our model is implemented with Theano (Theano Development Team, 2016) and Keras (Chollet, 2015).
4.2
Results on Public Datasets
To verify the effectiveness of our proposed model, we first tested our model on public datasets. Our
evaluation is carried out on CNN news datasets (Hermann et al., 2015) and CBTest NE/CN datasets (Hill
et al., 2015), and the statistics of these datasets are listed in Table 3. No pre-processing is done with these
datasets. The experimental results are given in Table 4. We evaluate the model in terms of its accuracy.
Due to the time limitations, we did not evaluate our model in ensemble.
CNN News. The performance on CNN news datasets shows that our model is on par with the Attention
Sum Reader, with 0.4% decrease in validation and 0.5% improvements in the test set. But we failed to
outperform the Stanford AR model. While the Stanford AR utilized GloVe embeddings (Pennington et
Deep LSTM Reader†
Attentive Reader†
Impatient Reader†
Human (context+query)‡
LSTMs (context+query)‡
MemNN (window + self-sup.)‡
Stanford AR\
AS Reader]
CAS Reader (mode: avg)
CNN News
Valid Test
CBTest NE
Valid Test
CBTest CN
Valid Test
55.0
61.6
61.8
63.4
72.4
68.6
68.2
51.2
70.4
73.8
74.2
62.6
64.2
68.8
68.2
57.0
63.0
63.8
66.8
72.4
69.5
70.0
81.6
41.8
66.6
68.6
69.2
81.6
56.0
63.0
63.4
65.7
Table 4: Results on the CNN news, CBTest NE (named entity) and CN (common noun) datasets. Results
marked with † are taken from (Hermann et al., 2015), and ‡ are taken from (Hill et al., 2015), and \ are
taken from (Chen et al., 2016), and ] are taken from (Kadlec et al., 2016)
People Daily
Valid Test
AS Reader
CAS Reader (mode: avg)
CAS Reader (mode: sum)
CAS Reader (mode: max)
64.1
65.2
64.7
63.3
Children’s Fairy Tale
Test-auto Test-human
67.2
68.1
66.8
65.4
40.9
41.3
43.0
38.3
33.1
35.0
34.7
32.0
Table 5: Results on People Daily datasets and Children’s Fairy Tale (CFT) datasets.
al., 2014), and only normalized the probabilities over the named entities in the document, rather than all
the words, and this could make a difference in the results. But in our model, we do not optimize for a
certain type of dataset, which make it more general.
CBTest NE/CN. In CBTest NE dataset, our model gives slight improvements over AS Reader, where
0.4% improvements in the validation set and 0.6% improvements in the test set. In CBTest CN, though
there is a slight drop in the validation set with 0.6% declines, there is a boost in the test set with an
absolute improvements 2.3%, which suggest our model is effective, and it is beneficial to consider every
slices of the query when answering.
4.3
Results on Chinese Reading Comprehension Datasets
The results on Chinese reading comprehension datasets are listed in Table 5. As we can see that, the
proposed CAS Reader significantly outperform the AS Reader in all types of test set, with a maximum
improvements 2.1% on the CFT test-auto dataset. The results indicate that making a consensus attention
over multiple time steps are better than just relying on single attention (as AS Reader did). This is similar
to the use of “model ensemble”, which is also a consensus voting result by different models.
We also evaluated different merging functions. From the results, we can see that the avg and sum
methods significantly outperform the max heuristics, and the max heuristics failed to outperform the
AS Reader. A possible reason can be explained that the max operation is very sensitive to the noise. If
a non-answer word is given to a high probability in one time step of the query, the avg and sum could
easily diminish this noise by averaging/summing over other time steps. But once there is a higher value
given to a non-answer word in max situation, the noise can not be removed, and will preserve till the
end of final attentions, which will influence the predictions a lot.
Also, we have noticed that, though we have achieved over 65% in accuracy among People Daily
datasets, there is a significant drop in the two CFT test sets. Furthermore, the the human evaluated test
set meets a sharp decline over 8% accuracy to the automatically generated test set. The analyses can be
concluded as
• As we regard the CFT datasets as the out-of-domain tests, there is a gap between the training data
and CFT test data, which poses declines in these test sets. Such problems can be remedied by
introducing the similar genre of training data.
• Regardless of the absolute accuracies in CFT datasets, the human test set is much harder for the
machine to read and comprehend as we discussed before. Through these results, we can see that
there is a big gap between the automatically generated queries and the human-selected questions.
Note that, in our human-evaluated test set, the query is also formulated from the original sentence in
the document, which suggest that if we use more general form of queries, there should be another rise in
the comprehension difficulties. For example, instead of asking “I went to the XXXXX this morning .”,
we change into a general question form of “Where did I go this morning ?”, which makes it harder for
the machine to comprehend, because there is a gap between the general question form and the training
data.
5
Related Work
Many NN-based reading comprehension models have been proposed, and all of them are attention-based
models, which indicate that attention mechanism is essential in machine comprehensions.
Hermann et al. (2015) have proposed a methodology for obtaining a large quantities of hD, Q, Ai
triples. By using this method, a large number of training data can be obtained without much human
intervention, and make it possible to train a reliable neural network to study the inner relationships inside
of these triples. They used attention-based neural networks for this task. Evaluation on CNN/DailyMail
datasets showed that their approach is effective than traditional baselines.
Hill et al. (2015) also proposed a similar approach for large scale training data collections for children’s book reading comprehension task. By using window-based memory network and self-supervision
heuristics, they have surpass all other methods in predicting named entities(NE) and common nouns(CN)
on both the CBT and the CNN QA benchmark.
Our CAS Reader is closely related to the work by Kadlec et al. (2016). They proposed to use a
simple model that using the attention result to directly pick the answer from the document, rather than
computing the weighted sum representation of document using attention weights like the previous works.
The proposed model is typically motivated by Pointer Network (Vinyals et al., 2015). This model aims
to solve one particular task, where the answer is only a single word and should appear in the document
at least once. Experimental results show that their model outperforms previously proposed models by a
large margin in public datasets (both CBTest NE/CN and CNN/DailyMail datasets).
Liu et al. (2016) proposed an effective way to generate and exploit large-scale pseudo training data for
zero pronoun resolution task. The main idea behind their approach is to automatically generate largescale pseudo training data and then using the neural network model to resolve zero pronouns. They also
propose a two-step training: a pre-training phase and an adaptation phase, and this can be also applied to
other tasks as well. The experimental results on OntoNotes 5.0 corpus is encouraging and the proposed
approach significantly outperforms the state-of-the-art methods.
In our work, we proposed an entirely new Chinese reading comprehension dataset, which add the diversity to the existing Cloze-style reading comprehension datasets. Moreover, we propose a refined neural network model, called Consensus Attention-based Sum Reader. Though many impressive progress
has been made in these public datasets, we believe that the current machine comprehensions are still
in the pre-mature stage. As we have discussed in the previous section, to answer a pseudo query to
the document is not enough for machine comprehension. The general question form can be seen as a
comprehensive processing of our human brains. Though our human-evaluated test set is still somewhat
easy for machine to comprehend (but harder than the automatically generated test set), releasing such
dataset will let us move a step forward to the real-world questions, and becomes a good bridge between
automatic questions and real-world questions.
6
Conclusion
In this paper, we introduce the first Chinese reading comprehension datasets: People Daily and Children’s
Fairy Tale. Furthermore, we also propose a neural network model to handle the Cloze-style reading comprehension problems. Our model is able to take all question words into accounts, when computing the
attentions over the document. Among many public datasets, our model could give significant improvements over various state-of-the-art baselines. And also we set up a baseline for our Chinese reading
comprehension datasets, that we hopefully make it as a starter in future studies.
The future work will be carried out in the following aspects. First, we would like to work on another
human-evaluated dataset, which will contain the real-world questions and is far more difficult than the
existing datasets publicly available. Second, we are going to investigate hybrid reading comprehension
models to tackle the problems that rely on comprehensive induction of several sentences.
Acknowledgements
We would like to thank the anonymous reviewers for their thorough reviewing and proposing thoughtful
comments to improve our paper. This work was supported by the National 863 Leading Technology
Research Project via grant 2015AA015407, Key Projects of National Natural Science Foundation of
China via grant 61632011, and National Natural Science Youth Foundation of China via grant 61502120.
References
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to
align and translate. arXiv preprint arXiv:1409.0473.
Wanxiang Che, Zhenghua Li, and Ting Liu. 2010. Ltp: A chinese language technology platform. In Proceedings
of the 23rd International Conference on Computational Linguistics: Demonstrations, pages 13–16. Association
for Computational Linguistics.
Danqi Chen, Jason Bolton, and Christopher D. Manning. 2016. A thorough examination of the cnn/daily mail
reading comprehension task. In Association for Computational Linguistics (ACL).
Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk,
and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine
translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing
(EMNLP), pages 1724–1734. Association for Computational Linguistics.
François Chollet. 2015. Keras. https://github.com/fchollet/keras.
Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and
Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems, pages 1684–1692.
Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. 2015. The goldilocks principle: Reading children’s
books with explicit memory representations. arXiv preprint arXiv:1511.02301.
Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. 2016. Text understanding with the attention
sum reader network. arXiv preprint arXiv:1603.01547.
Diederik Kingma and Jimmy Ba.
arXiv:1412.6980.
2014.
Adam: A method for stochastic optimization.
arXiv preprint
Ting Liu, Yiming Cui, Qingyu Yin, Shijin Wang, Weinan Zhang, and Guoping Hu. 2016. Generating and exploiting large-scale pseudo training data for zero pronoun resolution. arXiv preprint arXiv:1606.01603.
Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. 2013. On the difficulty of training recurrent neural networks. ICML (3), 28:1310–1318.
Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP),
pages 1532–1543. Association for Computational Linguistics.
Andrew M Saxe, James L McClelland, and Surya Ganguli. 2013. Exact solutions to the nonlinear dynamics of
learning in deep linear neural networks. arXiv preprint arXiv:1312.6120.
Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout:
a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–
1958.
Wilson L Taylor. 1953. Cloze procedure: a new tool for measuring readability. Journalism and Mass Communication Quarterly, 30(4):415.
Theano Development Team. 2016. Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.02688, May.
Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In Advances in Neural Information
Processing Systems, pages 2692–2700.
| 9cs.NE
|
DNA Self-Assembly For Constructing 3D Boxes
Ming-Yang Kao1,⋆ and Vijay Ramachandran2,⋆⋆
arXiv:cs/0112009v1 [cs.CC] 8 Dec 2001
1
2
Department of Computer Science, Northwestern University, Evanston, IL 60201,
USA, [email protected]
Department of Computer Science, Yale University, New Haven, CT 06520-8285,
USA, [email protected]
Abstract. We propose a mathematical model of DNA self-assembly using 2D tiles to form 3D nanostructures. This is the first work to combine
studies in self-assembly and nanotechnology in 3D, just as Rothemund
and Winfree did in the 2D case. Our model is a more precise superset of their Tile Assembly Model that facilitates building scalable 3D
molecules. Under our model, we present algorithms to build a hollow
cube, which is intuitively one of the simplest 3D structures to construct.
We also introduce five basic measures of complexity to analyze these algorithms. Our model and algorithmic techniques are applicable to more
complex 2D and 3D nanostructures.
1
Introduction
DNA nanotechnology and DNA self-assembly are two related technologies with
enormous potentials.
The goal of DNA nanotechnology is to construct small objects with high precision. Seeman’s visionary work [8] in 1982 pioneered the molecular units used
in self-assembly of such objects. More than a decade later, double-crossover
(DX) molecules were proposed by Fu and Seeman [3] and triple-crossover (TX)
molecules by LaBean et al. [5] as DNA self-assembly building blocks. Laboratory efforts have been successful in generating interesting three-dimensional (3D)
molecular structures, including the small cube of Chen and Seeman [1]. However,
these are immutable and limited in size, mainly because their fabrication is not
based on a mathematical model that can be extended as necessary.
In parallel to DNA nanotechnology, studies on self-assembly of DNA tiles
have focused on using local deterministic binding rules to perform computations. These rules are based on interactions between exposed DNA sequences
on individual tiles; tiles assemble into a particular 1D or 2D structure when
in solution, encoding a computation. Winfree [10] formulated a model for 2D
computations using DX molecules. Winfree et al. [11] used 1D tiles for 1D computations and 2D constructions with DX molecules. LaBean et al. [4] were the
first to compute with TX molecules.
⋆
⋆⋆
Supported in part by NSF Grants CCR-9531028 and EIA-0112934. Part of this work
was performed while this author was visiting the Department of Computer Science,
Yale University, New Haven, CT 06520-8285, USA, [email protected].
Supported by a 2001 National Defense Science and Engineering Graduate Fellowship.
Combining these two technologies, several researchers have demonstrated
the power of DNA self-assembly in nanostructure fabrication. Winfree et al. [13]
investigated how to use self-assembly of DX molecules to build 2D lattice DNA
crystals. Rothemund and Winfree [7] further proposed a mathematical model
and a complexity measure for building such 2D structures.
A natural extension of the seminal 2D results of Winfree et al. [13] and
Rothemund and Winfree [7] would be the creation of 3D nanostructures using
tiling. To initiate such an extension, this paper (1) proposes a general mathematical model for constructing 3D structures from 2D tiles; (2) identifies a set
of biological and algorithmic issues basic to the implementation of this model;
and (3) provides basic computational concepts and techniques to address these
issues. Under the model, the paper focuses on the problem of constructing a
hollow cube, which is intuitively one of the simplest 3D structures to construct.
We present algorithms for the problem and analyze them in terms of five basic
measures of complexity.
There are three natural approaches to building a hollow cube. The first approach uses 1D tiles to form 2D DX-type tiles as in [11], and then uses these tiles
to construct a cube. Our paper does not fully investigate this possibility because
of the incovenient shape of these molecules (see Sect. 2.1), but our algorithms can
be modified to accommodate these DX-type tiles. The second approach builds a
cube from genuine 2D tiles, which is the focus of this paper. The third approach
is perhaps the most natural: build a cube from genuine 3D tiles. It is not yet
clear how such 3D tiles could be created; conceivably, the cube of Chen and
Seeman [1] may lead to tiles of this form. This paper does not fully investigate
this possibility, either, because this approach is algorithmically straightforward
and similar to the 2D case.
The basic idea of our algorithms is to use 2D tiles to form a shape on the
plane that can fold into a box, as illustrated in Fig. 1(a)–(b). We can easily
synthesize a set of tiles to create the intitial 2D shape. To overcome a negligible
probability of success due to biochemical factors, we must put many copies of
these tiles into solution at once; but we must then worry about multiple copies
of the shape interfering with each other, preventing folding, as in Figure 1(c).
To avoid this problem, we introduce randomization, so that different copies of
the shape have unique sticky ends. The growth of tiles into a complete structure
must still be deterministic (as it is based on Watson-Crick hybridization), but
we randomize the computation input — the seed tiles from which the rest of the
shape assembles. The edges then still relate to each other, but depend on the
random input that is different for each shape in solution. If each input can form
with only low probability, interference with another copy of the shape will be
kept to a minimum.
This raises another important issue — that of using self-assembly to communicate information from one part of the shape to another. Since the edges
must relate to each other and the random input, designing local rules becomes
nontrivial. In this paper, we explore and formalize patterns used in completing
this task. In addition, we formalize biological steps that allow a specific subset
1
4
1
2
3
4
5
1
5
3
2
6
6
7
7
2
2
4
3
3
1
1
5
2 2
6
1
7
(a)
(b)
(c)
Fig. 1. (a) 2D planar shape that will fold into a box. Each section is formed from
many smaller 2D DNA tiles. Edges with the same number have complementary
sticky ends exposed so they can hybridize. (b) Folding of the shape in (a) into
a box. Here, edges 4, 5, 6, and 7 have all hybridized. Hybridization of edges 2
and 3, whose two complements are now in close proximity, will cause edge 1 to
hybridize and form the complete box. (c) Multiple copies of the 2D shape in
solution. Copies of the shape can interfere and attach infinitely without control
as long as edges have matching sticky ends.
of tiles to be added in an isolated period of time, thus allowing better control of
growth. We couple this with the use of temperature to improve the probability
of a successful construction.
The remainder of this paper is organized as follows. Section 2 describes the
model of computation, including notation for DNA tiles and definitions of complexity measures. Section 3 describes the algorithms in detail, and Section 4
discusses future research possibilities.
2
Model of Computation
In this section we formally introduce our model of self-assembly, the Generalized
Tile Assembly Model, on both the mathematical and biological level. It is an
extension of the model presented by Rothemund and Winfree in [7].
2.1
Molecular Units of Self-Assembly
We begin with the biological foundation for our model. We intend to build 3D
structures using the folding technique shown in Fig. 1 and allow construction of
all 2D structures possible with the Tile Assembly Model.
Our model relies on using the molecular building block of a DNA tile. Tiles
can naturally hybridize to form stable shapes of varying sizes, and the individual
tiles can easily be customized and replicated (via synthesis and PCR before the
procedure) for a specific algorithm.
DNA tiles are small nucleotides with exposed action sites (also known as
sticky ends of a DNA strand) consisting of a single-stranded sequence of base
pairs. When this sequence matches a complementary sequence on an action site
of another tile, the Watson-Crick hybridization property of DNA causes these
two molecules to bind together, forming a larger structure. A tile can be synthesized in the laboratory to have specific sticky ends. Different combinations
of sticky ends on a tile essentially yield uniquely-shaped puzzle pieces. The tiles
will automatically hybridize when left in solution.
Most work in self-assembly uses DX and TX molecules for tiles, but the shape
of these molecules causes a problem for 3D construction. Since the sticky ends
are on diagonally opposite ends (see [3] and [5]), these tiles form structures with
ragged edges when they hybridize, as in Figure 2(a). Our algorithms can easily
be modified to use these tiles by adjusting for proper alignment before folding
into a box.
(a)
(b) sticky end
(c)
sticky end
individual DX molecules
Fig. 2. (a) 2D structure formed from DX molecules. The left and right sides
cannot hybridize because they are aligned improperly; the same is true for the
top and bottom sides. (b) Branched-molecule DNA tile from [8]. (c) Snynthetic
DNA tile derived from the structure of tRNA.
However, we propose a simpler alternative, which is using the branched
molecules of Seeman [8] or a variant derived from the structure of tRNA. These
molecules, sketched in Figures 2(b) and (c), are truly 2D with sticky ends on four
sides. The structure is stable while the sticky ends are free-floating in solution
— so the molecules have flexibility to align properly during folding.
Such molecules offer a natural motivation for modeling them using Wang’s
theory of tiling [9], which allows us to abstract construction using these molecules
to a symbolic level.
2.2
Symbolic Representation of Tiles
Definition 1. A DNA sequence of length n is an ordered sequence of base pairs
5′ − b1 b2 · · · bn − 3′ where the sequence has a 5-prime and 3-prime end, and
bi ∈ B = {A, T, C, G}, the set of base pairs. We will assume that if the directions
are not explicity written, the sequence is written in the 5′ → 3′ direction.
1. The Watson-Crick complement of sequence s = 5′ − b1 b2 · · · bn − 3′ , denoted
s, is the sequence1 3′ − b1 b2 · · · bn − 5′ , where A = T, C = G. Define s = s.
2. The concatenation of two sequences s = s1 · · · sn and t = t1 · · · tm , denoted
s · t, or simply st, is the sequence s1 · · · sn t1 · · · tm .
3. The subsequence from i to j of sequence s = 5′ − b1 b2 · · · bn − 3′ , denoted
s[i : j], is the sequence 5′ − bi bi+1 · · · bj−1 bj − 3′ , where 1 ≤ i < j ≤ n.
Given the above definitions, two DNA strands can hybridize if they have complementary sequences. Formally, s = s1 · · · sn and t = t1 · · · tm can hybridize if
there exist integers hs1 , hs2 , ht1 , ht2 such that s[hs1 : hs2 ] = t[ht1 : ht2 ]. We assume there are no misbindings, that is, the above condition must be met exactly
with no errors in base-pair binding.
Remark 2. Note that st (or (s · t) ) 6= st (or s · t); rather, st = t · s.
Definition 3. The threshold temperature for a DNA sequence is a temperature t
in some fixed set T such that the sequence is unable to remain stably hybridized
to its complement when the solution is at a temperature higher than t′ ∈ (t −
ǫ, t + ǫ) for ǫ > 0.2 (Heating a solution generally denatures strands, so this
definition has strong biological foundation. The consequences and methodology
of using temperature in designing DNA sequences for tiles is discussed in [2].) If
s has a lower threshold temperature than t, we say s binds weaker than t.
As with most work in DNA computing, our model uses DNA sequences to
encode information3 — in our case, an identifier specifying what kinds of matches
are allowed between tiles on a given side. Since there are no misbindings, these
identifiers map uniquely to DNA sequences present on the sides of tiles that can
bind to each other. Formally, we have the following.
Definition 4. Let S be the set of symbols used to represent the patterns on
the sides of our tiles. We assume S is closed under complementation, that is, if
s ∈ S then there exists some s′ ∈ S such that s′ = s where
S s is the complement
of s (the purpose of this will be clear below). Let W ⊂ i B i be the set of DNA
sequences called DNA words such that the words do not interfere with each
other or themselves (i.e., bind inappropriately). We then define the injective
map enc : S → W that is the encoding of a symbol into a DNA word. This map
obeys complementation: enc (s) = enc(s).
1
2
3
The assumption that sequences written without directions are given 5′ → 3′ means
that the complement of b1 b2 · · · bn is bn · · · b2 b1 , which is not standard convention
but is technically correct.
A fixed set of threshold temperatures simplifies the model and corresponds to the
temperature parameter in [7]. To compensate we allow the actual threshold temperature to deviate slightly from the fixed point.
Condon, Corn, and Marathe [2] have done work on designing good DNA sequences
for problems like this one.
Definition 5. A DNA tile is a 4-tuple of symbols T = (sN , sE , sS , sW ) such
that si ∈ S and enc(si ) is the exposed DNA sequence at the north, east, south,
or west action site of the tile, for i = N, E, S, W . Given two tiles T1 and T2 , they
will bind if two sides have complementary symbols. Properties of hybridization,
including threshold temperature, carry over to the hybridization of tiles. We
make a stronger no-misbinding assumption for tiles, requiring that the sticky
ends on the tiles match exactly and fully.
At this stage, our model exactly matches that of Rothemund and Winfree in
[7], except that our tiles can “rotate”; that is, (sN , sE , sS , sW ) = (sE , sS , sW , sN ).
This corresponds more closely to tile structure. The model, at this point, could
require many symbols to express different tile types, and possibly an exponential
number of DNA words. Ideally, we would like to arbitrarily extend the symbolic
or informational content of each side of a tile. Therefore we make the following
generalization.
Definition 6. Let Σ be a set of symbols closed under complementation, and let
Ω be a set of corresponding DNA words. A k-level generalization of the model
defines a map g : Σ k → S and a corresponding encoding genc : Σ k → W, where
genc(σ) = enc(g(σ)) for σ ∈ Σ k such that an abstract tile definition, which is a
4-tuple of k-tuples of symbols in Σ, is equivalent to a DNA tile.
We define complementation for a k-tuple in Σ k as follows: let σ = (σ1 , . . . , σk )
be (σ1 , . . . , σk ) so that genc(σ) = enc(g(σ)) = enc(g(σ)) = enc(g(σ)). This
makes the hybridization condition equivalent to having complementary symbols
in k-tuples for sides that will bind.
The definition is purposefully broad in order to allow different algorithms to
define the encoding based on the number of words and tiles needed. A 1-level
generalization with Σ = S and Ω = W where g(s) = s and genc = enc is the
original Rothemund-Winfree Tile Assembly Model. In this paper, we use the
following model.
Definition 7. The concatenation generalization is a k-level generalization where
W ⊂ Ω k and g maps every combination of symbols in Σ k to a unique symbol in
S. Partition S into S ′ and S ′ such that each set contains the complement symbols of the other, and S ′ ∩ S ′ = ∅.4 For σ ∈ S ′ , define genc(g −1 (σ)) = enc(σ) =
ω1 ω2 · · · ωk , where enc(σi ) = ωi ∈ Ω and g −1 (σ) = (σ1 , σ2 , . . . , σk ). Then for
σ ∈ S ′ , let enc(σ) = ωk · ωk−1 · · · ω1 , so that genc(g −1 (σ)) = genc(g −1 (σ)).
In other words, the concatenation generalization model is a straightforward extension of the tile model where each side of a tile corresponds to a k-tuple of
symbols, where the DNA sequence at the corresponding action site is simply
4
The map enc as defined will be one-to-one, and so is g, and so this can be done
since the DNA sequence corresponding to any symbol has a unique complement,
and therefore a unique complement symbol.
the concatenation of the encodings of the individual symbols.5 Using this simple
model, we can reduce the number of DNA words needed to |Σ| from |Σ|k , and
create simpler descriptions of our tiles.
2.3
Algorithmic Procedures
With the above models for tiles, we now discuss procedures for growing larger
structures.
We follow Rothemund and Winfree [7] and Markov [6] and use the common
self-assembly assumption that a structure begins with a seed tile and grows, at
each timestep, by hybridization with another free-floating tile.6 The new tile
hybridizes at a given position following one of two types of rules:
Deterministic Given the surrounding tiles at that position, only one tile type,
with specific sticky ends on the non-binding sides, can fit.
Randomized Multiple tile types (with different sticky ends on the non-binding
sides) could fit the position given the tiles present; a new action site is created
with probability proportional to the concentration of its tile type in solution.
Therefore, to grow a structure, an algorithm repeats steps until the structure is complete: add tiles to solution; wait for them to adhere to the growing
structure; optionally removes excess tiles from solution by “washing them away.”
Cycling temperature during these steps to prevent or induce binding (based on
threshold temperatures) can be done while waiting for hybridization, and is
called temperature-sensitive binding.
2.4
Complexity
We consider five basic methods of analyzing algorithms using our model.
Time complexity Each algorithm is a sequence of self-assembly steps described
above, thus the natural measure of time complexity in our model is the number of steps required (which describes laboratory time).
Space complexity The number of distinct physical tile types (not the actual
number of molecules produced) is space complexity. Introduced by [7], this
describes the amount of unique DNA synthesis necessary.
Alphabet size The number of DNA words, or |Ω| or |W|, has a rough laboratory limit [2], and so the size of the symbol set used (|Σ| or |S|), which
corresponds directly to the number of words, has practical significance.
5
6
Potentially, the concatenation model could cause interference among tiles. If we
maintain no misbindings, however, our model removes this from analysis. In addition,
it is theoretically possible to design DNA words so interference does not occur,
depending on the algorithm.
In reality, multiple tiles can hybridize at once and structures consisting of more than
one tile can hybridize to each other, but we lose no generality with the Markov
assumption.
Generalization level The generalization level is the amount of information on
a side of a tile. This is related to the length of the sticky ends (and thus has
biological consequences) and the number of actual DNA words (via |S|).
Probability of misformation Misformed structures contain tiles that are not
bound properly on all sides. Assuming the Markov model, consider adding
tile T to a partial structure S. If complete hybridization requires binding
on two sides, but T manages to hybridize only on one side (while the other
action site does not match), S + T has a misformation. We quantify this
probability with the following.
Definition 8. Let the success probability at step t be the probability that a
free-floating tile in solution binds at all possible sides to a partial structure
at a given spot. (Step t is the addition of a tile to that spot on structure St ,
resulting in St+1 .) This is
Pr(St+1 is correct | St is correct) =
Ncorrect
,
Nall
where Ncorrect is the number of tile types that can correctly bind, while
Nall is the number of tile types in solution that could bind, possibly even
incompletely. Call this qt . Then the misformation probability at step t is
pt = 1−qt . If the algorithm has k additions, then the misformation probability
for the algorithm is 1 − q0 q1 · · · qk−1 . Then an algorithm is misformationproof if its misformation probability at every step is zero, yielding a zero
total probability of misformation.
3
Hollow Cube Algorithms
In this section, we examine algorithms designed to use our model to build a 3D
hollow cube using the folding technique shown in Fig. 1. Let the length of a side
of the cube, n, be the input to the algorithms. We present the most interesting
algorithm in detail and discuss some others briefly.
3.1
Overview
Figure 3(a) illustrates the planar shape our algorithms construct. We will reference the labels and shading of regions in the figure during our discussion.
As stated earlier in Sect. 1, we must make each shape unique so different
partial structures in solution do not bind to and interfere with each other. Once
we have a unique seed structure, we can then use self-assembly with basic rules
to make the edges of the shape correspond so folding will occur.
There are three basic self-assembly patterns used to construct different parts
of the shape.
Random assembly Implements a random rule (see Sect. 2.3). Formally, add
all tiles in a set of distinct tiles R in equal concentrations so each could
base strip
R Q
L
H P
G B
K E C
F
D
J
I
(a)
O
(before
growth)
pattern
pattern
pattern
(after
growth)
(b)
pattern
M N A
(arrow shows
growth direction)
(c)
Fig. 3. (a) Regions of the 2D planar shape. (b) A straight-copy pattern. (c)
A turn-copy pattern.
potentially hybridize completely at a given position. Thus the information
at that position is completely random. The tiles differ by a component of
their exposed k-tuples, assuming a k-level generalization.
Straight copy See Figure 3(b). Tiles are added to copy the pattern along one
end of a region through to a parallel end of an adjacent region being constructed. This rule is deterministic.
Turn copy See Figure 3(c). Tiles are added to copy the pattern along one end
of a region to a perpendicular end of an adjacent region being constructed.
Counters will be required to position the tiles appropriately to complete this
deterministic rule.
The algorithm begins by assembling a random pattern string that will be copied
to the top and bottom of the box. Then random patterns are added for the
remaining edges of the box, and these are copied to the corresponding edges
accordingly. (Refer to Fig. 1 for corresponding edges.) Finally, the shape that
will fold is cut out by raising the temperature, assuming that the bonds between
tiles along the region-borderline have weak threshold temperatures. The regions
that are shaded in Fig. 3(a) are cut away.
3.2
Notation
All of our algorithms will use a 3-level concatenation generalization model; thus
a tile is a 4-tuple of triplets that we write TN × TS × TW × TE , with each
Ti = (σ1 , σ2 , σ3 ). We change the order of the tuple to more easily identify tiles
that will bind, since most binding will be north-south or west-east. (This decision
is arbitrary and purely notational.) We assume all tiles are oriented so that the
directions are clear.
We define the set Π ⊂ Σ to be the “random patterns” π1 , π2 , . . . , πp used as
components of exposed triplets for tiles used in random assembly; their use will
become clear when we discuss implementation of random assembly below.
We use counters to control the growth of of our planar shape. This concept
has been well explored in [7] and earlier papers. Each tile can be assigned a
position in the plane denoted by a horizontal and vertical coordinate. We create
symbols for position counters and then allow tiles to hybridize if the positions
match, creating a mechanism for algorithms to place tiles in absolute or relative
positions. Let H(i) and V (j) be the symbols denoting horizontal position i and
vertical position j, respectively.
3.3
Row-by-row Algorithm
The row-by-row algorithm sacrifices time complexity to reduce space complexity
and alphabet size. In addition, by using more steps and temperature-sensitive
binding, the algorithm eliminates the possibility of misformations.
Implementation of Random Assembly The base strip shown in Fig. 3(a)
is created via random assembly and represents the unique seed structure for the
shape. The pattern assembled on the strip will be copied to other edges. The
strip will have length n as it is an edge of the cube, and so we use horizontal
counters to control growth. We add the following tiles to solution:
(β, H(0), V (0)) × β, H(0), V (−1) × ρ′ , H(0), V (0) × κ1 , H(1), V (0)
′
(1)
(β, H(0), V (0)) × β, H(0), V (−1) × (κ1 , H(n − 1), V (0)) × ρ , H(0), V (0)
(α, κ2 , V (0)) × α, κ2 , V (−1) × (κ1 , H(1), V (0)) × κ1 , H(2), V (0)
(α, κ2 , V (0)) × α, κ2 , V (−1) × (κ1 , H(n − 2), V (0)) × κ1 , H(n − 1), V (0)
(πk , κ2 , V (0)) × πk , κ2 , V (−1) × (κ1 , H(i), V (0)) × κ1 , H(i + 1), V (0)
(2)
(3)
(4)
(5)
where the tiles in (5) vary over all i, 2 ≤ i ≤ n − 2, and all k, 1 ≤ k ≤ |Π|.
Since the ends, given by (1) and (2), the tiles bordering the ends, given by (3)
and (4), and the pattern tiles in the middle of the strip, given by (5), all have
appropriate horizontal counter markings H(i) for every position 0, . . . , n − 1,
every base strip must have the appropriate ends and length n. Note that the
pattern on the exposed north and south side of tiles in position 2, . . . , n − 2
given by (5), is completely random as any one of the tiles containing some πi
could hybridize there. Thus, this shape now has a unique identifier, namely, the
sequence of patterns {πi } along the base strip; the probability of any sequence
expressed on a given shape is |Π|n−2 .
Implementation of Straight Copy We must complete a straight-copy from
the base strip through regions A,B,C, and D, in addition to straight-copy patterns required elsewhere. This is done using 2n + 1 steps as the tiles for A and
C are done in parallel (the same for B and D). In the following example for a
straight-copy through A-D, assume that the triplet constant X = (ϕ1 , ϕ2 , ϕ3 ).
The following tiles are added for each step i, 1 ≤ i ≤ 2n − 2:
(6)
(7)
(α, κ2 , V (i)) × α, κ2 , V (i − 1) × (σ1 , σ2 , V (i)) × X
(α, κ2 , V (i)) × α, κ2 , V (i − 1) × X × (σ1 , σ2 , V (i))
(α, κ2 , V (−i)) × α, κ2 , V (−i − 1) × (σ1 , σ2 , V (−i)) × X
(8)
(α, κ2 , V (−i)) × α, κ2 , V (−i − 1) × X × (σ1 , σ2 , V (−i))
(9)
(πk , κ2 , V (i)) × πk , κ2 , V (i − 1) × X × X
(10)
(πk , κ2 , V (−i)) × πk , κ2 , V (−i − 1) × X × X
(11)
Tiles versions exist for all patterns in Π when πk is present in the tile description
above. We assume that X binds weaker than κ2 , so cycling the temperature
ensures the tiles are attached on the κ2 side.
The tiles are added sequentially to prevent misformations. It is evident that
the borders of these regions, given by (6)–(9) have symbols on the exposed sides
to hybridize with a strip of tiles that will form a folding edge of the box, and
the middle tiles copy the random pattern to another location.
After 2n − 2 of these steps, the middle regions are complete, except for two
rows on top (rows 2n − 1 and 2n) and one row on the bottom (2n − 1). The
order of these rows is important, as we add an extra set of tiles to prevent the
top and bottom from hybridizing before the folding is complete.
Implementation of Turn Copy The turn-copy step, for example, copying the
bottom edge of E to the left edge of D through I so the shape can fold, is done
using vertical and horizontal counters, which essentially places a tile in a specific
spot. Therefore we can add all the tiles at once to complete the region without
possibility of misformation. For the above example region, we would add the
following tiles. Let i and j vary such that −n ≤ i ≤ −1 and −2n + 1 ≤ j ≤ −n.
For all i, j, add:
(12)
(13)
(14)
i = j+(n−1), (πk , H(i), V (j))× κ3 , H(i), V (j − 1) ×(κ3 , H(i), V (j))× πk , H(i − 1), V (j)
i < j+(n−1), (πk , H(i), V (j))× πk , H(i), V (j − 1) ×(κ3 , H(i), V (j))× κ3 , H(i − 1), V (j)
i > j+(n−1), (κ3 , H(i), V (j))× κ3 , H(i), V (j − 1) ×(πk , H(i), V (j))× πk , H(i − 1), V (j)
The above copies the pattern through I, and the following adds the left edge to
D, where −2n + 1 ≤ i ≤ −n − 1:
(β, H(0), V (−n)) × κ3 , H(0), V (−n − 1) × (πk , H(0), V (−n)) × σ1 , σ2 , V (−n)
(κ3 , H(0), V (i)) × κ3 , H(0), V (i − 1) × (πk , H(0), V (i)) × σ1 , σ2 , V (i)
As an extra precaution we can set κ3 binding to be weaker than πk , and cycle
the temperature several times. In addition, we force the encoding of horizontal
and vertical counters at the edges where the folding occurs to be the same, so
that the sticky ends are in fact complementary.
Summary After random assembly of the base strip, we use a row-by-row
straight copy using a one-dimensional counter through regions A-D to copy the
base strip’s pattern. We then use straight copy to fill in the bodies of regions
E and F and add the edges using random assembly, as these will correspond
to other portions of the shape. We then use a turn copy through G-J to make
those edges correspond. Finally, we do a sequence of straight and turn copies
from E and F through K-N and O-R to complete the shape. We then raise the
temperature to cut away the shaded regions.
Analysis of the Row-by-row Algorithm
Theorem 9. |Σ| = 8n + |Π| + O(1).
Proof. The symbol set S consists of the following: pattern symbols in Π; 4n
horizontal and 4n vertical counters for positions throughout the shape; and some
specialized symbols, not dependent on the inputs.
⊓
⊔
Theorem 10. The algorithm has has time complexity approximately 5n.
Proof. The algorithm consists of the following steps: 1 step to build the base
strip; 2n+1 steps to complete A-D; n steps to grow E and F; 1 step to add random
patterns to the leaf borders; 1 step for the first set of turn-copy steps (G,I,H,J);
and 2n + 2 steps to complete regions K-N and O-R. The time complexity for the
basic version, therefore, is approximately 5n steps.
⊓
⊔
Theorem 11. The space complexity of the row-by-row algorithm is approximately 6|Π|n2 + 10|Π|n + 4|Π| + 8n tiles.
Proof. Clear from counting the number of tiles required, including all the necessary variations of the same tile for position and pattern.
⊓
⊔
Theorem 12. The number of distinct temperatures required is 3.
Proof. One temperature is required for detaching the excess portions of the 2D
shape, and the highest temperature level is for the main portion of the box that
remains intact. Adding tiles in rows involves cycling the temperature to prevent
misformations, by ensuring the “random-patterned” side of the tiles bind. This
requires a third temperature between the two discussed where potential misformations denature.
⊓
⊔
Theorem 13. The misformation probability of row-by-row is 0.
Proof. The random assembly steps grow the shape in one direction using a onedimensional counter, and only the correct tiles are marked with the appropriate
counters. Thus the probability for these steps is zero. The turn-copy steps use
both vertical and horizontal counters defining a two-dimensional position that
bind weaker than the random patterns. As these are the only three words on
tiles in this pattern, these also have a misformation probability of zero. Finally,
because the straight-copy steps are performed in a sequence and all the potential
spots for misformation are sealed with tiles one row at a time, the probability
of misformation here is also zero. Thus the misformation probability of the algorithm is 0.
⊓
⊔
3.4
All-together Algorithm
The row-by-row algorithm does not take full advantage of the parallelism of
self-assembly. Instead of adding one row after another, we can synthesize tiles
with both horizontal and vertical counters, so each tile type can fit at only one
absolute position. Then we can begin with virtually all our tiles in solution and
allow the assembly to proceed in constant time.
Random assembly is performed as above, with special one-dimensional counters to control length. However, these strips must be assembled separately and
independently to prevent interference. The sequences exposed opposite the random patterns will dictate where these strips can later hybridize on the larger
shape, and the random patterns will cause growth that makes the shape unique,
as needed. The turn copy implementation is unchanged, as counters are already
present. The straight copy, rather than using a “constant” symbol, must instead
use counters for growth in both directions.
Although this algorithm is straightforward, with all the tiles in solution at
once, we might get incomplete partial structures. Specifically, tiles can bind on
sides that have counters but do not have any πi component, thus creating shapes
with no information allowing them to fold. In this case, at any of the O(n2 )
positions, a misformation could occur if the counters bind but the patterns do
2
not, and so the misformation probability of all-together is 1 − |Π|−n . We can
prevent this by requiring the counters to bind weaker than the πi and cycle the
temperature (requiring 3, rather than 2, distinct temperature levels), giving a
misformation probability of 0, but practically, this has little meaning since we
may still get incomplete, partial structures and may have to wait a long time for
proper hybridization to occur.
Besides this drawback, this algorithm performs well by other complexity measures. Since counters are necessary for the row-by-row algorithm, the alphabet
size is not greatly increased. The algorithm has space complexity O(n2 |Π|), essentially because a different tile type is created for every position and every
pattern that could be present there.
3.5
Other Algorithms
We have considered several other algorithms using the same 3-level generalization
model, experimenting with different uses or omissions of counters and other
implementations of the self-assembly patterns.
The by-region algorithm keeps space complexity and time complexity low by
removing counters and controlling growth in only certain rows and columns of
a region. The probability of misformation is increased.
The border-first algorithm has the highest misformation probability, although
has a very low time and space complexity. In this algorithm, the frame of regions are constructed first, when possible, and then “filler tiles” are added in
to strengthen the structure later. There are potential stability problems during
shape growth.
Finally, another possibility is to build the six faces separately and allow them
to hybridize, but with this method there is little control over how the final shape
actually forms in solution.
4
Conclusion
Our paper introduces a precise extension to the Tile Assembly Model [7] that
allows greater information content per tile and scalability to three dimensions.
The model better formalizes the abstraction of DNA tiles to symbols, introduces five complexity measures to analyze algorithms, and is the first to extend
nanostructure fabrication to three dimensions.
In addition, our paper opens up wide-ranging avenues of research.
First of all, it may be possible to encode information on tiles more succintly
than our algorithms do to accomplish the copy patterns discussed. The existence
of good 2-level or 1-level generalization algorithms is unknown.
Algorithms to form other 3D structures, having various applications in biology and computation, can be studied. More work also must be done to quantify
the probabilities specified in the paper (possibly including a free-energy analysis
of tile binding).
Finally, there remain some important biological issues. In particular, design of
a strong tile suitable for our method of computation and design of a 3D building
block are two important steps to increasing the feasibility of 3D self-assembly.
The use of temperature may be further refined and exploited to improve some
complexity results and the number of steps needed in the lab.
References
1. J. Chen and N. C. Seeman. The synthesis from DNA of a molecule with the
connectivity of a cube. Nature, 350:631–633, 1991.
2. A. Condon, R. M. Corn, and A. Marathe. On combinatorial DNA word design. In
Winfree and Gifford [12].
3. T.-J. Fu and N. C. Seeman. DNA double-crossover molecules. Biochemistry,
32:3211–3220, 1993.
4. T. H. LaBean, E. Winfree, and J. H. Reif. Experimental progress in computation
by self-assembly of DNA tilings. In Winfree and Gifford [12].
5. T. H. LaBean, H. Yan, J. Kopatsch, F. Liu, E. Winfree, H. Reif, and N. Seeman.
The construction, analysis, ligation and self-assembly of DNA triple crossover complexes. J. Am. Chem. Soc., 122:1848–1860, 2000.
6. I. V. Markov. Crystal Growth for Beginners: Fundamentals of Nucleation, Crystal
Growth, and Epitaxy. World Scientific, Singapore, 1995.
7. P. Rothemund and E. Winfree. The program-size complexity of self-assembled
squares. In F. F. Yao, editor, Proceedings of the 32nd Annual ACM Symposium
on Theory of Computing, Portland, OR, 21–23 May 2000. ACM Special Interest
Group on Algorithms and Computation Theory.
8. N. C. Seeman. Nucleic-acid junctions and lattices. Journal of Theoretical Biology,
2:237–247, 1982.
9. H. Wang. Proving theorems by pattern recognition. Bell System Technical Journal,
40:1–42, 1961.
10. E. Winfree. On the computational power of DNA annealing and ligation. In E. B.
Baum and R. J. Lipton, editors, DNA Based Computers, DIMACS: Series in Discrete Mathematics and Theoretical Computer Science, pages 199–210. American
Mathematical Society, May 1995.
11. E. Winfree, T. Eng, and G. Rozenberg. String tile models for DNA computing by
self-assembly. In A. Condon and G. Rozenberg, editors, DNA Based Computers VI,
Leiden, The Netherlands, 13–17 June 2000. Leiden Center for Natural Computing.
12. E. Winfree and D. Gifford, editors. Preliminary Proceedings, Fifth International
Meeting on DNA Based Computers, Cambridge, Massachusetts, 14–15 June 1999.
DIMACS.
13. E. Winfree, F. Liu, L. A. Wenzler, and N. C. Seeman. Design and self-assembly of
two-dimensional DNA crystals. Nature, 394:539–544, 1998.
| 5cs.CE
|
Context Generation from Formal Specifications
for C Analysis Tools
Michele Alberti1? and Julien Signoles2
TrustInSoft, Paris, France
[email protected]
CEA LIST, Software Reliability and Security Laboratory
F-91191 Gif-sur-Yvette Cedex, France
[email protected]
1
arXiv:1709.04497v1 [cs.PL] 5 Sep 2017
2
Abstract. Analysis tools like abstract interpreters, symbolic execution tools and
testing tools usually require a proper context to give useful results when analyzing a particular function. Such a context initializes the function parameters
and global variables to comply with function requirements. However it may be
error-prone to write it by hand: the handwritten context might contain bugs or not
match the intended specification. A more robust approach is to specify the context in a dedicated specification language, and hold the analysis tools to support
it properly. This may mean to put significant development efforts for enhancing
the tools, something that is often not feasible if ever possible.
This paper presents a way to systematically generate such a context from a formal
specification of a C function. This is applied to a subset of the ACSL specification language in order to generate suitable contexts for the abstract interpretationbased value analysis plug-ins of Frama-C, a framework for analysis of code written in C. The idea here presented has been implemented in a new Frama-C plugin which is currently in use in an operational industrial setting.
Keywords: Formal Specification, Code Generation, Transformation, Code Analysis, Frama-C, ACSL
1
Introduction
Code analysis tools are nowadays effective enough to be able to provide suitable results
on real-world code. Nevertheless several of these tools including abstract interpreters,
symbolic execution tools, and testing tools must analyze the whole application from the
program entry point (the main function); or else either they just cannot be executed, or
they provide too imprecise results. Unfortunately such an entry point does not necessarily exist, particularly when analyzing libraries.
In such a case, the verification engineer must manually write the context of the
analyzed function f as a main function which initializes the parameters of f as well
as the necessary global variables. This mandatory initialization step must enforce the
function requirements and may restrict the possible input values for the sake of memory
footprint and time efficiency of the analysis. This approach is however error-prone:
?
This work was done when the first author was at CEA LIST, Software Reliability and Security
Laboratory.
II
additionally to usual pitfalls of software development (e.g. bugs, code maintenance,
etc.), the handwritten context may not match the function requirements, or be over
restrictive. Moreover this kind of shortcomings may be difficult to detect due to the fact
that the context is not explicitly the verification objective.
A valid and more robust alternative is to specify such a context in a dedicated specification language, and make the analysis tools handle it properly. This is often an arduous approach as the support for a particular specification language feature may entail
a significant development process, something that is often not feasible if ever possible.
Also, it requires to do so for every tool.
This paper presents a way to systematically generate an analysis context from a
formal specification of a C function. The function requirements as well as the additional restrictions over the input domains are expressed as function preconditions in the
ANSI/ISO C Specification Language (in short, ACSL) [2]. This specification S is interpreted as a constraint system, simplified as much as possible, then converted into a
C code C which exactly implements the specification S. Indeed not only every possible
execution of C satisfies S but conversely, there is an execution of C for every possible
input satisfying the constraints expressed by S. We present the formalization of this
idea for an expressive subset of ACSL including standard logic operators, integer arithmetic, arrays and pointers, pointer arithmetic, and built-in predicates for the validity
and initialization properties of memory location ranges.
We also provide implementation details about our tool, named CfP for Context from
Preconditions, implemented as a Frama-C plug-in. Frama-C is a code analysis framework for code written in C [11]. Thanks to the aforementioned technique, CfP generates
suitable contexts for two abstract interpretation-based value analysis tools, namely the
the Frama-C plug-in EVA [3] and TIS-Analyzer [8] from the TrustInSoft company.
Both tools are actually distinct evolved versions of an older plug-in called Value [6]. In
particular, TrustInSoft successfully used CfP on the mbed-TLS library (also known as
PolarSSL), an open source implementation of SSL/TLS3 , when building its verification kit [21]. It is worth noting that CfP revealed some mistakes in contexts previously
written by hand by expert verification engineers when comparing its results with these
pieces of code. Also, CfP generates code as close as possible to human-written code: it
is quite readable and follows code patterns that experts of these tools manually write.
Contributions The contributions of this paper are threefold: a novel technique to systematically generate an analysis context from a formal specification of a C function,
a precise formalization of this technique, and a presentation of a tool implementing
this technique which is used in an operational industrial setting.
Outline Section 2 presents an overview of our technique through a motivating example.
Section 3 details preconditions to constraints conversion, while Section 4 explains the C
code generation scheme for these latter. Section 5 evaluates our approach and Section 6
discusses related work. Section 7 concludes this work by also discussing future work.
3
https://tls.mbed.org/
III
2
Overview and Motivating Example
We illustrate our approach on context generation through the function aes crypt cbc,
a cryptographic utility implemented by the mbed-TLS library. Figure 1 shows its prototype and ACSL preconditions as written by TrustInSoft for its verification kit [21].
1
2
3
4
5
typedef struct {
int nr;
unsigned long *rk;
unsigned long buf[68];
} aes_context;
/*
/*
/*
number of rounds
AES round keys
unaligned data
*/
*/
*/
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
/*@
@
@
@
@
@
@
@
@
@
@
@
int
requires ctx_valid: \valid(ctx);
requires ctx_init: \initialized(ctx->buf + (0 .. 63));
requires ctx_rk: ctx->rk == ctx->buf;
requires ctx_nr: ctx->nr == 14;
requires mode: mode == 0 || mode == 1;
requires length: 16 <= length <= 16672;
requires length_mod: length % 16 == 0;
requires iv_valid: \valid(iv + (0 .. 15));
requires iv_init: \initialized(iv + (0 .. 15));
requires input_valid: \valid_read(input + (0 .. length - 1));
requires input_init: \initialized(input + (0 .. length - 1));
requires output_valid: \valid(output + (0 .. length - 1)); */
aes_crypt_cbc(aes_context *ctx,int mode,size_t length,unsigned char iv[16],
const unsigned char *input,unsigned char *output);
Fig. 1: ACSL preconditions of the mbed-TLS function aes crypt cbc.
Specification The function aes crypt cbc provides encryption and decryption of
a buffer according to the AES cryptographic standard and the CBC encryption mode.
The function takes six parameters. The last two are the input and the output strings. The
parameter ctx stores the necessary information to the AES substitution-permutation
network, in particular the number of rounds and the round keys defined in a dedicated
structure at lines 1–5. The parameter mode indicates whether the function should encrypt or decrypt the input. The parameter length indicates the length of the input
string. Finally the parameter iv provides an initialization vector for the output of
16 characters (unsigned char iv[16]). This declared length is actually meaningless for most C tools because an array typed parameter is adjusted to have a pointer
type [10, Section 6.9.1 and also footnote 79 at page 71], but CfP nevertheless considers
it as part of the specification in order to generate a more precise context.
ACSL annotations are enclosed in /*@ ... */ as a special kind of comments.
Therefore they are ignored by any C compiler. A function precondition is introduced
by the keyword requires right before the function declaration or definition. It must be
satisfied at every call site of the given function. Here the function aes crypt cbc has
12 precondition clauses, and the whole function precondition is the conjunction of all of
them. Clauses may be tagged with names, which are logically meaningless but provide a
way to easily refer to and to document specifications. For instance, the first precondition
(line 7) is named ctx valid while the second (line 8) is named ctx init.
We now detail the meaning of each precondition clause. All pointers must be valid,
that is properly allocated, and point to a memory block of appropriate length that
the program can safely access either in read-only mode (predicate \valid_read), or in
read-write mode (predicate \valid) . That is the purpose of preconditions ctx valid,
IV
iv valid, input valid and output valid: ctx must point to a memory block
containing at least a single aes context struct, iv must be able to contain at least 16
unsigned characters (ranging from 0 to 15), while input and output must be able to
contain at least length unsigned characters (ranging from 0 to length − 1). Memory locations, which are read by the function, must be properly initialized. That is the
purpose of the precondition clauses ctx init, iv init, and input init which
initialize the first 64 cells of ctx->buf as well as every valid cell of iv and input.
The specification clause mode specifies that the mode must be either 0 (encryption)
or 1 (decryption), while the specification clause length mod specifies that the length
should be a multiple of the block size (i.e. 16) as specified in mbed-TLS. The other
clauses restrict the perimeter of the analysis in order to make it tractable.
The clause ctx rk is a standard equality for an AES context, while the clause
ctx nr is true for 256-bit encryption keys. Finally the clause length aims to restrict
the analysis to buffers of size from 16 to 16672 unsigned characters.
Context Generation A naive approach for context generation would consider one precondition clause after the other and directly implement it in C code. However, this
would not work, in general, since requirements cannot be treated in any order. In our
running example, for instance, variables input and output depends on the variable
length: the precondition clauses over this latter must be treated before those over
the former, as well as the generated code for these variables must initialize the latter, first, and the former afterwards, to be sound. To solve such problems, one could
first record every dependency among the left-values involved in the specification, and
then proceed to generate C code accordingly. An approach based only on a dependency graph is nonetheless insufficient for those preconditions that need an inference
reasoning in order to be implemented correctly. As an example, treating the precondition /*@requires \valid(x+(0..3)) && *(x+4)==1;*/ demands to infer x as an array of 5
elements in order to consider the initialization x[4] = 1; correct.
We now give an overview on how we treat context generation by means of the
plug-in CfP of Frama-C. On the aes crypt cbc function contract, CfP provides
the result shown in Figure 2 (assuming that the size of unsigned long is 4 bytes4 ).
First note that every execution path ends by a call to the function aes crypt cbc.
Up to these calls, the code initializes the context variables (prefixed by cfp) in order
to satisfy the precondition of this function, while the different paths contribute to cover
all the cases of the specification. The initialization code is generated from sets of constraints that are first inferred for every left-value involved in the precondition. While
inferring these constraints from the precondition clauses, the implicit dependencies
among left-values are made explicit and recorded in a dependency graph. This latter
is finally visited to guide the code generation process in order to obtain correct C code.
Let us start detailing the generated code for both preconditions about length
(Figure 1, lines 12–13). First CfP declares a variable cfp length of the same type
as length (line 4). Then it initializes it by means of the Frama-C library function
Frama C unsigned int interval (line 7). It takes two unsigned int arguments and returns a random value comprised between the two. This allows to fulfill
4
This kind of system-dependent information is customizable within Frama-C.
V
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
int cfp_aes_crypt_cbc(void) {
unsigned char *cfp_output, *cfp_input;
unsigned char cfp_iv[16];
size_t cfp_length;
aes_context cfp_ctx;
int cfp_disjunction;
cfp_length = Frama_C_unsigned_int_interval(16, 16672);
if (cfp_length % 16 == 0) {
Frama_C_make_unknown((char *)cfp_ctx.buf,256);
cfp_ctx.nr = 14;
cfp_ctx.rk = cfp_ctx.buf;
Frama_C_make_unknown((char *)cfp_iv,16);
cfp_input = (unsigned char *)malloc(cfp_length);
if (cfp_input != (unsigned char *)0) {
Frama_C_make_unknown((char *)cfp_input, cfp_length);
cfp_output = (unsigned char *)malloc(cfp_length);
if (cfp_output != 0) {
cfp_disjunction = Frama_C_int_interval(0,1);
if (cfp_disjunction) {
int cfp_mode;
cfp_mode = 1;
aes_crypt_cbc(&cfp_ctx,cfp_mode,cfp_length,cfp_iv,cfp_input,cfp_output);
}
else {
int cfp_mode;
cfp_mode = 0;
aes_crypt_cbc(&cfp_ctx,cfp_mode,cfp_length,cfp_iv,cfp_input,cfp_output);
}
}
}
}
return 0;
}
Fig. 2: Slightly simplified version of the code generated by CfP for the specification in Figure 1.
Compared to the actual version, only a few integer casts have been removed for reasons of brevity.
the former requirement and to guarantee that Frama-C-based abstract interpreters will
interpret this result with exactly the required interval. Also, it corresponds to the way
that expert engineers would write a general context for such analyzers. Finally, the requirement length % 16 == 0 is implemented by the conditional at line 8.
Lines 9–11 implement the preconditions about ctx, a pointer to an aes context.
Instead of allocating such a pointer, the generated code just declares a local variable
cfp ctx and passes its address to the function calls. This automatically satisfies the
precondition on pointer validity. Line 9 initializes the 256 first bytes of the structure field
buf by using the Frama-C library function Frama C make unknown. Assuming
that the size of unsigned long is 4 bytes, 256 bytes is the size of 64 values of type
unsigned long. Again, an expert engineer would also use this library function.
Lines 10 and 11 initialize the fields ctx->nr and ctx->rk by single assignments.
Here CfP fulfills the equality requirement ctx->rk == ctx->buf with respect to
ctx->rk instead of ctx->buf because the latter already refers to a memory buffer.
The requirements on function arguments iv, input, and output are implemented by lines 12–17. Let us just point out how CfP defines the respective variables:
while ctx iv is as an array of 16 unsigned char, ctx input and ctx output
are just pointers to dynamically allocated memory buffers. Indeed, while CfP can infer
the exact dimension of the former from the specification, the dimension of these latter
depends on the value of ctx length, which is determined only at runtime.
VI
The last part of the generated code (lines 18–29) handles the requirement on mode,
which is either 0 or 1. Although the generated conditional may seem excessive in the
case of these particular values, it is nonetheless required in the general case (for instance, consider the formula mode == 5 || mode == 7).
3
Simplifying ACSL Preconditions into State Constraints
This section presents a way to systematically reduce a function precondition to a set of
constraints on the function context (i.e. function parameters and global variables).
We first introduce an ACSL-inspired specification language on which we shall formalize our solution. Then, we define the notion of state constraint as a form of requirement over a C left-value, which in turn we generate as C code for initializing it. In
order to simplify state constraints the most, we make use of symbolic ranges, originally
introduced by Blume and Eigenmann [4] for compiler optimization. We finally provide
a system of inference rules that formalizes such a simplification process.
3.1
Core Specification Language
In this work we shall consider the specification language in Figure 3. It is almost a subset of ACSL [2] but for the predicate defined, which subsumes the ACSL predicates
\initialized and \valid (see below).
Predicates
P ::= T cop T
| defined(M )
| P ∧ P | P ∨ P | ¬P
term comparison (cop ∈ {≡, ≤, <, ≥, >})
M is defined
logic formula
Terms
T ::= z
| M
| T bop T
integer constant (z ∈ Z)
memory value
arithmetic operation (bop ∈ {+, -, ×, /, %})
Memory Values M ::= L
| M ++ T
| M ++ T ..T
left-value
single displacement
displacement range
Left-Values
L ::= x
| ?M
C variable
dereference
Types
κ ::= ι
| κ?
integer
pointer
Fig. 3: Predicates, terms, and types.
Predicates are logic formulæ defined on top of typed term comparisons and predicates defined. Terms are arithmetic expressions combining integer constants and
memory values by means of the classic arithmetic operators. Memory values include
left-values, which are C variables and pointer dereferences (?), and memory displacements through the operator (++). In particular, M ++ T1 ..T2 defines the set of memory
values {M ++ T1 , . . . , M ++ T2 } and may only appear as the outermost construct in
a predicate defined. On integers, defined(L) holds whenever L is an initialized
left-value. On pointers, defined(M ) holds whenever M is a properly allocated and
initialized memory region.
VII
Term typing Terms of our language are typed. A left-value may take either an integer (ι)
or a pointer (κ?) type, while memory values are pointers. We omit the typing rules for
terms, which are quite standard. Let us just specify that memory values of the form
M ++ T have pointer type, as well as the recursive occurrence M , while T must have
integer type. (Memory values M ++ T ..T are typed as set of pointers [2].) Since we
do not consider any kind of coercion construct, terms of pointer type cannot appear
where integer terms are expected, that is, they cannot appear in arithmetic expressions.
It also follows that term comparisons only relate terms of the same type.
Term normal forms For the sake of concision and simplicity, the remainder of this work
assumes some simplifications to take place on terms in order to consider term normal
forms only. In particular, arithmetic expressions are maximally flattened and factorized
(e.g. by means of constant folding techniques, etc.). We will conveniently write single
displacements M ++ T as M ++ T ..T . We also assume memory values with displacement ranges to be either of the form x ++ T1 ..T2 or ?L ++ T1 ..T2 . To this end, terms
of the form (L ++ T1 ..T2 ) ++ T3 ..T4 simplify into L ++ (T1 + T3 )..(T2 + T4 ). Finally, memory values L ++ 0..0 normalize to L.
Disjunctive normal forms A precondition is a conjunction of predicate clauses, each
one given by an ACSL requires (cf. example in Figure 1). As a preliminary
step, we
W V
shall rewrite this conjunctive clause into its disjunctive normal form i j Pij , where
each Pij is a predicate literal (or simply literal), that is, a predicate without nested logic
formulæ. A negative literal is either of the form ¬defined(M ) or ¬(M1 ≡ M2 ), with
M1 , M2 pointers, as every other negative literal in the input predicates is translated into
a positive literal by applying standard arithmetic and logical laws. A non-negative literal
is called a positive literal. Most of the rest of this section focuses on positive literals:
negative literals and conjunctive clauses are handled in the very end, while disjunctive
clauses will be considered when discussing code generation in Section 4.
3.2
State Constraints
We are interested in simplifying a predicate literal into a set of constraints over C leftvalues, called state constraints. These are meant to indicate the minimal requirements
that the resulting C function context must implement for satisfying the function precondition. In Section 4, they will be, in turn, converted into C code.
We intuitively consider a state constraint to represent the domain of definition of a
C left-value of the resulting function context state. Since such domains might not be
determined in terms of integer constants only, we shall found their definition on the
notion of symbolic ranges [4]. As we want to simplify state constraints the most, we
define them in terms of the symbolic range algebra proposed by Nazaré et al. [14]. Our
definitions are nonetheless significantly different, even though inspired from their work.
Symbolic Expressions A symbolic expression E is defined by the following grammar,
where z ∈ Z, bop ∈ {+, -, ×, /, %}, and max and min are, respectively, the largest
and the smallest expression operators. We denote E the set of symbolic expressions.
E ::= z | x | ?E | E bop E | max(E, E) | min(E, E).
VIII
In the rest of this section, we assume a mapping from memory values to their respective symbolic expression, and let the context discriminate the former from the latter.
In Section 3.3 we shall simplify symbolic expressions. For this, we need a domain
structure. Let us denote E∞ = E ∪ {−∞; +∞} and Z∞ = Z ∪ {−∞; +∞}. We define
a valuation of a symbolic expression E every map V(E), from E∞ to Z∞ , obtained by
substituting every C variable in E with a distinct integer, the symbol ? with a natural
number strictly greater than 1 as a multiplicative coefficient, and interpreting the operators {bop, min, max} as their respective functions over Z∞ × Z∞ . If we denote ≤∞
the standard ordering relation on Z∞ , then the preorder 4 on E∞ is defined as follows:
E1 4 E2 ⇐⇒ ∀V, V(E1 ) ≤∞ V(E2 ).
The partial order over E∞ is therefore the one induced from 4 by merging in the
same equivalence class elements x and y of E∞ such that x 4 y and y 4 x. As an
example, the elements 0 and min(0, 0) are equivalent.
Lattice of Symbolic Expression Ranges A symbolic range R is a pair of symbolic expressions E1 and E2 , denoted [E1 , E2 ]. Otherwise said, a symbolic range is an interval
with no guarantee that E1 E2 . We denote R the set of symbolic ranges extended
with the empty range ∅ and v its partial ordering which is the usual partial order over
(possibly empty) ranges. Any symbolic range [E1 , E2 ] such that E2 ≺ E1 is therefore
equivalent to ∅. Consequently (R, v) is a domain. Its infimum is ∅ while its supremum
is [−∞, +∞]. We denote t and u its join and meet operators, respectively. It is worth
noting that, given (Ei )1≤i≤4 four symbolic expressions, the following equations hold:
[E1 , E2 ] t [E3 , E4 ] = [min(E1 , E3 ), max(E2 , E4 )]
[E1 , E2 ] u [E3 , E4 ] = [max(E1 , E3 ), min(E2 , E4 )] .
In words, min and max are compliant with our ordering relations. In Section 3.3,
when simplifying literals, they will be introduced as soon as incomparable formulæ will
be associated to the same left-value, resulting into an unsimplifiable constraint. Also,
it is worth noting that t and u are, in general, not statically computable operators. To
solve this practical issue, when these are not computable on some symbolic expressions, CfP relies on the above equations in order to delay their evaluations at runtime.
Eventually, the code generator will convert them into conditionals.
State Constraints as Symbolic Ranges with Runtime Checks Symbolic ranges capture
most minimal requirements over the C left-values of a function precondition: for integer
typed left-values, a symbolic range represents the integer variation domain, while for
pointer typed left-values, it represents a region of valid offsets. They are commonly
used in abstract interpreters for range [7,13] and region analysis [14,18], respectively.
However, some predicate literals cannot be simplified into symbolic ranges, requiring their encoding as runtime checks, that is, to be verified at runtime by means of
conditionals. We denote RTC(T1 cop T2 ) a runtime check between two terms T1 and
T2 . We then call state constraint any pair C = R ⊕ X given by a symbolic range R and
a set X of runtime checks. We denote π1 (C) (resp. π2 (C)) the first (resp. the second)
projection of C, that is, R (resp. X).
IX
3.3
Inferring State Constraints
We now formalize our solution for simplifying a positive literal into a set of state constraints as a system of inference rules. Negative literals, as well as conjunctive clauses,
are handled separately at the end of the section.
Simplification Judgments Simplification rules are given over judgments of the form
Σ ` P ⇒ Σ0,
where P is a predicate literal, and Σ, Σ 0 are maps from left-values to state constraints.
Each judgment associates a set of state constraints Σ and a literal P with the result of
simplifying P with respect to the left-values appearing in it, that is, an updated map Σ 0
equal to Σ but for the state constraints on these latter. Figures 4 shows the formalization
of the main literal simplifications. This system does not assume the consistency of the
precondition: if this is inconsistent, no rule applies and the simplification process fails.
Predicates defined Figure 4a provides the simplification rules for literal defined.
Rules VARIABLE and D EREFERENCE enforce the initialization of a left-value L in
terms of the symbolic range neutral ival(κ). This latter is respectively defined as ∅,
for κ a pointer type, and [−∞, +∞], for κ integer type. These are quite common initial
approximations when inferring variation domains of either memory or integer values.
Rules R ANGE -1 and R ANGE -2 enforce the validity of a memory region determined
by the displacement range L ++ (T1 ..T2 ). The first premise of these rules established
whether L is already enforced in Σ to be an alias of a memory value M , as indicated by
the singleton range [M ; M ]. If not, rule R ANGE -1 first enforces the initialization of L
and the soundness of the displacement bound determined by T1 and T2 , and then it updates the region of valid offsets pointed to by L to include the range [0; T2 ]. In practice,
predicates 0 ≤ T1 ≤ T2 are added only if not statically provable. Moreover, note that we
do not consider T1 as the lower bound of the symbolic range, because C memory regions
must start at index 0. Rule R ANGE -2 handles the case of L alias of M in Σ by enforcing
the validity of the memory region determined by M to take into account the displacement range (T1 ..T2 ). In particular, since single displacements only may appear in
memory equality predicates (cf. rule M EMORY-E Q), M is of the form L0 ++ (T3 ..T3 ),
and the validity of the alias L within the range (T1 ..T2 ) is obtained by requiring the
validity of the displacement range L0 ++ (min(T1 , T3 )..max(T2 , T3 )).
Rule I DEMPOTENCE is provided only to allow the inference process to progress.
Term comparison predicates Rules in Figure 4b formalize the simplification of integer term comparison and memory equality predicates. The first two are actually rule
schema, as C MP -1 and C MP -2 describe term comparison simplifications over the integer comparison operators {≡, ≤, ≥}. (Strict operators are treated in terms of nonstrict ones.) Let us detail rule C MP -1 with respect to a generic operator cop. The rule
applies whenever T1 cop T2 can be rewritten by means of classic integer arithmetic
transformations as L cop T3 , that is, as a left-value in relation cop with an integer
term T3 . If so, C MP -1 reduces the symbolic range of L with respect to the one given
by ival(cop, T3 ). This latter function takes a comparison operator cop and an integer
X
I DEMPOTENCE
L∈Σ
VARIABLE
x 6∈ Σ
type(x) = κ
Σ 0 = Σ ∪ {x 7→ neutral ival(κ)}
Σ ` defined(x) ⇒ Σ 0
Σ ` defined(L) ⇒ Σ
D EREFERENCE
?M 6∈ Σ
Σ ` defined(M ) ⇒ Σ 0
00
type(?M ) = κ
Σ = Σ 0 ∪ {?M 7→ neutral ival(κ)}
Σ ` defined(?M ) ⇒ Σ 00
R ANGE -1
π1 (Σ (L)) 6= [M ; M ]
Σ ` defined(L) ∧ 0 ≤ T1 ≤ T2 ⇒ Σ 0
Σ 00 = Σ 0 L ← π1 (Σ 0 (L)) t [0; T2 ]
Σ ` defined(L ++ (T1 ..T2 )) ⇒ Σ 00
R ANGE -2
π1 (Σ (L)) = [M ; M ]
base(M ) = L0
offset(M ) = T3
0
Σ ` defined(L ++ (min(T1 , T3 )..max(T2 , T3 ))) ⇒ Σ 0
Σ ` defined(L ++ (T1 ..T2 )) ⇒ Σ 0
(a) Simplification of literal defined.
C MP -1
T1 cop T2
L ∈ {T1 , T2 }
^
Σ ` defined(L) ∧
defined(L0 ) ⇒ Σ 0
L cop T3
L0 ∈T3
Σ 00 = Σ 0 L ← π1 (Σ 0 (L)) u ival(cop, T3 )
Σ ` T1 cop T2 ⇒ Σ 00
C MP -2
Σ`
L ∈ {T1 , T2 }
^
defined(L) ⇒ Σ 0
L∈{T1 ,T
2}
00
0
Σ = Σ L ← π2 (Σ 0 (L)) ∪ RTC(T1 cop T2 )
Σ ` T1 cop T2 ⇒ Σ 00
M EMORY-E Q
i, j ∈ {1, 2} ∧ i 6= j
base(M{i,j} ) = L{i,j}
offset(M{i,j} ) = T{i,j}
T3 = Tj + (−Ti )
0
M 0 = Lj ++ (T3 ..T3 )
Σ ` defined(Li ) ∧ defined(M
) ⇒ Σ0
0
0
0
00
0
0
π1 (Σ (Li )) v π1 (Σ (Lj ))
Σ = Σ Li ← M ; M
Σ ` M1 ≡ M2 ⇒ Σ 00
(b) Simplification of term comparison and memory equality literals.
N OT-D EFINED
M 6∈ Σ
Σ ` ¬ defined(M ) ⇒ Σ
M EMORY-N EQ
Σ ` defined(M1 ) ∧ defined(M2 ) ⇒ Σ 0
base(M{i,j} ) = L{i,j}
[Li ; Li ] 6v π1 (Σ 0 (Lj ))
i, j ∈ {1, 2} ∧ i 6= j
[Lj ; Lj ] 6v π1 (Σ 0 (Li ))
Σ ` M1 6≡ M2 ⇒ Σ 0
(c) Simplification of negative literals.
Fig. 4: Simplification of literals into state constraints.
XI
term T as arguments, and returns as result the symbolic range [T ; T ] when cop is ≡,
[−∞; T ] (resp. [T ; +∞]) when cop is ≤ (resp. ≥). Since both L and T3 are integer
typed terms, there is no aliasing issue here. Rule C MP -2 can always be applied, although
we normally consider it when C MP -1 cannot. In that case, rule C MP -2 conservatively
enforces the validity of the term comparison by means of a runtime check.
Aliasing Rule M EMORY-E Q handles aliasing between two pointers with single displacement M1 and M2 . Assuming both of the form L{i,j} ++ T{i,j} , with distinct
i, j ∈ {1, 2}, a pointer M 0 is first defined as Lj with single displacement T3 , this
latter determined by summing the offsets −Ti and Tj together. Such a pointer is then
enforced to be defined, and in the case that the actual region pointed by Lj is established
to be larger then the one pointed by Li , then Li is considered an alias of M 0 . Although
rather conservative, due to the fact that v is not statically computable in general, the
second to last premise is important for ensuring soundness.
Negative literals Figure 4c shows the rules for negative literals. These rules do not
simplify literals into state constraints, but rather ensure precondition consistency. For
instance, ¬defined(x) ∧ x == 0 is inconsistent as x should be defined with value 0
and undefined at the same time. In such a case, the system must prevent code generation.
Rule N OT-D EFINED just checks that the memory value M does not appear in the
map Σ, which suffices to ensure that M is not yet defined.
Rule M EMORY-N EQ applies under the hypothesis that both pointers M1 and M2
determine different memory regions. In particular, the two are not aliases whenever
each base address of one pointer does not overlap with the memory region of the other.
V
Conjunctive Clauses i Pi , on either positive or negative literals Pi , are handled sequentially through the following A ND rule. Given the definition of M EMORY-N EQ and
N OT-D EFINED, it assumes that negative literals are treated only after the positive ones,
by exhaustively applying rule M EMORY-N EQ first, and rule N OT-D EFINED afterwards.
A ND
Σ0 ` P 1 ⇒ Σ1
Σ1 ` P 2 ⇒ Σ2
···
^
Σ0 `
P i ⇒ Σn
Σn−1 ` Pn ⇒ Σn
i
Dependency Graph on Memory Values On a conjunctive clause, the system of inference rules in Figure 4 not only generates a map Σ, but it also computes a dependency
graph G on memory values. (Considering only the formalization of this section, the
memory values of the graph are actually left-values only. However, when considering
separately the ACSL predicates \initialized and \valid instead of defined, this is
not true anymore.) This graph is necessary for ensuring, first, the soundness of the rule
system with respect to mutual dependency on left-values in Σ, and, consequently, for
the correct ordering of left-value initializations when generating C code (cf. Section 4).
Generally speaking, each time a rule that needs inference is used in a state constraint
derivation for some left-value L (e.g. D EREFERENCE, R ANGE -1, C MP -1, etc.), edges
from L to every other left-value involved in some premise are added to the dependency
graph G. Such derivation fails as soon as this latter operation makes the graph G cyclic.
XII
Example When applying the inference system on our example in Figure 1, the final
map associates the integer length to [16, 16672] ⊕ {RTC(length%16 ≡ 0)} and
the array input to [0, length − 1] ⊕ ∅, along with the dependency graph in Figure 5.
ctx
iv
*ctx
iv + (0 .. 15)
input + (0 .. length - 1)
*(iv + (0 .. 15))
*(input + (0 .. length - 1))
ctx->nr
ctx->rk
ctx->buf
mode
input
length
output
output + (0 .. length - 1)
ctx->buf[0 .. 63]
Fig. 5: Dependency graph for the aes crypt cbc preconditions generated by CfP.
The system of inference rule in Figure 4 is sound: given a conjunctive clause P , the
simplification procedure on P always terminates, either with Σ or it fails. In the former
case, for each left-value L in P , state constraints in Σ satisfy respective literals in P
(that we denote as Σ |= P ).
Theorem 1. For all conjunctive clause P , either ∅ ` P ⇒ Σ and Σ |= P , or it fails.
4
Generating C Code from State Constraints
This section presents the general scheme for implementing preconditions, through state
constraints, in a C language enriched with one primitive function for handling ranges.
In practice, such primitive is meant to be analyzer-specific so as to characterize state
constraints as precisely as possible. As an example, we report on the case of our tool
CfP. However, for the sake of conciseness, we do neither detail nor formalize the code
generation scheme. We nevertheless believe that the provided explanation should be
enough to both understand and implement such a system in a similar setting.
Generating Code from a Conjunctive Clause Consider a conjunctive clause C and the
pair (Σ, G), respectively given by the map of state constraints and the dependency
graph of C, inferred by the system
Wn of rules in Figure 4. We shall show the general case
of disjunctive normal forms i=1 Ci later on.
To generate semantically correct C code, we topologically iterate over the leftvalues of G so as to follow the dependency ordering. For every visited left-value L,
we consider its associated state constraint C = R ⊕ X in Σ. Then, the symbolic range
R is handled by generating statements that initialize L. For most constructs, these statements are actually a single assignment, although a loop over an assignment may be
sometimes needed (e.g. when initializing a range of array cells). In particular, initializations of left-values L to symbolic ranges [T1 , T2 ] are implemented by means of the
primitive function make range(κ, T1 , T2 ), where κ is integer or pointer type. In practice, this function must be provided by the analyzer for which the context is generated,
so that, when executed symbolically, the analyzer’s abstract state will associate abstract
XIII
values [T1 , T2 ] to respective left-values L. Finally, conditionals are generated to initialize left-values with symbolic expressions involving min and max.
Once L has been initialized, the rest of the code is guarded by conditionals generated
from runtime checks in X. To resume, the generation scheme for L is the following:
1
/* initialization of L from R through assignments */
if (/* runtime checks from X */) {
/* code for initializing the next left-values */ ...; } }
2
3
After the initialization of the last left-value, the function under consideration (in our
running example, the function aes crypt cbc) is called with the required arguments.
Wn
Handling Disjunctions We rewrite preconditions into disjunctive normal form i=1 Ci
as a preliminary step. Then we process each disjunct Ci independently by applying the
inference system in Figure 4 and the code generation scheme previously W
described.
n
We now describe the code generation scheme of such a precondition i=1 Ci given
the code fragments for each and every of its disjunct Ci . If n = 1, then the code fragment
of C1 is directly generated. Otherwise, an additional variable cfp disjunction is
generated and initialized to the interval [1, n]. Then, a switch construct (or a conditional if n = 2) is generated, where each case contains the fragment Bi respective to Ci .
To resume, the context is generated as a function including the following code pattern:
1
cfp_disjunction = make_range(ι, 1, n);
switch (cfp_disjunction) {
case 1: { B_1; break; }
case 2: { B_2; break; }
...
case n: { B_n; break; }
}
2
3
4
5
6
7
Primitives in CfP Our tool CfP follows the generation scheme just described. It implements make range in terms of the Frama-C built-ins Frama C τ interval,
with τ a C integral type, and Frama C make unknown to handle symbolic ranges
for integers and pointers, respectively. These built-ins are properly supported by the
two abstract interpretation-based value analysis tools EVA [3] and TIS-Analyzer [8].
5
Implementation and Evaluation
We have implemented our context generation mechanism as a Frama-C plug-in, called
CfP for Context from Preconditions, written in approximately 3500 lines of OCaml.
(Although Frama-C is open source, CfP is not, due to current contractual obligations.)
CfP has been successfully used by the company TrustInSoft for its verification kit [21]
of the mbed-TLS library, an open source implementation of the SSL/TLS protocol.
We now evaluate our approach, and in particular CfP, in terms of some quite natural
properties, that is, usefulness, efficiency, and quality of the generated contexts.
This work provides a first formal answer to a practical and recurring problem when
analyzing single functions. Indeed, the ACSL subset considered is expressive enough
for most real-world C programs. Most importantly, CfP enables any tool to support a
compelling fragment of ACSL at the minor expense of implementing two Frama-C
built-ins, particularly so if compared to the implementation of a native support (if ever
possible). Finally, CfP has proved useful in an operational industrial setting in revealing
some mistakes in contexts previously written by hand by expert verification engineers.
XIV
Although we cannot disclose precise data about these latter, CfP revealed, most notably,
overlooked cases in disjunctions and led to fix incomplete specifications.
CfP is able to efficiently handle rather complex ACSL preconditions: the generation
of real-world contexts (e.g. the one of Figure 2) is usually instantaneous. Although
the disjunctive normal form can be exponentially larger than the original precondition
formula, such transformation is used in practice [17,12] and leads to better code in
terms of readability and tractability by the verification tools. This approach is further
justified by the fact that, in practice, just a small number of disjuncts are typically used
in manually-written ACSL specifications.
Our approach allows to generate contexts which are reasonably readable and follows code patterns that experts of the Frama-C framework use to manually write. In
particular, when handling disjunctions, CfP factorizes the generated code for a particular left-value as soon as the rule system infers the very same solution in each conjunctive clause. For instance, in our running example, only the initialization of the variable
mode depends on the disjunction mode == 0 || mode == 1. Hence all the other
left-values are initialized before considering cfp disjunction (cf. Figure 2).
We conclude by briefly discussing some current limitations. Our ACSL fragment
considers quantifier free predicate formulæ, and no coercion constructs are allowed.
Support for casts among integer left-values should be easy to add, whereas treating
memory addresses as integers is notoriously difficult. We leave these for future work.
6
Related Work
Similarly to our approach, program synthesis [12,20,16] automatically provides program fragments from formal specifications. However, the two approaches have different purposes. Once executed either symbolically or concretely, a synthesized program
provides one computational state that satisfies the specification, while a context must
characterize all such states. In particular, not only every state must satisfy the specification but, conversely, this set of states must contain every such possible one.
In software testing, contexts are useful for concentrating the testing effort on particular inputs. Most test input generation tools, like CUTE [19] and PathCrawler [5,9],
allow to express contexts as functions which, however, the user must manually write.
Some others, like Pex [1], directly compile formal preconditions for runtime checking.
The tool STADY [15] shares some elements of our approach. It instruments C functions with additional code for ensuring pre- and postconditions compliance, allowing
monitoring and test generation. However, the tool performs a simple ACSL-to-C translation, it does neither take into account dependencies among C left-values, nor it inferences their domain of definition.
7
Conclusion
This paper has presented a novel technique to automatically generate an analysis context
from a formal precondition of a C function. The core of the system has been formalized,
while we provide enough details about code generation to allow similar systems to
be implemented. Future work includes the formalization of code generation as well as
statements and proofs of the fundamental properties of the system as a whole. A running
example from the real world has also illustrated our presentation. The whole system is
XV
implemented in the Frama-C plug-in CfP. It generates code as close as possible to
human-written code. It is used in an operational industrial setting and already revealed
some mistakes in contexts previously written by hand by expert verification engineers.
Acknowledgments
Part of the research work leading to these results has received funding for the S3P
project from French DGE and BPIFrance. The authors thank TrustInSoft for the support and, in particular, Pascal Cuoq, Benjamin Monate and Anne Pacalet for providing
the initial specification, test cases and insightful comments. Thanks to the anonymous
reviewers for many useful suggestions and advice.
References
1. M. Barnett, M. Fähndrich, P. de Halleux, F. Logozzo, and N. Tillmann. Exploiting the synergy between automated-test-generation and programming-by-contract. In ICSE’09.
2. P. Baudin, J.-C. Filliâtre, C. Marché, B. Monate, Y. Moy, and V. Prevosto. ACSL: ANSI/ISO
C Specification Language. http://frama-c.com/acsl.html.
3. S. Blazy, D. Bühler, and B. Yakobowski. Structuring Abstract Interpreters through State and
Value Abstractions. In VMCAI’17.
4. W. Blume and R. Eigenmann. Symbolic Range Propagation. In IPPS’95.
5. B. Botella, M. Delahaye, S. H. T. Ha, N. Kosmatov, P. Mouy, M. Roger, and N. Williams.
Automating Structural Testing of C Programs: Experience with PathCrawler. In AST’09.
6. G. Canet, P. Cuoq, and B. Monate. A Value Analysis for C Programs. In SCAM’09.
7. P. Cousot and R. Cousot. Abstract Interpretation: A Unified Lattice Model for Static Analysis
of Programs by Construction or Approximation of Fixpoints. In POPL’77.
8. P. Cuoq and R. Rieu-Helft. Result graphs for an abstract interpretation-based static analyzer.
In JFLA’17.
9. M. Delahaye and N. Kosmatov. A Late Treatment of C Precondition in Dynamic Symbolic
Execution. In CSTVA’13.
10. ISO. The ANSI C standard (C99). Technical Report WG14 N1124, ISO/IEC, 1999. http:
//www.open-std.org/JTC1/SC22/WG14/www/docs/n1124.pdf.
11. F. Kirchner, N. Kosmatov, V. Prevosto, J. Signoles, and B. Yakobowski. Frama-C: A Software Analysis Perspective. Formal Aspects of Computing, 2015.
12. V. Kuncak, M. Mayer, R. Piskac, and P. Suter. Complete Functional Synthesis. In PLDI’10.
13. F. Logozzo and M. Fähndrich. Pentagons: A Weakly Relational Abstract Domain for the
Efficient Validation of Array Accesses. In SAC’08.
14. H. Nazaré, I. Maffra, W. Santos, L. Barbosa, L. Gonnord, and F. M. Quintão Pereira. Validation of Memory Accesses Through Symbolic Analyses. SIGPLAN Not., 49(10), 2014.
15. G. Petiot, B. Botella, J. Julliand, N. Kosmatov, and J. Signoles. Instrumentation of Annotated
C Programs for Test Generation. In SCAM’14.
16. N. Polikarpova, I. Kuraj, and A. Solar-Lezama. Program Synthesis from Polymorphic Refinement Types. In PLDI’16.
17. W. Pugh. A Practical Algorithm for Exact Array Dependence Analysis. Comm. ACM, 1992.
18. R. Rugina and M. Rinard. Symbolic Bounds Analysis of Pointers, Array Indices, and Accessed Memory Regions. In PLDI’00.
19. K. Sen, D. Marinov, and G. Agha. CUTE: A Concolic Unit Testing Engine for C. In FSE’13.
20. A. Solar-Lezama, G. Arnold, L. Tancau, R. Bodik, V. Saraswat, and S. Seshia. Sketching
Stencils. In PLDI’07.
21. TrustInSoft.
PolarSSL 1.1.8 verification kit, v1.0.
Technical report.
http://
trust-in-soft.com/polarSSL_demo.pdf.
| 6cs.PL
|
Eldan’s Stochastic Localization and the KLS Hyperplane Conjecture:
An Improved Lower Bound for Expansion
arXiv:1612.01507v2 [math.FA] 9 Dec 2016
Yin Tat Lee∗, Santosh S. Vempala†
December 12, 2016
Abstract
1/4
We show that the KLS constant
√ for n-dimensional isotropic logconcave measures is O(n ), improving on
the current best bound of O(n1/3 log n). As corollaries we obtain the same improved bound on the thin-shell
estimate, Poincaré constant and exponential concentration constant and an alternative proof of this bound for
the isotropic constant; it also follows that the ball walk for sampling from an isotropic logconcave density in Rn
converges in O∗ (n2.5 ) steps from a warm start.
1
Introduction
The isoperimetry of a subset is the ratio of the measure of the boundary of the subset to the measure of the subset
or its complement, whichever is smaller. The minimum such ratio over all subsets is the Cheeger constant, also
called expansion or isoperimetric coefficient. This fundamental constant appears in many settings, e.g., graphs and
convex bodies and plays an essential role in many lines of study.
In the geometric setting, the KLS hyperplane conjecture [21] asserts that for any distribution with a logconcave
density, the minimum expansion is approximated by that of a halfspace, up to a universal constant factor. Thus,
if the conjecture is true, the Cheeger constant can be essentially determined simply by examining hyperplane cuts.
More precisely, here is the statement. We use c, C for absolute constants, and kAk2 for the spectral/operator norm
of a matrix A.
Conjecture 1 ([21]). For any logconcave density p in Rn with covariance matrix A,
R
p(x)dx
c
1 def
∂S
o ≥p
nR
= inf n
.
R
S⊆R
ψp
kAk2
min S p(s)dx, Rn \S p(x)dx
For an isotropic logconcave density (all eigenvalues of its covariance matrix are equal to 1), the conjectured
isoperimetric ratio is an absolute constant. Note that the isoperimetric constant or KLS constant ψp is the reciprocal
of the minimum expansion or Cheeger constant (this will be more convenient for comparisons with other constants).
This conjecture was formulated by Kannan, Lovász and Simonovits in the course of their study of the convergence
of a random process (the ball walk) in a convex body and they proved the following weaker bound.
Theorem 2 ([21]). For any logconcave density p in Rn with covariance matrix A, the KLS constant satisfies
p
ψp ≤ C Tr(A).
√
For an isotropic distribution, the theorem gives a bound of O ( n), while the conjecture says O (1).
The conjecture has several important consequences. It implies that the ball walk mixes in O∗ n2 steps from a
warm start in any isotropic convex body (or logconcave density) in Rn ; this is the best possible bound, and is tight
e.g., for a hypercube.
The KLS conjecture has become central to modern asymptotic convex geometry. It is equivalent to a bound on
the spectral gap of isotropic logconcave functions [25]. Although it was formulated due to an algorithmic motivation,
it implies several well-known conjectures in asymptotic convex geometry. We describe these next.
The thin-shell conjecture (also known as the variance hypothesis) [29, 6] says the following.
∗ Microsoft
† Georgia
Research and University of Washington, [email protected] and [email protected]
Tech, [email protected]
1
Conjecture 3 (Thin-shell). For a random point X from an isotropic logconcave density p in Rn ,
√
def
σp2 = E((kXk − n)2 ) = O(1).
It implies that a random point X from an isotropic logconcave density lies in a constant-width annulus (a thin
shell) with constant probability. Noting that
√
1
σp2 = E((kXk − n)2 ) ≤ Var(kXk2 ) ≤ Cσp2 ,
n
2
the conjecture is equivalent to asserting that Var(kXk ) = O(n) for an isotropic logconcave density. The following
1
connection is well-known: σp ≤ Cψp . The current best bound is σp ≤ n 3 by Guedon and Milman [20], improving
on a line of work that started with Klartag [23, 24, 18]. Eldan [14] has shown that the reverse inequality holds
approximately, in a worst-case sense, namely the worst possible KLS constant over all isotropic logconcave densities
in Rn is bounded by the thin-shell estimate
to within roughly a logarithmic factor in the dimension. This results
1√
in the current best bound of ψp ≤ n 3 log n. A weaker inequality was shown earlier by Bobkov [5] (see also [30]).
The slicing conjecture, also called the hyperplane conjecture [8, 4] is the following.
Conjecture 4 (Slicing/Isotropic constant). Any convex body of unit volume in Rn contains a hyperplane section
of at least constant volume. Equivalently, for any convex body K of unit volume with covariance matrix L2K I, the
isotropic constant LK = O(1).
The isotropic constant of a general isotropic logconcave density p with covariance a multiple of the identity is
defined as Lp = p(0)1/n . The best current bound is Lp = O(n1/4 ), due to Klartag [22], improving on Bourgain’s
bound of Lp = O(n1/4 log n) [7]. The study of this conjecture has played an influential role in the development of
convex geometry over the past several decades. It was shown by Ball that the KLS conjecture implies the slicing
conjecture. More recently, Eldan and Klartag [15] showed that the thin shell conjecture implies slicing, and therefore
an alternative (and stronger) proof that KLS implies slicing: Lp ≤ Cσp .
The next conjecture is a bound on the Poincaré constant for logconcave distributions.
Conjecture 5 (Poincaré constant). For any isotropic logconcave density p in Rn , we have
Q2p =
def
Varp (g(x))
= O(1).
smooth Ep k∇g(x)k2
2
sup
g
It was shown by Maz’ja [34] and Cheeger [11] that this Poincaré constant is bounded by twice the KLS constant
(Qp ≤ 2ψp ). The current best bound is the same as the KLS bound.
Finally, it is conjectured that Lipschitz functions concentrate over isotropic logconcave densities.
Conjecture 6 (Lipschitz concentration). For any L-Lipschitz function g in Rn , and isotropic logconcave density
p,
Px∼p (|g(x) − Eg| > t) ≤ e−t/(Dp L)
where Dp = O(1).
Gromov and Milman [19] showed that Dp is also bounded by the KLS constant (see Lemma ??). For more
background on these conjectures, we refer the reader to [10, 2, 3].
1.1
Results
We prove the following bound, conjectured in this form in [33].
Theorem 7. For any logconcave density p in Rn , with covariance matrix A,
1/4
.
ψp ≤ C Tr A2
1
For isotropic p, this gives a bound of ψp ≤ Cn 4 , improving on the current best bound. The following corollary
is immediate.
Corollary 8. For any logconcave density p in Rn , the isotropic (slicing) constant Lp , the
Poincare constant Qp ,
the thin-shell constant σp and the concentration coefficient Dp are all bounded by O n1/4 .
We mention an algorithmic consequence.
Corollary 9. The mixing time of the ball walk to sample from an isotropic logconcave density from a warm start
is O∗ n2.5 .
2
1.2
Approach
The KLS conjecture is true for Gaussian distributions. More generally, for any distribution whose density function
is the product of the Gaussian density for N (0, σ 2 I) and any logconcave function, it is known that the expansion is
Ω(1/σ). This fact is used crucially in the Gaussian cooling algorithm of [13] for computing the volume of a convex
body by starting with a standard Gaussian restricted to a convex body and gradually making the variance of the
Gaussian large enough that it is effectively uniform over the convex body of interest. Our overall strategy is similar
in spirit — we start with an arbitrary isotropic logconcave density and gradually introduce a Gaussian term in
the density of smaller and smaller variance. The isoperimetry of the resulting distribution after sufficient time will
be very good since it has a large Gaussian factor. And crucially, it can be related to the isoperimetry of initial
distribution. To achieve the latter, we would like to maintain the measure of a fixed subset close to its initial value as
the distribution changes. For this, our proof uses the localization approach to proving high-dimensional inequalities
[27, 21], and in particular, the elegant stochastic version introduced by Eldan [14] and used in subsequent papers
[17, 16].
We fix a subset E of the original space with the original logconcave measure of measure 12 . This is without
loss of generality due to a result of [30]. In standard localization we then bisect space using a hyperplane that
preserves the volume fraction of E. The limit of this process is 1-dimensional logconcave measures (“needles”),
for which inequalities are much easier to prove. This approach runs into major difficulties for proving the KLS
conjecture. While the original measure might be isotropic, the 1-dimensional measures could, in principle, have
variances roughly equal to the trace of the original covariance (i.e., long thin needles), for which only much weaker
inequalities hold. Stochastic localization can be viewed as the continuous time version of this process, where at
each step, we pick a random direction and multiply the current density with a linear function along the chosen
direction. Over time, the distribution can be viewed as a spherical Gaussian times a logconcave function, with
the Gaussian gradually reducing in variance. When the Gaussian becomes sufficiently small in variance, then the
overall distribution has good isoperimetric coefficient, determined by the inverse of the Gaussian standard deviation
(such an inequality can be shown using standard localization, as in [12]). An important property of the infinitesimal
change at each step is balance – the density at time t is a martingale and therefore the expected measure of any
subset is the same as the original measure. Over time, the measure of a set E is a random quantity that deviates
from its original value of 21 over time. The main question is then what direction to use at each step so that (a)
the measure of E remains bounded and (b) the Gaussian part of the density has small variance. We show that
the simplest choice, namely a pure random direction chosen from the uniform distribution suffices. The analysis
needs a potential function that grows slowly but still maintains good control over the spectral norm of the current
covariance matrix. The direct choice of kAt k2 , where At is the covariance matrix of the distribution at time t, is
hard to control. We use Tr(A2t ). This gives us the improved bound of O(n1/4 ). In the appendix, we show that a
third moment assumption implies further improvement via the same localization technique1 .
2
Preliminaries
In this section, we review some basic definitions and theorems that we use.
2.1
Stochastic calculus
In this paper, we only consider stochastic processes given by stochastic differential equations. Given real-valued
stochastic processes xt and yt , the quadratic variations [x]t and [x, y]t are real-valued stochastic processes defined
by
[x]t = lim
|P |→0
[x, y]t = lim
|P |→0
∞
X
n=1
∞
X
n=1
xτn − xτn−1
2
,
xτn − xτn−1
yτn − yτn−1 ,
where P = {0 = τ0 ≤ τ1 ≤ τ2 ≤ · · · ↑ t} is a stochastic partition of the non-negative real numbers, |P | =
maxn (τn − τn−1 ) is called the mesh of P and the limit is defined using convergence in probability. Note that [x]t
1 In
the first arXiv version of this paper, we incorrectly claimed the assumption as a lemma
3
is non-decreasing with t and [x, y]t can be defined via polarization as
[x, y]t =
1
([x + y]t − [x − y]t ) .
4
For example, if the processes xt and yt satisfy the SDEs dxt = µ(xt )dt + σ(xt )dWt and dyt = ν(yt )dt + η(yt )dWt
Rt
Rt
where Wt is a Wiener process, we have that [x]t = 0 σ 2 (xs )ds and [x, y]t = 0 σ(xs )η(ys )ds and d[x, y]t =
σ(xt )η(yt )dt; for a vector-valued SDE dxt = µ(xt )dt + Σ(xt )dWt and dyt = ν(yt )dt + M (yt )dWt , we have that
Rt
Rt
[xi , xj ]t = 0 (Σ(xs )ΣT (xs ))ij ds and d[xi , y j ]t = 0 (Σ(xs )M T (ys ))ij ds.
Lemma 10 (Itô’s formula). Let x be a semimartingale and f be twice continuously differentiable function, then
df (xt ) =
X df (xt )
i
dxi
dxi +
1 X d2 f (xt ) i j
d[x , x ]t .
2 i,j dxi dxj
The next two lemmas are well-known facts about Wiener processes; first the reflection principle.
Lemma 11 (Reflection principle). Given a Wiener process W (t) and a, t ≥ 0, then we have that
P( sup W (s) ≥ a) = 2P(W (t) ≥ a).
0≤s≤t
Second, a decomposition lemma for continuous martingales.
Theorem 12 (Dambis, Dubins-Schwarz theorem). Every continuous local martingale Mt is of the form
Mt = M0 + W[M]t for all t ≥ 0
where Ws is a Wiener process.
2.2
Logconcave functions
Lemma 13 (Dinghas; Prékopa; Leindler). The convolution of two logconcave functions is also logconcave; in
particular, any linear transformation or any marginal of a logconcave density is logconcave.
The next lemma about logconcave densities is folklore, see e.g., [28].
Lemma 14 (Logconcave moments). Given a logconcave density p in Rn , and any positive integer k,
k/2
2
k
.
Ex∼p kxk ≤ (2k)k Ex∼p kxk
The following elementary concentration lemma is also well-known (this version is from [28]).
Lemma 15 (Logconcave concentration). For any isotropic logconcave density p in Rn , and any t > 0,
√
Px∼p kxk > t n ≤ e−t+1 .
A much stronger concentration bound was shown by Paouris [32].
Lemma 16 ([32]). For any isotropic logconcave distribution and any t > C,
√
√
Px∼p kxk ≥ t n ≤ e−c nt .
The following inequality bounding the small ball probability is from [3].
Lemma 17 ([3, Thm. 10.4.7]). For any isotropic logconcave density p, and any ǫ < ǫ0 ,
√
√
Px∼p kxk ≤ ǫ n ≤ ǫc n
were ǫ0 , c are absolute constants.
Definition 18. We define ψn as the supremum of the KLS constant over all isotropic logconcave distributions in
Rn .
4
The next lemma follows from the fact that the Poincaré constant is bounded by the KLS constant and Lemma
13.
Lemma 19. For any matrix A, and any isotropic logconcave density p,
2
Varx∼p xT Ax ≤ O(ψr2 )Ex∼p kAxk
where r = rank(A + AT ).
To prove a lower bound on the expansion, it suffices to consider subsets of measure 1/2. This follows from the
concavity of the isoperimetric profile. We quote a theorem from [30, Thm 1.8], which applies even more generally
to Riemannian manifolds under suitable convexity-type assumptions.
Theorem 20. The Cheeger constant of any logconcave density is achieved by a subset of measure 1/2.
2.3
Matrix inequalities
√
For any symmetric matrix B, we define |B| = B 2 , namely, the matrix formed by taking absolute value of all
eigenvalues of B. For any matrix A, we define R(A) to be the span of the rows of A and N (A) to be the null space
of A. For any vector x and any positive semi-definite matrix A, we definekxk2A = xT Ax.
Lemma 21 (Matrix Hölder inequality). Given a symmetric matrices A and B and any s, t ≥ 1 with s−1 + t−1 = 1,
we have
1/t
t
s 1/s
Tr |B|
.
Tr(AB) ≤ (Tr |A| )
Lemma 22 (Lieb-Thirring Inequality [26]). Given positive semi-definite matrices A and B and r ≥ 1, we have
Tr((B 1/2 AB 1/2 )r ) ≤ Tr(B r/2 Ar B r/2 ).
Since the following lemma is stated differently in [14, 1], we show the proof from Eldan [14] here for completeness.
Lemma 23 ([14, 1]). Given a symmetric matrix B, a positive semi-definite matrix A and α ∈ [0, 1], we have
Tr(Aα BA1−α B) ≤ Tr(AB 2 ).
Proof. Without loss of generality, we can assume A is diagonal. Hence, we have that
X
1−α 2
Aα
Tr(Aα BA1−α B) =
ii Ajj Bij
i,j
≤
X
i,j
=α
2
(αAii + (1 − α)Ajj ) Bij
X
i,j
3
2
Aii Bij
+ (1 − α)
X
2
Ajj Bij
= Tr(AB 2 ).
i,j
Eldan’s stochastic localization
In this section, we consider the stochastic localization scheme introduced in [14] in slightly more general terms. In
discrete localization, the idea would be to restrict the distribution with a random halfspace and repeat this process.
In stochastic localization, this discrete step is replaced by infinitesimal steps, each of which is a renormalization
with a linear function in a random direction. One might view this √
informally as an averaging over infinitesimal
needles. The discrete time equivalent would be pt+1 (x) = pt (x)(1 + h(x − µt )T w) for a sufficiently small h and
1 2
random Gaussian vector w. Using the approximation 1 + y ∼ ey− 2 y , we see that over time this process introduces
a negative quadratic factor in the exponent, which will be the Gaussian factor. As time tends to ∞, the distribution
tends to a more and more concentrated Gaussian and eventually a delta function, at which point any subset has
measure either 0 or 1. The idea of the proof is to stop at a time that is large enough to have a strong Gaussian
factor in the density, but small enough to ensure that the measure of a set is not changed by more than a constant.
5
3.1
The process and its basic properties
Given a distribution with logconcave density p(x), we start at time t = 0 with this distribution and at each time
t > 0, we apply an infinitesimal change to the density. This is done by picking a random direction from a Gaussian
with a certain covariance matrix Ct , called the control matrix. In Section 4 and 6, we use this process with Ct = I.
In Section 7, we use a varying Ct to get a bound for non-isotropic distributions.
In order to construct the stochastic process, we assume that the support of p is contained in a ball of radius
R > n. There is only exponentially small probability outside this ball, at most e−cR by Lemma 16. Moreover, since
by Theorem 20, we only need to consider subsets of measure 1/2, this truncation does not affect the KLS constant
of the distribution.
Definition 24. Given a logconcave distribution p, we define the following stochastic differential equation:
1/2
c0 = 0,
dct = Ct
B0 = 0,
(3.1)
dWt + Ct µt dt,
dBt = Ct dt,
where the probability distribution pt , the mean µt and the covariance At are defined by
T
pt (x) = R
2
1
ect x− 2 kxkBt p(x)
Rn
T
1
2
ect y− 2 kykBt p(y)dy
,
At = Ex∼pt (x − µt )(x − µt )T ,
µt = Ex∼pt x,
and the control matrices Ct are symmetric matrices to be specified later.
In Section 4 and 6, we only consider the process with Ct = I for all t ≥ 0. In this case, we have that
T
Bt = tI,
pt (x) = R
2
t
ect x− 2 kxk2 p(x)
Rn
T
t
2
ect y− 2 kyk2 p(y)dy
.
Also, since µt is a bounded function that is Lipschitz with respect to c and hence standard existence and uniqueness
theorems (e.g. [31, Sec 5.2]) show the existence and uniqueness of the solution on time [0, T ] for any T > 0. In
general, we have the following result:
1/2
Lemma 25 (Existence and Uniqueness). If p(x) has compact support and if Ct
are bounded and Lipschitz
functions of At and Bt , the stochastic differential equation (3.1) has a unique solution.
We defer all proofs for statements in this section, considered standard in stochastic calculus, to Section 5. Now
we proceed to analyzing the process and how its parameters evolve. Roughly speaking, the first lemma below says
that the stochastic process is the same as continuously multiplying pt (x) by a random infinitesimally small linear
function.
Lemma 26 ([14, Lem 2.1]). We have that
1/2
dpt (x) = (x − µt )T Ct
n
for any x ∈ R .
dWt pt (x)
By considering the derivative d log pt (x), we see that applying dpt (x) as in the lemma above results in the
distribution pt (x), with the Gaussian term in the density:
d log pt (x) =
dpt (x) 1 d[pt (x)]t
−
pt (x)
2 pt (x)2
1
1/2
= (x − µt )T Ct dWt − (x − µt )T Ct (x − µt )dt
2
1
1
1/2
1/2
T
= x Ct dWt + Ct µt dt − xT Ct xdt − (µTt Ct dWt + µTt Ct µt dt)
2
2
1
= xT dct − xT dBt xdt + g(t)
2
where the last term is independent of x and the first two terms explain the form of pt (x) and the appearance of the
Gaussian.
Next we analyze the change of the covariance matrix.
Lemma 27 ([14]). We have that
Z
1/2
(x − µt )(x − µt )T (x − µt )T Ct dWt pt (x)dx − At Ct At dt.
dAt =
Rn
6
3.2
Bounding expansion
Our plan is to bound the expansion by the spectral norm of the covariance matrix at time t. First, we bound the
measure of a set of initial measure 21 .
R
Lemma 28. For any set E ⊂ Rn with E p(x)dx = 21 and t ≥ 0, we have that
P(
Proof. Let gt =
R
E
1
≤
4
Z
E
Z t
3
9
)≥
− P(
Cs1/2 As Cs1/2
4
10
0
pt (x)dx ≤
2
ds ≥
1
).
64
pt (x)dx. Then, we have that
Z
1/2
(x − µt )T Ct dWt pt (x)dx
dgt =
E
Z
1/2
=
(x − µt )pt (x)dx, Ct dWt
E
where the integral might not be 0 because it is over the subset E and not all of Rn . Hence, we have,
Z
d[gt ]t =
1/2
E
Ct
2
(x − µt )pt (x)dx
dt
2
2
= max
ζ
− µt )pt (x)dx dt
kζk2 ≤1
E
Z
Z
2
1/2
≤ max
ζ T Ct (x − µt ) pt (x)dx
pt (x)dxdt
kζk2 ≤1 Rn
Rn
1/2
1/2
= max ζ T Ct At Ct ζ dt
Z
T
1/2
Ct (x
kζk2 ≤1
1/2
= Ct
Hence, we have that
d[gt ]t
dt
≤
1/2
Ct
1/2
At Ct
1/2
At Ct
2
2
dt.
. By the Dambis, Dubins-Schwarz theorem, there exists a Wiener
process W̃t such that gt − g0 has the same distribution as W̃[g]t . Using g0 = 21 , we have that
P(
1
3
−1
1
≤ gt ≤ ) = P(
≤ W̃[g]t ≤ )
4
4
4
4
1
1
≥ P( min1 W̃s ≤ and [g]t ≤
)
4
64
0≤s≤ 64
1
1
)
≥ 1 − P( max1 W̃s > ) − P([g]t >
4
64
0≤s≤ 64
1
1
1
1 >
) − P([g]t >
)
≥ 1 − 4P(W̃ 64
4
64
2 9
1
≥
− P([g]t >
)
10
64
where we used reflection principle for 1-dimensional Brownian motion in 1 and the concentration of normal distribution in 2 , namely Px∼N (0,1) (x > 2) ≤ 0.0228.
Theorem 29 (Brascamp-Lieb [9]). Let γ : Rn → R+ be the standard Gaussian density in Rn . Let f : Rn → R+ be
any logconcave function. Define the density function h as follows:
f (x)γ(x)
.
Rn f (y)γ(y) dy
h(x) = R
Fix a unit vector v ∈ Rn , let µ = Eh (x). Then, for any α ≥ 1, Eh (|v T (x − µ)|α ) ≤ Eγ (|v T x|α ).
Using this we derive the following well-known isoperimetric inequality that was proved in [12] and was also used
in [14].
7
R
1 T
1 T
Theorem 30 ([12, Thm. 4.4]). Let h(x) = f (x)e− 2 x Bx / f (y)e− 2 y By dy where f : Rn → R+ is an integrable
logconcave function and B is positive definite. Then h is logconcave and for any measurable subset S of Rn ,
(Z
)
Z
Z
1
−
h(x)dx = Ω B −1 2 2 min
h(x)dx .
h(x)dx,
∂S
Rn \S
S
In other words, the expansion of h is Ω
B −1
− 21
2
.
Proof. The proof uses the localization lemma to reduce the statement to a 1-dimensional statement
about a Gaussian
times a logconcave density, where the Gaussian is a projection of the Gaussian N 0, B −1 (but the logconcave
function might be different as the limit of localization is the original function along an interval times an exponential
function). We then apply the Brascamp-Lieb inequality in one dimension (Theorem 29) to prove that for the
resulting one-dimensional distribution, the variance is at most that of the Gaussian, therefore at most B −1 . The
isoperimetric constant is bounded by the inverse of the standard deviation times a constant. The complete proof,
in more general terms, is carried out in [12, Thm. 4.4].
We can now prove a bound on the expansion.
Lemma 31. Given a logconcave distribution p. Let At be defined by Definition 24 using initial distribution p.
Suppose that there is T > 0 such that
!
Z T
3
1
and BT uI ≥
P
Cs1/2 As Cs1/2 ds ≤
64
4
2
0
Then, we have that ψp = Ω u−1/2 .
Proof. By Milman’s theorem [30], it suffices to consider subsets of measure 12 . Consider any measurable subset E
of Rn of initial measure 21 . By Lemma 26, pt is a martingale and therefore
Z
Z
Z
p0 (x)dx = E
p(x)dx =
pt (x)dx .
∂E
∂E
∂E
cT x− 1 kxk2
Next, by the definition ofpT (3.1), we have that pT (x) ∝ e T 2 BT p(x) and Theorem 30 shows that the expansion
of E is Ω λmin (BT )−1/2 . Hence, we have
Z
Z
p(x)dx = E
pT (x)dx
∂E
∂E
Z
Z
−1/2
= Ω(u
)E 1BT uI min
pT (x)dx
pT (x)dx,
Ē
E
Z
3
1
−1/2
pT (x)dx ≤
≥ Ω(u
)P BT uI and ≤
4
4
E
Z
3
1
−1/2
− P(BT uI is false)
pT (x)dx ≤
≤
= Ω(u
) P
4
4
E
Z t
9
1
−1/2
1/2
1/2
≥ Ω(u
)
− P(
) − P(BT uI is false)
(Lem 28)
Cs As Cs
ds ≥
10
64
2
0
= Ω(u−1/2 )
where we used the assumption at the end. Using Theorem 20, this shows that ψp = Ω u−1/2 .
4
Controlling At via the potential Tr(A2t )
In this section, we only use Ct = I for the control matrix.
8
4.1
Third moment bounds
Here are two key lemmas about the third-order tensor of a log-concave distribution. A special case of the first
inequality was used in [14]. For our main theorem, we only the first lemma with B = I, but we need the general
case for the proof in Section 7.
Lemma 32. Given a logconcave distribution p with mean µ and covariance A. For any B 0, symmetric matrix
C, we have that
1/2
Ex∼p B 1/2 (x − µ)(x − µ)T C(x − µ) = O A1/2 BA1/2
Tr A1/2 CA1/2 .
2
2
Proof. We first consider the case C = vv T . Taking y = A−1/2 (x − µ) and w = A1/2 v. Then, y follows an isotropic
log concave distribution p̃ and the statement becomes
1/2
2
2
Ey∼p̃ B 1/2 A1/2 y y T w
= O A1/2 BA1/2
kwk2 .
2
2
Then, we calculate that
Ey∼p̃ B 1/2 A1/2 y y T w
2
2
= max Ey∼p̃ (B 1/2 A1/2 y)T ζ y T w
kζk2 ≤1
2
q
2 q
4
≤ max
Ey∼p̃ (B 1/2 A1/2 y)T ζ
Ey∼p̃ (y T w)
kζk2 ≤1
q
2
=O
A1/2 BA1/2 2 kwk2
where we used the fact that for a fixed w, y T w has a one-dimensional logconcave distribution (Lemma 13) and
hence Lemma 14 shows that
4
2 2
4
= O(kwk2 ).
Ey∼p̃ y T w = O(1) Ey∼p̃ y T w
For a general symmetric matrix C, we write C =
Hence, we have that
Ex∼p B 1/2 (x − µ)(x − µ)T C(x − µ)
2
≤
P
λi vi viT where λi , vi are eigenvalues and eigenvectors of C.
X
i
|λi | Ex∼p B 1/2 (x − µ)(x − µ)T vi viT (x − µ)
≤ O(1)
X
i
|λi | A1/2 BA1/2
= O(1) A1/2 BA1/2
= O(1) A1/2 BA1/2
1/2
2
1/2
2
X
i
1/2
2
A1/2 vi
2
2
Tr A1/2 |λi |vi viT A1/2
Tr A1/2 CA1/2 .
Lemma 33. Given a logconcave distribution p with mean µ and covariance A. We have
3/2
3
.
Ex,y∼p |hx − µ, y − µi| = O Tr A2
Proof. Without loss of generality, we assume µ = 0. For a fixed x and random y, hx, yi follows a one-dimensional
logconcave distribution (Lemma 13) and hence Lemma 14 shows that
3/2
3/2
3
= O xT Ax
Ey∼p |hx, yi| ≤ O(1) Ey∼p hx, yi2
.
Next, we note that A1/2 x follows a logconcave distribution (Lemma 13) and hence Lemma 14 shows that
2 3/2
3
3/2
3
.
= O Tr A2
Ex,y∼p |hx, yi| = O(1)Ex∼p A1/2 x ≤ O(1) Ex∼p A1/2 x
9
4.2
Analysis of At
Using Itô’s formula and Lemma 27, one can compute the derivatives of TrA2t . Since a similar calculation appears
in Sections 4, 6 and 7, we prove a common generalization in Lemma 36.
Lemma 34. Let At be defined by Definition 24. We have that
3
dTrA2t =2Ex∼pt (x − µt )T At (x − µt )(x − µt )T dWt − 2Tr(A3t )dt + Ex,y∼pt (x − µt )T (y − µt ) dt.
Lemma 35. Given a logconcave distribution p with covariance matrix A s.t.TrA2 = n. Let At defined by Definition
24 using initial distribution p. There is a universal constant c1 such that
P( max Tr A2t ≥ 8n) ≤ 0.01
t∈[0,T ]
with
c1
T = √ .
n
Proof. Let Φt = TrA2t . By Lemma 34, we have that
3
dΦt = − 2Tr(A3t )dt + Ex,y∼pt (x − µt )T (y − µt ) dt + 2Ex∼pt (x − µt )T At (x − µt )(x − µt )T dWt
def
=δt dt + vtT dWt .
(4.1)
For the drift term δt dt, Lemma 33 shows that
3/2
3
3/2
≤ C ′ Φt
δt ≤ Ex,y∼pt (x − µt )T (y − µt ) = O Tr A2t
(4.2)
for some universal constant C ′ . Note that we dropped the term −2Tr(A3t ) since At is positive semidefinite and
therefore the term is negative.
For the martingale term vtT dWt , we note that
kvt k2 = Ex∼pt (x − µt )T At (x − µt )(x − µt )
≤
≤
1/2
kAt k2 Tr
5/4
O(Φt ).
2
A2t
(Lem 32)
5/4 √
So the drift term grows roughly as Φ3/2 t while the stochastic
t. Thus, both bounds (on the drift
grows as Φt
term
1
term and the stochastic term) suggest that for t up to O √n , the potential Φt remains O(n). We now formalize
this, by decoupling the two terms.
Let
1
.
f (a) = − √
a+n
By (4.1) and Itô’s formula, we have that
1
df (Φt ) = f ′ (Φt )dΦt + f ′′ (Φt )d[Φ]t
2
!
2
kvt k2
1 vtT dWt
3
δt
1
dt
+
−
=
2 (Φt + n)3/2
8 (Φt + n)5/2
2 (Φt + n)3/2
≤ C ′ dt + dYt
(4.3)
v T dW
t
where dYt = 21 (Φtt+n)3/2
, Yt = 0 and C ′ is the universal constant in (4.2).
Note that
2
d[Y ]t
Φ5/2
C
1 kvt k2
= O(1)
≤√ .
=
3
3
dt
4 (Φt + n)
(Φt + n)
n
By Theorem 12, there exists a Wiener process W̃t such that Yt has the same distribution as W̃[Y ]t . Using the
reflection principle for 1-dimensional Brownian motion, we have that
√
γ2 n
).
P( max Yt ≥ γ) ≤ P( max W̃t ≥ γ) = 2P(W̃ √C T ≥ γ) ≤ 2 exp(−
n
2CT
t∈[0,T ]
t∈[0, √Cn T ]
10
Since Φ0 = kAp k2F = n, we have that f (Φ0 ) = − √12n and therefore (4.3) shows that
√
γ2 n
1
′
).
P( max f (Φt ) ≥ − √ + C T + γ) ≤ 2 exp(−
2CT
t∈[0,T ]
2n
Putting T =
1
√
256(C ′ +C) n
and γ =
1
√
,
4 n
we have that
1
P( max f (Φt ) ≥ − √ ) ≤ 2 exp(−8)).
3 n
t∈[0,T ]
Note that f (Φt ) ≥ − 3√1 n implies that Φt ≥ 8n. Hence, we have that
P( max Φt ≥ 8n) ≤ 0.01.
t∈[0,T ]
4.3
Proof of Theorem 7
Proof of Theorem 7. By rescaling, we can assume TrA2 = n. By Lemma 35, we have that
c1
P( max Tr A2s ≤ 8n) ≥ 0.99 with t = √ .
n
s∈[0,t]
√
Since Tr A2t ≤ 8n implies that kAt k2 ≤ 8n, we have that
Z
P(
where T = min
5
n
1√
,c
64 8 1
T
0
kAs k ds ≤
1
and BT T · I) ≥ 0.99
64
o √
/ n. Now the theorem follows from Lemma 31.
Localization proofs
We begin with the proof of existence of a unique solution for the SDE.
Proof of Lemma (25). We can write the stochastic differential equation as
dct = C 1/2 (At , Bt )dWt + C(At , Bt )µ(ct , Bt )dt
and dBt = C(At , Bt )dt where
Z
xq(c, B, x)dx
µ(c, B) =
and At (c, B) =
Z
Rn
Rn
and
q(c, B, x) = R
ec
Rn
T
(x − µ(c, B))(x − µ(c, B))T q(c, B, x)dx
x− 12 kxk2B
p(x)
2
1
T
ec y− 2 kykB p(y)dy
.
Since p has compact support, we have that q is Lipschitz in c and B variables, so are the functions µ and A. Next,
we note that both µ and A are bounded since p has compact support. Since C 1/2 is bounded and Lipschitz function
in c and B variables, so is C. Therefore, we can use a standard existence and uniqueness theorem (e.g. [31, Sec
5.2]) to show the existence and uniqueness of the solution on time [0, T ] for any T > 0.
Next is the proof of the infinitesimal change in the density.
T
1
2
Proof of Lemma (26). Let qt (x) = ect x− 2 kxkBt p(x). By Itô’s formula, applied to f (c, B) = ec
have that
1
1
1
d[cTt x]t − d[hBt x, xi]t
qt (x)
dqt (x) = dcTt x − hdBt x, xi +
2
2
2
def
11
T
x− 21 kxk2B
p(x), we
Note that
Hence, the quadratic variations of cTt x is
D
E
1/2
dcTt x = Ct dWt + Ct µt dt, x .
d[cTt x]t = hCt x, xi dt.
Also, dBt is a predictable process (namely, does not have a stochastic term) and hence d[hBt x, xi]t = 0. Therefore,
this gives
D
E
1/2
dqt (x) = Ct dWt + Ct µt dt, x qt (x).
(5.1)
Let Vt =
R
Rn
qt (y)dy. Then, we have
dVt =
Z
dqt (y)dy
n
ZR D
E
1/2
=
Ct dWt + Ct µt dt, y qt (y)dy
Rn
E
D
1/2
= Vt Ct dWt + Ct µt dt, µt .
By Itô’s formula, we have that
1
1
dVt + 3 d[V ]t
2
Vt
Vt
D
E
1/2
−1
= −Vt
Ct dWt + Ct µt dt, µt + Vt−1 hCt µt , µt i dt
E
D
1/2
= −Vt−1 Ct dWt , µt .
dVt−1 = −
(5.2)
Combining (5.1) and (5.2), we have that
dpt (x) =d(Vt−1 qt (x))
=qt (x)dVt−1 + Vt−1 dqt (x) + d[Vt−1 , qt (x)]t
E
D
E
D
E
D
1/2
1/2
1/2
1/2
= − qt (x)Vt−1 Ct dWt , µt + Vt−1 Ct dWt + Ct µt dt, x qt (x) − Vt−1 Ct µt , Ct x qt (x)dt
D
E
1/2
=pt (x) Ct dWt , x − µt .
The next proof is for the change in the covariance matrix.
Proof of Lemma (27). Recall that
At =
Z
Rn
(x − µt )(x − µt )T pt (x)dx.
Viewing At = f (µt , pt ), i.e., as a function of the variables µt and pt , we apply Itô’s formula. In the derivation below,
we use [µt , µTt ]t to denote the matrix whose i, j coordinate is [µt,i , µt,j ]t . Similarly, [µt , pt (x)]t is a column vector
and [µTt , pt (x)]t is a row vector.
Z
(x − µt )(x − µt )T dpt (x)dx
dAt =
n
R
Z
Z
(x − µt )(dµt )T pt (x)dx
dµt (x − µt )T pt (x)dx −
−
Rn
Rn
Z
Z
1
1
T
− ·2
(x − µt )d[µt , pt (x)]t dx − · 2
d[µt , pt (x)]t (x − µt )T dx
2
2
Rn
Rn
Z
1
pt (x)dx
+ · 2d[µt , µTt ]t
2
Rn
12
where the factor 2 comes from the Hessians of x2 and xy. Now the second term vanishes because
Z
Z
T
(x − µt )pt (x)dx)T = 0.
dµt (x − µt ) pt (x)dx = dµt (
Rn
Rn
Similarly, the third term also vanishes:
Z
Rn
(x − µt )(dµt )T pt (x)dx = 0.
To compute the last 3 terms, we note that
Z
xpt (x)dx
dµt = d
Rn
Z
1/2
x(x − µt )T Ct dWt pt (x)dx
=
Rn
Z
Z
T 1/2
=
(x − µt )(x − µt ) Ct dWt pt (x)dx +
Rn
1/2
Rn
1/2
= At Ct
µt (x − µt )T Ct
dWt pt (x)dx
dWt .
Therefore, we have for the last term
X
1/2
1/2
1/2
1/2
d[µt , µTt ]t ij =
At Ct
At Ct
dt = (At Ct (Ct )T ATt )ij dt = (At Ct At )ij dt
iℓ
ℓ
jℓ
which we can simply write as d[µt , µTt ]t = At Ct At dt. Similarly, we have
d[µt , pt (x)]t = pt (x)At Ct (x − µt )dt.
This gives the fourth term
Z
Z
(x − µt )d[µTt , pt (x)]t dx =
Rn
Rn
Similarly, we have the fifth term
(x − µt )(x − µt )T Ct At pt (x)dtdx = At Ct At dt.
d[µt , pt (x)]t (x − µt )T dx = At Ct At dt. Combining all the terms, we have that
Z
(x − µt )(x − µt )T dpt (x)dx − At Ct At dt.
dAt =
R
Rn
Rn
Next is the proof of stochastic derivative of the potential Φt = Tr((At − γI)q ).
Lemma 36. Let At be defined by Definition 24. For any integer q ≥ 2, we have that
1/2
dTr((At − γI)q ) =qEx∼pt (x − µt )T (At − γI)q−1 (x − µt )(x − µt )T Ct dWt − qTr((At − γI)q−1 A2t Ct )dt
q X
Ex,y∼pt (x − µt )T (At − γI)α (y − µt )(x − µt )T (At − γI)β (y − µt )(x − µt )T Ct (y − µt )dt.
+
2
α+β=q−2
Proof. Let Φ(X) = Tr((X − γI)q ). Then the first and second-order directional derivatives of Φ at X is given by
∂Φ
∂X
H
= qTr((X − γI)q−1 H) and
∂2Φ
∂X∂X
=q
H1 ,H2
q−2
X
k=0
Tr((X − γI)k H2 (X − γI)q−2−k H1 ).
Using these and Itô’s formula, we have that
dTr((At − γI)q ) = qTr((At − γI)q−1 dAt ) +
q
2
X
X
α+β=q−2 ijkl
Tr((At − γI)α eij (At − γI)β ekl )d[Aij , Akl ]t
where eij is the matrix that is 1 in the entry (i, j) and 0 otherwise, and Aij is the real-valued stochastic process
defined by the (i, j)th entry of At .
13
Using Lemma 27 and Lemma 26, we have that
1/2
dAt = Ex∼pt (x − µt )(x − µt )T (x − µt )T Ct
T
T
= Ex∼pt (x − µt )(x − µt ) (x − µt )
dWt − At Ct At dt
1/2
Ct ez dWt,z
− At Ct At dt
(5.3)
where Wt,z is the z th coordinate of Wt . Therefore,
X
1/2
1/2
d[Aij , Akl ]t =
Ex∼pt (x − µt )i (x − µt )j (x − µt )T Ct ez Ex∼pt (x − µt )k (x − µt )l (x − µt )T Ct ez dt
z
= Ex,y∼pt (x − µt )i (x − µt )j (y − µt )k (y − µt )l (x − µt )T Ct (y − µt )dt.
(5.4)
Using the formula for dAt (5.3) and d[Aij , Akl ]t (5.4), we have that
dTr((At − γI)q )
1/2
=qEx∼pt (x − µt )T (At − γI)q−1 (x − µt )(x − µt )T Ct dWt − qTr((At − γI)q−1 A2t Ct )dt
q X X
+
Tr((At − γI)α eij (At − γI)β ekl )Ex,y∼pt (x − µt )i (x − µt )j (y − µt )k (y − µt )l (x − µt )T Ct (y − µt )dt
2
α+β=q−2 ijkl
1/2
=qEx∼pt (x − µt )T (At − γI)q−1 (x − µt )(x − µt )T Ct dWt − qTr((At − γI)q−1 A2t Ct )dt
q X
Ex,y∼pt (x − µt )T (At − γI)α (y − µt )(x − µt )T (At − γI)β (y − µt )(x − µt )T Ct (y − µt )dt.
+
2
α+β=q−2
Acknowledgement. We thank Ravi Kannan, Laci Lovász, Assaf Naor and Nisheeth Vishnoi for their continuous
support and encouragement. We also thank Sébastien Bubeck, Ben Cousins, Ronen Eldan, Bo’az Klartag, Anup
B. Rao and Ting-Kam Leonard Wong for helpful discussions.
References
[1] Zeyuan Allen-Zhu, Yin Tat Lee, and Lorenzo Orecchia. Using optimization to obtain a width-independent,
parallel, simpler, and faster positive sdp solver. In Proceedings of the Twenty-Seventh Annual ACM-SIAM
Symposium on Discrete Algorithms, pages 1824–1831. SIAM, 2016.
[2] David Alonso-Gutiérrez and Jesús Bastero. Approaching the Kannan-Lovász-Simonovits and variance conjectures, volume 2131. Springer, 2015.
[3] Shiri Artstein-Avidan, Apostolos Giannopoulos, and Vitali D Milman. Asymptotic Geometric Analysis, Part
I, volume 202. 2015.
[4] K. M. Ball. Logarithmically concave functions and sections of convex sets in rn. Studia Mathematica, 88:69–84,
1988.
[5] S. Bobkov. On isoperimetric constants for log-concave probability distributions. Geometric aspects of functional
analysis, Lect. notes in Math., 1910:81–88, 2007.
[6] S. G. Bobkov and A. Koldobsky. On the Central Limit Property of Convex Bodies, pages 44–52. Springer
Berlin Heidelberg, Berlin, Heidelberg, 2003.
[7] J. Bourgain. On the distribution of polynomials on high dimensional convex sets. 1469:127–137, 1991.
[8] Jean Bourgain. On high dimensional maximal functions associated to convex bodies. American Journal of
Mathematics, 108(6):1467–1476, 1986.
[9] H. J. Brascamp and E. H. Lieb. On extensions of the brunn-minkowski and prekopa-liendler theorems, including inequalities for log-concave functions, and with application to the diffusion equationslogarithmic concave
measures and functions. J. Functional Anal., 22:366–389, 1976.
14
[10] Silouanos Brazitikos, Apostolos Giannopoulos, Petros Valettas, and Beatrice-Helen Vritsiou. Geometry of
isotropic convex bodies, volume 196. American Mathematical Society Providence, 2014.
[11] Jeff Cheeger. A lower bound for the smallest eigenvalue of the Laplacian, pages 195–199. Princeton Univ.
Press, 1969.
[12] B. Cousins and S. Vempala. A cubic algorithm for computing Gaussian volume. In SODA, pages 1215–1228,
2014.
[13] B. Cousins and S. Vempala. Bypassing KLS: Gaussian cooling and an O∗ (n3 ) volume algorithm. In STOC,
pages 539–548, 2015.
[14] R. Eldan. Thin shell implies spectral gap up to polylog via a stochastic localization scheme. Geometric and
Functional Analysis, 23:532–569, 2013.
[15] R. Eldan and B. Klartag. Approximately gaussian marginals and the hyperplane conjecture. Contermporary
Mathematics, 545, 2011.
[16] Ronen Eldan and James R. Lee. Talagrand’s convolution conjecture on gaussian space. In IEEE 56th Annual
Symposium on Foundations of Computer Science, FOCS 2015, Berkeley, CA, USA, 17-20 October, 2015, pages
1395–1408, 2015.
[17] Ronen Eldan and Joseph Lehec. Bounding the norm of a log-concave vector via thin-shell estimates. 2013.
[18] B. Fleury. Concentration in a thin Euclidean shell for log-concave measures. J. Funct. Anal., 259(4):832–841,
2010.
[19] M. Gromov and V. D. Milman. A topological application of the isoperimetric inequality. Amer. J. Math.,
105(4):843–854, 1983.
[20] Olivier Guedon and Emanuel Milman. Interpolating thin-shell and sharp large-deviation estimates for isotropic
log-concave measures. Geometric and Functional Analysis, 21(5):1043–1068, 2011.
[21] R. Kannan, L. Lovász, and M. Simonovits. Isoperimetric problems for convex bodies and a localization lemma.
Discrete & Computational Geometry, 13:541–559, 1995.
[22] B. Klartag. On convex perturbations with a bounded isotropic constant. Geom. and Funct. Anal., 16(6):1274–
1290, 2006.
[23] B. Klartag. A central limit theorem for convex sets. Invent. Math., 168:91–131, 2007.
[24] B. Klartag. Power-law estimates for the central limit theorem for convex sets. J. Funct. Anal., 245:284–310,
2007.
[25] M. Ledoux. Spectral gap, logarithmic sobolev constant, and geometric bounds. In Surveys in Diff. Geom.,
Vol. IX, pages 219–240. Int. Press, 2004.
[26] E Lieb and W Thirring. Inequalities for the moments of the eigenvalues of the Schrödinger equation and their
relation to Sobolev inequalities. Studies in Mathematical Physics: Essays in honor of Valentine Bargman,
Lieb, E., Simon, B., Wightman, AS (eds.), pages 269–303, 1976.
[27] L. Lovász and M. Simonovits. Random walks in a convex body and an improved volume algorithm. In Random
Structures and Alg., volume 4, pages 359–412, 1993.
[28] L. Lovász and S. Vempala. The geometry of logconcave functions and sampling algorithms. Random Structures
and Algorithms, 30(3):307–358, 2007.
[29] Irini Perissinaki Milla Anttila, Keith Ball. The central limit problem for convex bodies. Transactions of the
American Mathematical Society, 355(12):4723–4735, 2003.
[30] E. Milman. On the role of convexity in isoperimetry, spectral gap and concentration. Invent. Math., 177(1):1–43,
2009.
15
[31] Bernt Oksendal. Stochastic differential equations: an introduction with applications. Springer Science &
Business Media, 2013.
[32] G. Paouris. Concentration of mass on convex bodies. Geometric and Functional Analysis, 16:1021–1049, 2006.
[33] S. S. Vempala. Algorithmic aspects of convexity. Lecture notes from the Institut Henri Poincare Winter School,
January 19-23, 2015.
[34] V.G.Maz’ja. Classes of domains and imbedding theorems for function spaces. Dokl. Acad. Nauk SSSR (Engl.
transl. Soviet Math. Dokl., 1 (1961) 882-885), 3:527–530, 1960.
6
A reduction to a third moment assumption
In this section, we use the following assumption. In the first arXiv version of this paper, we claimed this assumption
as a lemma. While it might be true, our proof was not correct.
3
Assumption (Third Moment). For an isotropic logconcave distribution p ∈ Rn , Ex,y∼p hx, yi = O (nψn ).
P
3
Note that Ex,y∼p hx, yi = i,j,k Ex (xi xj xk )2 . Under this assumption we prove the following.
Theorem 37. Under
the third moment assumption above, any isotropic logconcave density p in Rn , the KLS
√
O( log n log log n)
constant is ψp = e
.
The proof will use the same process with Ct = I, but with a more sensitive potential function Φt = Tr((At − I)q )
for even integers q.
6.1
Tensor bounds
Definition 38 (3-Tensor). For any isotropic logconcave distribution p in Rn and any symmetric matrices A, B and
C, we define
Tp (A, B, C) = Ex,y∼p (xT Ay)(xT By)(xT Cy).
Often, we drop the subscript p to indicate the worst case bound
def
T (A, B, C) =
sup
isotropic logconcave p
Tp (A, B, C).
Remark. It is clear from the definition that T is symmetric, namely T (A1 , A2 , A3 ) = T (Aσ(1) , Aσ(2) , Aσ(3) ) for any
permutation σ.
We first start with some simple equalities
about a 3-tensor. Here we repeatedly use the elementary facts
Tr(AB) = Tr(BA), xT Ay = Tr AyxT .
Lemma 39. For any isotropic logconcave distribution p and any symmetric matrices A, B and C, we have that
X
Tr(A∆i B∆i )
Tp (A, B, I) =
i
and
Tp (A, B, I) =
X
Aij Tr(∆i B∆j )
i,j
where ∆i = Ex∼p xxT xi .
Proof. Direct calculation shows that
Tp (A, B, I) = Ex,y∼p xT AyxT ByxT y =
X
Ex,y∼p xT AyxT Byxi yi
i
=
X
i
X
Tr(A∆i B∆i )
Ex,y∼p Tr Axx Byy xi yi =
T
T
i
16
and
Tp (A, B, I) = Ex,y∼p xT AyxT ByxT y =
X
Aij Ex,y∼p xi yj xT ByxT y
i,j
=
X
i,j
X
Aij Tr(∆i B∆j ).
Aij Ex,y∼p Tr xx Byy T xi yj =
T
i,j
Lemma 40. For any A1 , A2 , A3 0, we have that
T (A1 , A2 , A3 ) ≥ 0
and for any symmetric matrices B1 , B2 , B3 , we have that
T (B1 , B2 , B3 ) ≤ T (|B1 | , |B2 | , |B3 |).
1/2
Proof. Fix any isotropic logconcave distribution p. We define ∆i = Ex∼p xxT xT A3 ei which is well defined since
A3 0. Then, we have that
Tp (A1 , A2 , A3 ) =Ex,y∼p xT A1 yxT A2 yxT A3 y
X
Tr(A1 ∆i A2 ∆i ).
=
i
1/2
1/2
Since ∆i is symmetric and A1 , A2 0, we have that A1 ∆i A2 ∆i A1 0 and Tr(A1 ∆i A2 ∆i ) ≥ 0. Therefore,
T (A1 , A2 , A3 ) ≥ Tp (A1 , A2 , A3 ) ≥ 0.
(1)
(2)
(1)
(2)
(1)
(2)
For the second part, we write B1 = B1 − B1 where B1 0, B1 0 and |B1 | = B1 + B1 . We define
(1)
(2)
(1)
(2)
B2 , B2 , B3 , B3 similarly. Note that
(1)
(1)
(1)
(1)
(1)
(2)
(1)
(2)
(1)
(1)
(2)
(2)
T (B1 , B2 , B3 ) =T (B1 , B2 , B3 ) − T (B1 , B2 , B3 ) − T (B1 , B2 , B3 ) + T (B1 , B2 , B3 )
(2)
(1)
(1)
(2)
(1)
(2)
(2)
(2)
(1)
(2)
(2)
(2)
− T (B1 , B2 , B3 ) + T (B1 , B2 , B3 ) + T (B1 , B2 , B3 ) − T (B1 , B2 , B3 ).
(i)
(i)
(j)
(k)
Since Bj 0, the first part of this lemma shows that every term T (B1 , B2 , B3 ) ≥ 0. Hence, we have that
(1)
(1)
(1)
(1)
(1)
(2)
(1)
(2)
(1)
(1)
(2)
(2)
T (B1 , B2 , B3 ) ≤T (B1 , B2 , B3 ) + T (B1 , B2 , B3 ) + T (B1 , B2 , B3 ) + T (B1 , B2 , B3 )
(2)
(1)
(1)
(2)
(1)
(2)
(2)
(2)
(1)
(2)
(2)
(2)
+ T (B1 , B2 , B3 ) + T (B1 , B2 , B3 ) + T (B1 , B2 , B3 ) + T (B1 , B2 , B3 )
=T (|B1 | , |B2 | , |B3 |).
Lemma 41. Suppose that ψk ≤ αk β for all k ≤ n for some 0 ≤ β ≤ 21 and α ≥ 1. Given an isotropic logconcave
distribution p and an unit vector v, we define ∆ = Ex∼p xxT xT v. Then, we have that
1. For any orthogonal projection matrix P with rank r, we have that
2
Tr(∆P ∆) ≤ O(ψmin(2r,n)
).
2. For any symmetric matrix A, we have that
2β
Tr(∆A∆) ≤ O(α2 log n) Tr |A|1/(2β)
.
Proof. We first bound Tr(∆P ∆). This part of the proof is generalized from a proof by Eldan [14]. Note that
Tr(∆P ∆) = Ex∼p xT P ∆xxT v. Since ExT v = 0, we have that
q
q
Lem 19
q
p
2
T
2
T
Tr (∆P ∆).
Tr(∆P ∆) ≤ E(x v) Var (x P ∆x) ≤ O ψrank(P ∆+∆P )
Ex kP ∆xk = O ψrank(P ∆+∆P )
17
2
This gives Tr(∆P ∆) ≤ O(ψmin(2r,n)
).
Now we bound
P Tr(∆A∆). Since Tr(∆A∆) ≤ Tr(∆ |A| ∆), we can assume without loss of generality that A 0.
We write A = i Ai + B where each Ai has eigenvalues between (kAk2 2i /n, kAk2 2i+1 /n] and B has eigenvalues
smaller than or equals to kAk2 /n. Clearly, we only need at most ⌈log(n) + 1⌉ many such Ai . Let Pi be the
orthogonal projection from Rn to the span of the range of Ai . Using kAi k2 Pi Ai , we have that
X
2
2
rank(Ai )2β kAi k2
kA
k
≤
O(α
)
Tr(∆Ai ∆) ≤ kAi k2 Tr(∆Pi ∆) ≤ O ψmin(2rank(A
i
),n)
2
i
i
where we used the first part of this lemma in the last inequality.
Similarly, we have that
Tr(∆B∆) ≤ O ψn2 kBk2 ≤ O(n kBk2 ) ≤ O(1) kAk2 .
Combining the bounds on Tr(∆Ai ∆) and Tr(∆B∆), we have that
X
rank(Ai )2β kAi k2 + O(1) kAk2
Tr(∆A∆) ≤ O(α2 )
i
≤ O(α2 )
X
i
rank(Ai ) kAi k1/(2β)
2
!2β
log(n)1−2β
2β
1/(2β)
≤ O(α2 log n) Tr |A|
.
In the next lemma, we collect tensor related inequalities that will be useful.
Lemma 42. Suppose that ψk ≤ αk β for all k ≤ n for some 0 ≤ β ≤
distribution p in Rn and symmetric matrices A and B , we have that
1
2
and α ≥ 1. For any isotropic logconcave
1. Ex (xT Ax)2 ≤ O(1)(Tr |A|)2 ,
2
2. Ex∼p xT Ax − TrA ≤ O(ψn2 )TrA2 ,
3. T (A, I, I) ≤ O(ψn ) kAk2 n,
4. T (A, I, I) ≤ O(ψn2 )Tr |A|,
5. T (A, B, I) ≤ O(ψr2 ) kBk2 Tr |A| where r = min(2rank(B), n),
2β
6. T (A, B, I) ≤ O(α2 log n) Tr |B|1/(2β)
Tr |A|,
1/s
s
7. T (A, B, I) ≤ (T (|A| , I, I))
1/t
t
for any s, t ≥ 1 with s−1 + t−1 = 1.
T (|B| , I, I)
Proof. Without loss of generality, we can assume A is diagonal by rotating space. In particular, if we want to prove
something for Tr(Aα ∆Aβ ∆) where A, ∆ are symmetric matrices, we use the spectral decomposition A = U ΣU T to
rewrite this as
Tr U Σα U T ∆U Σβ U T ∆ = Tr Σα (U T ∆U )Σβ (U T ∆U )
which puts us back in the same situation, but with a diagonal matrix A.
Let ∆i = Ex∼p xxT xi . For inequality 1, we note that
Ex (xT Ax)2 =
X
ij
Aii Ajj Ex x2i x2j ≤
X
ij
q
|Aii | |Ajj | Ex4i Ex4j
Lem
≤
14
O(1)
X
i
|Aii |
For inequality 2, we note that Ex∼p xT Ax = TrA. Hence,
2
Lem 19
Ex∼p xT Ax − TrA = Var xT Ax
≤ O(ψn2 )Ex kAxk2 = O(ψn2 )TrA2 .
18
!2
.
For remaining inequalities, it suffices to upper bound T by upper bounding Tp for any isotropic logconcave
distribution p.
For inequality 3, we note that
Tp (A, I, I)
X
39 X
Lem 39
Tr(∆2i ) = kAk2 T (I, I, I) ≤ O(nψn ) kAk2
=
Aii Tr(∆2i ) ≤ kAk2
Lem
i
i
where the last inequality is from the third moment assumption.
For inequality 4, we note that
Lem 41 X
39 X
Aii Tr(∆2i ) ≤
|Aii | O(ψn2 ) = O(ψn2 )Tr |A| .
=
Lem
Tp (A, I, I)
i
i
For inequality 5, we let P be the orthogonal projection from Rn to the span of the range of B. Let r = rank(P ).
Then, we have that
(Lem 40)
Tp (A, B, I) ≤ Tp (|A|, |B|, I)
X
|Aii | Tr(∆i |B|∆i )
=
(Lem 39)
i
X
1
|Aii | Tr(∆i P ∆i )
≤ kBk2
i
≤ O(ψr2 )Tr|A| kBk2 .
(Lem 41)
where we used that |B| kBk2 P in 1 .
For inequality 6, we note that
2β
Lem 41
39 X
1/(2β)
Aii Tr(∆i B∆i ) ≤ O(α2 log n)Tr |A| Tr |B|
=
.
Lem
Tp (A, B, I)
i
For inequality 7, we note that
(Lem 40)
Tp (A, B, I) ≤ Tp (|A| , |B| , I)
X
=
Tr(|A| ∆i |B| ∆i )
(Lem 39)
i
≤
=
X
i
X
i
Tr(|A| |∆i | |B| |∆i |)
Tr(|∆i |
≤
X
≤
X
=
X
i
i
|A| |∆i |
|∆i |
1/s
1/s
X
i
s
Tr |A|
s
s
1/t
|B| |∆i |
s 1/s
1/t
Tr
!1/s
1/s
)
|∆i |
1/t
|B| |∆i |
1/t
1/t
t
Tr |∆i | |B| |∆i |
1/s t 2 1/t
Tr |B| ∆i
∆2i
= (Tp (|A| , I, I))
1/s
1/s
s
Tr |A| ∆2i
|∆i |
|A| |∆i |
(Tr (|∆i | |A| |∆i |))
i
≤
Tr
1/s
X
i
t
Tr |B| ∆2i
t 1/t
(Lem 22)
!1/t
1/t
t
.
Tp (|B| , I, I)
Lemma 43. For any positive semi-definite matrices A, B, C and any α ∈ [0, 1], then
T (B 1/2 Aα B 1/2 , B 1/2 A1−α B 1/2 , C) ≤ T (B 1/2 AB 1/2 , B, C).
19
(Lem 21)
(Lem 39)
Proof. Fix any isotropic logconcave distribution p. Let ∆i = Ex∼p B 1/2 xxT B 1/2 xT C 1/2 ei . Then, we have that
Tp (B 1/2 Aα B 1/2 , B 1/2 A1−α B 1/2 , C) = Ex,y∼p xT B 1/2 Aα B 1/2 yxT B 1/2 A1−α B 1/2 yxT Cy
X
E y T B 1/2 Aα B 1/2 x xT B 1/2 A1−α B 1/2 y xT C 1/2 ei y T C 1/2 ei
=
i
=
X
i
=
X
E Tr Aα B 1/2 xxT B 1/2 A1−α B 1/2 yy T B 1/2 xT C 1/2 ei y T C 1/2 ei
Tr(Aα ∆i A1−α ∆i )
i
Using Lemma 23, we have that
X
X
Tr(A∆2i ) = Ex,y∼p xT B 1/2 AB 1/2 yxT ByxT Cy = Tp (B 1/2 AB 1/2 , B, C).
Tr(Aα ∆i A1−α ∆i ) ≤
i
i
Taking the supremum over all isotropic logconcave distributions p, we get the result.
6.2
Derivatives of the potential
Lemma 44. Let At be defined by Definition 24. For any integer q ≥ 2, we have that
dTr((At − I)q ) =qEx∼pt (x − µt )T (At − I)q−1 (x − µt )(x − µt )T dWt − qTr((At − I)q−1 A2t )dt
q X
+
Ex,y∼pt (x − µt )T (At − I)α (y − µt )(x − µt )T (At − I)β (y − µt )(x − µt )T (y − µt )dt.
2
α+β=q−2
We give the proof in Section 5. The next lemma bounds the stochastic process that controls this potential
function.
Lemma 45. Let At and pt be defined by Definition 24. Let Φt = Tr((At − I)q ) for some even integer q ≥ 2, then
we have that dΦt = δt dt + vtT dWt with
δt ≤
1+ 1
1− 1 1
1
q(q − 1)T (At (At − I)q−2 , At , At ) + q(Φt q + Φt q n q )
2
and
kvt k2 ≤ q Ex∼p (x − µt )T (A − I)q−1 (x − µt )(x − µt )T
2
.
Proof. By Lemma 44, we have
dΦt =qEx∼pt (x − µt )T (At − I)q−1 (x − µt )(x − µt )T dWt − qTr((At − I)q−1 A2t )dt
q X
+
Ex,y∼pt (x − µt )T (At − I)α (y − µt )(x − µt )T (At − I)β (y − µt )(x − µt )T (y − µt )dt
2
α+β=q−2
=qEx∼p (x − µt )T (A − I)q−1 (x − µt )(x − µt )T dWt − qTr((At − I)q−1 A2t )dt
q X
Ex,y∼p̃t xT At (At − I)α yxT At (At − I)β yxT At ydt
+
2
α+β=q−2
def
=δt dt + vtT dWt .
1/2
where p̃t is the isotropic version of pt defined by p̃t (x) = p(At x + µt ), δt dt is the drift term in dΦt and vtT dWt is
the martingale term in dΦt .
For the drift term αt dt, we have
q X
T (At (At − I)α , At (At − I)β , At ) − qTr((At − I)q−1 A2t ).
δt ≤
2
α+β=q−2
The first term in the drift is
q X
q
T (At (At − I)α , At (At − I)β , At ) ≤
2
2
α+β=q−2
X
α+β=q−2
α
β
T (At |At − I| , At |At − I| , At )
≤
q
2
=
q(q − 1)
T (At (At − I)q−2 , At , At ).
2
X
α+β=q−2
20
T (At |At − I|
q−2
, At , At )
(Lem 40)
(Lem 43)
For the second term in drift, since q is even, we have that
−qTr((At − I)q−1 A2t ) = − qTr((At − I)q−1 (At − I + I)2 )
≤ − qTr((At − I)q+1 ) − qTr((At − I)q−1 )
1+ 1q
≤qΦt
1− 1q
+ qΦt
1
nq .
For the martingale term vtT dWt , we note that
kvt k2 = q Ex∼p (x − µt )T (A − I)q−1 (x − µt )(x − µt )T .
6.3
Analysis of At
We first bound the drift term from Lemma 44.
Lemma 46. Suppose that ψk ≤ αk β for all k ≤ n for some 0 ≤ β ≤ 21 and α ≥ 1. Let Φ = Tr((A − I)q ) for some
even integer q ≥ 2 and A 0. If βq ≥ 2, then
2β− 2β + 2 1− 2
2β 2
2β− β
+ q1 1− 1q
2β− 2βq+1
2β− q1 1+ q1
q−2
2
q
q
q
q
.
Φ
+n
Φ
+n
Φ + (log n)n
Φ
n
T (A(A − I) , A, A) ≤ O α
Proof. We have that
T (A(A − I)q−2 , A, A)
=T ((A − I)q−1 + (A − I)q−2 , (A − I) + I, (A − I) + I)
=T ((A − I)q−1 , A − I, A − I) + 2T ((A − I)q−1 , A − I, I) + T ((A − I)q−1 , I, I)
+ T ((A − I)q−2 , A − I, A − I) + 2T ((A − I)q−2 , A − I, I) + T ((A − I)q−2 , I, I)
≤T (|A − I|q−1 , |A − I|, |A − I|) + 2T (|A − I|q−1 , |A − I|, I) + T (|A − I|q−1 , I, I)
q−2
+ T ((A − I)
≤T (|A − I|
q−1
q−2
, |A − I|, |A − I|) + 2T ((A − I)
q−1
, |A − I|, |A − I|) + 3T (|A − I|
+ 3T (|A − I|
q−1
q−2
, I, I) + T ((A − I)
, I, I)
q−2
, |A − I|, I) + T ((A − I)
(Lem 40)
, I, I)
(6.1)
, |A − I|, I)
where we used Lemma 43 at the end.
For the first term in (6.1), we have that
T (|A − I|q−1 , |A − I|, |A − I|) ≤ T (|A − I|q , |A − I|, I)
2β
≤ O(α2 log n)Φ Tr|A − I|1/2β
2β
= O(α2 log n)Φ (Tr|A − I|q )1/(2βq) n1−1/(2βq)
(Lem 43)
(Lem 42.6)
1
≤ O(α2 log n)n2β− q Φ1+1/q
where we used 2βq ≥ 1 at the last line.
For the second term in (6.1), we write
|A − I| = B1 + B2
where B1 consists of the eigen-components of |A − I| with eigenvalues ≤ η and B2 is the remaining part where we
will pick η ≥ 0 later. Then, we have that
T (|A − I|
q−1
, |A − I|, I) = T (B1q−1 , B1 , I) + T (B1q−1 , B2 , I) + T (B2q−1 , B1 , I) + T (B2q−1 , B2 , I).
(6.2)
For the first term in (6.2), we note that
T (B1q−1 , B1 , I) ≤ T (B1q , I, I)
q
≤ O(ψn n) kB1 k
≤ O(η q ψn n).
21
(Lem 43)
(Lem 42.3)
For the second term in (6.2), we note that
T (B1q−1 , B2 , I) ≤ T (B1q , I, I)
≤ O(η q ψn n)
where we used kB1 k2 ≤ η and
have
TrB2q
q−1
q
q−1
q
1
T (B2q , I, I) q
O(ψn2 Φ)
(Lem 42.7)
1
q
(Lem 42.3 and Lem 42.4)
q
≤ Tr |A − I| = Φ at the last line. For the third term in (6.2), similarly, we
T (B2q−1 , B1 , I) ≤ T (B2q , I, I)
≤
O(ψn2 Φ)
q−1
q
q−1
q
1
T (B1q , I, I) q
(Lem 42.7)
1
q
q
(Lem 42.3 and Lem 42.4)
O(η ψn n) .
For the fourth term in (6.2), we let P be the orthogonal projection from Rn to the range of B2 . We have that
T (B2q−1 , B2 , I) = T (P B2q−1 P, P B2 P, I)
≤ T (P B2q P, P, I)
≤
≤
where r ≤ 2rank(P ) ≤
T (|A − I|
q−1
2Φ
ηq .
(Lem 43)
O(ψr2 )TrB2q
O(ψr2 )Φ
(Lem 42.5)
Using ψk ≤ αk β and combining all four terms, we have that
q−1
q−1
1
1
, |A − I|, I) ≤ O(η q ψn n) + O(η q ψn n) q O(ψn2 Φ) q + O(ψn2 Φ) q O(η q ψn n) q + O(ψr2 )Φ
1
1
Φ1+2β
+ β(q+1)
+β(2− 1q ) q−1
2
q 1+β
q−1 q−1
q
q
q
q
q
Φ + ηn
Φ
+ 2βq .
≤ O(α ) η n
+η
n
η
1
β
1
Balancing the last two terms and setting η = Φ q n− q + q(2βq+1) , we get that
2β 2 (q−1)
2β 2
2β 2
2β 2 q
q−1
T (|A − I|
, |A − I|, I) ≤ O(α2 Φ) n2β− 1+2βq + n2β− 1+2βq + n2β− 1+2βq + n2β− 1+2βq
β
≤ O(α2 n2β (1− 2βq+1 ) Φ)
where we used q ≥ 2 and β ≥ 0.
For the third term in (6.1), we have that
T (|A − I|
q−1
, I, I) = T (B1q−1 , I, I) + T (B2q−1 , I, I)
Φ
≤ O αη q−1 n1+β + α2 n2β
η
≤ O(α2 n
β−1
1−β
q +2β
(Lem 42.3 and Lem 42.4)
1
Φ1− q )
1
where we set η = n q Φ q at the last line.
For the fourth term in (6.1), we have that
T (|A − I|q−2 , I, I) = T (B1q−2 , I, I) + T (B2q−2 , I, I)
Φ
≤ O(αη q−2 n1+β + α2 n2β 2 )
η
≤ O(α2 n
β−1
2(1−β)
+2β
q
(Lem 42.3 and Lem 42.4)
2
Φ1− q )
1
where we set η = n q Φ q at the last line.
Combining all terms, we have the result.
Next we bound the martingale term.
Lemma 47. Let p be a logconcave distribution with covariance matrix A. Let Φ = Tr((A − I)q ) for some even
integer q ≥ 2. Then,
Ex∼p (x − µt )T (A − I)q−1 (x − µt )(x − µt )T
22
1
2
1
1
1
≤ O(Φ1− 2q n q + Φ1+ 2q + n q ).
Proof. Note that
Ex∼p (x − µt )T (A − I)q−1 (x − µt )(x − µt )T
1/2
≤ O (1) kAk1/2
(A − I)q−1 A1/2
2 Tr A
2
1/2
≤ O (1) kAk2
1/2
Tr|A − I|q−1 + O (1) kAk2
1
1
(Lem 32)
1
1
≤ O(1 + Φ 2q )Φ1− q n q + O(1 + Φ 2q )Φ
1
1
Tr|A − I|q
1
1
≤ O(Φ1− 2q n q + Φ1+ 2q + n q ).
Using Lemma 46 and Lemma 47, we know that Φt = Tr((At − I)q ) satisfies the stochastic equation dΦt =
δt dt + vtT dWt with
β
2β 2
2β
2
1
1
1− 2
1− 1
1+ 1
δt ≤O α2 q 2 n2β− q + q Φt q + n2β− q + q Φt q + n2β− 2βq+1 Φt + (log n)n2β− q Φt q
1+ q1
+ q(Φt
≤O α2 q 2
and
1− 1q
+ Φt
n2β−
2β
q
1
nq )
+ q2
1− q2
Φt
2
kvt k2
β
1
1− 1q
+ n2β− q + q Φt
2
≤ O(q )
2− 1 2
Φt q n q
2β 2
1
1+ q1
+ n2β− 2βq+1 Φt + (log n)n2β− q Φt
+
2+ 1
Φt q
+n
2
q
(6.3)
(6.4)
where we used α ≥ 1, 2βq ≥ 1 and q ≥ 1 in (6.3).
Using these, one can bound the growth of Φt using a stochastic Grönwall’s inequality. For completeness, we
bound Φt directly below.
Lemma 48. Suppose that ψk ≤ αk β for all k ≤ n for some 0 ≤ β ≤ 12 and α ≥ 1. Given an isotropic logconcave
distribution p. Let At be defined by Definition 24 using initial distribution p. Let Φt = Tr((At − I)q ) for some even
integer q ≥ 2. If βq ≥ 2 and n ≥ q ηq for some large constant η, then there is a universal constant c such that
β
β
q
P( max Φt ≥ n1− 2 log− 2 n) ≤ 0.01
t∈[0,T ]
with
T =
cn−2β+ 2q
qα2 log n
Proof. The idea is to choose a a function Ψt = f (Φt , t) so that the resulting stochastic equation for Ψt effectively
decouples the drift and martingale terms. We use
2q1
q
f (a, t) = a + 1 + Etq + F t 2
with
q
E = q q α2q n1−β+2βq and F = q 2 αq n1−β+βq .
By 45 and Itô’s formula, we have that
df
1 d2 f
df
dt +
dΦt +
d[Φ]t
dt
dΦ
2 dΦdΦ
df
df
df T
≤ dt +
δt dt +
v dWt
dt
dΦ
dΦ t
where we used f (a, t) is concave in a in the last line and dropped the second derivative term. The rationale for our
q
choice of f (a, t) is that 1 + Etq + F t 2 is our guess for the solution of the SDE for Φt , and the power 1/2q is chosen
so that Φt can be eliminated from the stochastic term in the bound for Ψt above.
df
δt dt, we use (6.3) and get that
For the term dΦ
dΨt =
df
δt
1
δt =
dΦ
2q (Φt + 1 + Etq + F t q2 )1− 2q1
2β
β
2
1
1− 2
1− 1
n2β− q + q Φt q
n2β− q + q Φt q
2
≤O(qα )
1 +
1
q
q
(Φt + 1 + Etq + F t 2 )1− 2q
(Φt + 1 + Etq + F t 2 )1− 2q
1
2β 2
2β− q1 1+ q
2β− 2βq+1
Φt
Φt
(log n)n
n
.
+
+ O(qα2 )
1
q 1− 1
q
q
q
2q
(Φt + 1 + Et + F t 2 )
(Φt + 1 + Et + F t 2 )1− 2q
23
(6.5)
For the first term in (6.5), we note that
n2β−
2β
2
q +q
1− 2q
Φt
1
q
(Φt + 1 + Etq + F t 2 )1− 2q
n2β−
≤
2β
2
q +q
2
1
q
(Φt + 1 + Etq + F t 2 )1− 2q −(1− q )
≤
n2β−
2β
2
q +q
q
1
=
3
(F t 2 ) 2q
F 2q
3
qα2 t 4
q
where we used F = q 2 αq n1−β+βq at the end. For the second term in (6.5), we note that
β
1− q1
1
β
n2β− q + q Φt
1
q
(Φt + 1 + Etq + F t 2 )1− 2q
≤
1
1
n2β− q + q
1
(Etq ) 2q
=
E 2q
√
qα2 t
β
q
where we used E = q q α2q n1−β+2βq at the end. For the third term in (6.5), assuming Φt ≤ n1− 2 log− 2 n, we have
that
2β 2
β
β
2β 2
2β 2
1
1
1
1/q
n2β− 2βq+1 Φt
n2β− 2βq+1 Φt
n2β− 2βq+1 n q − 2q 1 n2β− q + q 2 E 2q
√
≤
≤
=
1
1
1
1 ≤
q
qα2 t
(Φt + 1 + Etq + F t 2 )1− 2q
(Etq ) 2q
(Etq ) 2q
(Etq ) 2q
where we used that 2qβ ≥ 1 implies 2β −
the fourth term in (6.5), assuming Φt ≤ n
1
2β 2
1
2βq+1 + q
q
β
−2
1− 2
log
1+ 1q
β
2q
≤ 2β −
1
(Φt + 1 + Etq + F t 2 )1− 2q
≤
β
q
+
1
q
at 1 and E = q q α2q n1−β+2βq at 2 . For
n, we have that
1
(log n)n2β− q Φt
q
−
2
(log n)n2β− q Φtq
1
(Etq ) 2q
β
≤
1
n2β− q + q
1
(Etq ) 2q
1
E 2q
√
=
qα2 t
where we used E = q q α2q n1−β+2βq at the end. Combining all four terms in (6.5), we have that
1 1
1
3
df
δt ≤ O E 2q t− 2 + F 2q t− 4 .
dΦ
For the term
df
dt dt,
we have that
q
1 1
qEtq−1 + 2q F t 2 −1
1
3
df
1
= O E 2q t− 2 + F 2q t− 4 .
=
q 1− 1
dt
2q (Φt + 1 + Etq + F t 2 ) 2q
For the term
df T
dΦ vt dWt ,
q
β
using (6.4) and assuming Φt ≤ n1− 2 log− 2 n, we have that
2
2− 1 2
2
kvt k2 ≤ O(q 2 ) Φt q n q + n q .
Hence, we have that
df
vt
dΦ
Combining the terms
df
df
dΦ δt dt, dt dt
2
2
2
≤
2− 1 2
Φt q n q
2
q
O(q )
+n
2
1
≤ O(n q ).
q 2− 1
2
4q (Φt + 1 + Etq + F t 2 ) q
df T
dΦ vt dWt ,
and
β
q
we have that, when Φt ≤ n1− 2 log− 2 n,
df
df
1 d2 f
dt +
dΦt +
d[Φ]t
dt dΦ
2 dΦdΦ
1
1
1
3
≤ C1 E 2q t− 2 + F 2q t− 4 dt + dYt
dΨt =
(6.6)
]t
2/q
where Yt is a martingale with Y0 = 0 and d[Y
for some universal constant C1 , C2 ≥ 1.
dt ≤ C2 n
By Theorem 12, there exists a Wiener process W̃t such that Yt has the same distribution as W̃[Y ]t . Using the
reflection principle for 1-dimensional Brownian motion, we have that
P( max Yt ≥ γ) ≤ P(
t∈[0,T ]
1
β
max
t∈[0,C2 n2/q T ]
W̃t ≥ γ) = 2P(W̃C2 n2/q T ≥ γ) ≤ O(exp(−
1
γ2
)).
2C2 n2/q T
Let Ψu = n 2q − 4q log− 4 n. As long as Ψt ≤ Ψu , the estimate (6.6) is valid and hence
Z T
1 1
1
3
P( max Ψt ≥ Ψu ) ≤ P( max Yt ≥ Ψu − 1 −
C1 E 2q t− 2 + F 2q t− 4 dt)
t∈[0,T ]
t∈[0,T ]
0
24
where we used that Ψ0 = 1 at the last line. Note that
Z T
1
1 1
√
1−β
1−β
β
1
1 √
1
1
1
3
C1 E 2q t− 2 + F 2q t− 4 dt = 2C1 E 2q T + 4C1 F 2q T 1/4 ≤ 4C1 q 2 αn 2q +β T + q 4 α 2 n 2q + 2 T 1/4 .
0
Setting T =
−2β+
β
2q
n
216 C14 qα2 log n
and using n ≥ q ηq for some large constant η, we have that
Z
T
0
Hence, we have that
1 1
1
3
Ψu
− 1.
C1 E 2q t− 2 + F 2q t− 4 dt <
2
Ψu
P( max Ψt ≥ Ψ ) ≤ P( max Yt ≥
) = O(exp(−
2
t∈[0,T ]
t∈[0,T ]
u
Ψu
2
2
1
)).
2C2 n2/q T
Note that
Ψu
2
2
1
β
1
1
1+β p
n q − 2q log− 2 n 216 C14 qα2 log n
1
=
log n ≥ C2−1 nβ
= 213 C2−1 C14 qα2 n2β− q
β
2/q
2/q
2C2 n T
8n C2
n−2β+ 2q
1
where we used qβ ≥ 2, 0 ≤ β ≤ 21 , C1 ≥ 1, α ≥ 1, q ≥ 1 at 1 . Using the fact that Ψt ≥ Φt2q , we have that
q
β
P( max Φt ≥ n1− 2 log− 2 n) = O(exp(−C2−1 nβ )) ≤ 0.01
t∈[0,T ]
where we used that qβ ≥ 2 and n ≥ q ηq for some large constant η.
Lemma 49. Suppose that ψk ≤ αk β for all k ≤ n for some 0 ≤ β ≤ 21 and α ≥ 1. For any even integer q ≥ 2 with
βq ≥ 2, if n ≥ q ηq for some large constant η, we have that
p
β
ψn ≤ Cα q log nnβ− 4q
for some universal constant C.
Proof. By Lemma 48, for t up to T =
−2β+
β
2q
cn
qα2 log n
, with probability 0.99,
q
β
Tr((At − I)q ) = Φt ≤ n1− 2 log− 2 n
for all 0 ≤ t ≤ T.
Assuming this event, we have
1
β
1
kAt k2 ≤ 1 + n q − 2q log− 2 n
and
Z
which is less than
get that
1
64
1
T
0
for all 0 ≤ t ≤ T.
kAt k2 dt ≤ T · (1 + n
β
1
q − 2q
log
− 21
n) ≤
c1 n q −2β
qα2 log3/2 n
when n is large enough. Also, we have that BT = T · I. Hence, we can apply Lemma 31 and
p
β
ψp = O α q log nnβ− 4q .
Since this argument holds for any isotropic logconcave distribution, this gives the bound for ψn .
Proof of Theorem 37. Fix a large enough n. We start with a known bound:
ψk ≤ α1 k β1 for all k ≤ n
wherelα1m is some universal constant larger than 1 and β1 =
q = 2 β1 . Hence, we have that
− 12 p
ψk ≤ 4Cα1 β1
1
2.
2
β1
Now, we apply Lemma 49 for every k ≤ n with
log nk β1 − 16 for all k ≤ n.
25
Repeating this process, we have that ψk ≤ αℓ k βℓ for all k ≤ n with
− 12 p
αℓ+1 = 4Cαℓ βℓ
βℓ+1 = βℓ −
By induction, we have that αℓ = O(ℓ log n)ℓ/2 and βℓ ≤
βℓ2
16
16
ℓ .
log n,
.
Hence, we have that
ψn ≤ O(ℓ log n)ℓ n
for all ℓ ≥ 1. Setting ℓ =
lp
m
log n/ log log n , we have that
ψn = n
7
16
ℓ
O
q
log log n
log n
p
log n log log n .
= exp O
Adaptive localization for anisotropic distributions
In this section, we show that the same third moment assumption gives the following bound on the KLS constant
for arbitrary logconcave distributions.
Theorem 50. Under the third moment assumption of Section 6, for any logconcave density p in Rn with covariance
matrix A, for any integer q ≥ 1, the KLS constant is bounded as
1
ψp ≤ Cq O(q) (Tr (Aq )) 2q .
7.1
Controlled stochastic localization
Definition 51. Given a symmetric matrix B, let E<u (B) be the span of all eigenvectors in B with eigenvalues less
than u and Λ<u (B) = dim E<u (B). We define E≥u , Λ≥u , etc similarly.
For this reduction, we apply localization only in the subspace where the matrix Bt controlling the Gaussian has
small eigenvalues. At time t, the control matrix is chosen so that it is the inverse of the projection of the current
covariance matrix At to the subspace of the small eigenvalue of Bt . This is captured in the next definition.
Definition 52. Given a logconcave distribution p and a threshold u, we define the following process: p0 = p, c0 =
0, B0 = 0, τ0 = 0 and for k ≥ 1,
1. τk = inf{t > τk−1 such that Λ<u (Bt ) 6= Λ<u (Bτk−1 )}.
2. pt , ct , Bt are defined by Definition 24 on [τk−1 , τk ) with the initial data cτk−1 , Bτk−1 , pτk−1 (instead of 0, 0 and
p) and with the control matrix Ct given by
−2
(k) −1
Ct = ψ2r
)
(k) lim (I + At + sP
s→∞
where At is the covariance matrix of pt , r(k) = Λ<u (Bτk−1 ), P (k) is the orthogonal projection onto E≥u (Bτk−1 )
and ψr is any known bound on the KLS constant for isotropic logconcave densities. Let rt = rank(Ct ).
The following lemma gives an alternative definition of Ct .
Lemma 53. For any A ≻ 0 and an orthogonal projection matrix P , we let T = lims→∞ (A + sP )−1 . Then, we
have that
T = ((I − P )A(I − P ))†
where † denotes pseudoinverse. Furthermore, we have that R(P ) = N (T ), i.e., the rowspace of P equals the nullspace
of T .
26
Proof. By taking P = U T ΣU , we can see that
lim (A + sP )−1 = U T lim (U AU T + sΣ)−1 U
s→∞
and
s→∞
†
((I − P )A(I − P ))† = U T (I − Σ)U AU T (I − Σ) U.
Hence, it suffices to prove the case P is a diagonal matrix whose first r diagonal entries are 0 and the remaining
diagonal entries are 1. Write
A1 A2
A=
AT2 A3
where A1 is a r × r matrix. Then, we have that
−1
(A + sP )
=
A1
AT2
A2
A3 + sI
−1
=
−1
T −1
A−1
1 + A1 A2 F A2 A1
T −1
−F A2 A1
−A−1
1 A2 F
F
.
−1
where F = (A3 + sI − AT2 A−1
. As s → ∞, we have that F → 0 and hence
1 A2 )
−1
A1
0
†
= ((I − P )A(I − P )) .
lim (A + sP )−1 =
s→∞
0
0
For any x ∈ R(P ), we have that x is 0 in the first r coordinates and hence
2
0 ≤ xT T x ≤ xT (I + sP )−1 x =
kxk
.
1+s
Taking s → ∞, we have that xT T x = 0. Using T 0, this shows that R(P ) ⊂ N (T ). On the another hand, we
have that
A + sP kAk2 (I − P ) + sP
and hence
−1
(A + sP )−1 kAk2 (I − P ) + s−1 P.
Taking the limit, we have that T kAk−1
2 (I − P ). This shows that N (T ) ⊂ R(P ). Hence, we have R(P ) =
N (T ).
The specific formula above is not important and the reduction of this section uses only the following properties
of the control matrix Ct .
Lemma 54. For any t ≥ 0, we have that
−2 −1
1. (Focus on small values in At ) 0 Ct ψ2r
At .
t
2. (Focus on small values in Bt ) Bt uI and rt = r(k) = Λ<u (Bt ) for t ∈ [τk−1 , τk ).
1+ q1
3. (Large step size) TrCt ≥ rt
1/q
2
(Φt
/(ψ2r
t
+ 1)).
−2
−1
Proof. For the first part, since I + At + sP (k) At , we have that ψ2r
Ct 0.
(k) At
2
For Bt uI in the second part, we prove it by a continuous induction. Let t′ = inf t≥0 {xT Bt x > u for some kxk =
′
′
1}. Suppose that t < +∞. By the definition, we have that Bt uI for t ≤ t . Fix any t ∈ [τk−1 , τk ). Since
t
Bt uI, we can write Bt = Bt′ + Bt′′ where all eigenvalues of Bt′ are u and Bt′′ ≺ uI. Since dB
dt = Ct 0 and since
′
′
the number of eigenvalues being u is unchanged during [τk−1 , τk ), we know that Bt = Bτk−1 . By the definition of
P (k) , we have that
R(P (k) ) = E≥u (Bτ′ k−1 ) = E≥u (Bt′ ).
For any xT Bt x ≥ u kxk2 , we have that x ∈ E≥u (Bt′ ) because Bt uI. Hence, we have that x ∈ R(P (k) ) and
d
Lemma 53 shows that R(P (k) ) = N (Ct ) and hence xT Ct x = 0. Since dt
Bt = Ct , we have that
xT
dBt
2
x = 0 for any xT Bt x ≥ u kxk and any t ≤ t′ .
dt
27
This contradicts the definition of t′ . Therefore, we have that Bt uI for all t.
For rt = r(k) = Λ<u (Bt ) in the second part, Lemma 53 shows that R(P (k) ) = N (Ct ). Therefore, we have that
n − rt = dim N (Ct ) = rank(P (k) ) = n − r(k) = n − Λ<u (Bτk−1 ) = n − Λ<u (Bt )
where we used that the number of eigenvalues being u is unchanged during [τk−1 , τk ) at the end.
For the third part, we use the inequality
r≤
r
X
xqi
i=1
1
! 1+q
r
X
x−1
i
i=1
q
! 1+q
and the fact that Ct is a rank rt matrix. We have that
1
1+q
q
rt ≤ TrCt†q
(TrCt ) 1+q
(7.1)
where we use Ct†q to denote the q th power of pseudo inverse of Ct . By Lemma 53, we have that
2
(I − P (k) )(I + At )(I − P (k) ).
Ct† = ψ2r
t
Hence, we have that
TrCt†q
Lem
≤
22
2q
2q
Tr(I + At )q .
Tr(I − P (k) )q (I + At )q (I − P (k) )q ≤ ψ2r
ψ2r
t
t
Putting this in (7.1) gives that
1+ q1
TrCt ≥
7.2
rt
TrCt†q
1+ 1q
1q ≥
rt
1/q
2 (Φ
ψ2r
t
t
+ 1)
.
Analysis of At
Lemma 55. Let At be defined by Definition 52. For any integer q ≥ 2, we have that
1/2
Ct )dt
dTr(Aqt ) =qEx∼pt (x − µt )T Atq−1 (x − µt )(x − µt )T Ct dWt − qTr(Aq+1
t
q X
T α
T β
+
Ex,y∼pt (x − µt ) At (y − µt )(x − µt ) At (y − µt )(x − µt )T Ct (y − µt )dt.
2
α+β=q−2
Proof. Note that At is defined by concatenating solutions of finitely many SDEs. Therefore, it suffices to prove this
equality for each SDE solution and this follows from Lemma 36.
In Section 6 that At − I may not be positive semi-definite and hence we need to take q to be even integer. But
in this section, we analyze the process by the potential Tr(Aqt ) and hence we do not require that q to be even.
Lemma 56. Let At be defined by Definition 52. Let Φt = Tr(Aqt ) for some integer q ≥ 2, then we have that
dΦt = δt dt + vtT dWt with
δt ≤ O(q 2 )Φt and kvt k2 ≤ O(q)Φt .
Proof. By Lemma 55, we have
1/2
Ct )dt
dΦt =qEx∼pt (x − µt )T Atq−1 (x − µt )(x − µt )T Ct dWt − qTr(Aq+1
t
q X
T α
T β
+
Ex,y∼pt (x − µt ) At (y − µt )(x − µt ) At (y − µt )(x − µt )T Ct (y − µt )dt
2
α+β=q−2
1/2
=qEx∼p (x − µt )T Atq−1 (x − µt )(x − µt )T Ct dWt − qTr(Aq+1
Ct )dt
t
q X
1/2
T α+1
T β+1
T 1/2
Ex,y∼p̃t x At yx At yx At Ct At ydt
+
2
α+β=q−2
=δt dt + vtT dWt .
def
28
1/2
where p̃t is the isotropic version of pt defined by p̃t (x) = p(At x + µt ), δt dt is the drift term in dΦt and vtT dWt is
the martingale term in dΦt .
For the drift term δt dt, using that Ct 0, we have
q X
1/2
1/2
T (Aα+1
, Aβ+1
, At Ct At ) − qTr(Aq+1
Ct )
δt ≤
t
t
t
2
α+β=q−2
q X
1/2
1/2
≤
T (Aα+1
, Aβ+1
, At Ct At )
t
t
2
α+β=q−2
q X
1/2
1/2
≤
T (Aqt , I, At Ct At )
(Lem 43)
2
α+β=q−2
1/2
1/2
≤ q T (Aqt , At Ct At , I)
2
1/2
1/2
2
)Φt At Ct At
≤ O(q 2 ψ2r
t
(Lem 42.5)
2
≤ O(q 2 Φt ).
(Lem 54)
For the martingale term vtT dWt , we note that
1/2
kvt k2 = q Ex∼pt (x − µt )T Atq−1 (x − µt )(x − µt )T Ct
1/2
1/2
≤ O (q) At Ct At
1/2
2
1/2
1/2
Tr At Atq−1 At
(Lem 32)
−2
≤ O(qψ2r
)Φt ≤ O(q)Φt .
t
where we used that ψ2rt = Ω(1) at the last line.
Using these, one can bound the growth of Φt .
Lemma 57. Let At be defined by Definition 52 using initial distribution p. Let Φt = Tr(Aqt ) for some integer q ≥ 2.
Suppose that Φ0 = n, there is a universal constant c such that
P( max
Φt ≥ 2n) ≤ 0.01
t∈[0,Tmax ]
with
Tmax =
64(C ′
1
.
+ C)q 2
Proof. By Lemma 56 and Itô’s formula, we have that
1 −2
d log Φt = Φ−1
t dΦt − Φt d[Φ]t
2
δt dt + vtT dWt
≤ Φ−1
t
T
≤ O(q 2 )dt + Φ−1
t vt dWt
≤ Cq 2 dt + dYt
(7.2)
T
where dYt = Φ−1
t vt dWt , Yt = 0 and C is some universal constant.
Note that
d[Y ]t
2
2
′ 2
= Φ−2
t kvt k2 = O(q ) ≤ C q
dt
for some universal constant C ′ . By Theorem 12, there exists a Wiener process W̃t such that Yt has the same
distribution as W̃[Y ]t . Using the reflection principle for 1-dimensional Brownian motion, we have that
P( max Yt ≥ γ) ≤ P(
t∈[0,T ]
max
′ 2
t∈[0,C q T ]
W̃t ≥ γ) = 2P(W̃C ′ q2 T ≥ γ) ≤ 2 exp(−
Since Φ0 = n, we have that log Φ0 = log n and therefore (7.2) shows that
P( max log Φt ≥ log n + Cq 2 T + γ) ≤ 2 exp(−
t∈[0,T ]
Putting T =
1
64(C ′ +C)q2
γ2
).
2C ′ q 2 T
and γ = 12 , we have that
2
P( max log Φt ≥ log n + ) ≤ 2 exp(−8)) ≤ 0.01.
3
t∈[0,T ]
29
γ2
).
2C ′ q 2 T
7.3
Proof of Theorem 50
Lemma 58. Let At be defined by Definition 52 using initial distribution p. Suppose that Tr(Aqt ) = n. Then, we
have that
1
and u = q −c2 q n−1/q
P(BTmax = uI) ≥ 0.99 with Tmax =
c1 q 2
where c1 and c2 are universal constants.
Proof. By Lemma 57 and Lemma 54, we have that Φt ≤ 2n for all 0 ≤ t ≤ Tmax with 0.99 probability, subject to
1+ q1
−2
r
this event, we have that TrCt ≥ cψ2r
t t
n−1/q . Let Ψt = TrBt . Then this shows that
1
dΨt
−2 1+ q −1/q
≥ cψ2r
r
n
≥ 0.
t t
dt
By Theorem ??, we have that ψn−2 ≥ e−C
√
log n log log n
for some universal constant C ≥ 0. Hence, we have that
√
1+ 1
ce−C log rt log log rt rt q n−1/q
dΨt
≥
≥ 0.
dt
Also, Lemma 54 shows that rt = Λ<u (Bt ). Therefore, we have un ≥ Ψt ≥ u(n − rt ).
Let Tk = inf t≥0 {Ψt ≥ un(1 − 2−k )}. For any t ∈ [Tk−1 , Tk ), we have that
Since e−C
√
1
log r log log r 1+ q
r
rt ≥ n −
Ψt
≥ n2−k .
u
is an increasing function in r, for t ∈ [Tk−1 , Tk ), we have that
√
1
dΨt
−k
−k
≥ ce−C log(n2 ) log log(n2 ) (n2−k )1+ q n−1/q
dt
√
−k
−k
= n2−k · c2−k/q e−C log(n2 ) log log(n2 ) .
From the definition of Tk , we see that
Tk − Tk−1 ≤
un2−k
√
−C
n2−k · c2−k/q
√e
log(n2−k ) log log(n2−k )
log(n2−k ) log log(n2−k )
= O(u)2k/q eC
.
Therefore, we have that
log2 n
Tlog2 n ≤ O(u)
X
2k/q eC
√
log(n2−k ) log log(n2−k )
= uq O(q) n1/q
k=0
where the last inequality can be seen by noting that the sequence is exponentially increasing with rate 1/q until
2
the maximal at 2k = n/q Θ(q ) .
Setting u = qO(q)1n1/q , we have that Tlog2 n ≤ Tmax and hence Ψt increases to un and Bt = uI before time Tmax .
After Bt has increased to uI, Ct = 0 and the localization process freezes.
Proof of Theorem 50. The case q = 1 is proven in Theorem 2. So, we assume q ≥ 2. By rescaling, we can assume
TrAq = n. To apply Lemma 31, we note that by Lemma 58, we have that
1
and u = q −c2 q n−1/q .
P(BTmax = uI) ≥ 0.99 with Tmax =
c1 q 2
Furthermore, Lemma 54 shows that
1/2
Ct
1/2
At Ct
1/2
2
1/2
= At Ct At
2
−2
= O(1).
≤ ψ2r
t
Therefore,
P
Z
Tmax
0
Cs1/2 As Cs1/2
1
and BTmax uI
ds ≤
64
2
!
≥ 0.99.
Hence, Lemma 31 shows that
1/2q
.
ψp = Ω u−1/2 = q O(q) n1/2q = q O(q) (TrAq )
30
| 8cs.DS
|
Image Registration Techniques: A Survey
1
Sayan Nag*
Department of Electrical Engineering
Jadavpur University
1
Abstract— Image Registration is the process of aligning two or
more images of the same scene with reference to a particular
image. The images are captured from various sensors at different
times and at multiple view-points. Thus to get a better picture of
any change of a scene/object over a considerable period of time
image registration is important. Image registration finds
application in medical sciences, remote sensing and in computer
vision. This paper presents a detailed review of several
approaches which are classified accordingly along with their
contributions and drawbacks. The main steps of an image
registration procedure are also discussed. Different performance
measures are presented that determine the registration quality
and accuracy. The scope for the future research are presented as
well.
Keywords—Image registration, classification, contribution,
drawback, performance measures, registration quality, accuracy,
future research.
I. INTRODUCTION
Image Registration is interpreted as the process of
overlaying two or more images of the same scene with respect
to a particular reference image. The images may be taken at
various circumstances (time-points), from various perspectives
(view-points), and additionally by various sensors. The
reference image is generally one of these captured images. It
geometrically transforms different sets of data into a particular
reference co-ordinate system. The discrepancies among these
images are interposed owing to the disparate imaging
conditions. Image acquisition devices underwent rapid
modifications and proliferating amount and diversity of
acquired images elicited the research on automatic image
registration.
In image analysis ventures, one of the most significant step
is Image Registration. It is a necessary step to obtain the final
information from a combination of a multitude of divergent
sources capturing the same information in varied circumstances
and diverse manners. Essentially the objective is to detect the
concealed relationship existing between the input and the
reference images which is usually indicated by a coordinate
transformation matrix. Accordingly, an image registration can
be essentially devised as an optimization problem. Image
registration plays a crucial role in many real-world
applications.
Image registration finds applications in remote sensing [13] involving multispectral classification, environmental
__________________________________________________
*
Corresponding author mail id: [email protected]
monitoring, change detection, image mosaicing, weather
forecasting, creating super-resolution images and integrating
information into geographic information systems (GIS), in
medicine [4-8] including fusion of computer tomography (CT)
and NMR data to obtain more complete information about the
patient, multi-modal analysis of different diseases like epilepsy
where the protocols incorporate functional EEG/MEG data
along with anatomical MRI, monitoring tumor evolution,
treatment verification, juxtaposition of the patient’s data with
anatomical atlases, in cartography for map updating, and in
computer vision for target localization, automatic quality
control and motion tracking.
According to the manner of image acquisition the
application of Image Registration can be segregated into the
following groups.
1. Multi-view Analysis: Images of the similar object or
scene are captured from multiple viewpoints to gain a
better representation of the scanned object or scene.
Examples include mosaicing of images and shape
recovery from the stereo.
2. Multi-temporal Analysis: Images of the same object/
scene are captured at various times usually under
dissimilar conditions to notice changes in the object/
scene which emerged between the successive images
acquisitions. Examples include motion tracking,
tracking the growth of tumors.
3. Multi-modal Analysis: Different sensors are used to
acquire the images of the same object/scene to merge
the information obtained from various sources to
obtain the minutiae of the object/scene. Examples
include integration of information from sensors with
disparate characteristics providing better spatial and
spectral resolutions independent of illumination-this
depends upon the robustness of the registration
algorithm, combination of sensors capturing the
anatomical information like magnetic resonance
image (MRI), ultrasound or CT with sensors acquiring
functional information like positron emission
tomography (PET), single photon emission computed
tomography (SPECT) or magnetic resonance
spectroscopy (MRS) to study and analyze seizure
disorders, Alzheimer’s disease, depression and other
diseases. Figure 1 shows a MEG-MRI co-registration,
an example of Multi-Modal Registration.
Section 2 presents steps involved in Image registration,
Section 3 contains classification criteria, Registration methods
are presented in Section 4, Transform Model Estimation and
Performance Analysis are discussed in Sections 5 and 6
respectively while Section 7 contains the conclusion.
Fig. 1. Multimodal MRI-MEG Co-registration. Top-Yellow dots
represent anatomical landmarks or fiducial points in the axial view
of the brain image (anatomical information). Bottom- Pink dots
represent the MEG sensors locations and the Green dots represent
the scalp-EEG sensors locations. These MEG and EEG data contain
the functional information and the bottom picture shows the coregistered brain image (sagittal view).
II. STEPS INVOLVED IN IMAGE REGISTRATION
An Image Registration task involves the following steps as
follows:
1. Feature detection: This is an important task of the
Image Registration process. The detection process can
be manual or automatic depending upon the
complexity though automatic detection of features is
preferred. Closed-boundary regions [9-16], edges,
contours [17-26], line intersections, corners [27] along
with their point representatives like center of gravity or
line endings (collectively known as Control Points) can
serve as features. These features consisting of
distinctive objects must be easily detectable, that is, the
features will be physically interpretable and
identifiable. The feature set of the reference image
must be sharing sufficient common features with the
non-aligned image(s) irrespective of any undesired
occlusions or unexpected changes for proper
registration. The algorithm for detection should be
robust enough to be able to detect the same features in
all projections of the scene without being affected by
any specific image deformation or degradation.
2. Feature matching: This step essentially establishes the
correspondence between the features detected in the
non-aligned sensed image and those detected in the
reference image [28-36]. Different feature descriptors
and similarity measures besides spatial relationships
among the features are adopted to set up an accurate
accordance. The feature descriptors must so formulated
such that they remain unchanged in spite of any
degradations and concurrently they must be able to
properly discriminate among diverse features while
remaining unaffected by noise.
3. Transform model assessment: For alignment of the
sensed image with the reference image the parameters
of the mapping functions are to be estimated [37-43].
These parameters are computed with the established
feature correspondence obtained from the previous
step. The selectivity of a mapping function depends on
a priori knowledge regarding the acquisition process
and expected image deformations. In absence of any a
priori information the flexibility of the model must be
ensured to tackle image deformations.
4. Image transformation: The sensed image is
transformed for alignment employing the mapping
functions.
The above mentioned image registration steps are generally
followed. Figure 2 shows a pictorial representation of the steps
involved in image registration. Though it is noteworthy to
mention that it is difficult to fabricate a universal method
applicable to all registration assignments the reason attributed
to the diversity of images to be registered obtained from a
miscellany of sources and the several types of degradations
introduced in the images. Besides geometric deformation
between the images, the radiometric deformations and noise
corruptions should be taken into account for proper registration
of images.
Fig. 2. Steps Involved in Image Registration
Fig. 3. Steps Involved in Image Registration. Top Left- Reference Image. Top Right- Non-Aligned Image. Bottom- Aligned Image.
Yellow dots represent the extracted features and there are enough common features in both the images. A mapping function is
established which gives the bottom image as the final output.
III. CLASSIFICATION CRITERIA OF IMAGE REGISTRATION
TECHNIQUES
Image registration techniques can be classified based on
some criteria [44-45]. These are as follows:
1. Dimensionality: This specifies the dimensions of
different possible registrations. It may be 2D-2D, 2D3D or 3D-3D based on the requirement.
2. Domain of transformation: It may be global when the
entire image is to be registered or it may be local when
a portion of the image is taken into consideration for
registration purpose.
3. Type of transformation: It may be rigid (translation,
rotation, reflection), affine (translation, rotation,
scaling, reflection, shearing), projective or non-linear.
4. Registration Quality: Depending on the data or the
features extracted several measures can be adopted and
applied.
5. Parameters of Registration: These are obtained
employing search oriented methods. The optimum
parameters found from a search method (e.g., a
heuristic search method) determines the quality of
transformation and hence the registration.
6. Subject of Registration: Same subject is considered for
intra-subject registration. If the subjects are different
then it is known as inter-subject registration.
7. Object of Registration: Different objects include head,
abdomen, thorax, knee, etc.
8.
Nature of Registration basis: It may be extrinsic (based
on foreign objects which are easily detectable, e.g.,
markers glued to skin), intrinsic (based on image
information) or non-image based (where imaging coordinates of the two devices are matched).
9. Interaction: It may be interactive, semi-automatic or
entirely automatic.
10. Modalities involved: It may be mono-modal (which is
also termed as intra-modal) using modalities like
Computed Tomography (CT), Magnetic Resonance
Imaging (MRI), Positron Emission Tomography
(PET),
Single
Photon
Emission
Computed
Tomography (SPECT), Ultra Sound (US), or Xray or
Digital Subtraction Angiography (DSA) or multimodal
(which is also known as inter-modal image) employing
two or more modalities mentioned above.
IV. METHODS OF IMAGE REGISTRATION
Various methods of Image Registration are as follows.
1. Extrinsic Methods
In this method artificial foreign objects which are easily
detectable are attached to the patient body [46-53]. They serve
as external features to be used for feature matching. The
complexity is lessened and hence computational is fast and
accuracy is also maintained. Examples are markers glued to
patient’s skin or stereo-tactic frame attached rigidly to the
patient’s outer skull for invasive neurosurgery related purposes.
2. Surface Methods
Surfaces or boundaries or contours are generally distinct in
medical images unlike landmarks. For example, surface-based
approach is employed for registering multimodality brain
image. These surface matching algorithms are generally
applied to rigid body registration. A collection of points,
generally called a point set is extracted from the contours in an
image. If two surfaces are considered for registration then there
will be two such sets. The surface covering the larger volume
of the patient, or that having a higher resolution if volume
coverage is comparable, is generally considered for generation
of the surface model. Iterative Closest Point Algorithm and
Correspondence Matching Algorithm are successfully applied
as registration algorithms for surface-based techniques [54-63].
Meta-heuristics and Evolutionary Optimization are also seen to
solve these high dimensional optimization problems of surface
registrations.
3. Moments and Principle Axes Methods
The orthogonal axes about which the moments of inertia
are minimized are known as the principle axes. Two identical
objects can be registered accurately by bringing their principal
axes into concurrence without employing any rigid/affine
transformations. If the objects are not identical but similar in
appearance then they can be approximately registered by this
technique [16, 64]. For moment based methods presegmentation is done in many cases to engender satisfactory
outcomes.
4. Correlation Based Methods
This method is essentially useful for registration of monomodal images and for comparison of several images of the
similar object [65]. It has immense usage in the field of
medical sciences for analyzing and treatment of disease.
Extracted features from the images are also used to obtain the
cross-correlation coefficients for image registration. [66-69].
Cross-correlation and Phase-correlation techniques based on
Fourier domain are also used for image registration. Successful
yet complex ventures have been significantly made using
subspace-based frequency estimation approach for the Fourier
based image registration problem employing multiple signal
classification algorithm (MUSIC) to proliferate robustness
eventually yielding accurate results [70]. Normalized mutual
information between the images have been used for image
registration purposes adopting an Entropy Correlation
Coefficient (ECC) [71]. Fourier-based techniques accompanied
by search algorithms have been exploited to evaluate the
transformation between two input images [72].
5. Mutual Information Based Methods
In mutual information-based registration methods the joint
probability of the intensities of comparable voxels in the
images under consideration are estimated. Mutual information
based measures are utilized to aid Voxel-based Registration.
Mutual information can be fruitfully utilized for establishing
the correspondence between the features of the reference and
the sensed images as mentioned in the step of featurematching. Correlation methods have proved inefficient for
multi-modal registration. But, the mutual information based
methods do not suffer from such a problem, rather they are
found to perform effectively in multi-modal registration tasks.
Gradient descent optimization methods have been employed to
maximize mutual information [73]. Window and pyramid
based approaches are used to achieve image registration using
mutual information [74]. Other methods used include
hierarchical search strategies along with simulated annealing
[35] and the Powell’s multi-dimensional direction set method
[66]. Recently various optimization methods and multiresolution strategies are adopted for mutual information
maximization.
6. Wavelet Based Methods
Wavelet Transform was introduced to get an idea of the
time instant at which a particular frequency exists. The width
of the window is altered as the transform is computed for each
spectral component- the most important characteristic of the
multi-resolution wavelet transform. It offers both time and
frequency selectivity, that is, it is able to localize properties in
both temporal and frequency domains. The wavelet-based
image registration can be effectively. After choosing several
wavelet coefficients by selection rules like the maximum
absolute wavelet coefficient in the multi-spectral image and the
high-resolution image for individual band the partial wavelet
coefficients of the high-resolution image are replaced with
those of the multi-spectral low-resolution image. The
pyramidal approaches also use wavelet decomposition owing
to its intrinsic multiresolution properties. Different types of
wavelets like the Haar, Symlet, Daubechies [75] and Coiflets
are applied for finding the correspondence with different sets
of wavelet coefficients. Wavelet-based feature extraction
techniques along with normalized cross-correlation matching
and relaxation-based image matching techniques are used
thereby incorporating sufficient control points to reduce the
local degradations, for image registration [76].
7. Soft Computing Based Methods
These methods are comparatively recent and advanced and
are successfully applied to image registration tasks. They
include Artificial Neural Networks, Fuzzy Sets and several
Optimization Heuristics.
a) Artificial Neural Networks:
An artificial neural network (ANN) is a computational
model which is formulated based on biological neural
networks. It is also known as Multi-Layer Perceptron (MLP)
since it contains a number of hidden layers. These layers
consist of an interconnected group of artificial neurons an
information is passed on from one layer to the next layer.
Artificial Neural Networks or simply Neural Networks
learns adaptively in the learning phase when information
flows through the network and updates the neuron-links
accordingly by assigning various weights to them. Neural
Networks can be viewed upon as non-linear statistical data
modeling tools employed to model complex relationships
between inputs and outputs or to recognize patterns in data,
also called Pattern Recognition. There are two types of
schemes: (1) feed-forward networks, where the links are
devoid of any loop (e.g., multilayer perceptron (MLP) and
radial basis function neural networks (RBF) and (2)
recurrent networks which include loops (e.g., selforganizing maps (SOM) and Hopfield Neural Networks). A
priori information about the output is an essential
requirement for training feed forward networks, on the other
hand, recurrent neural networks generally do not require any
such previous knowledge regarding the expected output. The
rigorous training process in an ANN modifies and
adaptively updates the network architecture abreast the
connection weights or link weights to be able to learn
complex non-linear input-output relationships thereby
parlaying the robustness and efficacy of performance. Multilayer Perceptron, Radial basis functions, self-organizing
maps and Hopfield networks have been utilized for different
computational and optimization aspects and for designing
registration matrices in Image Registration problems [77].
Neural Networks have also been used for solving monomodal and multi-modal medical image registration problems
[78].
b) Fuzzy Sets:
A fuzzy set is a collection of elements having a
continuous sequence of membership grades or degrees.
Fuzzy sets was introduced by L. A. Zadeh in 1965. Fuzzy
sets follow the properties of inclusion, union, complement,
intersection, etc. In classical set theory, the membership
values of elements in a set are decided in binary terms
depending upon whether an element belongs or does not
belong to the set. In contrast, fuzzy set theory allows the
grading of the membership of elements in a fuzzy set as
decided with the assistance of a membership function which
assigns values residing in the interval [0, 1]. Fuzzy Sets
manifest the perception of partial membership of an element
within the set- this permits Fuzzy sets to tackle uncertainty
and inaccuracies. Fuzzy Sets have been explicitly applied to
Image registration techniques [79-80]. It has also been
utilized to choose and pre-process the extracted features to
be registered. Fuzzy logic is used to enhance the precision in
the transformation parameters as estimated approximately
previously eventually leading to accurate registration
estimates [81].
Optimization Heuristics:
Optimization problems applied in several domains of
Engineering Design and Optimization have some
mathematical models and objective functions. They may be
unconstrained (without constraints) or constrained (with
constraints) having both continuous as well as discrete
variables. The task of finding the optimal solutions is
difficult with numerous curtailments being active at the
points of global optima. Traditional methods including
Gradient Descent, Dynamic Programming and Newton
Methods are computationally less efficient, whereas provide
feasible solutions in a stipulated time. The list of metaheuristics include Genetic Algorithm (GA) [82], Particle
Swarm Optimization (PSO) [83], Gravitational Search
Algorithm (GSA) [84], Ant Colony Optimization (ACO)
[85-86], Stimulated Annealing (SA) [87-88], and Plant
Propagation Algorithm (PPA) [89-90] and so on.
GA is a relatively old, approximate search technique
used in computing. These global search heuristics form an
important class of evolutionary algorithms that mimics
evolutionary biological processes such as mutation,
selection, and crossover and abandonment. Likewise,
Particle Swarm Optimization and Differential Evolution
along with their existing variants are relatively advanced
heuristics than can efficiently solve Optimization problems.
These optimization heuristics are applied to image
registration problems for finding the optimal parameters
necessary for designing a transformation model [91].
V. TRANSFORM MODEL ESTIMATION
A transformation is expounded as the process of mapping a
set of points to various other locations. The objective is to
design a proper transformation model which transforms the
sensed image with respect to the original image with maximum
accuracy. The transformations that may be performed are
translation, rotation, scaling, shearing and reflection. These are
collectively known as affine transformation. Also there are
projective and non-linear transformations as well.
1. Translation
Let a point x is to be translated by t units, then the matrix
representation of this transformation is given as:
(1)
where, y1, y2 = new point, x1, x2 = old point, t1, t2 = translation
value.
2. Rotation
A point with co-ordinate P1(x1, x2) on a 2-D plane is rotated
by an angle θ with respect to origin then the relationship
between the final point P2(y1, y2) and the initial point is given
as:
(2)
c)
where, y1, y2 = new point, x1, x2 = old point, θ = rotational
parameter.
3. Scaling
Scaling is required to resize an image, or to work with
images whose voxel sizes differ between images. It is
represented as:
(3)
where, y1, y2 = new point, x1, x2 = old point, s1, s2 = scaling
parameters.
4. Shearing
In shearing the parallel lines are only preserved. It may be
represented as:
(4)
where, y1, y2 = new point, x1, x2 = old point, a11, a12, a13, a21, a22,
a23, = shearing parameters. Fig. 4. shows an example of
shearing transformation.
selection of similarity measures depends on modality of images
to be registered. Correlation based metrics like Correlation
Coefficient is applicable to mono-modal registration and
Mutual Information is utilized for multi-modal image
registration purposes.
a) Correlation Coefficient (CC):
CC is essentially a similarity measure which gives an
idea of how well the reference and transformed images are
identical [32-34]. If two images are perfectly identical, CC
gives a value equal to 1, whereas, if the two images are
completely uncorrelated CC value is equal to 0 and CC
value equal to -1 indicates that the images are completely
anti-correlated, which means one image is the negative of
the other. It gives satisfactory results with mono-modal
registration. It is represented as:
(5)
where xi , yi = intensity of ith pixel in the reference and
sensed image respectively, and xm , ym = mean intensity of
reference and sensed image respectively.
b) Mutual Information (MI):
MI is yet another measure determining the degree of
similarity measured between the image intensities of
corresponding voxels in both images [35-36]. MI is
maximized when both the images are accurately aligned.
The values of MI are non-negative and symmetric. The
range of MI values starts from zero and can vary up to a
high value. High MI value depicts large reduction in
uncertainty whereas zero MI value is clear indication that
the two variables are independent. It is represented as:
(6)
where p(x, y) = joint distribution function and p1(x) , p2(y) =
marginal distribution functions.
Fig. 4. Example of Affine Transformations (Translation, Rotation, Scaling and
Shearing)
VI. PERFORMANCE ANALYSIS
It is required to estimate how accurate the registration
actually is. Also to qualitatively analyze the performance of the
algorithms some metrics are used. They also serve as the basis
for improvement in the registration in each iteration. The
Fig. 5. Example of Multi-modal Image registration using Mutual Information
as similarity measure. Top Left- Reference MRI brain image (axial view), Top
Right- Non- Aligned PET brain image (axial view), Bottom- Transformed PET
brain image (axial view).
VII. CONCLUSION
This paper tries to present a survey of the registration
methods along with the detailed classifications among various
approaches. Image registration is an essential step for
integrating or fusing and analyzing information from various
sensors (sources). It has immense applications in fields of
medical sciences, computer vision and remote sensing. Image
registrations with complex nonlinear distortions, multi-modal
registration and registrations of occluded images despite being
affected by illumination factors among others thus contributing
to the robustness of the approaches belong to the most
challenging tasks at the present scenario. Generation of
features or control points and the mapping or transformation
functions are essential steps and a lot of research work needs to
be done to enhance the accuracy. In multimodal registration,
MI technique has gained popularity in particular whereas for
mono-modal images correlation based similarity metrics are
preferred. Robustness and Reliability can be proliferated by
hybrid approaches combining MI based techniques with
feature-based measures. Several soft computing methods
including the optimization heuristics are applied to find the
optimum parameters mostly in case of affine transformations
based registration. No gold standard algorithms or approaches
can be developed for image registration purposes because of
the dependency on the images under consideration. Thus,
despite a lot of work has been done, automatic image
registration is still considered as an open problem. The future
works will be introducing new feature-based methods, where
apt modality-insensitive features can provide robust as well as
accurate outcomes for the registration.
ACKNOWLEDGMENT
I would like to extend my sincere gratitude to Professor
Sugata Munshi, Professor Amitava Chatterjee and Professor
Mita Dutta for their support and guidance.
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
REFERENCES
[1]
[2]
[3]
[4]
[5]
L.M.G. Fonseca, B.S. Manjunath, Registration techniques for
multisensor remotely sensed imagery, Photogrammetric
Engineering and Remote Sensing 62 (1996) 1049–1056.J. Clerk
Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol.
2. Oxford: Clarendon, 1892, pp.68–73.
E. Gu¨lch, Results of test on image matching of ISPRS WG,
ISPRS Journal of Photogrammetry and Remote Sensing 46
(1991) 1–18
J. le Moigne, First evaluation of automatic image registration
methods, Proceedings of the International Geoscience and
Remote Sensing Symposium IGARSS’98, Seattle, Washington,
1998, pp. 315–317.
D.L.G. Hill, P.G. Batchelor, M. Holden, D.J. Hawkes, Medical
image registration, Physics in Medicine and Biology 46 (2001)
R1–R45.
H. Lester, S.R. Arridge, A survey of hierarchical non-linear
medical image registration, Pattern Recognition 32 (1999) 129–
149.
[18]
[19]
[20]
[21]
[22]
P.A. van den Elsen, E.-J.D. Pol, M.A. Viergever, Medical image
matching-a review with classification, IEEE Engineering in
Medicine and Biology 12 (1993) 26–39.
J.B.A. Maintz, M.A. Viergever, A survey of medical image
registration, Medical Image Analysis 2 (1998) 1–36.
Damas, Sergio, Oscar Cordón, and Jose Santamaría. "Medical
image registration using evolutionary computation: An
experimental survey." IEEE Computational Intelligence
Magazine 6.4 (2011): 26-42.
A. Goshtasby, G.C. Stockman, C.V. Page, A region-based
approach to digital image registration with subpixel accuracy,
IEEE Transactions on Geoscience and Remote Sensing 24
(1986) 390–399
A. Goshtasby, G.C. Stockman, Point pattern matching using
convex hull edges, IEEE Transactions on Systems, Man and
Cybernetics 15 (1985) 631–637.
Y.C. Hsieh, D.M. McKeown, F.P. Perlant, Performance
evaluation of scene registration and stereo matching for
cartographic feature extraction, IEEE Transactions on Pattern
Analysis and Machine Intelligence 14 (1992) 214–237
M. Holm, Towards automatic rectification of satellite images
using feature based matching, Proceedings of the International
Geoscience and Remote Sensing Symposium IGARSS’91,
Espoo, Finland, 1991, pp. 2439–2442
P.A. Brivio, A.D. Ventura, A. Rampini, R. Schettini, Automatic
selection of control points from shadow structures, International
Journal of Remote Sensing 13 (1992) 1853–1860.
M. Sester, H. Hild, D. Fritsch, Definition of ground control
features for image registration using GIS data, Proceedings of
the Symposium on Object Recognition and Scene Classification
from Multispectral and Multisensor Pixels, CD-ROM,
Columbus, Ohio, 1998, 7 pp
M. Roux, Automatic registration of SPOT images and digitized
maps, Proceedings of the IEEE International Conference on
Image Processing ICIP’96, Lausanne, Switzerland, 1996, pp.
625–628
J. Flusser, T. Suk, A moment-based approach to registration of
images with affine geometric distortion, IEEE Transactions on
Geoscience and Remote Sensing 32 (1994) 382–387.
Y.C. Hsieh, D.M. McKeown, F.P. Perlant, Performance
evaluation of scene registration and stereo matching for
cartographic feature extraction, IEEE Transactions on Pattern
Analysis and Machine Intelligence 14 (1992) 214–237.
S. Moss, E.R. Hancock, Multiple line-template matching with
EM algorithm, Pattern Recognition Letters 18 (1997) 1283–
1292
W.H. Wang, Y.C. Chen, Image registration by control points
pairing using the invariant properties of line segments, Pattern
Recognition Letters 18 (1997) 269–281.
X. Dai, S. Khorram, Development of a feature-based approach
to automated image registration for multitemporal and
multisensor remotely sensed imagery, International Geoscience
and Remote Sensing Symposium IGARSS’97, Singapore, 1997,
pp. 243–245.
V. Govindu, C. Shekhar, R. Chellapa, Using geometric
properties for
correspondence-less image
alignment,
Proceedings of the International Conference on Pattern
Recognition ICPR’98, Brisbane, Australia, 1998, pp. 37–41.
G.P. Penney, J. Weese, J.A. Little, P. Desmedt, D.L.G. Hill, D.J.
Hawkes, A comparison of similarity measures for use in 2D–3D
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
medical image registration, IEEE Transactions on Medical
Imaging 17 (1998) 586–595.
G. Medioni, R. Nevatia, Matching images using linear features,
IEEE Transactions on Pattern Analysis and Machine
Intellingence 6 (1984) 675–685.
D. Shin, J.K. Pollard, J.P. Muller, Accurate geometric correction
of ATSR images, IEEE Transactions on Geoscience and Remote
Sensing 35 (1997) 997–1006
S.Z. Li, J. Kittler, M. Petrou, Matching and recognition of road
networks from aerial images, Proceedings of the Second
European Conference on Computer Vision ECCV’92, St
Margherita, Italy, 1992, pp. 857–861
N. Vujovic, D. Brzakovic, Establishing the correspondence
between control points in pairs of mammographic images, IEEE
Transactions on Image Processing 6 (1997) 1388–1399.
A. Noble, Finding corners, Image and Vision Computing 6
(1988) 121–128.
R.J. Althof, M.G.J. Wind, J.T. Dobbins, A rapid and automatic
image registration algorithm with subpixel accuracy, IEEE
Transactions on Medical Imaging 16 (1997) 308–316.
D.I. Barnea, H.F. Silverman, A class of algorithms for fast
digital image registration, IEEE Transactions on Computing 21
(1972) 179–186.
W.K. Pratt, Correlation techniques of image registration, IEEE
Transactions on Aerospace and Electronic Systems 10 (1974)
353–358.
H. Hanaizumi, S. Fujimura, An automated method for
registration of satellite remote sensing images, Proceedings of
the International Geoscience and Remote Sensing Symposium
IGARSS’93, Tokyo, Japan, 1993, pp. 1348–1350
R. Berthilsson, Affine correlation. Proceedings of the
International Conference on Pattern Recognition ICPR’98,
Brisbane, Australia, 1998, p. 1458–1461.
A. Simper, Correcting general band-to-band misregistrations,
Proceedings of the IEEE International Conference on Image
Processing ICIP’96, Lausanne, Switzerland, 1996, 2, pp. 597–
600
W.K. Pratt, Digital Image Processing, 2nd ed., Wiley, New
York, 1991.
N. Ritter, R. Owens, J. Cooper, R.H. Eikelboom, P.P. van
Saarloos, Registration of stereo and temporal images of the
retina, IEEE Transactions on Medical Imaging 18 (1999) 404–
418.
C. Studholme, D.L.G. Hill, D.J. Hawkes, An overlap invariant
entropy measure of 3D medical image alignment, Pattern
Recognition 32 (1999) 71–86.
J.M. Fitzpatrik, J.B. West, The distribution of target registration
error in rigid-body point-based registration, IEEE Transactions
on Medical Imaging 20 (2001) 917–927
J. Flusser, An adaptive method for image registration, Pattern
Recognition 25 (1992) 45–54.
R.B. Huseby, O.M. Halck, R. Solberg, A model-based approach
for geometrical correction of optical satellite images,
Proceedings of the International Geoscience Remote Sensing
Symposium IGARSS’99, Hamburg, Germany, 1999, pp. 330–
332.
O. Thepaut, K. Kpalma, J. Ronsin, Automatic registration of
ERS and SPOT multisensor images in a data fusion context,
Forest Ecology and Management 128 (2000) 93–100.
[41] A. Goshtasby, Piecewise linear mapping functions for image
registration, Pattern Recognition 19 (1986) 459–466.
[42] A. Goshtasby, Piecewise cubic mapping functions for image
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
registration, Pattern Recognition 20 (1987) 525–533.
A. Goshtasby, Image registration by local approximation
methods, Image and Vision Computing 6 (1988) 255–261.
P. A. Van Den Elsen, E. J. D. Pol and M. A. Viergever,
“Medical image matching: a review with classification”, IEEE
Engineering in medicine and biology, 12(1):1993,26-39
J. B. Antoine Maintz and Max A. Viergever, “A Survey of
Medical Image Registration”, Medical Image Analysis (1998)
Volume2, number 1, pp 1-36, Oxford University Press
L. Lemieux, N. D. Kitchen, S. W. Hughes and D. G. T. Thomas,
“Voxel-based localization in frame-based and frameless
stereotaxy and its accuracy”. Medical physics, 21(8):1301–1310,
1994.
L. Lemieux and R. Jagoe, “Effect of fiducial marker localization
on stereotactic target coordinate calculation in CT slices and
radiographs” Physics in medicine and biology, 39:1915–1928,
1994.
K. P. Gall and L. J. Verhey, “Computer-assisted positioning of
radiotherapy patients using implanted radioopaque fiducials”,
Medical physics, 1993, 1152–1159
C. R. Maurer, G. B. Aboutanos, B. M. Dawant, R. A. Margolin,
R. J. Maciunas and J. M. Fitzpatrick., “Registration of CT and
MR brain images using a combination of points and surfaces”,
Medical imaging: image processing, volume 2434, Bellingham,
WA, 1995. SPIE Press, 109-123
R. E. Ellis, S. Toksvig-Larsen, M. Marcacci, D. Caramella, and
M. Fadda., “A biocompatible fiducial marker for evaluating the
accuracy of CT image registration”, Computer assisted
radiology, volume 1124 of Excerpta medica - international
congress series, Amsterdam, 1996,693-698
C. R. Maurer, R. Calvin, J. J. McCrory, and J. M. Fitzpatrick,
“Estimation of accuracy in localizing externally attached
markers in multimodal volume head images”, In M. H. Loew,
editor, Medical imaging: image International Journal of Signal
Processing, Image Processing and Pattern Recognition Vol. 2,
No.3, September 2009 22 processing, volume 1898, 1993, SPIE
Press, 43-54
A. C. Evans, S. Marrett, J. Torrescorzo, S. Ku, and L.Collins,
“MRI-PET correlation in three dimensions using a volume of
interest (VOI) atlas”, Journal of cerebral blood flow and
metabolism , 11, A69–A78, 1991
W. D. Leslie, A. Borys, D. McDonald, J. O. Dupont and A. E.
Peterdy, “External reference markers for the correction of head
rotation in brain single-photon emission tomography”, European
journal of nuclear medicine, 22(4):351–355, 1995.
C. Dorai, G. Wang, A.K. Jain, C. Mercer, “Registration and
integration of multiple object views for 3D model construction”,
IEEE Trans. Pattern Anal. Machine Intelligence. 20 (1),(1998),
83–89.
T. Masuda, K. Sakaue, N. Yokoya, “Registration and integration
of multiple range images for 3-D model construction”,
Proceedings of the 13th International Conference on Pattern
Recognition, Vol. 1, 1996, 879–883.
P.J. Besl, N.D. McKay, “A method for registration of 3-D
shapes”, IEEE Trans. Pattern Analysis Machine Intelligence 14
(2) (1992), 239–256.
[57] G. Blais, M.D. Levine, “Registering multiview range data to
[74] P. Thevenaz, M. Unser, Spline pyramids for inter-modal image
create 3D computer objects”, IEEE Trans. Pattern Analysis and
.Machine Intelligence 17 (8) (1995) 820–824.
T. Masuda, N. Yokoya, “A robust method for registration and
segmentation of multiple range images”, Proceedings of the
Second Workshop on CAD-Based Vision, 1994, 106–113.
S.M. Yamany, A.A. Farag, “Free-form surface registration using
surface signatures” Proceedings of the Seventh IEEE
International Conference on Computer Vision, Vol. 2, 1999,
1098–1104.
C. Schutz, T. Jost, H. Hugli, “Multi-feature matching algorithm
for free-form 3D surface registration”, Proceedings of the 14th
International Conference on Pattern Recognition, Vol. 2, 1998,
982–984.
A.E. Johnson, M. Hebert, “Surface registration by matching
oriented points”, Proceedings of the International Conference on
Recent Advances in 3-D Digital Imaging and Modeling, 1997,
121–128.
W.R. Fright, A.D. Linney, “Registration of3-D head surfaces
using multiple landmarks”, IEEE Trans. on Medical Imaging 12
(3) (1993), 515–520.
Chi Kin Chow, Hung Tat Tsui, Tong Lee, “Surface registration
using a dynamic genetic algorithm”, Pattern Recognition 37,
(2004), 105-117
J. Flusser, T. Suk, “Pattern recognition by affine moment
invariants”, Pattern Recognition 26, (1993), 167– 174
B.K. Ghaffary, A.A. Sawchuk, “A survey of new techniques for
image registration and mapping”, Proceedings of the SPIE:
Applications of Digital Image Processing 432 (1983) 222–239.
F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, and
Suetens, “Multimodality image registration by maximization of
mutual information,” IEEE Transactions on Medical Imaging,
vol. 16, no. 2, (1997), 187–198
S. Sanjay-Gopal, H. P. Chan, T. Wilson, M. Helvie, N. Petrick
and B. Sahiner, “A regional registration technique for automated
interval change analysis of breast lesions on mammograms,”
Medical Physics, vol. 26, no. 12, (1999), 2669–2679
L. Junck, J. G. Moen, G. D. Hutchins, M. B. Brown, and D. E.
Kuhl, “Correlation methods for the centering, rotation, and
alignment of functional brain images”, Journal of nuclear
medicine, 31,(1990), 1220–1276
S. L. Bacharach, M. A. Douglas, R. E. Carson, P. J. Kalkowski,
N. M. T. Freedman, P. Perrone-Filardi and R.O.Bonow, “Threedimensional registration of cardiac positron emission
tomography attenuation scans”, Journal of nuclear medicine,
34(2), (1993), 311–321
Min Xu Varshney P.K. ,“ A Subspace Method for Fourier-Based
Image Registration”, Geoscience and Remote Sensing Letters,
IEEE 33, Volume 6, Issue 3 (July 2009) , 491-494
F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, and
Suetens, “Multimodality image registration by maximization of
mutual information,” IEEE Transactions on Medical Imaging,
vol. 16, no. 2, (1997), 187–198
Samritjiarapon O. Chitsobhuk O. , “An FFT-Based Technique
and Best-first Search for Image Registration”, International
Symposium on Communications and Information Technologies,
ISCIT 2008.
P.Viola, W.M. Wells, “Alignment by maximization of mutual
information”, International Journal of Computer Vision 24,
(1997), 137–154.
registration using mutual information, Proceedings of SPIE:
Wavelet Applications in Signal and Image Processing, San
Diego, CA, (1997), 236– 247.
J. le Moigne, “Parallel registration of multi-sensor remotely
sensed imagery using wavelet coefficients”, Proceedings of the
SPIE: Wavelet Applications, Orlando, Florida, 2242, (1994),
432–443
Gang Hong Yun hang , “Combination of feature-based and areabased image registration technique for high resolution remote
sensing image”, IEEE international symposium on Geoscience
and remote sensing, IGARSS 2007 (July 2007), 377-380
]Heng Liua,, Jingqi Yan,, David Zhang, “Three-dimensional
surface
registration:
A
neural
network
strategy”
Neurocomputing 70, (2006), 597–602
]Lifeng Shang, Jian Cheng Lv, Zhang Yi, “Rigid medical image
registration using PCA neural network”, Neurocomputing 69
(2006), 1717–1722
G. Berks, A. Ghassemi, and D. G. von Keyserlingk, "Spatial
registration of digital brain atlases based on fuzzy set theory,"
Comput. Med. Imaging Graph., vol. 25, (Jan. 2001), 1-10
Y. Hata, S. Kobashi, S. Hirano, and M. Ishikawa, "Registration
of multi-modality medical images by soft computing approach,"
in Proc. ICONIP' 99, (1999), 878-883.
Ramirez L. Durdle N.G. Raso V.J. “A Parameters Selection
Scheme for Medical Image Registration”, Fuzzy Information
Processing Society, 2006. NAFIPS 2006. Annual meeting of the
North American ,(June2006), 505-510
S.-J. Wu and P.-T. Chow, Genetic algorithms for nonlinear
mixed discrete-integer optimization problems via meta-genetic
parameter optimization, Engineering Optimization, vol. 24, no.
2, pp. 137–159, 1995.
[27] R. Eberhart and J. Kennedy, A new optimizer using particle
swarm theory, in Proceedings of the 6th International
Symposium on Micro Machine and Human Science (MHS ’95),
pp. 39– 43, IEEE, Nagoya, Japan, October 1995.
[28] Rashedi, E., Nezamabadi-pour, H., Saryazdi, S., GSA: A
Gravitational Search Algorithm 179(13), 2232–2248 (2009).
[29] M. Dorigo and G. D. Caro, Ant algorithms for discrete
optimization, Artificial Life, vol. 5, no. 3, (1999), pp. 137-172.
[30] M. Dorigo and L. M. Gambardella, Ant colony system: a
cooperative learning approach to the traveling salesman
problem, IEEE Transactions on Evolutionary Computation, vol.
1, no. 1, (1997), pp. 53-66.
C. Zhang and H.-P. Wang, Mixed-discrete nonlinear
optimization
with
simulated
annealing,
Engineering
Optimization, vol. 21, no. 4, pp. 277–291, 1993.
E. H. L. Aarts, J. H. M. Korst, and P. J. M. van Laarhoven,
Simulated annealing, in Local Search in Combinatorial
Optimization, pp. 91–120, 1997.
Nag S. Adaptive Plant Propagation Algorithm for Solving
Economic Load Dispatch Problem. arXiv preprint
arXiv:1708.07040. 2017 Aug 4.
Nag S. A Type II Fuzzy Entropy Based Multi-Level Image
Thresholding Using Adaptive Plant Propagation Algorithm.
arXiv preprint arXiv:1708.09461. 2017 Aug 23.
J.M. Rouet, J.J. Jacq, and C. Roux, "Genetic algorithms for a
robust 3-D MR-CT registration," IEEE Trans. Inform. Technol.
Biomed., vol. 4, (Jun. 2000), 126-136
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
[91]
| 1cs.CV
|
Unique Parallel Decomposition for the π -calculus ∗
Matias David Lee
Univ. Lyon, ENS de Lyon, CNRS, UCB Lyon 1, LIP, France.
[email protected]
Bas Luttik
Eindhoven University of Technology, The Netherlands.
[email protected]
A (fragment of a) process algebra satisfies unique parallel decomposition if the definable behaviours
admit a unique decomposition into indecomposable parallel components. In this paper we prove
that finite processes of the π -calculus, i.e. processes that perform no infinite executions, satisfy this
property modulo strong bisimilarity and weak bisimilarity. Our results are obtained by an application
of a general technique for establishing unique parallel decomposition using decomposition orders.
1
Introduction
A (fragment of a) process algebra has unique parallel decomposition (UPD) if all definable behaviours
admit a unique decomposition into indecomposable parallel components. In this paper we prove that
finite processes definable in the π -calculus satisfy this property modulo strong bisimilarity and modulo
weak bisimilarity.
From a theoretical point of view, this property is interesting because it can be used to prove other
theoretical properties about process calculi. For instance, relying on unique parallel decomposition,
Moller proves in [18, 19] that PA and CCS cannot be finitely axiomatized without auxiliary operations,
and Hirshfeld and Jerrum prove in [12] that bisimilarity is decidable for normed PA. Unique parallel
decomposition can be also used to define a notion of normal form. Such a notion of normal form is
useful in completeness proofs for equational axiomatizations in settings in which an elimination theorem
for parallel composition is lacking (see, e.g., [1, 2, 3, 9, 11]). In [13], UPD is used to prove complete
axiomatisation and decidability results in the context of a higher-order process calculus.
From a practical point of view, unique parallel decomposition can be used to devise methods for
finding the maximally parallel implementation of a behaviour [6], or for improving verification methods
[10]. In [8], a unique parallel decomposition result is used as a tool in the comparison of different security
notions in the context of electronic voting.
The UPD property has been widely studied for different process calculi and variants of the parallel
operator. Milner and Moller were the first to establish a unique parallel decomposition theorem; they
proved the property for a simple process calculus that allows the specification of all finite behaviours
up to strong bisimilarity and includes parallel composition in the form of pure interleaving without interaction between its components [16]. Moller, in his dissertation [17], extended this result replacing
interleaving parallel composition by CCS parallel composition, and then also considering weak bisimilarity. Christensen, also in his dissertation [5], proved unique decomposition for normed behaviours
recursively definable modulo strong bisimilarity, and for all behaviours recursively definable modulo
∗
M.D. Lee has been supported by the project ANR 12IS02001 PACE
D. Gebler and K. Peters (Eds.): Combined Workshop on Expressiveness in
Concurrency and Structural Operational Semantics (EXPRESS/SOS 2016).
EPTCS 222, 2016, pp. 45–59, doi:10.4204/EPTCS.222.4
46
Unique Parallel Decomposition for the π-calculus
distributed bisimilarity; the proof of the latter result relies on a cancellation law for parallel composition
up to distributed bisimilarity, first established by Castellani as [4, Lemma 4.14].
Most of the aforementioned unique parallel decomposition results were established with subsequent
refinements of an ingenious proof technique attributed to Milner. In [15], the notion of decomposition
order is introduced in order to formulate a sufficient condition on commutative monoids that facilitates
an abstract version of Milner’s proof technique. It is then proved that if a partial commutative monoid
can be endowed with a decomposition order, then it has unique decomposition. Thus, an algebraic tool
is obtained that allows one to prove UPD for a process calculus by finding a decomposition order.
The tool can deal with most of the settings aforementioned. In this paper, we show how the tool can
also be applied to obtain unique parallel decomposition results for finite processes of the π -calculus w.r.t.
strong bisimilarity and w.r.t. weak bisimilarity. But, to this end, we do face two complications: The first
complication, in the context of the π -calculus is that, as opposed to previous settings, the decomposition
order is not directly induced on the commutative monoid of processes by the transition relation. The
culprit is that, in general, two parallel components may fuse into a single indecomposable process as a
result of scope extrusion. To define the decomposition order we consider a fragment of the transition
relation that avoids this phenomenon. The second complication, which arises only in the case of weak
bisimilarity, is that certain transitions are deemed unobservable, and that, as a consequence, there are
transitions that do not change state (are between weakly bisimilar processes). We demonstrate that a
decomposition order can, nevertheless, be obtained by ignoring such stuttering transitions.
The paper [7] studies unique parallel decomposition w.r.t. both strong bisimilarity and weak bisimilarity for the applied π -calculus. The applied π -calculus is a variant of the π -calculus that was designed
for the verification of cryptographic protocols. Its main feature is that channels can only transmit variables and the values of the variables are set using active substitutions. Roughly, active substitution is an
extension of the grammar of the π -calculus that works as a ‘memory’ that save the value of a variable.
Because the variables in a transition are observable but the ‘memories’ are not, it is possible to mask
sensitive information. The proof of the result for the strong case in [7] relies on induction over the norm
of a process and the fact that the norm of the arguments of a parallel composition is less than the norm of
the parallel composition. Unfortunately, this property is not true because of the restriction operator (see
Section 4 for a counter example). This is the reason why we restrict ourselves to finite processes in the
strong setting. The proof of the weak case in [7] follows the proof technique attributed to Milner. The
general techniques from [15] cannot be applied directly in the setting of the applied π -calculus due to
the active substitutions.
In [14], the second author presented an adaptation of the general result of [15] in order to make
it suitable for establishing unique parallel decomposition in settings with a notion of unobservable behaviour. The ensued technique amounts to showing that the transition relation induces a so-called weak
decomposition order satisfying a property that is called power cancellation. In the present paper, we
show how, instead of using the adapted technique from [14], the original technique from [15] may be
applied in settings with a notion of unobservable behaviour, considering a stutter-free fragment of the
transition relation. This method appears to be simpler than the method suggested by the result in [14].
The paper is organized as follows. In Section 2, we briefly recall the abstract framework introduced
in [15] to prove UPD results. In Section 3 we recall the syntax and different semantics of the π -calculus.
Section 4 is composed of two subsections. In Section 4.1 we introduce the notion of depth of a process
and we prove some properties of this notion. In Section 4.2 we use these results and the result in Section 2 to prove that finite processes of the π -calculus satisfy unique parallel decomposition w.r.t. strong
bisimilarity. Section 5 follows a similar structure. In Section 5.1 we introduce the notion of processes
without stuttering transitions and we prove some properties of this kind of processes. These properties
M.D. Lee and B. Luttik
47
and the result in Section 2 are used in Section 5.2 to prove that finite processes of the π -calculus satisfy
unique parallel decomposition w.r.t. weak bisimilarity. In Section 6 we present some final remarks.
2
Decomposition orders
In this section, we briefly review the theory of unique decomposition for commutative monoids that we
shall apply in the remainder of the paper to prove UPD results in the context of the π -calculus.
Definition 1. A commutative monoid is a set M with a distinguished element e and a binary operation
on M denoted by ⋅ such that for all x,y,z ∈ M :
• x ⋅ (y ⋅ z) = (x ⋅ y) ⋅ z (associativity);
• x ⋅ y = y ⋅ x (commutativity);
• x ⋅ e = e ⋅ x = x (identity).
In the remainder of the paper we often suppress the symbol ⋅ or use ∥.
Definition 2. An element p of a commutative monoid M is called indecomposable if p ≠ e and p = xy
implies x = e or y = e.
Definition 3. Let M be a commutative monoid. A decomposition in M is a finite multi-set *p1 ,... , pk + of
indecomposable elements of M such that p1 ⋅ p2 ⋯pk is defined. The element p1 ⋅ p2 ⋯pk in M will be called
the composition associated with the decomposition *p1 ,... , pk +, and, conversely, we say that *p1 ,..., pk +
is a decomposition of the element p1 ⋅ p2 ⋯pk of M. Decompositions d = *p1 ,..., pk + and d ′ = *p′1 ,... , p′l +
are equivalent in M (notation d ≡ d ′ ) if they have the same composition, i.e. p1 ⋅ p2 ⋯pk = p′1 ⋯p′l . A
decomposition d in M is unique if d ≡ d ′ implies d = d ′ for all decompositions d ′ in M. We say that an
element x of M has a unique decomposition if it has a decomposition and this decomposition is unique.
If every element of M has a unique decomposition, then we say that M has unique decomposition.
Theorem 1 below gives a sufficient condition to ensure that a commutative monoid M has unique
decomposition. It requires the existence of a decomposition order for M.
Definition 4. Let M be a commutative monoid; a partial order ≼ on M is a decomposition order if
1. it is well-founded, i.e., for every non-empty subset M̂ ⊆ M there is m ∈ M̂ such that for all m′ ∈ M,
m′ ≼ m implies m′ = m. In this case, we say that m is a ≼-minimal element of M̂;
2. the identity element e of M is the least element of M with respect to ≼, i.e., e ≼ x for all x in M;
3. ≼ is strictly compatible, i.e., for all x,y,z ∈ M if x ≺ y (i.e. x ≼ y and x ≠ y) and yz is defined, xz ≺ yz;
4. it is precompositional, i.e., for all x,y,z ∈ M x ≼ yz implies x = y′ z′ for some y′ ≼ y and z′ ≼ z; and
5. it is Archimedean, i.e., for all x,y ∈ M xn ≼ y for all n ∈ N0 implies that x = e.
Theorem 1 ([15]). Every commutative monoid M with a decomposition order has unique decomposition.
3
The π -calculus
We recall the syntax of the π -calculus and the rules to define the transition relation [20]. We assume a
set of names or channels V. We use a,b,c,x,y,z to range over V.
Unique Parallel Decomposition for the π-calculus
48
α
(Out)
xy
xy.P Ð→ P
xy
x(z).P Ð→ P{y/z}
(Inp)
P Ð→ P′
(Sum-L)
α
P + Q Ð→ P′
xy
x(z)
τ
z ∈/ fn(Q)
α
z /∈ n(α )
(Res)
P Ð→ P′
!P Ð
→ (P′ ∣ P′′ ) ∣!P
z≠x
x(z)
(Open)
ν z(P) ÐÐ→ P′
xy
P Ð→ P′′
x(z)
(Close-L)
P Ð→ P′
α
!P Ð→ P′ ∣!P
(Rep-Act)
xz
P ÐÐ→ P′
(Rep-Comm)
(Par-L)
α
xz
ν z(P) Ð→ ν z(P′ )
τ
Q Ð→ Q′
P∣QÐ
→ ν z(P′ ∣ Q′ )
α
P Ð→ P′
(Mat)
xz
P ÐÐ→ P′
(Comm-L)
τ
xy
bn(α ) ∩ fn(Q) = ∅
α
P ∣ Q Ð→ P′ ∣ Q
xy
Q Ð→ Q′
P∣QÐ
→ P′ ∣ Q′
P Ð→ P′
α
[x = x]π .P Ð→ P
α
α
P Ð→ P′
P Ð→ P′
τ
τ .P Ð
→P
π .P Ð→ P
(Tau)
P Ð→ P′′
τ
!P Ð
→ (ν z(P′ ∣ P′′ )) ∣!P
z /∈ fn(P)
(Rep-Close-L)
Table 1: Transition rules for the π -calculus
Definition 5. The processes, summations and prefixes of the π -calculus are given respectively by
P ∶∶= M
∣ P ∣ P′
∣ ν z.P
∣ !P
π .P ∣ M + M
π ∶∶= xy ∣ x(z) ∣ τ ∣ [x = y]π
′
M ∶∶= 0 ∣
We denote with Π the set of processes of the π -calculus.
An occurrence of a name z ∈ V is bound in a process P if it is in the scope of a restriction ν z or of an
input a(z). A name a ∈ V is free in a process P if there is at least one occurrence of a that is not bound.
We write bn(P) and fn(P) to denote, respectively, the set of bound names and free names of a process P.
The set of names of a process P is defined by n(P) = bn(P)∪ fn(P). We employ the following convention
of the π -calculus w.r.t. names.
Convention 1. [20, P.47] In any discussion, we assume that the bound names of any processes or actions
under consideration are chosen to be different from the names free in any other entities under consideration, such as processes, actions, substitutions, and sets of names. This convention is subject to the
x(z)
limitation that in considering a transition P ÐÐ→ Q, the name z that is bound in x(z) and in P may occur
free in Q. This limitation is necessary for expressing scope extrusion.
The transition relation associated to a term is defined by the rules in Table 1, where we have omitted
the symmetric version of the rules (Sum-L), (Par-L), (Comm-L) and (Close-L). We denote with A the set
of visible actions that can be executed by a process P ∈ Π, i.e. A = {xy ∣ x,y ∈ V} ∪ {xy ∣ x,y ∈ V} ∪ {x(z) ∣
x,z ∈ V}. The action τ is the internal action. We define Aτ = A ∪ {τ }. For P,P′ ∈ Π and α ∈ Aτ , we write
α
α
P Ð→ P′ if there is a derivation of P Ð→ P′ with rules in Table 1.
Definition 6. Strong bisimilarity is the largest symmetric relation over Π, notation ∼, such that whenever
α
α
P ∼ Q, if P Ð→ P′ then there is Q′ s.t. Q Ð→ Q′ and P′ ∼ Q′ .
M.D. Lee and B. Luttik
49
The relation ∼ is not compatible with input prefix: We have that
zx ∣ a(y) ∼ zx.a(y) + a(y).zx ,
whereas
b(a).(zx ∣ a(y)) ∼/ b(a).(zx.a(y) + a(y).z),
because when z is received over the channel a, we have
(zx ∣ a(y)){a/z} ∼/ (zx.a(y) + a(y).z){a/z} .
Hence, ∼ is not a congruence for the full syntax of the π -calculus. It is, however, a so-called non-input
congruence (see [20, Theorem 2.2.8]): it is compatible with all the other constructs in the syntax. In the
present paper we shall only use the fact that ∼ is compatible with parallel composition, i.e., if P1 ∼ Q1
and P2 ∼ Q2 , then P1 ∣ P2 ∼ Q1 ∣ Q2 .
We recall now the weak variant of bisimilarity. We write P Ô⇒ P′ if P = P′ or if there are P0 ,... ,Pn
α
τ
τ
→ ... Ð
→ Pn = P′ . We write P Ô⇒ Q with α ∈ Aτ if there are P′ ,Q′ s.t. P Ô⇒
with n > 0 s.t. P = P0 Ð
τ
α
P′ Ð→ Q′ Ô⇒ Q. Notice the difference between P Ô⇒ P′ and P Ô⇒ P′ , in the second case, at least one
τ -transition is executed.
Definition 7. Weak bisimilarity is the largest symmetric relation over Π, notation ≈, such that whenever
α
α
τ
P ≈ Q, (i) if P Ð→ P′ with α ∈ A then there is Q′ s.t. Q Ô⇒ Q′ and P′ ≈ Q′ and (ii) if P Ð
→ P′ then there is
Q′ s.t. Q Ô⇒ Q′ and P′ ≈ Q′ .
Like strong bisimilarity, it is only possible to prove that ≈ is a congruence for non-input contexts (see
[20, Theorem 2.4.22]).
4
Unique decomposition with respect to strong bisimilarity
In this section, we shall use the result presented in Section 2 to prove that every finite π -calculus process
has a unique parallel decomposition w.r.t. strong bisimilarity. In Section 4.1 we introduce the definition
of depth of a process and some of its properties. We also explain why we restrict our development to
finite processes. In Section 4.2 we present the unique decomposition result.
4.1 The depth of a process
Given a set X , we denote with X ∗ the set of finite sequences over X , where ε ∈ X ∗ is the empty sequence.
α1
ω
For ω = α1 α2 ⋯αn ∈ Aτ∗ with n > 0, we write P Ð→ P′ if there are processes P0 ,P1 ,... ,Pn s.t. P = P0 Ð→
α2
αn
ω
ω
P1 Ð→ ... Ð→ Pn = P′ . If ω = ε , then P Ð→ P′ implies P′ = P. If we are not interested in P′ , we write P Ð→.
α
/.
In addition, we write P ↓ if for all α ∈ Aτ , P Ð→
Definition 8. Let length ∶ A∗τ → N0 be the function defined by
⎧
0
⎪
⎪
⎪
⎪
length(ω ) = ⎨length(ω ′ ) + 1
⎪
⎪
⎪
′
⎪
⎩length(ω ) + 2
if ω = ε ,
if ω = αω ′ and α ≠ τ
if ω = αω ′ and α = τ
Unique Parallel Decomposition for the π-calculus
50
ω
Definition 9. A process P ∈ Π is normed if there is ω ∈ Aτ∗ such that P Ð→ P′ ↓. We denote with Πn the set
of normed processes. The depth ∶ Πn → N0 ∪ {∞} and the norm ∶ Πn → N0 of a normed process P ∈ Π are
defined, respectively, by
ω
depth(P) =sup({length(ω ) ∣ P Ð→ P′ and P′ ↓})
ω
norm(P) =inf({length(ω ) ∣ P Ð→ P′ and P′ ↓})
Where sup(X ) = ∞ whenever X is an infinite set, and inf(∅) = ∞.
We remark that we have assigned a higher weight to occurrences of the label τ in the definition of the
length of a sequence ω ∈ A∗τ . This is to ensure that depth is additive w.r.t. parallel composition (i.e., the
depth of a parallel composition is the sum of the depths of its components), as we shall prove in Lemma 6)
below. As opposed to other process calculi for which unique decomposition has been established (see,
e.g. [15]), due to scope extrusion, norm is not additive for the π -calculus: Consider
P = P0 ∣ P1 = ν z(az) ∣ a(x).!xa
τ
→ ν z(0 ∣!za) ↓ but P1 is not because it only performs an execution of infinite
P is normed because P Ð
length. Then, to ensure this kind of properties, one approach could be just consider normed processes.
Unfortunately this is not enough. Consider
Q = Q0 ∣ Q1 = ν z(az) ∣ a(x).xa
Processes Q, Q0 and Q1 are normed and, moreover, they perform no infinite execution. Despite this,
τ
we have that norm(Q) = 2, because Q Ð
→ ν z(0 ∣ za) ↓, and norm(Q0 ) + norm(Q1 ) = 1 + 2 = 3. Moreover,
notice that the norm of the arguments of a parallel composition is not less than the norm of the parallel
composition, i.e. norm(Q) = 2 and norm(Q1 ) = 2. In particular, these examples show that item 4 in
Lemma 3 of [7] (norm is additive) is false, and, as a consequence, some proofs in [7] are flawed. The
authors of [7] proposed a solution to this problem that we discuss in the conclusion of this paper. So, to
facilitate inductive reasoning, we will consider finite processes and depth.
Definition 10. A process P ∈ Π is finite if there is n ∈ N0 s.t. there is no ω = α1 α2 ⋯αn+1 ∈ Aτ∗ such that
ω
P Ð→. We denote with Π f the set of finite processes of Π.
Following the last example, we have that Q, Q0 ,Q1 ∈ Π f and depth(Q) = 3 = depth(Q0 ) + depth(Q1 ).
To conclude this section we present a collection of results including lemmas and theorems. Most of
the lemmas are only needed to prove the theorems. The theorems and only few lemmas will be used in
the next section. Theorem 2 states that bisimilar processes have the same depth. Theorem 3 states that
the depth of a parallel composition of two processes not bisimilar to 0 is greater than the depth of each
process. Thanks to these results, we will able to extend the notion of depth to equivalence classes and
apply inductive reasoning.
Lemma 1. For all P ∈ Π f , P ∼/ 0 implies depth(P) > 0.
α
Lemma 2. If P ∈ Π f and P Ð→ P′ , α ∈ Aτ , then P′ ∈ Π f and depth(P) > depth(P′ ).
Theorem 2. If P ∼ Q then P ∈ Π f iff Q ∈ Π f ; moreover, depth(P) = depth(Q).
M.D. Lee and B. Luttik
51
ω
ω
Proof. Suppose that P ∼ Q. Then, clearly, P Ð→ iff Q Ð→, and hence P ∈ Π f iff Q ∈ Π f .
To prove that depth(P) = depth(Q), first note that if P,Q ∈/ Π f , then depth(P) = ∞ = depth(Q). In
the case that remains, both depth(P) and depth(Q) are natural numbers; we proceed by induction over
depth(P). If depth(P) = 0 then P ∼ 0 by Lemma 1 , so Q ∼ 0 and therefore depth(Q) = 0. Suppose
now depth(P) = n > 0. Assume that the statement holds for processes with depth less than n. Suppose
α
depth(Q) = m > n then there is Q′ s.t. Q Ð→ Q′ and m = length(α ) + depth(Q′ ). By definition of ∼, we get
α
P Ð→ P′ , P′ ∼ Q′ . By Lemma 2, depth(P′ ) < depth(P), therefore depth(P′ ) = depth(Q′ ) and depth(P) ≥
depth(P′ ) + length(α ) = m > n = depth(P), i.e. we get a contradiction. Similarly, for the case depth(Q) =
α
m < n, we can reach a contradiction by considering a transition P Ð→ P′ with n = length(α ) + depth(P′ ).
Then we can conclude depth(P) = depth(Q).
α
ω′
Lemma 3. Let P,P′ ∈ Π f , ω = αω ′ ∈ A∗τ be such that P Ð→ P′ Ð→ and depth(P) = length(ω ). Then
depth(P′ ) = length(ω ′ ) and therefore depth(P) = depth(P′ ) + length(α ).
Lemma 4. For all P ∈ Π f , depth(P) ≥ depth(ν z(P)) for all z ∈ V.
ω
Lemma 5. Let P,Q ∈ Π f and ω ∈ Aτ∗ be such that P ∣ Q Ð→ and length(ω ) = depth(P ∣ Q). Then, there
ω1
ω2
are ω1 , ω2 ∈ Aτ∗ such that P Ð→, Q Ð→ and length(ω1 ) + length(ω2 ) = length(ω ).
Proof. We proceed by complete induction on n = depth(P) + depth(Q). Suppose that the property holds
for parallel compositions of finite processes such that the sum of the depths is smaller than n > 0. Let
α
ω′
ω = αω ′ and R be such that length(ω ) = n and P ∣ Q Ð→ R Ð→. We analyse the different ways of deriving
the first transition (we omit the symmetric cases).
α
• Case (Par-L). Then P Ð→ P′ and R = P′ ∣ Q. By Lemma 3 and induction depth(P′ ∣ Q) = length(ω ′ ) <
ω1
ω2
n and then there are ω1 and ω2 s.t. P′ Ð→, Q Ð→ and length(ω1 ) + length(ω2 ) = length(ω ′ ). Then
αω1
P ÐÐ→ and length(αω1 ) + length(ω2 ) = length(ω ).
xy
xy
• Case (Comm-L). Then P Ð→ P′ , Q Ð→ Q′ and R = P′ ∣ Q′ and α = τ . By Lemma 3 and induction
ω1
depth(P′ ∣ Q′ ) = length(ω ′ ) < length(τ ) + length(ω ′ ) = n and then there are ω1 and ω2 s.t. P′ Ð→,
xyω1
ω2
xyω2
Q′ Ð→ and length(ω1 ) + length(ω2 ) = length(ω ′ ). Then P ÐÐ→ and Q ÐÐ→ and length(xyω1 ) +
length(xyω2 ) = length(τ ) + length(ω ′ ) = length(ω ).
x(z)
xz
• Case (Close-L). Then P ÐÐ→ P′ ,Q Ð→ Q′ , α = τ and R = ν z(P′ ∣ Q′ ). The side condition of
x(z)xz
(Close-L) allows us to use the rules (Par-L) and its symmetric version, then P ∣ Q ÐÐÐ→ P′ ∣ Q′ .
On one hand, by Lemma 4, depth(P′ ∣ Q′ ) ≥ depth(ν z(P′ ∣ Q′ )). On the other hand, depth(P′ ∣
Q′ ) ≤ depth(ν z(P′ ∣ Q′ )) because ω = τω ′ is a maximal execution, length(τ ) = length(xzxz),
ω′
ν z(P′ ∣ Q′ ) Ð→ and by Lemma 3. Then depth(P′ ∣ Q′ ) = depth(ν z(P′ ∣ Q′ )) < n. Moreover, there
ω ′′
x(z)
xzω ′′
is ω ′′ such that P′ ∣ Q′ ÐÐ→ and length(xzxzω ′′ ) = depth(P ∣ Q). Then P ∣ Q ÐÐ→ P′ ∣ Q ÐÐÐ→ with
length(xzω ′′ ) < n. From this point we can repeat the proof of the first case.
Lemma 6. For all processes P,Q ∈ Π f , depth(P ∣ Q) = depth(P) + depth(Q).
Proof. By Lemma 5 we can ensure depth(P ∣ Q) ≤ depth(P) + depth(Q). On the other hand, by Conω
ω
ω
vention 1, we have that P Ð→ or Q Ð→ implies P ∣ Q Ð→. This allows us to conclude depth(P ∣ Q) ≥
depth(P) + depth(Q) and therefore depth(P ∣ Q) = depth(P) + depth(Q).
Unique Parallel Decomposition for the π-calculus
52
Lemma 7. For all P,Q ∈ Π, P,Q ∈ Π f iff P ∣ Q ∈ Π f .
Theorem 3. If P,Q,R ∈ Π f , P ∼/ 0, Q ∼/ 0 and P ∣ Q ∼ R then depth(P) < depth(R) and depth(Q) < depth(R).
Proof. By Lemma 1, depth(P) > 0, depth(Q) > 0. By Theorem 2, depth(R) = depth(P ∣ Q). By Lemma 6,
depth(R) = depth(P) + depth(Q) and we conclude depth(P) < depth(R) and depth(Q) < depth(R).
4.2 Unique decomposition
The commutative monoid associated with Π f modulo ∼ is defined by
• P∼ = {[P]∼ ∶ P ∈ Π f } where [P]∼ = {P′ ∶ P′ ∼ P}
• e = [0]∼ ∈ P∼ .
• ∥ ∶ P∼ × P∼ → P∼ is such that [P]∼ ∥ [Q]∼ = [P ∣ Q]∼
By Lemma 7 we have that the definition of ∥ is sound. By Theorem 2 we have that all P′ ∈ [P]∼ have the
same depth. Then we can lift function depth to P∼ , i.e. depth([P]∼ ) = depth(P).
Lemma 8. P∼ with neutral element [0]∼ and binary operation ∥ is a commutative monoid. I.e., ∥ ⊆
P∼ × P∼ satisfies the associativity, commutativity and identity properties.
In order to use the Theorem 1 we need to define on P∼ a decomposition order. In [15, 14], it is
shown that the transition relation directly induces a decomposition order on a commutative monoid of
processes. In the case of the π -calculus, however, the order induced by the transition relation cannot
be directly used, as is illustrated by the following example: Define a binary relation ↝ ⊆ P∼ × P∼ by
α
[R]∼ ↝ [S]∼ if there is R′ ∈ [R]∼ and S′ ∈ [S]∼ such that R′ Ð→ S′ . We denote the inverse of the reflexivetransitive closure of ↝ by ≼↝ , i.e., ≼↝ = (↝∗ )−1 . The order ≼↝ is not precompositional. Consider the
processes P = ν z.(az.zc.ca) and Q = a(x).x(y).yb. Then
τ
• P ∣ Q = ν z.(az.zc.ca) ∣ a(x).x(y).yb Ð
→ ν z.(zc.ca ∣ z(y).yb) = R and therefore
• [P]∼ ∥ [Q]∼ = [P ∥ Q]∼ ↝ [R]∼ = [ν z.(zc.ca ∣ z(y).yb)]∼ .
τ
• Note that R executes only one transition, i.e. ν z.(zc.ca ∣ z(y).yb) Ð
→ ν z(ca ∣ cb), then it is clear that
′
′
there are no processes P and Q s.t.
[ν z.(az.zc.ca)]∼ ↝∗ [P′ ]∼
[a(x).x(y).yb]∼ ↝∗ [Q′ ]∼
[P′ ]∼ ∥ [Q′ ]∼ = [ν z(zc.ca ∣ z(y).yb)]∼
The particularity of this example is the scope extrusion. We need to define an order based on a
fragment of the transition relation that avoids this phenomenon. We shall define the partial order ≼ over
P∼ as the reflexive-transitive closure of the relation Ð
→ ⊆ P∼ × P∼ , which is, in turn, defined as follows:
α
Ð
→0 = {([P]∼ ,[Q]∼ ) ∶ P Ð→ Q, α ∈ Aτ and ∃/ P0 ,P1 ∈ Π f s.t. P0 ∼/ 0,P1 ∼/ 0,P0 ∣ P1 ∼ P}
Ð
→k+1 = {([P0 ∣ P1 ]∼ ,[Q0 ∣ P1 ]∼ ) ∶ [P0 ]∼ Ð
→k [Q0 ]∼ ,P1 ∈ Π f }
∪ {([P0 ∣ P1 ]∼ ,[P0 ∣ Q1 ]∼ ) ∶ [P1 ]∼ Ð
→k [Q1 ]∼ ,P0 ∈ Π f }
∞
Ð
→= ⋃ Ð
→n .
k=0
The partial order ≼ is defined as the inverse of the reflexive-transitive closure of Ð
→ i.e., ≼ = (Ð
→∗ )−1 .
We write [P]∼ ≺ [Q]∼ if [P]∼ ≼ [Q]∼ and [P]∼ ≠ [Q]∼ . Notice that the definition of Ð
→ avoids any kind of
communications between the arguments of the parallel operator, this ensures that the scope extrusion is
also avoided.
M.D. Lee and B. Luttik
53
Lemma 9. If [P]∼ Ð
→ [Q]∼ then depth([Q]∼ ) < depth([P]∼ ).
Lemma 10. ≼ is a partial order.
Proof. We have to prove that ≼ is reflexive, antisymmetric, and transitive. ≼ is reflexive and transitive
because it is the reflexive-transitive closure of Ð
→. To prove that ≼ is antisymmetric notice that [P]∼ ≺ [Q]∼
implies [Q]∼ = [Pn ]∼ Ð
→ ... Ð
→ [P1 ]∼ Ð
→ [P0 ]∼ = [P]∼ for n > 0 and then, by Lemma 9, depth([P]∼ ) <
depth([Q]∼ ). Therefore [P]∼ ≼ [Q]∼ and [Q]∼ ≼ [P]∼ implies [P]∼ = [Q]∼ .
In Lemma 12, we prove that ≼ is a decomposition order. To prove this result we need to add a last
auxiliary result, Lemma 11.
→ [Q]∼ .
Lemma 11. If P ∈ Π f and depth(P) > 0 then there is Q s.t. [P]∼ Ð
Proof. We proceed by complete induction over n = depth(P). Assume that the hypothesis holds for
values less than n ≥ 1. Suppose there are no P0 ,P1 ∈ Π f such that P ∼ P0 ∣ P1 , P0 ∼/ 0 and P1 ∼/ 0. Given that
α
→0 [P′ ]∼ and therefore [P]∼ Ð
→ [P′ ]∼ . Finally, we can
n ≥ 1 then there is α ∈ Aτ s.t. P Ð→ P′ . Then [P]∼ Ð
′
define [Q]∼ = [P ]∼ .
Suppose there are P0 ,P1 ∈ Π f such that P ∼ P0 ∣ P1 , P0 ∼/ 0 and P1 ∼/ 0, then [P]∼ = [P0 ∣ P1 ]∼ . By Theorem 3 depth(P0 ) < depth(P) and by Lemma 1 depth(P0 ) > 0. By induction there is Q0 s.t. [P0 ]∼ Ð
→ [Q0 ]∼
and therefore there is k s.t. [P0 ]∼ Ð
→k [Q0 ]∼ By definitions of Ð
→k+1 and Ð
→, [P0 ∣ P1 ]∼ Ð
→k+1 [Q0 ∣ P1 ]∼ and
[P0 ∣ P1 ]∼ Ð
→ [Q0 ∣ P1 ]∼ . Therefore if we define [Q]∼ = [Q0 ∥ P1 ]∼ the proof is complete.
Lemma 12. ≼ ⊆ P∼ × P∼ is a decomposition order.
Proof.
1. ≼ is well-founded. We have to prove that every non-empty subset of P∼ has a ≼-minimal
element. Let X ⊆ P∼ with X ≠ ∅. Let [P]∼ be s.t. [P]∼ ∈ X and depth([P]∼ ) = min{depth([Q]∼ ) ∣
[Q]∼ ∈ X }, then [P]∼ is a minimal element of X by Lemma 9 and definition of ≼.
2. [0]∼ is the least element of P∼ w.r.t. ≼. We consider [P]∼ and proceed by induction on depth(P). If
depth(P) = 0, then P ∼ 0 and therefore [P]∼ = [0]∼ . Suppose that depth(P) = n > 0. By Lemma 11
there is Q s.t. [P]∼ Ð
→ [Q]∼ . By Lemma 9, depth(Q) < depth(P). By induction and definition of ≼,
[0]∼ ≼ [Q]∼ ≼ [P]∼ .
3. ≼ is strictly compatible. Suppose [Q]∼ ≺ [P]∼ and consider [P]∼ ∥ [S]∼ . By definition of ≺ there
are P0 ,...,Pn ∈ Π f , with n > 0, s.t. [P]∼ = [P0 ]∼ Ð
→ [P1 ]∼ Ð
→ ... Ð
→ [Pn ]∼ = [Q]∼ . By definition of Ð
→,
for each i = 0,... n − 1 there is ki s.t. [Pi ]∼ Ð
→ki [Qi ]∼ . Define k = max{ki ∶ i = 0,... n − 1}. Then
[P]∼ ∥ [S]∼ = [P0 ]∼ ∥ [S]∼ = [P0 ∣ S]∼ Ð
→k+1 ... Ð
→k+1 [Pn ∣ S]∼ = [Pn ]∼ ∥ [S]∼ = [Q]∼ ∥ [S]∼
By definition of Ð
→,
[P]∼ ∥ [S]∼ = [P0 ]∼ ∥ [S]∼ = [P0 ∣ S]∼ Ð
→ ... Ð
→ [Pn ∣ S]∼ = [Pn ]∼ ∥ [S]∼ = [Q]∼ ∥ [S]∼
By Lemma 9 and n > 0, depth([P]∼ ∥ [S]∼ ) > depth([Q]∼ ∥ [S]∼ ). Then [Q]∼ ∥ [S]∼ ≺ [P]∼ ∥ [S]∼ .
4. ≼ is precompositional. Suppose [P]∼ ≼ [Q]∼ ∥ [R]∼ , we have to prove there are [Q′ ]∼ ≼ [Q]∼ and
[R′ ]∼ ≼ [R]∼ s.t. [P]∼ = [Q′ ]∼ ∥ [R′ ]∼ . If Q ∼ R ∼ 0 then Q′ ∼ R′ ∼ 0 and the conditions are satisfied.
Suppose that only one of both processes is bisimilar to 0. W.l.o.g. suppose Q ∼/ 0 and R ∼ 0. In this
case, [P]∼ ≼ [Q]∼ ∥ [R]∼ = [Q]∼ , then if we define [Q′ ]∼ = [P]∼ and [R′ ]∼ = [0]∼ , the conditions
are also satisfied. Suppose now that Q ∼/ 0 and R ∼/ 0. By definition of ≼ there are n ≥ 0 and
Unique Parallel Decomposition for the π-calculus
54
processes Sn ,...,S0 s.t. [Q]∼ ∥ [R]∼ = [Q ∣ R]∼ = [Sn ]∼ Ð
→ ... Ð
→ [S0 ]∼ = [P]∼ . The proof proceed
by induction on n. Suppose that the hypothesis holds for n, we prove the case n + 1. Given that
[Sn+1 ]∼ = [Q ∣ R]∼ = [Q]∼ ∥ [R]∼ Ð
→ [Sn ]∼ , by definition of Ð
→, there is T s.t. either [Q]∼ Ð
→ [T ]∼
and [Sn ]∼ = [T ∣ R]∼ or, [R]∼ Ð
→ [T ]∼ and [Sn ]∼ = [Q ∣ T ]∼ . (We have omitted the sub-index of Ð
→
because it does not play any role.) W.l.o.g. suppose that [Q]∼ Ð
→ [T ]∼ and [Sn ]∼ = [T ∣ R]∼ . Then
[P]∼ ≼ [T ∣ R]∼ = [T ]∼ ∥ [R]∼ . By induction there are [T ′ ]∼ and [R′ ]∼ s.t. [T ′ ]∼ ≼ [T ]∼ , [R′ ]∼ ≼ [R]∼
and [P]∼ = [T ′ ]∼ ∥ [R′ ]∼ . Because [T ]∼ ≼ [Q]∼ and ≼ is a partial order, we have that [T ′ ]∼ ≼ [Q]∼
and we can conclude the proof.
5. ≼ is Archimedean. Suppose that [P]∼ ,[Q]∼ ∈ P∼ are s.t. [P]n∼ ≼ [Q]∼ for all n ∈ N0 . By Lemma 6,
depth(Pn ) = n ⋅ depth(P). Given that depth(Q) ∈ N0 we can conclude that depth(P) = 0 and therefore [P]∼ = [0]∼ .
By Theorem 1, it follows that P∼ has unique decomposition.
Corollary 1. The commutative monoid P∼ has unique decomposition.
5
Unique parallel decomposition with respect to weak bisimilarity
To prove the result of unique parallel decomposition w.r.t. strong bisimilarity, we relied on the definition
of depth and on the properties that are satisfied when we take into account strong bisimilarity. In particular, we proved that all strongly bisimilar processes have the same depth. For the weak bisimilarity we
do not have the same property. Consider the following processes
P = xy.0
P′ = τ .xy.0
P′′ = τ .τ .xy.0
Notice that P ≈ P′ ≈ P′′, despite this, depth(P) < depth(P′ ) < depth(P′′ ). To avoid this problem and
to adapt the ideas behind results in the previous section, we will consider processes without stuttering
α
transitions. A transition P Ð→ P′ is a stuttering transition if α = τ and P ≈ P′ .
We could not establish UPD for normed processes in the strong setting, because the norm of the
arguments of a parallel composition is not necessarily less than the norm of the parallel composition. In
the weak setting, it is known that normed processes of Π do not satisfy UPD w.r.t. bisimilarity. Consider
τ
the following counter example [7]: define P = ν z(zc ∣ z(x).!ab ∣ z(y)). P is normed because P Ð
→ ν z(0 ∣
z(x).!ab ∣ 0) ↓ but there is no a unique parallel decomposition of P because P ≈ P ∣ P.
We study processes without stuttering transitions in Section 5.1. Using the results developed in
that section and Theorem 1, in Section 5.2 we prove that for finite processes there is a unique parallel
decomposition w.r.t. weak bisimilarity.
5.1 Processes without stuttering steps
ω
a1
For ω = a1 a2 ⋯an ∈ Aτ∗ with n > 0, we write P Ô⇒ P′ if there are processes P0 ,P1 ,... ,Pn s.t. P = P0 Ô⇒
a2
an
ω
P1 Ô⇒ ... Ô⇒ Pn = P′ . If ω = ε , then P Ô⇒ P′ implies P Ô⇒ P′ .
Definition 11. A process P ∈ Π f is a process without stuttering transitions if there are no ω ∈ A∗ and
ω
τ
→ P′′ and P′ ≈ P′′ . We denote with Π↻
P′ ,P′′ ∈ Π f s.t. P Ô⇒ P′ Ð
/ the set of processes of Π f without
stuttering transitions.
M.D. Lee and B. Luttik
55
In Section 4.1 we discussed why we do not consider infinite processes, this discussion also applies
for weak bisimilarity. By definition, Π↻
/ ⊆ Π f . This fact and Lemma 14 ensure that we can use processes
in Π↻
/ to define properties over equivalence classes of processes in Π f w.r.t. weak bisimilarity.
To prove Lemma 14 we need to introduce some notation and Lemma 13. We write x(z).P to denote
ν z.xz.P. We call x(z) a bound-output prefix. We use λ to range over prefixes, including bound-outputs.
Lemma 13. For all P ∈ Π f there are prefixes λ1 ,..., λn and processes P1 ,... ,Pn such that P ∼ ∑ni=1 λi .Pi .
Proof. The proof proceeds by structural induction on P. Cases P = 0, P = π .P′ are straightforward by
definition. For the case P = P1 + P2, by induction hypothesis there are processes Q1 = ∑i∈I λi .Pi and Q2 =
∑ j∈J λ j .Pj and bisimulations R1 and R2 s.t. Pk Rk Qk for k = 1,2. It is easy to see that {(P,Q1 + Q2 )}∪ R1
∪ R2 is a bisimulation. Case P = P1 ∣ P2 is straightforward by induction and the Expansion Lemma for ∼
in the π -calculus [20, Lemma 2.2.14]. Thanks this lemma we can state that for all P = ∑i∈I λi Pi and
Q = ∑ j∈J λ j Q j there is R = ∑k∈K λk Rk s.t. P ∣ Q ∼ R. Case P = ν z.P′ proceeds by structural induction on
P′ . If P′ = 0 then ν z.P′ ≈ 0 = ∑0i=1 λi .Pi . If P′ = π .P′′ then there are three cases to analyse: (i) z ∈/ n(π ) then
P ∼ π .(ν z.P′′ ), (ii) π = xz then ν z.P′ can be denoted by x(z)P′′ , (iii) z ∈ n(π ) and π ≠ xz then ν z.π .P′′ ∼ 0.
If P′ = P0 + P1 then notice that ν z(P0 + P1 ) ∼ ν z.P0 + ν z.P1 ; by induction there are processes ∑i∈I λi Pi and
∑ j∈J λ j Pj s.t. ν z.P0 ∼ ∑i∈I λi Pi and ν z.P1 ∼ ∑ j∈J λ j Pj . From this point, the proof follows as in the case
P = P1 + P2 . Finally, case P′ = P0 ∣ P1 can be reduced to the previous case using the Expansion Lemma
for ∼ ([20, Lemma 2.2.14]).
Lemma 14. For every process P ∈ Π f there is Q ∈ Π↻
/ s.t. P ≈ Q.
Proof. The proof of the result follows by complete induction on n = depth(P). By Lemma 13 for P ∈ Π f
there are prefixes λ1 ,... , λn and processes P1 ,...,Pn such that P ∼ ∑ni=1 λi .Pi . By induction and Lemma 2,
for each i there is Qi ∈ Π↻
/ s.t. Pi ≈ Qi . Then if we define
Q = ∑i∈{1,...n} and P/≈Pi λi .Qi
we get Q s.t. Q ∈ Π↻
/ and P ≈ Q.
We cannot restrict our attention only to processes in Π↻
/ because the property of not executing
stuttering transitions is not preserved by parallel composition. Consider the processes P0 = ν z(az) and
P1 = a(x).(xb + τ .cb). Both P0 ,P1 ∈ Π↻
/ Π↻
/ but P0 ∥ P1 ∈
/ because
τ
→ ν z(0 ∥ (zb + τ .cb)) ∼ τ .cb ≈ cb
P0 ∥ P1 = ν z(az) ∥ a(x).(xb + τ .cb) Ð
If we compare this fact with the strong setting, we can say that it is not possible to prove a lemma similar
to Lemma 7 for processes in Π↻
/ .
As in Section 4.1, we conclude with a collection of theorems and lemmas. Theorems 4 and 5 are
equivalent, respectively, to Theorems 2 and 3 but w.r.t. processes in Π↻
/ and weak bisimilarity. Most of
the lemmas are needed to prove these results and only a few of them are used in the next section.
ω
∗
Lemma 15. If P ∈ Π↻
/ and P Ô⇒ Q for ω ∈ Aτ then Q ∈ Π↻
/ .
ω′
τ
′
∗
′
′′
′
→ Q′′ with Q′ ≈ Q′′ . Let
Proof. Suppose Q /∈ Π↻
/ , then there are ω ∈ A and Q ,Q ∈ Π f s.t. Q Ô⇒ Q Ð
ω̃ω ′
τ
ω̃ obtained from ω by removing τ ’s actions. Then P Ô⇒ Q′ Ð
→ Q′′ and therefore P ∈/ Π↻
/ , which is a
contradiction.
Unique Parallel Decomposition for the π-calculus
56
α
α
Lemma 16. If P,Q ∈ Π↻
→ P′ with α ∈ Aτ , then Q Ô⇒ Q′ , i.e. Q executes at least
/ are s.t. P ≈ Q and P Ð
a transition, and P′ ≈ Q′
τ
Proof. If α ≠ τ the result is straightforward by Def. 7. If α = τ and there is no transition Q0 Ð
→ Q1 s.t.
τ
α
′
′
′
′
′
Q Ô⇒ Q0 Ð
→ Q1 Ô⇒ Q and P ≈ Q then P ≈ Q since P ≈ Q. This implies that P ≈ Q ≈ P , i.e. P Ð→ P′
is a stuttering transition. This contradicts P ∈ Π↻
/ .
Theorem 4. If P,Q ∈ Π↻
/ and P ≈ Q then depth(P) = depth(Q).
Proof. We proceed by complete induction over n = depth(P). If depth(P) = 0, then P ∼ 0 and moreover
α
P ≈ 0. Because P ≈ Q and P ∼ 0, there is no α ∈ A s.t. Q Ô⇒. Taking into account this fact, if there
τ
τ
→ Q′ , Q′ is such that Q′ ≈ 0. This creates a contradiction because Q Ð
→ Q′ is a stuttering
is Q′ s.t. Q Ð
transition and Q ∈ Π↻
/ . Then Q ∼ 0 and therefore depth(Q) = 0 = depth(P). Suppose depth(P) = n + 1.
α
ω′
Let ω = αω ′ ∈ A∗τ and P′ be s.t. length(ω ) = n + 1 and P Ð→ P′ Ð→. Because P ≈ Q and Lemma 16
α
there are Q0 , Q1 , Q′ s.t. Q Ô⇒ Q0 Ð→ Q1 Ô⇒ Q′ and P′ ≈ Q′ . By induction depth(P′ ) = depth(Q′ ) and
therefore depth(Q) ≥ depth(P) = n + 1. We prove now that when we assume depth(Q) > n + 1 we reach
a contradiction; it then follows that depth(Q) = n + 1. Assume depth(Q) > n + 1 and let ω = αω ′ ∈ A∗τ
α
ω′
be such that Q Ð→ Q̃ Ð→ Q̃′ ↓ and length(ω ) = depth(Q). Because P ≈ Q and Lemma 16 there is P̃ s.t.
α
P Ô⇒ P̃, P̃ ≈ Q̃ and depth(P) = n+ 1 > depth(P̃). By the complete induction depth(P̃) = depth(Q̃). Then,
we reach a contradiction, n + 1 > depth(P̃) = depth(Q̃) ≥ n + 1.
α
Lemma 17. If P ∈ Π f and P Ð→ P′ with α ≠ τ then P ≈/ P′ .
ω
αω
Proof. Let ω ∈ A∗ be the largest sequence s.t. P′ Ô⇒ P′′ ↓. Then there is no Q s.t. P′ Ô⇒ Q. On the
αω
other hand P Ô⇒, therefore P ≈/ P′ .
α
α′
Lemma 18. If P ∈ Π↻
/ 0 and for all P′ ∈ Π f , α ∈ Aτ , P Ð→ P′ ↓, then there is α ′ ≠ τ s.t. P Ð→.
/ is s.t. P ∼
α
α
Proof. Because P ∼/ 0 there is α s.t. P Ð→. If for all P′ ∈ Π f , α ∈ Aτ , P Ð→ P′ ↓ and α = τ then P ≈ 0 and
therefore all transitions that can be executed by P are stuttering transitions. This contradicts P ∈ Π↻
/ .
Theorem 5. If P,Q,R ∈ Π↻
/ 0, Q ≈/ 0 and P ∥ Q ≈ R then depth(P) < depth(R) and depth(Q) <
/ , P ≈
depth(R).
Proof. We prove depth(P) < depth(R), the proof that depth(Q) < depth(R) is analogous. Note that, since
ω
Q ≈/ 0, there is Q′ with depth(Q′ ) = 1 that is reachable from Q, that is, there exists ω ∈ A∗ s.t. Q Ô⇒ Q′ and
α
Q′ ∼/ 0 (we remark the symbol ∼/ ) and for all Q′′ ∈ Π f , α ∈ Aτ , Q′ Ð→ Q′′ ↓. Then, by Lemma 15, Q′ ∈ Π↻
/
α
and, by Lemma 18, Q′ Ð→ with α ≠ τ . Furthermore, by Convention 1 and the symmetric version of rule
ω
α
(Par-L) we have that P ∥ Q Ô⇒ P ∥ Q′ . By Lemma 17, P ∥ Q′ Ð→ P ∥ 0 and α ≠ τ imply P ∥ Q′ ≈/ P ∥ 0 ≈ P.
′
Because P ∈ Π↻
/ , whenever S ∈ Π↻
/ and S ≈ P ∥ Q , depth(S) ≥ 1 + depth(P) (⋆). Given that R ≈ P ∥ Q,
ω
ω
P ∥ Q Ô⇒ P ∥ Q′ implies there is R′ s.t. R Ô⇒ R′ and R′ ≈ P ∥ Q′ . By Lemma 15, R′ ∈ Π↻
/ . In addition,
′
′
by (⋆), depth(R ) ≥ 1 + depth(P). Finally depth(R) ≥ depth(R ) ≥ 1 + depth(P) > depth(P).
α
α
′
′
′
Lemma 19. If P,Q ∈ Π↻
/ , P ≈ Q and P Ô⇒ P , with α ∈ Aτ , then there is Q s.t. Q Ô⇒ Q .
M.D. Lee and B. Luttik
57
5.2 Unique parallel decomposition
The development in this section is similar to the development in Section 4.2, for this reason in some
cases we use the same notation. This will not be a problem because both developments are independent.
In order to use Theorem 1 we need to define a commutative monoid with a decomposition order. The
commutative monoid is defined by
• P≈ = {[P]≈ ∶ P ∈ Π f } where [P]≈ = {P′ ∶ P′ ≈ P}
• e = [0]≈ ∈ P≈ .
• ∥ ⊆ P≈ × P≈ is s.t. [P]≈ ∥ [Q]≈ = [P ∣ Q]≈
Notice that we cannot ensure that for all P′ ,P′′ ∈ [P]≈ , depth(P′ ) = depth(P′′ ). Then, we extend the
/
′
notion of depth in the following way. Define [P]↻
/ . For [P]≈ ∈ P≈ , depth([P]≈ ) = depth(P )
≈ = [P]≈ ∩ Π↻
↻
/
′
with P ∈ [P]≈ . This definition is sound because of Lemma 14 and Theorem 4.
Lemma 20. P≈ with neutral element [0]≈ and binary operation ∥ is a commutative monoid. I.e., ∥ ⊆
P≈ × P≈ satisfies the associativity, commutativity and identity properties.
We shall define the partial order ≼ over P≈ using the relation Ð
→ ⊆ P≈ × P≈ defined as follows:
α
/
↻
/
′
′
′
Ð
→0 = {([P]≈ ,[Q]≈ ) ∶ ∃P′ ∈ [P]↻
≈ ,Q ∈ [Q]≈ ∶ P Ô⇒ Q , α ∈ Aτ
and ∃/ P0 ,P1 ∈ Π↻
/ 0,P1 ≈/ 0,P0 ∣ P1 ≈ P}
/ s.t. P0 ≈
Ð
→k+1 = {([P0 ∣ P1 ]≈ ,[Q0 ∣ P1 ]≈ ) ∶ [P0 ]≈ Ð
→k [Q0 ]≈ ,P1 ∈ Π f }
∪ {([P0 ∣ P1 ]≈ ,[P0 ∣ Q1 ]≈ ) ∶ [P1 ]≈ Ð
→k [Q1 ]≈ ,P0 ∈ Π f }
∞
Ð
→= ⋃ Ð
→k
k=0
The partial order ≼ is defined as the inverse of the reflexive-transitive closure of Ð
→ i.e., ≼ = (Ð
→∗ )−1 . We
write [P]≈ ≺ [Q]≈ if [P]≈ ≼ [Q]≈ and [P]≈ ≠ [Q]≈ .
Notice that in this case Ð
→ takes into account processes in Π↻
/ and weak transitions that execute at
least one transition. Also notice that we are avoiding communications between the arguments of the
parallel composition in order to avoid scope extrusion.
Similarly to the strong setting, we need two lemmas, Lemmas 21 and 22, to prove that ≼ is a partial
order (Lemma 23). In addition, to prove that ≼ is a decomposition order, we need the Lemma 24 that is
equivalent to Lemma 11. The proofs of these results follow similarly to their respective counterpart in
the strong setting. (The complete proofs are in the appendix.)
α
Lemma 21. If [P]≈ Ð
→ [Q]≈ then for all P̃ ∈ [P]≈ there are α ∈ Aτ and Q̃ ∈ [Q]≈ s.t. P̃ Ô⇒ Q̃.
Lemma 22. If [P]≈ Ð
→ [Q]≈ then depth([P]≈ ) > depth([Q]≈ ).
Lemma 23. ≼ is a partial order.
Lemma 24. If P ∈ Π↻
→ [Q]≈ .
/ and depth(P) > 0 then there is Q s.t. [P]≈ Ð
We are ready to prove that ≼⊆ P≈ × P≈ is a decomposition order. This proof does not present changes
w.r.t. proof of Lemma 12 except that for proving ≼ is Archimedean, we use Theorem 5. Notice that there
is no lemma equivalent to Lemma 6 in the weak setting.
Lemma 25. ≼ ⊆ P≈ × P≈ is a decomposition order.
By Theorem 1, it follows that P≈ has unique decomposition.
Corollary 2. The commutative monoid P≈ has unique decomposition.
Unique Parallel Decomposition for the π-calculus
58
6
Final Remarks
In this paper we have proved that finite processes of the π -calculus satisfy UPD w.r.t. both strong bisimilarity and weak bisimilarity. We have obtained these results using the technique presented in [15] (see
Theorem 1) and different properties that are satisfied in each setting. For the strong setting, we had to
prove properties related to the depth of processes. For the weak setting, we had to prove properties related to processes that execute no stuttering transitions. Our results show that the abstract framework of
[15] can be used in the context of the π -calculus, dealing with the complications that arise from scope
extrusion. In addition, the same framework can be used to deal with the weak setting if one considers
processes without stuttering transitions. In this way, we have avoided the abstract technique introduced
in [14] which is considerably more involved than the technique that we have used in this paper.
In Section 4 we showed with two examples that norm is not additive for the π -calculus and therefore
some proofs in [7] are flawed. After pointing out this problem to Dreier et al., they proposed us an
alternative definition of norm to solve it. Call this variant norm′ . Roughly, norm′ should not consider
traces where there is a scope extrusion of processes. We think this solution may work for the applied
π -calculus, but is not suitable for the variant of the π -calculus considered in the present paper. We
first explain what is the problem in our context, and then why this problem is not present in applied π .
In the first example in Section 4, we had P = P0 ∣ P1 = ν z(az) ∣ a(x).!xa. Process P is not normed, i.e.
τ
→ ν z(0 ∣!za) ↓, goes through a
norm′ (P) = ∞, because the only finite trace that the process executes, P Ð
scope extrusion. Now, consider the process
P′ = ν z(az).(0 ∣ a(x).!xa) + a(x).(ν z(az) ∣!xa) + τ .(ν z(0 ∣!za)) ;
it would be normed according to the alternative definition suggested above (the τ -transition from P′ is not
the result of a scope extrusion). Now, since P′ is just the expansion of P, it is clear that P ∼ P′ . Thus, we
find that the property of being normed is not compatible with bisimilarity. Since the applied π -calculus
does not include the construct for non-deterministic choice needed for the expansion, this problem is not
present there.
An open question that leaves this paper is related with UPD of the π -calculus w.r.t. strong full
bisimilarity[20]. Strong full bisimilarity is a stronger notion of bisimulation that is a congruence for all
constructs of the π -calculus. We have tried to apply the abstract technique in this setting so far without
success. When we tried to repeat the result in Section 4, taking into account the universal quantification
in the definition of strong full bisimilarity, a problem arose when we wanted to prove that the order
is a decomposition order. Particularly, we were not able to prove that the order is strict compatible.
Notice that this problem is not present in the asynchronous π -calculus[20], a well-known fragment of
the π -calculus, because (strong) bisimilarity and (strong) full bisimilarity coincide.
Acknowledgement. The authors thank Daniel Hirschkoff for discussions, comments and suggestions on
various drafts of this paper, and anonymous reviewers for their thorough reviews and good suggestions.
References
[1] L. Aceto, W. Fokkink, A. Ingólfsdóttir & B. Luttik (2005): CCS with Hennessy’s merge has no finiteequational axiomatization. Theor. Comput. Sci. 330(3), pp. 377–405, doi:10.1016/j.tcs.2004.10.003.
[2] L. Aceto, W. Fokkink, A. Ingólfsdóttir & B.. Luttik (2009): A finite equational base for CCS with left merge
and communication merge. ACM Trans. Comput. Log. 10(1), doi:10.1145/1459010.1459016.
M.D. Lee and B. Luttik
59
[3] L. Aceto, A. Ingólfsdóttir, B. Luttik & P. van Tilburg (2008): Finite Equational Bases for Fragments of CCS
with Restriction and Relabelling. In: IFIP 20th World Computer Congress, TC 1, Foundations of Computer
Science, pp. 317–332, doi:10.1007/978-0-387-09680-3 22.
[4] I. Castellani (1998): Bisimulations for Concurrency. Ph.D. thesis, University of Edinburgh. Also published
as LFCS-88-51.
[5] S. Christensen (1993): Decidability and Decomposition in Process Algebra. Ph.D. thesis, University of
Edinburgh.
[6] F. Corradini, R. Gorrieri & D. Marchignoli (1998): Towards parallelization of concurrent systems. Informatique théorique et applications 32(4-6), pp. 99–125.
[7] J. Dreier, C. Ene, P. Lafourcade & Y. Lakhnech (2016): On the existence and decidability of unique
Theor. Comput. Sci. 612, pp. 102–125,
decompositions of processes in the applied π -calculus.
doi:10.1016/j.tcs.2015.11.033.
[8] J. Dreier, P. Lafourcade & Y. Lakhnech (2012): Defining Privacy for Weighted Votes, Single and Multi-voter
Coercion. In: ESORICS 2012, pp. 451–468, doi:10.1007/978-3-642-33167-1 26.
[9] W. Fokkink & B. Luttik (2000): An omega-Complete Equational Specification of Interleaving. In:
Automata, Languages and Programming, 27th International Colloquium, ICALP 2000, pp. 729–743,
doi:10.1007/3-540-45022-X 61.
[10] J. Friso Groote & F. Moller (1992): Verification of Parallel Systems via Decomposition. In: CONCUR ’92,
Third International Conference on Concurrency Theory, pp. 62–76, doi:10.1007/BFb0084783.
[11] D. Hirschkoff & D. Pous (2008): A Distribution Law for CCS and a New Congruence Result for the π calculus. Logical Methods in Computer Science 4(2), doi:10.2168/LMCS-4(2:4)2008.
[12] Y. Hirshfeld & M. Jerrum (1999): Bisimulation Equivalence Is Decidable for Normed Process Algebra.
In: Automata, Languages and Programming, 26th International Colloquium, ICALP’99, pp. 412–421,
doi:10.1007/3-540-48523-6 38.
[13] I. Lanese, J. A. Pérez, D. Sangiorgi & A. Schmitt (2011): On the expressiveness and decidability of higherorder process calculi. Inf. Comput. 209(2), pp. 198–226, doi:10.1016/j.ic.2010.10.001.
[14] B. Luttik (2016): Unique parallel decomposition in branching and weak bisimulation semantics. Theor.
Comput. Sci. 612, pp. 29–44, doi:10.1016/j.tcs.2015.10.013.
[15] B. Luttik & V. van Oostrom (2005): Decomposition orders another generalisation of the fundamental theorem of arithmetic. Theor. Comput. Sci. 335(2-3), pp. 147–186, doi:10.1016/j.tcs.2004.11.019.
[16] R. Milner & F. Moller (1993): Unique Decomposition of Processes. Theor. Comput. Sci. 107(2), pp. 357–
363, doi:10.1016/0304-3975(93)90176-T.
[17] F. Moller (1989): Axioms for Concurrency. Ph.D. thesis, University of Edinburgh.
[18] F. Moller (1990): The Importance of the Left Merge Operator in Process Algebras. In: Automata, Languages
and Programming, 17th International Colloquium, ICALP90, pp. 752–764, doi:10.1007/BFb0032072.
[19] F. Moller (1990): The Nonexistence of Finite Axiomatisations for CCS Congruences. In: Proceedings of
LICS ’90, pp. 142–153, doi:10.1109/LICS.1990.113741.
[20] D. Sangiorgi & D. Walker (2001): π -Calculus: A Theory of Mobile Processes. Cambridge University Press,
New York, NY, USA.
| 6cs.PL
|
arXiv:1801.08207v1 [math.AC] 24 Jan 2018
MODULES OF INFINITE REGULARITY OVER COMMUTATIVE
GRADED RINGS
LUIGI FERRARO
Abstract. In this work, we prove that if a graded, commutative algebra
R over a field k is not Koszul then, denoting by m the maximal homogeneous ideal, the nonzero modules of the form mM have infinite CastelnuovoMumford regularity. We also prove that over complete intersections which are
not Koszul, a nonzero direct summand of a syzygy of k has infinite regularity. Finally, we relate the vanishing of the graded deviations of R to having a
nonzero direct summand of a syzygy of k of finite regularity.
Introduction
Let k[x1 , . . . , xn ] be a polynomial ring, graded by setting deg xi = di ≥ 1, and let
R = k[x1 , . . . , xn ]/I where I is a homogeneous ideal. We denote by m the maximal
homogeneous ideal of R. The size of a minimal free resolution of a graded Rmodule M is measured by its graded Betti numbers βi,j (M ) = rankk Exti,j
R (M, k).
Invariants arising from the Betti numbers of M are its projective dimension and
its Castelnuovo-Mumford regularity. In [4], Avramov and Eisenbud prove that the
regularity of k is finite if and only if the regularity of every finitely generated module
is finite. In [3], Avramov proves that nonzero modules of the form mM have infinite
projective dimension, provided that R is not regular. In this article, we prove that
nonzero modules of the form mM also have infinite regularity, provided that R is
not Koszul.
In section 2, we prove that a nonzero direct summand of a syzygy of k has infinite
regularity, provided that R is a complete intersection that is not Koszul. We ask
whether or not this holds true for any non-Koszul ring.
In section 3, we provide a connection between having a nonzero direct summand
of a syzygy of k of finite regularity and the vanishing of the graded deviations of
R.
1. Tightly embeddable modules
In this section we construct a class of modules, that have infinite regularity over
non-Koszul rings.
Definition. Let M be a graded R-module. We say that M is tightly embeddable if
there exists a finitely generated graded R-module L such that
mM & mL ⊆ M ⊆ L.
In this case, M ⊆ L is a tight embedding.
1.1. It follows from Nakayama’s Lemma that if L is finitely generated, then mi L ⊆
mi−1 L is a tight embedding for each i ≥ 1 such that mi L is not zero.
1
2
LUIGI FERRARO
Motivated by the results in [3], we explore the relation between tightly embeddable modules and regularity.
We establish the convention that if V is a graded k-vector space then V j = V−j :
1.2. The Hilbert series of a graded k-vector space V is
X
rankk Vj sj .
HV (s) =
j
1.3. The bigraded Poincaré series of a finitely generated graded R-module M is
X
R
R
βi,j
(M )sj ti ∈ Z[s±1 ][[t]].
PM
(s, t) =
1.4. Let
X
i
ai ti and
then we write
X
i,j
bi ti be formal power series in Z[[t]]. If ai ≤ bi for every i,
i
X
a i ti
X
b i ti
X
i
i
or
X
b i ti
a i ti .
i
i
1.5. If A is a graded k-algebra and X is a graded A-module, then the sth shift
of X is the graded A-module (Σs X)i = X i−s . Here, the A-action is defined as
a · x = (−1)|a|s ax, where |a| denotes the degree of a.
Remark 1.6. In accordance with a bigraded version of Gulliksen and Levin [6], the
algebra ExtR (k, k) is a (bi-)graded Hopf k-algebra. This algebra is the universal
enveloping algebra of a (bi-)graded Lie algebra π ∗,∗ (R) where the first degree is
homological and the second is the internal degree. This follows from bigraded
versions of the theorems in [8, (5.18)] (characteristic 0), [1, Theorem 17] (positive
odd characteristic) [9, Theorem 2] (characteristic 2).
Notation 1.7. We denote by E the algebra ExtR (k, k). For every R-module M , we
denote by E(M ) the left E-module ExtR (M, k).
The following theorem is a graded version of [3, Lemma 6] and has a similar
proof.
Theorem 1.8. If M ⊆ L is a tight embedding, then, with V =
coefficient-wise inequality
n
Y
R
HV (s)PkR (s, t) PM
(s, t) (1 + sdi t).
i=1
Proof. We set the following notation
M = M/mM,
L = L/mL,
N = L/M, W = L/mM.
Consider the following commutative diagram
/L
/N
/M
(1.8.1)
0
0
/M
/W
/N
/0
/0
mL
mM ,
there exists a
MODULES OF INFINITE REGULARITY OVER COMMUTATIVE GRADED RINGS
3
This induces a commutative diagram of homomorphisms of bigraded left E-modules.
(1.8.2)
E(L)
O
/ E(M )
O
ð
/ ΣE(N )
E(W )
/ E(M )
δ
/ ΣE(N ).
∨
As m annihilates M and N , there are natural isomorphisms E(M ) ∼
= E ⊗k M and
E(N ) ∼
= E ⊗k N ∨ . Let K denote the universal enveloping algebra of π ≥2,∗ (R).
Consider the following commutative diagram
E(M )
O
M
∨
δ
/ ΣE(N )
O
δ0
/ E 1 ⊗k N ∨
where the vertical maps are natural injective maps. Fix the k-basis
∨
ν1 , . . . , νn of N ∨ ,
µ1 , . . . , µm of M ,
ξ1 , . . . , ξe of E 1 .
The map δ is E-linear and so is also K-linear. Hence, if ε ∈ E
δ(ε ⊗ µh ) = εδ 0 (µh ) = Σi,j ai,j,h εξj ⊗ νi
where
δ 0 (µh ) = Σi,j ai,j,h ξj ⊗ νi
with ai,j,h ∈ k.
By the Poincaré-Birkhoff-Witt theorem, E is a free K-module with basis
(1.8.3)
{ξj1 · · · ξjk }j1 <···<jk .
Hence, E(N ) is a free K-module with basis
(1.8.4)
{ξj1 · · · ξjk ⊗ νi }j1 <···<jk ,i .
If κ ∈ K, then
δ(κ ⊗ µh ) = Σi,j ai,j,h κξj ⊗ νi .
This means that as a K-module Im δ|K⊗k M ∨ is generated by the elements δ 0 (µh ).
We can change the basis of M
∨
such that the coordinate vectors of
δ 0 (µ1 ), . . . , δ 0 (µm′ ),
with respect to the basis (1.8.4), are linearly independent over k and
δ 0 (µm′ +1 ), . . . , δ 0 (µm )
are all zero. Because the elements ξj ⊗ νi are part of a K-basis of E(N ) we deduce
that δ 0 (µ1 ), . . . , δ 0 (µm′ ) are linearly independent over K.
As a result
Im δ|K⊗k M ∨ = ΣK ⊗k Im δ 0 .
This means that E(N ) contains a copy of K ⊗k Im δ 0 , and based on the commutativity of the diagram (1.8.2), this gives
HE(M) (s, t) HK⊗k Im δ0 (s, t) = HK (s, t)HIm δ0 (s).
4
LUIGI FERRARO
According to (1.8.3), there is an equality of formal power series
HK (s, t) =
HE (s, t)
.
HV E 1 (s, t)
Because E 1 ∼
= (m/m2 )∨ and HX ∨ (t) = HX (t−1 ) for every R-module X, we deduce
Y
(1 + s−di t).
HV E 1 (s, t) =
i
Now, consider the following chain of equalities of Hilbert series
HIm δ0 (s) = HM ∨ (s) − HW ∨ (s) + HN ∨ (s)
= HM (s−1 ) − HW ∨ (s) + HN (s−1 )
= HW (s−1 ) − HW ∨ (s)
= HV (s−1 ) + HL (s−1 ) − HW ∨ (s)
= HV (s−1 ).
The first equality follows from the exact sequence in cohomology induced by the
first row in (1.8.1), i.e.,
0 → N ∨ → (W )∨ → M ∨ → Im δ 0 → 0,
the second equality follows from the already mentioned fact that HX ∨ (t) = HX (t−1 )
for every R-module X. The third equality follows from the second row in (1.8.1),
the fourth equality follows from
0 → V → W → L → 0,
and the last equality follows from HW ∨ (s) = HL (s−1 ).
Holistically we obtain
HE(M) (s, t) HK⊗k Im δ0 (s, t)
= HK (s, t)HIm δ0 (s, t)
HE (s, t)
H (s−1 ).
−di t) V
(1
+
s
i=1
= Qn
Now, we observe that
R −1
HE(M) (s, t) = PM
(s , t) and HE (s, t) = PkR (s−1 , t).
Hence, we obtain
P R (s−1 , t)
R −1
PM
(s , t) Qn k
H (s−1 ),
−di t) V
i=1 (1 + s
and replacing s−1 with s, we obtain the desired inequality.
Let M be a finitely generated graded R-module. The Castelnuovo-Mumford
regularity of M is
reg M = sup{j − i | βi,j (M ) 6= 0}.
Definition. A ring R is Koszul if the algebra ExtR (k, k) is generated by Ext1,∗
R (k, k).
According to [5], if reg k < ∞, then reg k = Σi (di − 1); this result can happen if
and only if R is Koszul.
MODULES OF INFINITE REGULARITY OVER COMMUTATIVE GRADED RINGS
5
Corollary 1.9. If M is a tightly embeddable module and reg M < ∞, then R is
Koszul.
R
Proof. Set r := reg M . If reg M < ∞, then βi,j
(M ) = 0 for j − i > r. If
Y
R
Σi,j ai,j sj ti = PM
(s, t) (1 + sdi t),
i
then ai,j = 0 for j − i > r′ . By Theorem 1.8, if
Σi,j ci,j sj ti = HV (s)PkR (s, t),
R
then ci,j = 0 if j − i > r′ for some r′ . Because mL/mM 6= 0 we deduce βi,j
(k) = 0
′′
′′
if j − i > r for some r , which is equivalent to reg k < ∞.
2. Direct summands of syzygies of the residue field
In this section, we prove that nonzero direct summands of syzygies of k have
infinite regularity if R is a complete intersection that is not generated by quadrics.
Special homological properties of this class of modules were already noticed in
[3, 7, 10].
Theorem 2.1. If R is a complete intersection, M = Ωm k, β : M → N is an Rmodule homomorphism of finitely generated graded R-modules such that for some
n, the map β n : ExtnR (N, k) → ExtnR (M, k) is not zero, then for some b ∈ Z, there
is a coefficient-wise inequality
n
Y
sb PkR (s, t) PΩRn N (s, t) (1 + sdi t).
i=1
Proof. We set the following notation
N ′ = Ωn N,
′
N = N ′ /mN ′ ,
M ′ = Ωm+n k.,
Let π be the canonical projection N ′ → N ′ /mN ′ and β ′ be a morphism M ′ → N ′
obtained by extending β to a morphism of free resolutions.
Let ξ denote the composed map
∗
′∗
∗
′
′ π
β
α
ξ : E ⊗k (N )∨ ∼
= E(N ) −→ E(N ′ ) −−→ E(M ′ ) −−→ Σm+n E ≥m+n,∗ ,
where α∗ is an iterated connecting homomorphism, and hence is an isomorphism.
By construction, ξ 0 6= 0.
V
As a result of the Poincaré-Birkhoff-Witt Theorem, E ∼
= K⊗k E 1 as (bi-)graded
left K-modules, if R is a complete intersection, then π i,∗ (R) = 0 if i ≥ 3 (see for
example, [2, Theorem 7.3.3 and Theorem 10.1.2] and the references given there).
Because π i,∗ (R) = 0 if i ≥ 3, K is a polynomial ring. Hence, Im ξ is a torsion-free
K-module (because E ≥m+n is a submodule of a free module over the polynomial
ring K). As a result, any nonzero element in the image of ξ 0 generates a copy of K
in Im ξ, whose internal degree might be shifted by some b ∈ Z. It follows that
HE (s, t)
.
−di t)
i=1 (1 + s
HE(N ′ ) (s, t) sb HK (s, t) = sb Qn
Now, we conclude as in Theorem 1.8.
Corollary 2.2. Let R be a complete intersection with reg k = ∞. For any n ≥ 0,
each nonzero direct summand of Ωn k has infinite regularity.
6
LUIGI FERRARO
We raise the following question:
Question 2.3. If β : Ωm k → N is an R-module homomorphism of finitely generated
graded R-modules such that for some n the map
β n : ExtnR (N, k) → ExtnR (Ωm k, k)
is not zero and N has finite regularity, then is R Koszul?
3. On the vanishing of the graded deviations
In this section, we relate Question 2.3 and the vanishing of the graded deviations
of R. Let RhXi be an acyclic closure of k (see [2, 6.3] for the definition). Over
graded rings, we require the differential of the acyclic closure to be a homogeneous
map. For the differential to be homogeneous, we must give the elements of X an
internal grading, making X a bigraded set.
Definition. The (i, j)th graded deviation of R is
εi,j (R) := Card(Xi,j ),
where Xi,j is the set of variables in the acyclic closure of homological degree i and
internal degree j.
Theorem 3.1. If M = Ωm k and β : M → N is a homomorphism of finitely
generated graded R-modules such that for some n the map
β n : ExtnR (N, k) → ExtnR (M, k)
is not zero and reg N < ∞, then εi,j (R) = 0 for i > m + n and i 6= j.
Proof. We set the following notation
′
N ′ = Ωn N,
N = N ′ /mN ′ ,
M ′ = Ωm+n k.
Let π be the canonical projection N ′ → N ′ /mN ′ and β ′ be a morphism M ′ → N ′
obtained by extending β to a morphism of free resolutions. Let V be the universal
enveloping algebra of π >m+n,∗ (R). Let ξ denote the composed map
β ′∗
π∗
′
α∗
ξ : E(N ) −→ E(N ′ ) −−→ E(M ′ ) −−→ Σm+n E ≥m+n,∗ .
Consider the following commutative diagram
ξ
′
E(N )
O
/ Σm+n E ≥m+n,∗
O
ξ0
′
(N )∨
/ E m+n,∗
By the Poincaré-Birkhoff-Witt theorem, a V-basis of E ≥m+n is given by
(3.1.1)
(i )
{θ1 1 · · · θn(in ) ||θ1 | ≤ · · · ≤ |θn |, ij ≤ 1 if |θj | odd ,
X
ij |θj | ≥ m + n, θj ∈ π <m+n,∗ (R)}.
j
The V-module Im ξ|V⊗k (N ′ )∨ is generated by the elements ξ 0 (νh ), where
ν1 , . . . , νm
MODULES OF INFINITE REGULARITY OVER COMMUTATIVE GRADED RINGS
7
′
is a k-basis of (N )∨ . We set
X
(i )
ξ 0 (νh ) =
ai1 ,...,in θ1 1 · · · θn(in ) ,
ai1 ,...,in ∈ k.
′
We can change the basis of (N )∨ so that the coordinate vectors of
ξ 0 (ν1 ), . . . , ξ 0 (νn′ ),
with respect to the basis (3.1.1), are linearly independent over k and
ξ 0 (νn′ +1 ), . . . , ξ 0 (νn )
(i )
(i )
are all zero. Because the elements θ1 1 · · · θn n form a basis of E ≥m+n,∗ over V, we
deduce that ξ 0 (ν1 ), . . . , ξ 0 (νn′ ) are also linearly independent over V.
Therefore
Im ξ|V⊗k (N ′ )∨ = Σm+n V ⊗ Im ξ 0 .
This means that Im ξ contains a copy of V and by the construction of ξ so does
E(N ′ ).
Now, we recall that by [2, Theorem 10.2.1] dimk π i,j (R) = εi,j (R).
If there is an even i > m + n such that εi,j (R) 6= 0 and i 6= j, then there is a
nonzero element x ∈ π i,j (R). The powers of x belong to V. However, a copy of
V is contained in E(N ′ ). Because bideg x = (i, j), then bideg x(l) = (li, lj) and
lj − li = l(j − i) goes to ∞ as l goes to ∞. This implies that reg N ′ = ∞, which is
a contradiction.
If there are infinitely many i with i > m + n, i odd, i 6= 0, and εi,j (R) 6= 0, then
there are infinitely many nonzero xt with t = 1, 2, . . . belonging to π it ,jt (R). The
products x1 x2 · · · xs belong to V. However, a copy of V is contained in E(N ′ ). The
bidegree of this product is
bideg x1 x2 · · · xs = (i1 + i2 + · · · + is , j1 + j2 + · · · + js ),
and because jt − it ≥ 1 for every t, the following inequality holds
j1 + j2 + · · · + js − (i1 + i2 + · · · + is ) ≥ s,
and as s goes to ∞, we obtain reg N ′ = ∞, which is a contradiction.
Now, we assume that εi,j (R) 6= 0 for finitely many i with i > m + n, i odd,
and i 6= j. Let RhXi be an acyclic closure of k. Then we can choose an element
y ∈ X that is of odd degree, off-diagonal, and of maximal homological degree.
We denote by |.| the homological degree of a homogeneous element in the acyclic
closure and by deg its internal degree. We claim that A = RhX\{y}i is a DG
subalgebra of the acyclic closure. Indeed, if x1 , . . . , xn , y are all the odd elements
and off-diagonal, then, by [5, Lemma 7], RhX\{x1 , . . . , xn , y}i is a DG subalgebra
of the acyclic closure. By the maximality of y, A is a DG subalgebra of the acyclic
closure. Hence, A is a DG algebra, and the acyclic closure of k can be written as
Ahyi, implying (by [5, Lemma 6]) that |y| = 1, which is a contradiction because
m ≥ 1.
Proposition 3.2. Let d ≥ 4 be an even number. If εi,j (R) = 0 for i ≥ d and i 6= j,
then εi,j (R) = 0 for i ≥ d − 1 and i 6= j.
8
LUIGI FERRARO
Proof. Let RhXi be an acyclic closure of k. Let y be an element of X of homological
degree d − 1, we need to prove that deg y = d − 1. We assume that y appears in a
boundary of a (bi)homogeneous element x with |x| ≥ d
X
dx =
ri ai + rwy.
The following string of (in)equalities is observed
deg w + deg y = deg(wy)
≥ |wy|
= |x| − 1 = deg x − 1
= deg w + deg y + deg r − 1
≥ deg w + deg y.
We deduce that
deg w + deg y = |w| + |y|.
If deg y > |y|, then deg w < |w|, which is not possible; hence, deg y = |y|.
If y does not appear in the boundary of any element, then we can write the
acyclic closure as Ahyi with A = RhX\{y}i, and by [5, Lemma 6] |y| = 1, hence
d = 2, which is not possible.
We set εi (R) = Σj εi,j (R). It is known that if εi (R) = 0 for i ≫ 0, then εi (R) = 0
for i ≥ 3, see [2, Theorem 7.3.3]. Motivated by this and by the previous proposition,
we raise the following question
Question 3.3. If εi,j (R) = 0 when i ≫ 0 and i 6= j, is it true that εi,j (R) = 0
when i ≥ 3 and i 6= j?
By [2, Theorem 7.3.3] and [2, Theorem 7.3.2],
1) εi (R) = 0 for i ≥ 2 if and only if R is regular, and
2) εi (R) = 0 for i ≥ 3 if and only if R is a complete intersection.
By [5, Theorem 2], if εi,j (R) = 0 when i ≥ 2 and i 6= j, then the ring R is of the
form Q ⊗k S with Q regular and S a standard graded Koszul ring. Motivated by
these results, we raise the following question
Question 3.4. If εi,j (R) = 0 when i ≥ 3 and i 6= j, then is the ring R of the form
Q ⊗k S with Q a complete intersection and S a standard graded Koszul ring?
Acknowledgements
The author would like to thank his advisors L. Avramov and S. Iyengar for
helpful conversations.
References
[1] M. André, Hopf algebras with divided powers, J. Algebra 18 1971 19-50.
[2] L. L. Avramov, Infinite free resolutions, Six Lectures on Commutative Algebra (Bellaterra,
1996), Progr. Math. 166, Birkhäuser, Basel, 1998; pp. 1–118.
[3] L. L. Avramov Modules with extremal resolutions, Math. Res. Lett. 3 (1996), 319-328.
[4] L. L. Avramov, D. Eisenbud, Regularity of modules over a Koszul algebra, J. Algebra 153
(1992), no. 1, 85-90.
[5] L. L. Avramov, I. Peeva, Finite regularity and Koszul algebras, Amer. J. Math. 123 (2001),
275–281
[6] T. H. Gulliksen, G. Levin, Homology of local rings, Queens Papers Pure Appl. Math. 20,
Queens Univ., Kingston, ON (1969).
MODULES OF INFINITE REGULARITY OVER COMMUTATIVE GRADED RINGS
9
[7] A. Martsinkovsky, A remarkable property of the (co) syzygy modules of the residue field of a
nonregular local ring, J. Pure Appl. Algebra 110 (1996), no. 1, 9-13.
[8] J. W. Milnor, J. C. Moore, On the structure of Hopf algebras, Ann. of Math. (2) 81 1965
211-264.
[9] G. Sjödin, Hopf algebras and derivations, J. Algebra 64 (1980), no. 1, 218-229.
[10] R. Takahashi, Syzygy modules with semidualizing or G-projective summands, J. Algebra 295
(2006), no. 1, 179-194.
| 0math.AC
|
arXiv:1802.01239v1 [cs.DS] 5 Feb 2018
Counting and Uniform Sampling from Markov
Equivalent DAGs
AmirEmad Ghassami,1 Saber Salehkaleybar,2 Negar Kiyavash1
1
Department of ECE, Coordinated Science Laboratory,
University of Illinois at Urbana-Champaign, Urbana, IL, USA
2
Electrical Engineering Department, Sharif University of Technology, Tehran, Iran
1
{ghassam2,kiyavash}@illinois.edu, 2 [email protected]
We propose an exact solution for the problem of finding the size of a Markov
equivalence class (MEC). For the bounded degree graphs, the proposed solution is capable of computing the size of the MEC in polynomial time. Our
proposed approach is based on a recursive method for counting the number
of the elements of the MEC when a specific vertex is set as the source variable.
We will further use the idea to design a sampler, which is capable of sampling
from an MEC uniformly in polynomial time.
1
Introduction
Directed acyclic graphs (DAGs) are the most commonly used graphical model to represent causal
relationships among a set of variables. In a DAG representation, a directed edge indicates a direct
causal relationship between the corresponding variables. Under Markov property and faithfulness
assumptions, conditional d-separation of variables in a DAG is in bijective correspondence with
conditional independencies of the variables in the underlying joint probability distribution (Spirtes
et al., 2000), and hence, a DAG representation demonstrates conditional independencies among
its variables.
The general approach for learning a causal structure is to use statistical data from the variables
to find a DAG which is the most consistent with the conditional independencies in the given
data. However, a DAG representation of a set of conditional independencies is not always unique.
1
This restricts the learning of the causal structure to Markov equivalence classes (MECs), where
elements of each class represent the same set of conditional independencies.
There is no preference between the elements of an MEC. Therefore, for a given set of data,
or more generally, for a given joint probability distribution on the variables, the size of the MEC,
i.e., the number of its elements, is a clear metric for measuring the success of an observational
structure learning approach, and indicates the necessity of using interventional experiments. This
indicates the importance of awareness about the size of the MEC for a given structure of interest.
Moreover, as mentioned earlier, one can compare different DAGs to find the most consistent
one with a given joint distribution over the variables. Here, the space of DAGs , or the space of
MECs can be chosen as the search space. Chickering showed that when the size of the MECs
are large, searching over the space of MECs is more efficient (Chickering, 2002). This again
indicates the importance of learning the size of MECs.
To the best of our knowledge, the main existing solution to the problem of finding the size of
MECs is to use Markov chain methods (He et al., 2013, Bernstein and Tetali, 2017). According
to this method, a Markov chain is constructed over the elements of the MEC whose properties
ensure that the stationary distribution is uniform over all the elements. The main problems with
the Markov chain method are its non-negligible time required for convergence and computational
issues, which in some cases prevents its use in practice.
In this paper, we propose an exact solution for the problem of finding the size of an MEC.
For the bounded degree graphs, the proposed solution is capable of computing the size of MEC
in polynomial time. Our proposed approach is based on a recursive method for counting the
number of the elements of the MEC when a specific vertex being set as the source variable. As
mentioned earlier, in an observational setting, there is no preference between the elements of an
MEC. Therefore, a reasonable approach for testing the performance of an algorithm on an MEC
is to test it on uniformly drawn samples from the MEC. This indicates the importance of the
ability of sampling from an MEC uniformly. We will use the introduced idea to design a sampler,
which is capable of sampling from an MEC uniformly in polynomial time. This approach is
applied to finding the best set of variables to intervene on when we are restricted to a certain
budget for the number of experiments (Ghassami et al., 2017).
The rest of the paper is organized as follows. In Section 2, we present the problem description.
The method for counting the elements of an MEC is presented in Section 3. The counting idea is
then used in Section 4 to design a uniform sampler from Markov equivalent DAGs.
2
2
Problem Description
An MEC is usually represented by a mixed graph M = (V, A, E), which is a graph containing
both directed and undirected edges. V , A, and E denote the set of vertices, the set of directed
edges and the set of undirected edges, respectively. The directed edges are those that have the
same direction in all elements of the class, that is, their directions are essential. This is the reason
that these mixed graphs are referred to as essential graph in the literature (Andersson et al., 1997).
For the rest of the edges, which are undirected, there exist two elements of the class which do
not agree on the direction of that edge. Essential graphs are also called complete partial directed
acyclic graphs (PDAGs) (Chickering, 2002), and maximally oriented graphs (Meek, 1995).
We define the skeleton of a DAG as an undirected graph on the same set of vertices as the
original DAG, and with undirected version of the edges of the DAG. We also define v-structure as
a structure comprising two converging directed edges whose tails are not connected by an edge.
The following result due to Verma and Pearl (Verma and Pearl, 1991) characterizes Markov
equivalence.
Proposition 1 (Verma and Pearl, 1991). Two DAGs are Markov equivalent if and only if they
have the same skeleton and the same v-structures.
Proposition 1 indicates that all the edges which are a part of a v-structure are directed in
the essential graph, i.e., in none of the members of an MEC, the subgraph corresponding to the
undirected part of the essential graph is allowed to contain a v-structure.
Let M be an essential graph. Consider the undirected graph resulted from removing all the
directed edges from M . This graph consists of disconnected chain components {G1 , ..., Gc }.
Note that each of these components will be an essential graph themselves. When the entire
essential graph is undirected, it forms a single chain component. Andersson et al. showed that
each of the chain components are chordal graphs (Andersson et al., 1997). Let Size(M ) denote
the size of the MEC corresponding to M . Gillispie and Perlman showed that Size(M ) can be
calculated as follows (Gillispie and Perlman, 2002).
c
Y
Size(M ) =
Size(Gi ).
i=1
This result suggests that one can study each chain component separately.
Therefore, we will study the problem of finding the size of an MEC for which the essential
graph is a connected undirected chordal graph, and hence, none of the members of the MEC are
allowed to have any v-structures.
3
3
Counting the Members of an MEC
The goal in this section is to count the members of an MEC, for which the essential graph is a
connected undirected chordal graph G = (V (G), E(G)).
Define a source variable in a digraph, as one for which the incoming degree is zero. Also, let
dG (v, u) be the distance between vertices v and u on G. The following result from (Bernstein
and Tetali, 2017) is used in our proposed method.
Proposition 2 (Bernstein and Tetali, 2017). Any acyclic and v-structure-free orientation of a
connected graph G has a unique source variable. The unique source v, determines the orientation
of all edges {u, w}, for which dG (v, u) 6= dG (v, w).
Proposition 2 determines that the direction of some edges will be resolved when the source
vertex is known. These are the edges with one endpoint closer to the source vertex than the
other. We note that removing the resolved edges, the obtained digraph will remain acyclic and
v-structure-free:
Lemma 1. If the original digraph is acyclic and v-structure-free with source vertex v, the graph
resulting from removing edges (u, w) with dG (v, u) 6= dG (v, w) is also acyclic and v-structurefree.
Proof. Acyclicity is obvious as we are removing edges. To show the v-structure-freeness,
consider a v-structure x → y ← z. We note that if a v-structure is created, it means that the
shielding vertex {x, z} has been removed, that is, dG (v, x) 6= dG (v, z). But since (x, y) and (z, y)
are unresolved, dG (v, x) = dG (v, y), and dG (v, y) = dG (v, z), which is a contradiction.
Due to Lemma 1, we can orient the edges of G and obtain a member of the corresponding
MEC as follows. First one vertex is chosen to be the source vertex. This will determine the
orientation of some edges according to Proposition 2. Removing the resolved edges and resulted
isolated vertices, we will have (possibly) disconnected components, where the components
should again not contain cycles or v-structures. In each component, one vertex is chosen as the
source vertex of that component. This procedure is repeated until no unresolved edge is left. The
idea in the aforementioned procedure can be used in a recursive manner to count the members of
the MEC. The main tool in our approach is presented in Theorem 1.
Theorem 1. Let W(v, G) be the number of elements of the MEC with connected and undirected
essential graph G, in which v is the source vertex; and let H be the undirected graph remained
4
after removing edges with one endpoint closed to v than the other, as well as removing the
resulting isolated vertices. We have
W(v, G) =
Y
X
W(u, G0 ),
G0 ∈comp(H) u∈V (G0 )
where comp(H) denotes the set of the components of H, and by convention, empty product
(nullary product) is assumed to be the multiplicative identity 1.
Proof. For each component of H, choosing different variables of that component as the source
partitions the set of elements of the MEC. That is, two different choices for the source do not
lead to the same DAG (due to Proposition 2), and no elements will be remained uncounted, as
each variable is chosen as the source once. Therefore, we have the inner summation.
To justify the outer product, we need to show that orientation of one component of H does
not cause any restrictions on the possible orientations in any other component. This is due to
the fact that each component is at a certain depth level from the source and they cannot create
v-structures pointing to lower levels. Also, for the same reason and the fact that the edges from
the source are towards each component, the edges in two component cannot create a cycle.
Therefore, using Theorem 1, in order to count the members of an MEC, it suffices to calculate
W(v, G) for all v ∈ V (G) and sum them up, that is,
X
Size(G) =
W(v, G).
(1)
v∈V (G)
The pseudo-code of our approach is presented in Algorithm 1. This code consists of two
functions F LOWED and W. The function F LOWED takes an undirected and connected graph with
one of its vertices as the input, and returns the set of edges whose orientation is resolved when
the input vertex is set as the source vertex. Function W applies Theorem 1 in a recursive manner
to calculate the number of elements of the MEC with connected and undirected essential graph
G, when input vertex v is set as the source vertex.
Theorem 2. Let n and ∆ denote the order and the maximum degree of the graph G, respectively.
The computational complexity of Algorithm 1 is O(n∆ ).
Proof. We consider the operation of Algorithm 1 in rounds. In the first round, F LOWED is
called for all pairs (G, v1 ), v1 ∈ V (G). For any specific v1 ∈ V (G), we obtain undirected
components Gv11 , Gv21 , ..., GvK11 . In the second round, for each component Gvk1 , 1 ≤ k ≤ K1 , we
5
Algorithm 1 MEC Size Calculator
Input: P
Undirected graph G = (V (G), E(G)).
Size = v∈V (G) W(v, G).
Output: Size.
function F LOWED(v,G)
Initiate: A = ∅.
Set v as the source variable in G.
for {u, w} ∈ E(G) do
if dG (v, u) < dG (v, w) then A = A ∪ (u, w).
end if
if dG (v, u) > dG (v, w) then A = A ∪ (w, u).
end if
end for
return A.
end function
function W(v,G)
F = Undirected version of elements of F LOWED(v, G).
G0 = G\F .
Remove isolated vertices from G0 .
Y
X
return
W(u, G)
G∈comp(G0 ) u∈V (G)
end function
call F LOWED for all pairs (Gvk1 , v2 ), v2 ∈ V (Gvk1 ). For any specific v2 ∈ V (Gvk1 ), we obtain
undirected components Gv11 v2 , Gv21 v2 , ..., GvK12v2 , and we continue this procedure. It suffices to
show that this procedure could be repeated only for ∆ rounds. Therefore, since in each round
the function F LOWED is called at most n times, we get the desired result.
Suppose that in a round, we obtain an undirected component G0 and execute F LOWED on it
for a pair (G0 , v), where v ∈ V (G0 ). Let G01 , · · · , G0K be the resulting undirected components.
We first show that for any vertex u ∈ V (G0i ), (1 ≤ i ≤ K), we have obtained a directed edge in
G0 from some vertex w to u after applying F LOWED. By contradiction, suppose that there is no
such directed edge. Consider the shortest path from v to u. This path must pass through one of
neighbors of u in G0i such as w. Since d(v, w) < d(v, u), (u, w) should be oriented as w → u by
calling F LOWED procedure. But it is in contradiction with the fact that u and w are in G0i and
thus there should be such a directed edge in G0 .
For any vertex u ∈ V (G), consider the set of all edges in E(G) which intersect with u.
6
𝑣1
𝐺
𝑣2
𝑣1
(𝑏)
𝑣4
𝑣4
𝑣3
(𝑐)
𝑣2
(𝑎)
𝐺′
𝑣4
𝑣3
𝑣2
𝑣3
𝑣2
𝑣1
𝑣3
(𝑑)
𝑣4
𝐺′′
𝑣1
𝑣3
(𝑒)
Figure 1: Graphs related to Example 1.
Based on the reasoning above, at least one of these edges will be oriented in a round if vertex u
is still in one of undirected components before that round. Since the degree of u is at most ∆, all
of these edges will be oriented after at most ∆ rounds.
Example 1. Consider the undirected graph presented in Figure 1(a) as the essential graph given
as the input to Algorithm 1.
Setting vertex v1 as the source vertex, the direction of edges (v1 , v2 ), (v1 , v3 ), (v2 , v4 ), and
(v3 , v4 ) will be obtained using function F LOWED(v1 , G), as shown in Figure 1(b). Therefore,
the remaining undirected graph is G0 , shown in Figure 1(c). Setting each vertex in G0 as the
source vertex, no undirected component will be remained. Hence, W(v2 , G0 ) = W(v3 , G0 ) = 1.
Therefore, W(v1 , G) = W(v2 , G0 ) + W(v3 , G0 ) = 2. Due to the symmetry in the structure, we
have W(v4 , G) = 2 as well.
Setting vertex v2 as the source vertex, the direction of edges (v2 , v1 ), (v2 , v3 ), and (v2 , v4 ) will
be obtained using function F LOWED(v2 , G), as shown in Figure 1(d). Therefore, the remaining
undirected graph is G00 , shown in Figure 1(e). Setting each vertex in G00 as the source vertex,
no undirected component will be remained. Hence, W(v1 , G00 ) = W(v3 , G00 ) = W(v4 , G00 ) = 1.
Therefore, W(v2 , G) = W(v1 , G00 ) + W(v3 , G00 ) + W(v4 , G00 ) = 3. Due to the symmetry in the
structure, we have W(v3 , G) = 3 as well.
Finally, using equation (1), we obtain that the number of the elements of the MEC corresponding to essential graph G is equal to 10.
7
Algorithm 2 Uniform Sampler
Input: Undirected graph G = (V (G), E(G)).
Output: Directed graph (V (G), R AND E DGE(G, ∅)).
function R AND E DGE(G,A)
Set v ∗ ∈ V (G) as the source variable of G with probability P
W(v ∗ , G)
.
v∈V (G) W(v, G)
A = A ∪ F LOWED(v ∗ , G).
F = Undirected version of elements of F LOWED(v ∗ , G).
G0 = G\F .
Remove isolated
vertices from G0 .
[
A=A
R AND E DGE(G, A).
G∈comp(G0 )
return A.
end function
4
Uniform Sampling from an MEC
As mentioned earlier in Section 1, in some applications, the performance of an action is needed
to be tested on an MEC with large size. In these cases, one can optimize the behavior on uniform
samples drawn from the MEC. In this section we introduce two samplers for providing random
DAGs from an MEC. The first sampler, presented in Subsection 4.1, is based on the counting idea
in Section 3. In the case of bounded degree graphs, this sampler is capable of producing uniform
samples in polynomial time. If the degree is not bounded, in the worst case, this sampler has high
computational complexity. Therefore, in Subsection 4.2, we provide another sampler, originally
introduced in (Ghassami et al., 2017), for the case of unbounded degree graphs. Although
this sampler is not unbiased, our extensive experimental results confirms that the sampling
distribution of the second sampler is very close to uniform.
4.1
Uniform Sampler
The pseudo-code of our uniform sampler is presented in Algorithm 2, where functions F LOWED
and W are defined in the pseudo-code for Algorithm 1. In this sampler, first the weights W(v, G)
is calculated for all vertices v ∈ V (G) and vertex v ∗ is chosen as the source variable with
P
probability W(v ∗ , G)/ v∈V (G) W(v, G). The source variable will determine the orientation
of some edges according to Proposition 2. Removing the directed edges and resulted isolated
vertices, we will have (possibly) disconnected components. Choosing a source variable in each
8
resulting component is repeated until all the edges are directed. Since we choose each vertex as
the source variable according to the portion of members of the MEC having that vertex as the
source, we will have the following result.
Theorem 3. The sampler proposed in Algorithm 2 is uniform.
Proof. Let g be a member of the MEC. For the clarity of the proof, we assume that g is obtained
form three rounds of choosing source vertices, otherwise, the idea is easily extendable to the
case with arbitrary many rounds. Therefore, assume g is obtained by having v1 as the chosen
source vertex in G in round 1, v21 , ..., v2K1 as the sources in components G11 , ..., G1K1 in round 2,
and v31 , ..., v3K2 as the sources in components G21 , ..., G2K2 in round 3. We have
W (v21 , G11 ) · · · W (v2K1 , G1K1 )
W (v1 , G)
× QK1 P
1
v∈V (G) W (v, G)
k=1
v∈V (G1 ) W (v, Gk )
P (g) = P
k
W (v31 , G21 ) · · · W (v3K2 , G2K2 )
.
× QK1 Q
P
2
k=1
G2 ∈comp(G1 )
v∈V (G2 ) W (v, Gj )
j
j
k
Since the process is terminated in three rounds, we have W (v3k , G2k ) = 1, for all 1 ≤ k ≤ K2 .
Therefore,
1
W (v1 , G)
× QK1 P
1
v∈V (G) W (v, G)
k=1
v∈V (G1 ) W (v, Gk )
P (g) = P
k
×
K1
Y
W (v2k , G1k )
Q
k=1
G2j ∈comp(G1k )
P
v∈V (G2j )
W (v, G2j ))
.
By Theorem 1, we have the second fraction and all the fractions in the product are equal to 1.
Therefore,
1
1
=
,
Size(G)
v∈V (G) W (v, G)
P (g) = P
where the last equality follows from equation (1).
Regarding the complexity of the sampler in Algorithm 2 we have the following result as a
corollary of Theorem 2.
Corollary 1. Let n and ∆ denote the order and the maximum degree of the graph G, respectively.
The computational complexity of Algorithm 2 is O(n∆ ).
9
Algorithm 3 Heuristic Sampler
Input: Undirected graph G = (V (G), E(G)).
Uniformly shuffle the order of the elements of V (G).
while the induced subgraph on any subset of size 3 of the variables is not directed, or a directed
cycle, or a v-structure do
for all {vi , vj , vk } ⊆ V (G) do
Orient the undirected edges among {vi , vj , vk } independently according to Bern( 12 )
until it becomes a directed structure which is not a directed cycle or a v-structure.
end for
end while
Output: Resulted directed graph.
4.2
Heuristic Sampler
For the case of unbounded degree, the computational complexity of Algorithm 2 will be high
in the worst case. Therefore, we also provide a heuristic sampler for this case, which was first
introduced in (Ghassami et al., 2017). Although not unbiased, the experimental results indicate
that this sampler has a performance very close to the uniform sampler. The pseudo-code of
the heuristic sampler is presented in Algorithm 3. In this algorithm we consider all subsets
of size 3 from V (G) in a uniformly random order (achieved by uniformly shuffling the labels
of elements of V (G)). For each subset {vi , vj , vk }, we orient the undirected edges among
{vi , vj , vk } independently according to Bernoulli( 12 ) distribution. If the resulting orientation
on the induced subgraph on {vi , vj , vk } became a directed cycle or a v-structure, we redo the
orienting. We keep checking all the subsets of size 3 until the induced subgraph on all of them
are directed and none of them is a v-structure or a directed cycle.
Theorem 4. Each generated directed graph in Algorithm 3 belongs to the Markov equivalence
class corresponding to the essential graph G.
We require the following lemma for the proof.
Lemma 2. A chordal graph has a directed cycle only if it has a directed cycle of size 3.
Proof. If the directed cycle is of size 3 itself, the lemma is trivial. Suppose the cycle Cn is
of size n > 3. Relabel the vertices of Cn to have Cn = (v1 , ..., vn , v1 ). Since the graph is
chordal, Cn has a chord and hence we have a triangle on vertices {vi , vi+1 , vi+2 } for some i.
If the direction of {vi , vi+2 } is (vi+2 , vi ), we have the directed cycle (vi , vi+1 , vi+2 , vi ) which
is of size 3. Otherwise, we have the directed cycle Cn−1 = (v1 , ..., vi , vi+2 , .., vn , v1 ) on n − 1
10
vertices. Relabeling the vertices from 1 to n − 1 and repeating the above reasoning concludes
the lemma.
Proof of Theorem 4. The undirected graph G is chordal (Hauser and Bühlmann, 2012). Therefore, by Lemma 2, to insure that a generated directed graph is a DAG, it suffices to make sure
that it does not have any directed cycles of length 3, which is one of the checks that we do in
the proposed procedure. For checking if the generated DAG is in the Markov equivalence class
corresponding to G, by Proposition 1, it suffices to check if they have the same set of v-structures,
which is the other check that we do in the proposed procedure.
11
References and Notes
Andersson et al., 1997. Andersson, S. A., Madigan, D., Perlman, M. D., et al. (1997). A
characterization of markov equivalence classes for acyclic digraphs. The Annals of Statistics,
25(2):505–541.
Bernstein and Tetali, 2017. Bernstein, M. and Tetali, P. (2017). On sampling graphical markov
models. arXiv preprint arXiv:1705.09717.
Chickering, 2002. Chickering, D. M. (2002). Optimal structure identification with greedy search.
Journal of machine learning research, 3(Nov):507–554.
Ghassami et al., 2017. Ghassami, A., Salehkaleybar, S., Kiyavash, N., and Bareinboim,
E. (2017).
Budgeted experiment design for causal structure learning.
arXiv preprint
arXiv:1709.03625.
Gillispie and Perlman, 2002. Gillispie, S. B. and Perlman, M. D. (2002). The size distribution
for markov equivalence classes of acyclic digraph models. Artificial Intelligence, 141(1-2):137–
155.
Hauser and Bühlmann, 2012. Hauser, A. and Bühlmann, P. (2012). Characterization and greedy
learning of interventional markov equivalence classes of directed acyclic graphs. Journal of
Machine Learning Research, 13(Aug):2409–2464.
He et al., 2013. He, Y., Jia, J., Yu, B., et al. (2013). Reversible mcmc on markov equivalence
classes of sparse directed acyclic graphs. The Annals of Statistics, 41(4):1742–1779.
Meek, 1995. Meek, C. (1995). Causal inference and causal explanation with background
knowledge. In Proceedings of the Eleventh conference on Uncertainty in artificial intelligence,
pages 403–410. Morgan Kaufmann Publishers Inc.
Spirtes et al., 2000. Spirtes, P., Glymour, C. N., and Scheines, R. (2000). Causation, prediction,
and search. MIT press.
Verma and Pearl, 1991. Verma, T. and Pearl, J. (1991). Equivalence and synthesis of causal
models. In Proceedings of Sixth Conference on Uncertainty in Artificial Intelligence, pages
220–227.
12
| 2cs.AI
|
arXiv:1410.5850v2 [math.OC] 27 Apr 2017
A Fast Hybrid Primal Heuristic for Multiband Robust
Capacitated Network Design with Multiple Time Periods✩
Fabio D’Andreagiovannia,b,c,∗, Jonatan Krolikowskia , Jonad Pulaja
a
Department of Optimization, Zuse-Institut Berlin (ZIB), Takustr. 7, 14195 Berlin,
Germany
b
DFG Research Center MATHEON, Technical University Berlin, Straße des 17. Juni
135, 10623 Berlin, Germany
c
Einstein Center for Mathematics Berlin (ECMath), Straße des 17. Juni 135, 10623
Berlin, Germany
Abstract
We investigate the Robust Multiperiod Network Design Problem, a generalization of the Capacitated Network Design Problem (CNDP) that, besides
establishing flow routing and network capacity installation as in a canonical
CNDP, also considers a planning horizon made up of multiple time periods
and protection against fluctuations in traffic volumes. As a remedy against
traffic volume uncertainty, we propose a Robust Optimization model based
on Multiband Robustness (Büsing and D’Andreagiovanni, 2012), a refinement of classical Γ-Robustness by Bertsimas and Sim (2004) that uses a
system of multiple deviation bands. Since the resulting optimization problem may prove very challenging even for instances of moderate size solved by
This is the authors’ final version of the paper published in Applied Soft Computing
26, 497-507, 2015, DOI: 10.1016/j.asoc.2014.10.016 . The final publication is available
at Elsevier ScienceDirect via http://dx.doi.org/10.1016/j.asoc.2014.10.016.
Please note that this paper is an extended, improved journal version of the paper
“A hybrid primal heuristic for Robust Multiperiod Network Design” presented at EvoComNet2014 and published in: EvoApplications 2014: Applications of Evolutionary
Computation, Granada, Spain, April 23-25, 2014, LNCS 8602, pp. 15-26, Springer, 2014.
DOI: 10.1007/978-3-662-45523-4 2.
The paper was awarded the EvoComNet Best Paper Award at EvoStar 2014.
✩
∗
Corresponding author
Email addresses: [email protected] (Fabio D’Andreagiovanni),
[email protected] (Jonatan Krolikowski), [email protected] (Jonad Pulaj)
URL: http://www.dis.uniroma1.it/~fdag/ (Fabio D’Andreagiovanni)
a state-of-the-art optimization solver, we propose a hybrid primal heuristic
that combines a randomized fixing strategy inspired by ant colony optimization and an exact large neighbourhood search. Computational experiments
on a set of realistic instances from the SNDlib (2010) show that our original heuristic can run fast and produce solutions of extremely high quality
associated with low optimality gaps.
Keywords: Capacitated Network Design, Multiperiod Design, Multiband
Robust Optimization, Traffic Uncertainty, Metaheuristic, Ant Colony
Optimization, Exact Large Neighborhood Search.
1. Introduction
In the last two decades, telecommunications have increasingly pervaded
our everyday life and the volume of traffic sent and exchanged over networks
has astonishingly increased: major companies like Nokia Siemens Networks
expect that the increase in the amount of traffic will strongly continue,
reaching a volume of more than 1000 exabyte per year in fixed networks by
2015 (Theimer, 2010). This dramatic growth that telecommunications have
experienced has greatly compounded the challenge for network professionals,
who are facing design problems of increasing complexity and difficulty. In
order to cope with traffic growth, the professionals have to plan much in
advance how the network will be expanded in topology and capacity to
accommodate the new traffic. This is especially important in the case of
fixed networks, which require (costly) digging operations for the installation
of cables in areas with possibly high population density.
To make the design task even more complicated, the future behaviour of
traffic over a network is not exactly known when the network is designed and
thus the decision problem is also affected by tricky data uncertainty. Until
recent times, data uncertainty has been generally neglected in real studies.
However, as indicated by recent industrial cooperations between industry
and academia, (e.g., Bauschert et al. 2014; Belotti et al. 2011; Bley et al.
2013; Koster et al. 2013), professionals are not only becoming aware of the
importance of adopting mathematical optimization to take better decisions,
but are also understanding the necessity of considering data uncertainty, in
order to avoid unpleasant surprises like infeasibility of implemented solutions
due to data deviations.
The task of designing a telecommunication network essentially consists
in establishing the topology of the network and the technological features
(e.g., transmission capacity and rate) of its elements, namely nodes and
2
links. One of the most studied problem in network design is the Capacitated
Network Design Problem (CNDP): the CNDP consists in minimizing the
total installation cost of capacity modules in a network to route traffic flows
generated by users. The CNDP is a central problem in network optimization, which appears in many real-world applications. For an exhaustive introduction to it, we refer the reader to Ahuja et al. (1993); Bertsekas (1998);
Bienstock et al. (1998).
In this paper, we focus on the development of a new Robust Optimization
model to tackle traffic uncertainty in the Multiperiod Capacitated Network
Design Problem (MP-CNDP). This problem constitutes a natural extension
of the classical CNDP, where, instead of a single design period, we consider
the design over a time horizon made up of multiple periods. Moreover,
traffic uncertainty is taken into account to protect design solutions against
deviations in the traffic input data, which may compromise feasibility and
optimality of solutions.
We immediately stress that, though the problem of optimally designing networks over multiple time periods is not new and can be traced back
at least to the seminal work by Christofides and Brooker (1974), to the
best of our knowledge, the MP-CNDP has received very little attention and
just a few works have investigated it - essentially, Lardeux et al. (2007) and
Gavros and Raghavan (2012). Checking literature, several other works dealing with multiperiod design of networks can be found (just to make a couple
of examples, Gendrau et al. 2006 and Gupta and Grossmann 2012): however, all these works consider problems that are application-specific or are
sensibly different from the more general setting that we consider here and we
thus avoid a more detailed discussion of them (Gendrau et al. (2006) study
capacity expansion problem in access networks with tree topology, whereas
Gupta and Grossmann (2012) consider the design of utility networks modeled by non-linear mathematical programs).
Our main references for this work, namely Lardeux et al. (2007) and
Gavros and Raghavan (2012), point out the difficulty of solving multiperiod
CNDP problems already for just two periods and even in (easier) contexts:
Lardeux et al. (2007) consider the CNDP when traffic flows may be split,
whereas Gavros and Raghavan (2012) consider a pure routing problem in
satellite communications. Our direct and more recent computational experience have confirmed the challenging nature of the MP-CNDP, even for
instances of moderate size with a low number of time periods and solved by
a state-of-the-art optimization solver.
Uncertain versions of the MP-CNDP where traffic uncertainty is considered have also been neglected, even though especially in the last years there
3
has been an increasing interest in network design under traffic uncertainty
for the single design period case (e.g., Bauschert et al. 2014; Belotti et al.
2011; Koster et al. 2013).
In this work, our main original contributions are:
1. the first Robust Optimization model for tackling traffic uncertainty in
Multiperiod CNDP. Specifically, we adopt Multiband Robustness, a
new model for Robust Optimization recently introduced by Büsing and D’Andreagiovanni
(2012);
2. a hybrid primal heuristic, based on the combination of a randomized
rounding heuristic resembling ant colony optimization (Dorigo et al.,
1999) with an exact large neighborhood search called RINS (Danna et al.,
2005). We stress here that our aim was not to use a standard implementation of an ant colony algorithm: we wanted instead to strengthen
the performance of the ant algorithm using highly valuable information from linear relaxations of the considered optimization problems.
Using this information allowed us to define a very strong ant construction phase, which produces very high quality solutions already before
the execution of any local search;
3. analytically proving how to solve the linear relaxation of the Multiperiod CNDP in closed form, thus obtaining a substantial reduction in
solution times w.r.t. our first algorithm presented in (D’Andreagiovanni et al.,
2014);
4. computational experiments over a set of realistic instances derived
from the Survivable Network Design Library (SNDlib, 2010), showing
that our hybrid algorithm is able to produce solutions of extremely
high quality associated with very small optimality gap.
The remainder of this paper is organized as follows: in Section 2, we review
a canonical model for the CNDP; in Section 3, we introduce the Multiperiod
CNDP and we study its linear relaxation; in Section 4, we introduce the new
formulation for Robust Multiperiod CNDP; in Sections 5 and 6, we present
our hybrid heuristic and computational results.
2. The Capacitated Network Design Problem
The CNDP is a central and highly studied problem in Network Optimization that appears in a wide variety of real-world applications (see
Ahuja et al. (1993) and Bertsekas (1998) for an exhaustive introduction to
it) and can be essentially described as follows: given a network and a set
4
of demands whose flows must be routed between vertices of the network,
we want to install capacities on network edges and route the flows through
the network, so that the capacity constraint of each edge is respected and
the total cost of installing capacity is minimized. More formally, we can
characterize the CNDP through the following definition.
Definition 1 (The Capacitated Network Design Problem - CNDP).
Given
• a network represented by a graph G(V, E), where V is the set of vertices and E the set of edges,
• a set of commodities C, each associated with a traffic flow dc to route
from an origin sc to a destination tc ,
• a set of admissible paths Pc for routing the flow of each commodity c
from sc to tc ,
• a cost γe for installing one module of capacity φ > 0 on edge e ∈ E,
the CNDP consists in establishing the number of capacity modules installed
on each edge e ∈ E such that the resulting capacity installation has minimum
cost and supports a feasible routing of the commodities. A feasible routing
assigns each commodity c ∈ C to exactly one feasible path p ∈ Pc .
Referring to the notation introduced above and introducing the following
two families of decision variables:
• binary path assignment variables xcp ∈ {0, 1} ∀ c ∈ C, p ∈ Pc such
that:
1 if the entire traffic of commodity c is routed on path p
xcp =
0 otherwise,
• integer capacity variables ye ∈ Z+ , ∀ e ∈ E, representing the number
of capacity modules installed on edge e,
5
we can model the CNDP as the following integer linear program:
X
min
γe y e
(CNDP-IP)
e∈E
X
X
X
xcp = 1
dc xcp ≤ φ ye
e∈E
(1)
c∈C
(2)
c∈C p∈Pc : e∈p
p∈Pc
xcp ∈ {0, 1}
c ∈ C, p ∈ Pc
ye ∈ Z+
e∈E.
The objective function minimizes the total cost of capacity installation. Capacity constraints (1) impose that the summation of all flows routed through
an edge e ∈ E must not exceed the capacity installed on e (equal to the
number of installed modules represented by ye multiplied by the capacity φ
granted by a single module). Constraints (2) impose that the flow of each
commodity c ∈ C must be routed through a single path.
Remark 1. This is an unsplittable version of the CNDP, namely the traffic
flow of a commodity c ∈ C cannot be split over multiple paths going from
sc to tc , but must be routed on exactly one path. Moreover, the set of
feasible paths Pc of each commodity is preset and constitutes an input of the
problem. This is in line with other works based on industrial cooperations
(e.g., Bley et al. 2000) and with our experience (Bauschert et al. 2014), in
which a network operator typically considers just a few paths that meet its
own specific business and quality-of-service considerations and uses stateof-the-art routing schemes based on an Open Shortest Path First protocol.
3. The Multiperiod Capacitated Network Design Problem
We define the multiperiod generalization of the CNDP by introducing
a time horizon made up of a set of elementary periods T = {1, 2, . . . , |T |}.
From a modeling point of view, this generalization requires to add a new index t ∈ T to the decision variables, to represent routing and capacity installation decisions taken in each period. This is a simple modeling operation,
which, however, greatly increases the size and complexity of the problem
as pointed out in our computational section and by Gavros and Raghavan
(2012) and Lardeux et al. (2007). The introduction of the new index leads
6
to the new integer linear program:
XX
min
γet yet
(MP-CNDP-IP)
e∈E t∈T
X
X
X
xtcp = 1
c∈C p∈Pc : e∈p
dtc
xtcp
≤φ
t
X
yeτ
e ∈ E, t ∈ T
(3)
τ =1
c ∈ C, t ∈ T
p∈Pc
xtcp ∈ {0, 1}
yet ∈ Z+
c ∈ C, p ∈ Pc , t ∈ T
e ∈ E, t ∈ T .
Besides including the new index t ∈ T , the program presents a modified
right-hand-side in the capacity constraint (3): for each time period t, we
must consider the presence of all the capacity modules that are installed on
an edge e ∈ E from period τ = 1 to period τ = t.
Concerning costs and traffic demands, in what follows, we realistically
assume that the cost per unit of capacity is non-increasing over time and
that the demand associated with each commodity is non-decreasing over
time (i.e., γet ≥ γet+1 , ∀ t = 1, . . . , |T | − 1 and dtc ≤ dt+1
c , ∀ t = 1, . . . , |T | − 1).
3.1. Computing the linear relaxation of MP-CNDP-IP
After having introduced the problem MP-CNDP-IP, we proceed to formally characterize in closed-form the value and optimal solution of its linear
relaxation, namely the problem that we obtain if we relax the integrality requirements on variables x and y (so we have 0 ≤ xtcp ≤ 1, ∀c ∈ C, p ∈ Pc , t ∈
T and yet ≥ 0, ∀e ∈ E, t ∈ T ). We refer to such relaxation as MP-CNDP-LP.
Characterizing the optimal value and the structure of an optimal solution of MP-CNDP-LP has been a crucial objective for us. In our preliminary
computational experience presented in the proceeding version of this paper
(D’Andreagiovanni et al., 2014), solving MP-CNDP-LP by using the stateof-the-art optimization solver IBM ILOG CPLEX (2014) proved indeed slow.
Using a solver like CPLEX to solve linear relaxations thus constituted a bottleneck that limited the possibility of running the heuristic a large number
of times within the time limit. By using instead the closed-form expression
coming from our new theoretical results presented in this section, we were
able to efficiently solve MP-CNDP-LP without using CPLEX, run the algorithm an incredibly higher number of times in the same time limit and gain
a determinant speed up in the overall execution of the algorithm.
7
In the case of an MP-CNDP including a single time period (i.e., |T | = 1), it
is well known that an optimal solution of the linear relaxation MP-CNDPLP can be obtained by just considering the shortest path p∗ ∈ Pc for each
commodity c ∈ C and then by installing on each edge of p∗ the smallest
number of capacity modules needed to support the traffic request dc of c
(see Ahuja et al. 1993).
Through the following propositions, we investigate how the optimal value
and an optimal solution of the MP-CNDP-LP with |T | > 1 look like, proving that they can be characterized efficiently. Note that in order to keep
the exposition light, we decided to move the more complex proofs of the
statements to the Appendix A.
As first step, we characterize the relation between flows in consecutive time
periods.
Proposition 1. Let x̄tcp := 0, dtcp := 0 and γet := ∞ for t = 0. There exists
an optimal solution x̄ of MP-CNDP-LP such that:
≤ dtc x̄tcp
dt−1
x̄t−1
cp
c
∀ c ∈ C, p ∈ Pc , t ∈ T
(4)
Proof. See the Appendix A.
Using the previous Proposition, we are able to derive the following Theorem,
which characterizes an optimal solution for the linear relaxation of MPCNDP-IP. In particular, we can show that to get an optimal solution we
just need to identify for each commodity c ∈ C and time period t ∈ T a
shortest path in edge length p∗c ∈ Pc among the feasible paths Pc of c. Of
course, this operation can be done very fast and efficiently.
Remark 2. We stress that in what follows, we will just focus on the determination of the optimal flows, since once we have an optimal flow, the
capacity installation can be immediately derived over the network.
Theorem 1. Consider problem MP-CNDP-LP, namely the linear relaxation
of MP-CNDP-IP, and let p∗c ∈ Pc be a shortest path in edge length for each
commodity c ∈ C. An optimal solution of MP-CNDP-LP can be defined by
routing in each time period t ∈ T the entire flow of each commodity c ∈ C on
the shortest path p∗c and by installing on each edge of p∗c the exact capacity
needed to route the traffic flow dtc .
Proof. See the Appendix A.
8
The previous theorem efficiently characterizes an optimal solution for MPCNDP-LP. However, our hybrid algorithm requires a feasible solution of the
linear relaxation when the routing and capacity installation has been established for a number of consecutive time periods (see Section 5 for details).
The following Corollary shows how to characterize a feasible solution for
the modified linear relaxation and identifies the conditions under which this
solution becomes optimal.
Corollary 1. Consider problem MP-CNDP-IP and suppose that for time
periods t = 1, . . . , τ − 1, τ ≤ |T | and for all the commodities c ∈ C, feasible
path assignments and capacity installations have been established (thus fixing
the values of the corresponding variables x and y). Moreover, suppose that
for period t = τ feasible path assignments and capacity installations have
been established for a subset of commodities C ′ ⊆ C.
Denote now by MP-CNDP-IPF IX the version of MP-CNDP-IP obtained
from this variable fixing and consider the corresponding linear relaxation
MP-CNDP-LPF IX . Let p∗c ∈ Pc be a shortest path in edge cost for each
commodity c ∈ C. A feasible solution for MP-CNDP-LPF IX can be defined
by:
• routing in time period t = τ the entire flow of each c ∈ C \ C ′ on
the shortest path p∗c and by installing on each edge of p∗c the minimum
number of capacity modules needed to route the traffic flow dτc ;
• routing in each time period t = τ + 1, . . . , |T | the entire flow of each
c ∈ C on the shortest path p∗c and by installing on each edge of p∗c the
minimum number of capacity modules needed to route the traffic flow
dtc ;
Furthermore, the solution determined above is optimal when for all the commodities c ∈ C, the path p∗c is chosen in every t ∈ T .
Proof. The feasibility of the solution built as specified above is clear. The
optimality condition is instead a straightforward consequence of Theorem 1
and we omit the proof.
The new theoretical results that we have introduced provide an alternative
way to compute an (optimal) feasible solution of MP-CNDP-LP, which prove
dramatically faster than the direct use of CPLEX. This allows us to greatly
increase the number of executions of the ant construction phase in the same
time limit.
9
4. Robust Optimization for traffic-uncertain Multiperiod Network
Design
After having introduced the multiperiod generalization MP-CNDP, we
can proceed to consider its version that takes into account traffic uncertainty.
To this end, in this section, we first state what we mean by traffic uncertainty, then we present fundaments of Robust Optimization, the methodology that we adopt to tackle data uncertainty, and finally we present a Robust
Optimization model for the traffic-uncertain version of the MP-CNDP.
4.1. Traffic uncertainty and Robust Optimization
Uncertainty of traffic is naturally present in telecommunications network
design, since the future behaviour of customers is not known in advance: the
number of users and the traffic generated by them can just be estimated and
these estimates can deeply differ from actual traffic conditions that will occur
in the future (see Bauschert et al. 2014). In what follows, we thus assume
that the demands dc are uncertain for all the commodities c ∈ C, i.e. their
value is not known exactly when the optimization problem is solved. In
order to clarify the concept of traffic uncertainty, we anticipate here that
we will model data uncertainty by a refined interval deviation model. In
an interval model, we assume to know a nominal value of traffic d¯c and
maximum negative and positive deviations δc− ≤ 0, δc+ ≥ 0 from it. The
(unknown) actual value dc is thus assumed to belong to the interval:
dc ∈ [d¯c + δc− , d¯c + δc+ ] .
In our direct experience with real-world network design, we have observed
that professionals often identify d¯c with the value of forecast traffic volume
(e.g., an expected value derived from historical data), whereas the deviations
δc− , δc+ are identified as the maximum deviations from the forecast considered
relevant by the network designer, again using historical data as reference.
Example 1 (traffic uncertainty). Consider two commodities c1 , c2 associated with nominal traffic demands d¯c1 = 200 Mb, d¯c2 = 300 Mb and suppose that these values may deviate up to 10%. So the maximum negative and
positive deviations for c1 , c2 are δc−1 = −20, δc+1 = 20 Mb, δc−2 = −30, δc+2 = 30
Mb, respectively. The actual values of traffic are therefore dc1 ∈ [180, 220]
Mb, dc2 ∈ [270, 330] Mb.
As it is well-known from sensitivity analysis, dealing with data uncertainty in
optimization problems may result really tricky: small variations in the value
10
of input data may completely compromise the optimality and feasibility of
produced solutions. Solutions that are supposed to be optimal may reveal to
be heavily suboptimal, whereas solutions supposed to be feasible may reveal
to be infeasible and thus meaningless when implemented. For a detailed
discussion of the issues associated with data uncertainty, we refer the reader
to (Ben-Tal et al., 2009; Bertsimas et al., 2011). The following example can
immediately help to visualize the possibly catastrophic effects of neglecting
data uncertainty.
Example 2 (infeasibility caused by deviations). Consider again the commodities of Example 1 and suppose that in some link we have installed exactly the capacity to handle the sum of their nominal values (i.e., we have
installed 200 + 300 Mb of capacity). This capacity installation neglects the
fact that the demands may deviate up to 10%. So, it is sufficient that
just one commodity experiences a positive deviation to violate the capacity
constraint of the link, thus making the design solution infeasible in practice.
The previous example makes clear that we cannot afford to neglect traffic uncertainty and therefore risk that our design solution will turn out to
be infeasible or of bad quality when implemented. As a consequence, we
have decided to tackle data uncertainty by adopting a Robust Optimization
(RO) approach. RO is a methodology for dealing with data uncertainty
that has received a lot of attention in recent times and has been preferred
to traditional methodologies like Stochastic Programming, especially thanks
to its accessibility and computational tractability. We refer the reader to
Ben-Tal et al. (2009) and Bertsimas et al. (2011) for an exhaustive introduction to theory and applications of RO and for a discussion about its
determinant advantages over Stochastic Programming.
RO is founded on two main facts:
• the decision maker must define an uncertainty set, which reflects his
risk aversion and identifies the deviations of coefficients against which
protection must be guaranteed;
• protection against deviations included in the uncertainty set is guaranteed through hard constraints that cut off all those feasible solutions
that may become infeasible for some deviations of the uncertainty set.
In a more formal way, suppose that we are given a generic integer linear
program:
v = max c′ x
subject to x ∈ F = {A x ≤ b, x ∈ Zn+ }
11
whose coefficient matrix A is uncertain, namely we do not know the exact value of its entries. However, we can identify a family A of coefficient
matrices that constitute possible valorizations of the uncertain matrix A.
The family A is the uncertainty set of the robust problem. A robust optimal solution, i.e. a solution that is protected against data deviations, can
then be found by considering the following robust counterpart of the original
problem:
v R = max c′ x
subject to x ∈ R = {Ā x ≤ b ∀Ā ∈ A, x ∈ Zn+ }
The feasible set R of the robust counterpart includes only the solutions of
F that are feasible for all the coefficient matrices in the uncertainty set
A. Consequently, R is a subset of the feasible set of the original problem,
i.e. R ⊆ F. We stress that the definition of robust counterpart can be
extended to any mixed-integer linear program, involving continuous and
integer variables at the same time. Additionally, we remark that the decision
maker should include in A coefficient matrices that reflect his specific risk
aversion.
Ensuring protection against data deviations according to an RO paradigm
comes at a price: the so-called price of robustness (PoR) (Bertsimas and Sim,
2004). The PoR is the deterioration of the optimal value of the robust counterpart w.r.t. the optimal value of the original problem (we have v R ≤ v)
and is caused by the hard constraints imposing robustness, which restrict
the feasible set F to the (in general smaller ) set of robust solutions R. The
PoR depends upon the features of the uncertainty set: uncertainty sets reflecting higher levels of risk aversion of the decision maker will include more
unlikely and extreme deviations, leading to higher protection yet associated
with higher PoR; uncertainty sets reflecting low risk aversion will instead
tend to neglect unlikely deviations, thus guaranteeing lower protection yet
associated with lower PoR.
Example 3 (protection against deviations). Following Example 2, a
simple way to grant protection would be to install sufficient capacity to deal
with the peak deviations of each commodity. So we should install 220+330
Mb of capacity. We note that in practice it is unlikely that all coefficients
will experience the worst deviation, so one of the aim of “smart” RO models
is to define appropriate uncertainty sets that result not too conservative,
while guaranteeing a satisfying level of protection (for example, we could
assume that at most one of the two demands will deviate from its nominal
value).
12
In the next paragraph, we provide a description of the model of uncertainty
that we adopt.
4.2. A concise introduction to Multiband Robust Optimization
In this work, we tackle uncertainty through Multiband Robust Optimization (MB), a new robust optimization model based on cardinalityconstrained uncertainty set that was proposed by Büsing and D’Andreagiovanni
(2012) and then extended and applied in a series of successive works (e.g,
Büsing and D’Andreagiovanni 2012, 2013, 2014; Büsing et al. 2014; Bauschert et al.
2014). MB represents a refinement and generalization of the well-known
Γ-Robustness (Γ-Rob) by Bertsimas and Sim (2004) that we developed to
satisfy practical needs of our industrial partners in real-world applications
(see Büsing and D’Andreagiovanni 2013; Bauschert et al. 2014).
We recall here the main results of MB, referring to the following generic
uncertain Mixed-Integer Linear Program (MILP):
X
max
cj xj
(M ILP )
j∈J
X
aij xj ≤ bi
i ∈ I = {1, . . . , m}
j∈J
xj ≥ 0
j ∈ J = {1, . . . , n}
xj ∈ Z +
j ∈ JZ ⊆ J .
where w.l.o.g we assume that the uncertainty only affects the coefficients aij
(uncertainty affecting cost coefficients or the right-hand-sides can be easily
reformulated as coefficient matrix uncertainty).
Moving from the hypothesis that the actual value aij of the coefficients
is unknown, the multiband uncertainty model at the basis of MB assumes
that:
1. for each coefficient aij , the decision maker knows its nominal value
−
K+
āij as well as the maximum negative and positive deviations dK
ij , dij
−
K+
from āij (so aij ∈ [āij + dK
ij , āij + dij ]);
K+
2. the overall single deviation band [dK−
ij , dij ] of each coefficient aij is
partitioned into K bands, defined on the basis of K deviation values:
+
−
−1
0
1
K
−∞ < dK
ij < · · · < dij < dij = 0 < dij < · · · < dij < +∞;
3. through these deviation values, K deviation bands are defined, namely:
a set of positive deviation bands k ∈ {1, . . . , K + } and a set of negative
13
deviation bands k ∈ {K − + 1, . . . , −1, 0}, such that a band k ∈ {K − +
k−1 k
1, . . . , K + } corresponds to the range (dij
, dij ], and band k = K −
−
corresponds to the single value dK
ij ;
4. for each constraint i ∈ I and each band k ∈ K, a lower bound lik and
an upper bound uik on the number of deviations that may fall in k are
defined, so 0 ≤ lik ≤ uik ≤ n;
5. the number of coefficients that take their nominal value is not limited,
i.e. u = n for all i ∈ I;
P i0
6.
k∈K lik ≤ n for all i ∈ I, so that there always exists a feasible
realization of the coefficient matrix.
We call this typology of uncertainty set a multiband uncertainty set.
MB thus generalizes the uncertainty definition of the Γ-Rob model by
Bertsimas and Sim (2004): the single deviation band is partitioned into multiple bands and each band k ∈ K is associated not only with an upper bound
uik , but also with a lower bound lik on the number of coefficients deviating
in that band. The lower bound improves the modeling power of the decision
maker and, more importantly, allows to take into account the presence of
negative value deviations, that are neglected in Γ-Rob. Of course, taking
into account negative deviations allows to improve the modeling of deviations commonly found in real-world problems and, even more critically,
reduces the value of the overall worst deviation and thus the price of robustness. The multiband model results particularly suitable to model histograms
that are commonly adopted by practitioners to visualize and analyze data
deviations (see Büsing and D’Andreagiovanni 2012; Bauschert et al. 2014).
Since the robust optimization paradigm entails that we must be protected against any possible deviation considered in the uncertainty set, the
robust counterpart of MILP under multiband uncertainty is:
X
cj xj
(5)
max
j∈J
X
āij xj + DEVi (x, M B) ≤ bi
i∈I
j∈J
xj ≥ 0
j∈J
xj ∈ Z +
j ∈ JZ ⊆ J ,
where an additional term DEVi (x, M B) is introduced in every feasibility
constraint to represent the maximum total deviation that could be incurred
by constraint i under the multiband uncertainty set for a solution x. This
14
problem is actually non-linear since the term DEVi (x, M B) hides the following maximization problem:
XX
k
DEVi (x, M B) = max
dkij xj yij
(6)
j∈J k∈K
lik ≤
X
k
yij
≤ uik
k∈K
(7)
j∈J
(8)
j ∈ J, k ∈ K .
(9)
j∈J
X
k
yij
≤1
k∈K
k
yij
∈
{0, 1}
k of the problem is equal to 1 when the coefficient j
A binary variable yij
in constraint i falls in deviation band k and is equal to 0 otherwise. Each
coefficient j in the constraint must fall in at most one deviation band, thus
requiring
the introduction of the family of constraints (7) (note that when
P
k = 0 then it is like having the coefficient falling in the zero deviation
y
k∈K ij
band). Constraints (8) impose the bounds on the number of coefficients that
may deviate in each band k ∈ K. Finally, the objective function (6) aims
at maximizing the deviation allowed by the multiband uncertainty set for a
given solution x and constraint i.
The robust counterpart (5) is a (non-linear) max-max problem, since it
includes the binary program (6-9). This is anyway not a real issue, since
we have proved that the robust counterpart is equivalent to a compact and
linear mixed-integer linear program, as stated in the following theorem.
Theorem 2 (Büsing and D’Andreagiovanni 2012). The robust counterpart of problem (MILP) for the multiband uncertainty set M B is equivalent
to the following compact and linear mixed-integer program:
X
max
cj xj
(Rob-MILP)
(10)
j∈J
X
āij xj +
X
θk wik +
X
i∈I
(11)
wik + zij ≥ dkij xj
i ∈ I, j ∈ J, k ∈ K
(12)
wik
i ∈ I, k ∈ K
(13)
zij ≥ 0
i ∈ I, j ∈ J
(14)
xj ≥ 0
j∈J
xj ∈ Z +
j ∈ JZ ⊆ J .
j∈J
k∈K
zij ≤ bi
j∈J
∈R
15
This problem includes K · m + n · m additional continuous variables (1314) and K · n · m additional constraints (12) to clear up the non-linearity
of the trivial robust counterpart (5). Moreover, constraints (11) include
additional terms and involve the values θk ≥ 0 that express the number of
deviations that occur in each deviation band k ∈ K (these values constitute
the so-called profile of the multiband uncertainty set and are derived on
the basis of the bounds lk , uk - see Büsing and D’Andreagiovanni (2013) for
details). The resulting formulation has thus the nice properties of being
compact and linear. The proof of the theorem is based on pointing out
the integrality of the polyhedron associated with and on exploiting strong
duality. We refer the reader to Büsing and D’Andreagiovanni (2012, 2013)
for the formal complete statement and proof of the presented theorem.
Theorem (2) is a central result from the theory of Multiband Robust
Optimization that we will use to derive a robust model for the MP-CNDP.
4.3. Multiband-Robust Multiperiod Network Design
We proceed now to use Multiband Robust Optimization and the related
Theorem (2) to tackle traffic uncertainty in the MP-CNDP. If we denote by
D the uncertainty set associated with the demands of the commodities, we
can write the general form of the robust counterpart of the MP-CNDP-IP
as follows:
XX
min
γet yet
(15)
e∈E t∈T
X
X
X
xtcp = 1
d¯tc xtcp + DEVet (x, D) ≤ φ
t
X
yeτ
e ∈ E, t ∈ T
(16)
τ =1
c∈C p∈Pc : e∈p
c ∈ C, t ∈ T
p∈Pc
xtcp ∈ {0, 1}
yet ∈ Z+
c ∈ C, p ∈ Pc , t ∈ T
e ∈ E, t ∈ T .
This robust counterpart differs from the MP-CNDP-IP in the capacity constraints (16): these constraints indeed consider the nominal traffic demands
values d¯tc and include the terms DEVet (x, D) to represent the total maximum
positive deviation that demands may experience on edge e in period t for a
routing vector x and the uncertainty set D.
We structure the uncertainty set D according to the principles of Multiband Robust Optimization introduced in the previous subsection. Specifically, we build a multiband uncertainty set for the MP-CNDP as follows:
16
1. for each commodity c ∈ C and time period t ∈ T , we know the nominal
value d¯tc of the traffic coefficient and maximum negative and positive
deviations δct− ≤ 0, δct+ ≥ 0 from it. The actual value dtc is then such
that dtc ∈ [d¯tc + δct− , d¯tc + δct+ ];
2. the overall deviation range [d¯tc + δct− , d¯tc + δct+ ] of each coefficient dtc is
partitioned into K bands, defined on the basis of K deviation values:
+
−∞ < δct− = δctK < · · · < δct−1 < δct0 = 0 < δct1 < · · · < δctK = δct+ < +∞;
−
3. through these deviation values, K deviation bands are defined, namely:
a set of positive deviation bands k ∈ {1, . . . , K + } and a set of negative
deviation bands k ∈ {K − + 1, . . . , −1, 0}, such that a band k ∈ {K − +
−
1, . . . , K + } corresponds to the range (dctk−1 , dtk
c ], and band k = K
−
;
corresponds to the single value dtK
c
4. for each capacity constraint (16) defined for an edge e ∈ E and period
t ∈ T and for each band k ∈ K, we introduce two values letk , utk
e :
t to represent the lower and upper bound on the
≤
n
0 ≤ letk ≤ utk
e
e
number of traffic coefficients whose value deviates in band k (nte is the
number of uncertain coefficients in the constraint). These bounds can
be used to derive the profile of the uncertainty set, namely the values
θetk ≥ 0 indicating the exact number of coefficients deviating in each
band (see Büsing and D’Andreagiovanni (2013) for details about the
definition of profile);
5. the number of coefficients that take their nominal value is not limited,
i.e. ut0 = nt for all e ∈ E, t ∈ T ;
P e tk e t
6.
k∈K le ≤ ne for all e ∈ E, t ∈ T , so that there always exists a
feasible realization of the coefficient matrix.
By using the previous characterization of the multiband uncertainty set and
Theorem (2), we can reformulate the non-linear robust counterpart (15) as
17
the following linear and compact robust counterpart:
XX
min
γet yet
(Rob-MP-CNDP)
e∈E t∈T
X
X
d¯tc xtcp +
c∈C p∈Pc : e∈p
+
X
θetk wetk
k∈K
+
X
X
t
zecp
≤φ
t
X
yeτ
e ∈ E, t ∈ T
τ =1
c∈C p∈Pc : e∈p
t
wetk + zecp
≥ δctk xtcp
e ∈ E, c ∈ C, p ∈ Pc : e ∈ p,
(17)
t ∈ T, k ∈ K
wetk ∈ R
t
zecp
≥0
X
e ∈ E, t ∈ T, k ∈ K (18)
e ∈ E, c ∈ C, p ∈ Pc : e ∈ p, t ∈ T
(19)
xtcp = 1
c ∈ C, t ∈ T
p∈Pc
xtcp ∈ {0, 1}
yet ∈ Z+
c ∈ C, p ∈ Pc , t ∈ T
e ∈ E, t ∈ T .
This formulation includes the additional constraints (17) and variables (18),
(19) to linearly reformulate the original (non-linear) problem including the
term DEVet (x, D) in each capacity constraint. Rob-MP-CNDP is the problem that we want to solve in the computational section to get robust solutions to the MP-CNDP.
5. A hybrid primal heuristic for the Rob-MP-CNDP
In principle, we can get a robust optimal solution for Rob-MP-CNDP by
using any commercial mixed-integer programming software, such as CPLEX.
However, as showed in the computational section, solving Rob-MP-CNDP
constitutes a difficult task even when considering a small number of time
periods and using a state-of-the-art solver like CPLEX: after several hours of
computation, solutions are still typically of low quality and far away from the
optimum. As a remedy, we were attracted by the effectiveness of (hybrid)
MIP-based and bio-inspired heuristics in hard network design problems.
Valid examples of such effectiveness are provided by Crainic and Gendreau
18
(2002), proposing a cooperative parallel tabu search algorithm for the singleperiod CNDP tested on transshipment networks, by Dely et al. (2013), proposing a linear relaxation-based decomposition method for a variant of the
single-period CNDP related to fair routing in wireless mesh networks, and
Kleeman et al. (2012), proposing a multiobjective evolutionary algorithm to
solve a single-period CNDP arising in the design of large communications
networks. Concerning hybrid heuristics, we refer the reader to Blum et al.
(2011) for a recent survey. In the case of Rob-MP-CNDP, we developed
a fast hybrid primal heuristic based on the combination of a randomized
variable-fixing strategy resembling Ant Colony Optimization (ACO) and an
exact large neighbourhood search.
It is widely known that ACO is a metaheuristic that was inspired by
the foraging behaviour of ants. The seminal work by Dorigo et al. (1996)
presenting an ACO algorithm for combinatorial problems has later been
extended to integer and continuous problems (e.g., Dorigo et al. 1999) and
has been followed by hundreds of other papers proposing refinements of
the basic algorithms (e.g., Gambardella et al. 2012; Maniezzo 1999) and investigating applications to relevant optimization problems (see Blum 2005
for an overview). An ACO algorithm presents the general structure specified in Algorithm 1. A loop consisting of two phases is executed until an
arrest condition is satisfied: in a first phase, an ant builds up a solution
under the guidance of probabilistic functions of variable fixing that resemble pheromone trails; then the pheromone trails are updated on the basis of
how effective the adopted variable fixing has resulted. Once that the arrest
condition is reached, a daemon action phase takes place and some solution
improvement strategy is applied to bring the feasible solution built by the
ants to a (local) optimum.
Algorithm 1 General structure of an ACO algorithm (Gen-ACO)
1: while an arrest condition is reached do
2:
ant-based solution construction
3:
pheromone trail update
4: end while
5: daemon actions
We now proceed to detail each phase of the previous sketch for our hybrid ACO-exact algorithm for the Rob-MP-CNDP. Our algorithm is defined
hybrid since after having passed the ACO construction phase, the daemon
action phase operates an exact large neighborhood search that is formulated
as an integer linear program and is solved exactly relying on the power of
19
modern state-of-the-art commercial solvers.
5.1. Ant-based solution construction.
In the first step of the inner cycle of (Gen-ACO), a number m ≥ 0 of ants
are defined and each ant iteratively builds a feasible solution for the problem.
At a generic iteration of the construction, the ant is in a state corresponding
with a partial solution and can make a further step towards completing the
solution by making a move: a move corresponds to fixing the value of some
not-yet-fixed variable. The move that is executed is chosen according to a
probability function: the function specifies the probability of implementing
a move and thus of fixing a variable and is derived by combining an a priori
measure of the efficacy of the move and an a posteriori measure of the
efficacy of the move. More in detail, the probability of choosing a move that
fixes a variable j after having fixed a variable i is specified by the following
canonical formula:
δ
τijβ + ηij
,
(20)
pij = P
β
δ
f ∈F τif + ηif
where τij represents the a priori measure of efficacy, commonly called pheromone
trail value, and ηij represents the a posteriori measure of efficacy, commonly
called attractiveness. Note that this formula is influenced by the two parameters β, δ that appear as exponents of the measures and should be chosen
by the decision maker on the basis of the specific problem considered. For
a more detailed description of the elements and actions of this phase, we
refer the reader to the paper (Maniezzo, 1999), which presents ANTS, an
improved ant algorithm used for solving the quadratic assignment problem,
which we have taken as reference for our work. We considered ANTS particularly attractive as it proposes a series of refinements for classical ACO that
allow to better exploit polyhedral information about the problem. Specifically, (Maniezzo, 1999) sketches some ideas about how alternative formulations of the original problem could be exploited to define the pheromone
trail and the attractiveness values. As additional desirable feature, ANTS
makes use of a reduced number of parameters and adopts more efficient
mathematical operations w.r.t. the canonical ant algorithms (products instead of exponentiations). For an exhaustive description of ANTS, we refer
the reader to the paper by Maniezzo (1999).
Before describing how our ANTS implementation is structured, we make
some preliminary considerations. The formulation Rob-MP-CNDP is based
on four families of variables: 1) the path assignment variables xtcp ; 2) the
t
coming from
capacity variables yet ; 3-4) the auxiliary variables wetk , zecp
20
robust dualization. Though we have to deal with four families, we can
notice that once routing decisions are taken over the entire time horizon,
then we can immediately derive the capacity installation of minimum cost.
Indeed, once the values of all path assignment variables are fixed, the routing
is completely established and the worst traffic deviation term DEVet (x, D)
can be efficiently derived without need of using the auxiliary variables wetk ,
t
(this can be indeed traced back to solving a min-cost flow problem, as
zecp
explained in Büsing and D’Andreagiovanni 2012, 2013). So we can easily
derive the total traffic Det sent over an edge e in period t in the worst
case. We can then derive the minimum cost installation by a sequential
evaluation
from period 1 to period |T |, keeping in mind that we must have
m
l
Det
capacity modules on e in t to accommodate the traffic. In the antφ
construction phase, we can consequently limit our attention to the binary
assignment variables and we introduce the concept of routing state.
S
Definition 2 (Routing state - RS). Let P = c∈C Pc and let R ⊆ C ×
P ×T be the subset of triples (c, p, t) representing the assignment of path p ∈
Pc to commodity c ∈ C in period t ∈ T . A routing state is an assignment of
paths to a subset of commodities in a subset of time periods which excludes
that multiple paths are assigned to a single commodity. Formally:
RS ⊆ R : 6 ∃(c1 , p1 , t1 ), (c2 , p2 , t2 ) ∈ RS : c1 = c2 ∧ p1 , p2 ∈ Pc1 ∧ t1 = t2 .
We say that a routing state RS is complete when it specifies the path used
by each commodity in each time period (thus |RS| = |C||T |). Otherwise
the RS is called partial and we have |RS| < |C||T |).
In the ANTS algorithm that we propose, we decided to assign paths considering time periods and commodities in a pre-established order. Specifically, we establish the routing in each time period separately, starting from
t = 1 and continuing up to t = |T |, and in each time period commodities are
sorted in descending order w.r.t. their nominal traffic demand. Formally,
this is operated through the cycle specified in Algorithm 2 that builds a
complete routing state.
For an iteration (t, c) of the nested cycles of Algorithm 2, the assignment
of a path to a commodity is equivalent to an ant that moves from a partial
routing state RSi to a partial routing state RSj such that:
RSj = RSi ∪ {(c, p, t)} with p ∈ Pc .
We note that, by the definition of routing state, a sequence of moves is
actually a sequence of fixings of decision variables, as in (Maniezzo, 1999).
21
Algorithm 2 Construction of a complete routing state
1: for t := 1 to |T | do
2:
sort c ∈ C in descending order of d¯ct
3:
for (sorted c ∈ C) do
4:
assign a single path p ∈ Pc to c
5:
end for
6: end for
The probability that an ant k moves from a routing state i to a more
complete routing state j, chosen among a set of feasible routing states, is
defined by the improved formula proposed by Maniezzo (1999):
pkij = P
α τij + (1 − α) ηij
,
f ∈F α τif + (1 − α) ηif
where α ∈ [0, 1] is a parameter assessing the relative importance of trail
and attractiveness. This formula presents two peculiar advantages over the
canonical formula (20): it adopts the single parameter α in place of the two
parameters β, δ and uses α as coefficient of a product instead of an index of
an exponentiation.
As discussed in Maniezzo (1999), the trail values τij and the attractiveness values ηij should be provided by suitable lower bounds of the considered
optimization problem. In our particular case: 1) τij is derived from the values of the variables in the solution associated with the linear relaxation of
the robust counterpart Rob-MP-CNDP; 2) ηij is equal to the value of a
(good) feasible solution of the linear relaxation MP-CNDP-LP where a subset of the decision variables have fixed values as effect of the fixing decision
taken in previous steps of the algorithm.
5.2. Daemon actions: Relaxation Induced Neighborhood Search
At the end of the ant-construction phase, we attempt at improving the
quality of the feasible solution found by executing an exact local search in a
large neighborhood. In particular, we adopt a modified Relaxation Induced
Neighborhood Search (RINS) (see Danna et al. (2005) for an exhaustive description of the method).
In an integer linear program, we look for a solution that is integral and
at the same time guarantees the best objective value. In RINS, it is observed
that an optimal solution of the linear relaxation of the problem provides an
objective value that is better than that of any feasible solution. However,
at the same time, such an optimal solution is fractional and therefore does
22
not guarantee integrality. On the contrary, a feasible solution guarantees
integrality, but provides a worse objective value. The fact that a variable is
fixed to the same value in the optimal solution of the linear relaxation and in
a feasible solution is a good indication that the fixing of the variable is good
and should thus be maintained. Given a feasible solution, RINS profits from
these observations by defining search neighborhoods, where those variables
that have the same value in the feasible solution and in the optimal solution
of the linear relaxation are fixed, while the other are free to vary their value.
The neighborhood is then explored exhaustively by formulating the search
as an integer linear program, which is solved exactly, possibly including an
arrest condition such as a solution time limit.
In our specific case, let (x̄, ȳ) be a feasible solution of Rob-MP-CNDP
found by an ant and (xLR , y LR ) be an optimal (continuous) solution of the
linear relaxation of Rob-MP-CNDP. Our modified RINS (mod-RINS) entails
to solve exactly through an optimization solver like CPLEX a subproblem
of Rob-MP-CNDP where:
1. we fix the variables x whose value in (x̄, ȳ) and (xLR , y LR ) differs of
at most ǫ > 0, i.e.:
x̄j = 0 ∩ xLR
≤ ǫ =⇒ xj = 0
j
x̄j = 1 ∩ xLR
≥ 1 − ǫ =⇒ xj = 1
j
2. impose a solution time limit of τ to the optimization solver.
A time limit is imposed since the subproblem may be difficult to solve, so
the exploration of the large neighbourhood of x̄, ȳ may need to be truncated.
Note that in point 1 we generalize the fixing rule of RINS, in which ǫ = 0.
We thus allow that fixed variables may differ up to ǫ in value, in contrast
to canonical RINS where they must have exactly the same value in the
relaxation and in the feasible solution.
5.3. Pheromone trail update
At the end of each ant-construction phase h, we update the pheromone
trails of a move τij (h − 1) according to an improved formula proposed by
Maniezzo (1999):
τij (h) = τij (h − 1) +
m
X
k=1
τijk
z k − LB
,
with τijk = τij (0) · 1 − curr
z̄ − LB
(21)
where the values τij (0) and LB are set by using the linear relaxation of
Rob-MP-CNDP: τij (0) is set equal to the values of the corresponding optimal decision variables and LB equal to the optimal value of the relaxation.
23
k
Additionally, zcurr
is the value of the solution built by ant k and z̄ is the
moving average of the values of the last ψ feasible solutions built. The
formula (21) has the desirable property of replacing the pheromone evaporation factor, a parameter whose setting may result tricky, with the moving
average ψ, whose setting has been proved to be much less critical.
Algorithm 3 shows the structure of our original hybrid exact-ACO algorithm.
The algorithm is based on the execution of two nested loops: the outer
loop is repeated until a time limit is reached and, during each execution
of it, an inner loop defines m ants to build the solutions. Pheromone trail
updates are done at the end of each execution of the inner loop. Once the
ant construction phase is over, mod-RINS is applied so to try to get an
improvement by exact large neighborhood search.
Algorithm 3 Hybrid ACO-exact algorithm for (Rob-MP-CNDP)
1: compute the linear relaxation of (Rob-MP-CNDP) and initialize the
values of τij (0) by it.
2: while the time limit is not reached do
3:
for µ := 1 to m do
4:
build a complete routing state
5:
derive a complete feasible solution for (Rob-MP-CNDP)
6:
end for
7:
update τij (t) according to (21)
8: end while
9: apply mod-RINS to the best feasible solution
6. Experimental results
In order to assess the performance of our hybrid algorithm, we executed computational tests on a set of 30 instances, based on realistic network topologies from the SNDlib (2010) and defined in collaboration with
industrial partners from former and ongoing industrial projects (see e.g.,
Bley et al. 2011, 2013; Bauschert et al. 2014). The 30 instances consider 10
network topologies, whose main features are presented in Table 1. For each
instance, in Table 1 we report its identification code (ID ) and features (|V |
= no. vertices, |E| = no. edges, |C| = no. commodities). For each network
topology, we defined three instances of Rob-MP-CNDP, each considering a
distinct number of time periods, namely 5, 7, and 10. We performed the
experiments on a machine with a 2.40 GHz quad-core processor and 16 GB
24
Table 1: Features of the instances
Name
ID
|V |
|E|
|C|
Germany50
Pioro40
Norway
Geant
France
Dfn-Gwin
Pdh
Ta1
Polska
Cost266
I1
I2
I3
I4
I5
I6
I7
I8
I9
I10
50
40
27
22
25
11
11
24
12
37
88
89
51
36
45
47
34
55
18
57
662
780
702
462
300
110
24
396
66
1337
of RAM and using the mixed-integer commercial solver IBM ILOG CPLEX
version 12.4.
All the instances lead to (very) large and hard to solve Rob-MP-CNDP.
We observed that even a state-of-the-art solver like CPLEX had big difficulties in identifying good feasible solutions and in the majority of cases the
final optimality gap was over 95%. In contrast, as clear from Table 2, in most
cases our hybrid primal heuristic was able to find very high quality solutions
associated with very low optimality gaps. The optimality gap indicates how
far the best feasible solution found of value v ∗ is from the best lower bound
LB available on the optimal value (formally gap% = |v ∗ − LB|/v ∗ · 100).
We note that using CPLEX is essential for providing a lower bound in our
problem instances and thus a quality guarantee for the given solutions. In
the case of solutions not produced by CPLEX, we computed the optimality
gap referring to the lower bound produced by CPLEX.
On the basis of preliminary tests, we found that an effective setting of the
parameters of our heuristic was: α = 0.5 (we thus balanced attractiveness
and trail level), m = 10000 ants, ψ = m/10 (width of the moving average
equal to 1/10 the number of ants), ǫ = 0.1 (tolerance of fixing in mod-RINS),
T = 30 minutes (time limit imposed to the execution of mod-RINS). Each
commodity admits 5 feasible paths, i.e. |Pc | = 5, ∀c ∈ C and 5 deviations
bands (2 positive, 2 negative and the null deviation band). In contrast to
our first computational experience presented in (D’Andreagiovanni et al.,
2014), where linear relaxations were solved exactly by CPLEX requiring a
non-negligible amount of time, we now 1) compute the linear relaxations
of nominal problems by the results presented in Section 3 and 2) we solve
25
the linear relaxation of Rob-MP-CNDP by a time-truncated primal simplex
method as implemented by CPLEX. This allowed to greatly reduce the time
of execution of the ant construction phase and hugely increase the number
of defined ants.
The complete set of results is presented in Table 2, where we show the
performance of the hybrid solution approach, that is denoted by the three
measures c∗ (ACO), c∗ (ACO+RINS), gapAR%, which respectively represent
the value of the best solution found by pure ACO, the value of the best solution found by ACO followed by RINS and the corresponding final optimality
gap. Moreover, we show the performance of CPLEX, which is denoted by
measures c∗ (IP) and gapIP% representing the value of the best solution
found and the corresponding final optimality gap. The value of the best
solutions found for each instance is highlighted in bold type. The overall
time limit for the execution of the heuristic was 1 hour. The same time limit
was imposed on CPLEX when used to solve the robust counterpart Rob-MPCNDP. We stress that increasing the time limit did not bring any remarkable
benefit to CPLEX: even when letting CPLEX run for many hours, the only
effects were getting negligible improvements in the best lower bound (we
observed an extremely slow improvement rate of the bound) and running
out of memory because of the huge size of the search trees generated. In
the case of two instances of I10, CPLEX was not even able to find a feasible
solution within the time limit (cases denoted by ∗).
The best solutions found by our hybrid algorithm have in most cases a
value that is at least one order of magnitude better than those found by
CPLEX (3700% better on average, excluding of course the cases for which
CPLEX did not find any feasible solutions). The results are of very high
quality and, given the very low optimality gap, we can suppose that some
of these solutions are actually optimal. The much higher performance is
particularly evident for instances I5, I8 and I9. We observe that, in contrast
to our conference paper (D’Andreagiovanni et al., 2014) where RINS was
able to improve the value of the best solution found by the ant construction
phase, in these new computations the role of RINS was reduced: thanks
to the dramatic speed-up in solving the linear relaxations, we were able to
really implement a swarm exploration of the feasible set, getting solutions
that we believe to be optimal or very close to the optimum. So finding an
improvement through RINS was impossible or really unlikely. We also think
that the fact that a powerful local search like RINS has a minor or null role
in finding higher quality solutions is an indication that we have defined a
very strong and effective ant construction phase: in contrast to common
experience in ACO, we indeed do not really require a final daemon/local
26
search phase to get solution of sufficiently high quality.
If we focus on a specific network topology, it may be noticed that the
optimality gaps produced by our algorithm tend to become smaller as the
number of time periods grows. This may result surprising and counterintuitive at first sight, as one would expect exactly the opposite behaviour.
Yet, considering our assumption of non-increasing demands for each commodity, it is reasonable to think that our primal hybrid heuristic discovers
converging near-optimal paths when the number of time periods is large,
thus performing better. We observe that it is likely that this behaviour
would not occur in the case that the non-increasing demand assumption is
dropped.
As further way to assess the validity and effectiveness of our new heuristic, we also used as benchmark a fast and easy method to define a possibly
good feasible solution to Rob-MP-CNDP, which is suggested by the theoretical results presented in Section 3: in each time period, we simply route
the entire flow of each commodity on its shortest path. However, as clear
from Table 2, where we report the value c(SP ) of the solution found by this
approach and the corresponding optimality gap gapSP %, in the vast majority of the instances (in particular instances I4, I6, I7 and I8, for which the
performance of our algorithm is far better), this simple approach performed
much worse than our heuristic and we believe that this provides further
evidence of the solidity of our algorithm.
7. Conclusion and future work
In this paper, we have introduced the first Robust Optimization model
to handle the uncertainty that affects traffic demands in the Multiperiod
Capacitated Network Design Problem (MP-CNDP). Data uncertainty may
compromise the quality and feasibility of produced solutions, so, as a remedy, we have proposed a Multiband Robustness model, following the wellestablished methodology of Robust Optimization. We have thus produced
robust solutions that are protected against deviations of input traffic data.
In general, the MP-CNDP already constitutes a challenging problem, even
for state-of-the-art commercial solvers like CPLEX. Accounting for robustness and considering multiple time periods has the effect of further increasing the complexity of the problem. As a matter of fact, solutions found by
CPLEX are of low quality and associated with very large optimality gaps.
To overcome these difficulties, we have defined a hybrid primal heuristic
based on the combination of a randomized fixing algorithm inspired by ant
27
Table 2: Experimental results
ID
I1
I2
I3
I4
I5
I6
I7
I8
I9
I10
|T |
c∗ (ACO)
c∗ (ACO+RINS)
gapAR%
c∗ (SP )
gapSP%
c∗ (IP)
gapIP%
5
7
10
5
7
10
5
7
10
5
7
10
5
7
10
5
7
10
5
7
10
5
7
10
5
7
10
5
7
10
4.26E+06
9.46E+06
3.07E+07
6.14E+06
1.34E+07
4.32E+07
3.14E+06
6.39E+06
2.02E+07
8.98E+05
2.15E+06
6.49E+06
1.34E+05
2.96E+05
9.43E+05
3.34E+05
6.54E+05
2.08E+06
1.19E+08
2.71E+08
8.29E+08
1.45E+08
2.83E+08
9.14E+08
2.12E+05
4.80E+05
1.60E+06
1.37E+08
2.78E+08
9.26E+08
4.26E+06
9.46E+06
3.07E+07
6.14E+06
1.34E+07
4.32E+07
3.14E+06
6.39E+06
2.02E+07
8.98E+05
2.15E+06
6.49E+06
1.34E+05
2.96E+05
9.43E+05
3.34E+05
6.54E+05
2.08E+06
1.19E+08
2.71E+08
8.29E+08
1.39E+08
2.83E+08
9.14E+08
2.12E+05
4.80E+05
1.60E+06
1.37E+08
2.78E+08
9.26E+08
3.8
2.5
2.0
11.6
11.5
8.4
14.6
6.5
2.2
2.6
11.6
8.8
3.0
1.4
0.7
12.2
5.6
1.7
1.1
0.5
0.2
9.0
3.8
2.1
1.1
0.3
0.1
20.7
12.7
14.9
4.35E+06
9.64E+06
3.10E+07
7.48E+06
1.64E+07
5.45E+07
3.61E+06
7.45E+06
2.37E+07
2.09E+06
4.94E+06
1.98E+07
1.35E+05
2.97E+05
9.42E+05
4.14E+05
8.37E+05
2.77E+06
1.50E+08
3.64E+08
1.27E+09
1.60E+08
3.36E+08
1.08E+09
2.22E+05
5.02E+05
1.68E+06
1.36E+08
3.02E+08
9.84E+08
5.9
4.3
3.0
27.4
27.6
27.3
25.6
19.7
16.8
58.0
61.6
70.1
3.7
1.7
0.6
29.1
26.2
25.9
21.7
25.8
34.8
21.2
19.1
17.2
5.5
4.9
4.9
20.1
19.8
19.8
3.42E+08
8.28E+08
2.89E+09
2.58E+08
6.19E+08
2.13E+09
9.56E+07
2.29E+08
7.88E+08
4.13E+07
9.58E+07
4.18E+08
5.09E+06
1.56E+07
4.46E+07
7.07E+05
6.69E+05
9.84E+07
1.19E+08
2.73E+08
8.34E+08
7.02E+09
1.86E+10
6.44E+10
5.13E+05
6.08E+05
9.80E+06
5.40E+09
*
*
98.8
98.9
99.0
97.9
98.1
98.1
97.2
97.4
97.5
97.9
98.0
98.6
97.5
98.1
97.9
58.5
7.7
97.9
1.1
0.9
0.7
98.2
98.5
98.6
59.2
21.5
83.7
98.0
*
*
28
colony optimization and an exact large neighborhood search. Computational experiments on a set of 30 realistic instances from the SNDlib have
confirmed that our heuristic drastically outperforms CPLEX, finding high
quality solutions associated with low optimality gaps in a short amount of
time. We believe that many of the best solutions found are actually optimal, so a future objective will be to characterize appropriate families of valid
inequalities for the problem, in the attempt to close the gaps and thus possibly prove the optimality of the found solutions. Furthermore, the excellent
computational performance suggests the possibility of using the heuristic,
conveniently adapted, for other applications and in more general settings,
where, for example, the paths of each commodity are not predetermined.
We expect our hybrid primal heuristic to perform well even in such different
contexts.
Acknowledgements
This work was partially supported by the German Research Foundation
(DFG), project Multiperiod Network Optimization, by the DFG Research
Center MATHEON (www.matheon.de), Project B3, by the Einstein Campus
for Mathematics Berlin (ECMath) Project ROUAN, and by the German
Federal Ministry of Education and Research (BMBF), Project ROBUKOM
(Bauschert et al., 2014), grant 05M10PAA.
Appendix A. Proofs of the Statements of Section 3
Appendix A.1. Proof of Proposition 1
Statement: Let x̄tcp := 0, dtcp := 0 and γet := ∞ for t = 0. There exists an
optimal solution x̄ of MP-CNDP-LP such that:
dt−1
x̄t−1
≤ dtc x̄tcp
c
cp
∀ c ∈ C, p ∈ Pc , t ∈ T
Proof. We prove the statement by induction, considering two consecutive
time periods t − 1 and t with t = {1, 2, . . . , |T |}. As basis step, we note that
relation (4) trivially holds for t = 1 by definition of x̄tcp := 0, dtcp := 0 for
t = 0. As inductive step, suppose that relation (4) holds for t ≤ |T | − 1.
We will show that it holds also for t ≤ |T | period. Suppose that relation
t−1
t t
(4) does not hold for t = |T |, i.e. ∃ c∗ ∈ C, p∗ ∈ Pc∗ : dt−1
c∗ x̄c∗ p∗ > dc∗ x̄c∗ p∗
for some commodity and its corresponding path when t = T . Without
29
loss of generality, we assume that there exists unique c∗ ∈ C, p∗ ∈ Pc∗
that does not respect relation (4) for period |T |, as the argument we use
below can be applied iteratively to any number of such deviating inequalities.
Moreover, suppose that p∗ ∈ Pc is a shortest path in edge cost for c∗ . Let
t−1
t t
∆ = dt−1
c∗ x̄c∗ p∗ − dc∗ x̄c∗ p∗ be the residual flow. For simplicity we assume that
∆ is routed over a single path as the argument similarly holds when it splits
over several paths. Given our assumption, it is easy to see that ∆ will be
routed on another
p′ ∈ PP
c∗ and that the objective function value will
P path
t
gain at most
γet ∆. When p′ is another shortest path,
γe ∆ ≥
e∈E:e∈p′
e∈E:e∈p∗
t−1
t
t
we get an equality and we see that xt−1
c∗ p∗ dc∗ ≤ xcp∗ dc∗ gives an equivalent
optimal solution in terms of the objective function value, otherwise we have
a contradiction.
Suppose
p∗ in xtc∗ p∗ is not a shortest path in edge cost for
P
P
γet ≤
γet then the same argument as above applies.
c∗ . If
e∈E:e∈p∗
e∈E:e∈p′
P
P
P
γet ∆ −
γet >
γet then the solution gains at most
If
∗
∗
′
e∈E:e∈p
e∈E:e∈p
e∈E:e∈p
P
γet ∆ > 0. If this is the case, we can reroute ∆ in period |T | − 1
e∈E:e∈p′
from p∗
t−1
= x̄tc∗ p∗ dtc∗ in period T, when in
to p′ . Then we have x̄t−1
c∗ p∗ dc∗
t−2
≤ x̄tc∗ p∗ dtc∗ holds. Then this new solution
the deviating inequality x̄t−2
c∗ p∗ dc∗ P
P
P
P
γet ∆ −
γet−1 ∆ −
γet ∆
γet−1 ∆ ≥
gains at most
e∈E:e∈p∗
e∈E:e∈p′
e∈E:e∈p∗
e∈E:e∈p′
by the non-increasing cost per unit of capacity. The same argument holds
t−2
t
t
when x̄t−2
c∗ p∗ dc∗ > x̄c∗ p∗ dc∗ .
Appendix A.2. Proof of Theorem 1
Statement: Consider problem MP-CNDP-LP, namely the linear relaxation
of MP-CNDP-IP, and let p∗c ∈ Pc be a shortest path in edge length for each
commodity c ∈ C. An optimal solution of MP-CNDP-LP can be defined by
routing in each time period t ∈ T the entire flow of each commodity c ∈ C on
the shortest path p∗c and by installing on each edge of p∗c the exact capacity
needed to route the traffic flow dct .
Proof. Before proceeding to the proof, we recall that in MP-CNDP-LP
the capacity variables lose their integrality constraint and thus the capacity
installation on each edge in each period will exactly equal the flow. It is
well known that the result stated in the theorem holds for one single period
(see Ahuja et al. 1993) and we propose a proof of the result, extending
below the reasoning to the case with multiple periods. Consider then a
30
single-period problem (in this case, the t apex is thus equal to 1 for all
the involved quantities) and a commodity c ∈ C. Suppose that x1cp =
k with 0 < k ≤ 1 is a non-zero variable in the optimal solution of the
problem
for some p ∈ Pc . Hence the contribution to thePobjective function is
P
γe1 d1c k. Let p′ be a shortest path for c. Then
γe1 d1c (x1cp′ +k) ≤
′
e∈E:e∈p
e∈E:e∈p
P
P
γe1 d1c x1cp . Thus the theorem holds for the single
γe1 d1c x1cp′ +
e∈E:e∈p′
e∈E:e∈p
period case. Assume now that the theorem holds for |T |−1 periods. We will
show that it holds for |T | periods. By Proposition 1, we may assume that
t−1 ≤ xt dt for all c ∈ C, p ∈ P , t ∈ T in an optimal solution of MPxt−1
c
cp dc
cp c
CNDP-LP with |T | periods. Let OP T (|T | − 1) denote the optimal objective
function value for |T |−1 periods according to the induction hypothesis. Also,
let OP Tt denote the contribution of the optimal solution of the MP-CNDPP
OP Tt =
LP for |T | periods to the objective function in period t. Then
t∈T
P
P
OP Tt + OP T|T | . Suppose
OP Tt is determined by, for each
t∈T \|T |
t∈T \|T |
commodity, the same shortest
P path in edge cost for each period up to |T |−1.
OP Tt . Using a similar argument as in the
Hence OP T (|T | − 1) =
t∈T \|T |
single period case, it is easy to see that the increase in demand for each
commodity in period |T | will be routed in the shortest paths that were used
in the previous periods. For this case, let OP T |T | denote the contribution
of the optimal solution
of MP-CNDP-LP to the objective function in period
P
OP Tt is not determined by, for each commodity, the
|T |. Suppose
t∈T \|T |
same shortest path in edge cost
P for each period up to |T | − 1. Then it is
clear that OP T (|T | − 1) ≤
OP Tt and OP T (|T |) ≤ OP T|T | . Hence
t∈T \|T |
P
OP T (|T | − 1) + OP T (|T |) ≤
OP Tt + OP T|T | .
t∈T \|T |
References
References
Ahuja, R.K., Magnanti, T., Orlin, J.: Network Flows: Theory, Algorithms,
and Applications. Prentice Hall, Upper Saddle River (1993)
Bauschert, T., Büsing, C., D’Andreagiovanni, F., Koster, A.M.C.A.,
Kutschka, M., Steglich, U.: Network Planning under Demand Uncertainty
with Robust Optimization. IEEE Communications Magazine, 52 (2), 178185 (2014) DOI: 10.1109/MCOM.2014.6736760
31
Belotti, P., Kompella, K., Noronha, L.: A comparison of OTN and
MPLS networks under traffific uncertainty. Submitted in 2011.
http://myweb.clemson.edu/∼pbelott/papers/robust-opt-network-design.pdf
(retrieved on 07.03.2014) (2011)
Ben-Tal, A., El Ghaoui, L., Nemirovski, A.: Robust Optimization. Springer,
Heidelberg (2009)
Bertsekas, D.P.: Network Optimization: Continuous and Discrete Models.
Athena Scientific, Belmont (1998)
Bertsimas, D., Brown, D., Caramanis, C.: Theory and Applications of Robust Optimization. SIAM Review 53 (3), 464-501 (2011)
Bertsimas, D., Sim, M.: The Price of Robustness. Oper. Res. 52 (1), 35-53
(2004)
Bienstock, D., Chopra, S., Günlük, O, Tsai, C.: Minimum cost capacity
installation for multicommodity network flows. Math. Prog. 81 (2), 177199 (1998)
Bley, A., D’Andreagiovanni, F, Hanemann, D.: Robustness in Communication Networks: Scenarios and Mathematical Approaches. In: Proc. of
the ITG Symposium on Photonic Networks 2011, 1-8, VDE Verlag, Berlin
(2011)
Bley, A., D’Andreagiovanni, F, Karch, D.: WDM Fiber Replacement
Scheduling. In: Proc. of INOC 2013, Electronic Notes in Discrete Mathematics 41 (5), 189-196 (2013) DOI: 10.1016/j.endm.2013.05.092
Bley, A., Grötschel, M., Wessäly, R.: Design of broadband virtual private
networks: Model and heuristic for the B-WiN. DIMACS Series in Discrete
Mathematics and Theoretical Computer Science 53, 1-16 (2000)
Blum, C.: Ant colony optimization: Introduction and recent trends. Physics
Life Rew. 2 (4), 353-373 (2005)
Blum., C., Puchinger, J., Raidl, G.R., Roli., A.: Hybrid metaheuristics in
combinatorial optimization: A survey. Appl. Soft Comp. 11 (6), 4135–
4151 (2011)
Büsing, C., D’Andreagiovanni, F.: New Results about Multi-band Uncertainty in Robust Optimization. In: Klasing, R. (ed.) Experimental Algorithms - SEA 2012, LNCS, vol. 7276, pp. 63-74. Springer, Heidelberg
(2012) DOI: 10.1007/978-3-642-30850-5 7
32
Büsing, C., D’Andreagiovanni, F.:
Robust Optimization under
Multi-band Uncertainty - Part I: Theory. CoRR abs/1301.2734,
http://arxiv.org/abs/1301.2734 (2013)
Büsing, C., D’Andreagiovanni, F.: A new theoretical framework for Robust
Optimization under multi-band uncertainty. In: Helber, S. et al. (eds.)
Operations Research Proceedings 2012, pp. 115-121. Springer, Heidelberg
(2014)
Büsing, C., D’Andreagiovanni, F., Raymond, A.: 0-1 Multiband Robust Optimization. In: Huisman, D. et al. (eds.) Operations Research Proceedings
2013, pp. 89-95. Springer, Heidelberg (2014)
Christofides, N., Brooker, P.: Optimal Expansion of an Existing Network.
Math. Program. 6, 197-211 (1974)
Crainic, T.G., Gendreau, M.: Cooperative Parallel Tabu Search for Capacitated Network Design. J. Heuristics 8 (6), 601-627 (2002)
Danna, E., Rothberg, E., Le Pape, C.: Exploring relaxation induced neighborhoods to improve MIP solutions. Math. Program. 102, 71-90 (2005)
D’Andreagiovanni, F., Krolikowski, J., Pulaj, J.: A hybrid primal heuristic
for Robust Multiperiod Network Design. To appear in Esparcia-Alcazar,
A. (ed.) Applications of Evolutionary Computation - 17th European
Conference, EvoApplications 2014, Lecture Notes in Computer Sciences.
Springer, Heidelberg (2014)
Dely, P., D’Andreagiovanni, F., Kassler, A.: Fair optimization of meshconnected WLAN hotspots. Wireless Communications and Mobile Computing (2013) DOI: 10.1002/wcm.2393
Dorigo, M., Di Caro, G., Gambardella, L.M.: Ant algorithms for discrete
optimization. Artificial Life 5 (2), 137-172 (1999)
Dorigo, M., Maniezzo, V., Colorni, A.: Ant System: Optimization by a
colony of cooperating agents. IEEE Trans. Syst. Man Cybern. B 26 (1),
29-41 (1996)
Gambardella, L.M., Montemanni, R., Weyland, D.: Coupling ant colony
systems with strong local searches. Europ. J. Oper. Res. 220 (3), 831-843
(2012)
33
Gamvros, I., Raghavan, S.: Multi-period traffic routing in satellite networks.
Europ. J. Oper. Res. 219 (3), 738-750 (2012)
Gendreau, M., Potvin, J., Smires A., Soriano P.: Multi-period capacity
expansion for a local access telecommunications network. Europ. J. Oper.
Res., 172 (3), 1051-1066 (2006)
Gupta, V., Grossmann, I.E.: An Efficient Multiperiod MINLP Model for
Optimal Planning of Offshore Oil and Gas Field Infrastructure. Ind. Eng.
Chem. Res., 51 (19), 68236840 (2012)
IBM ILOG CPLEX, http://www-01.ibm.com/software/integration/optimization/cplex-optimizer.
Kleeman, M.P., Seibert, B.A., Lamont, G.B., Hopkinson, K.M., Graham,
S.R.: Solving Multicommodity Capacitated Network Design Problems Using Multiobjective Evolutionary Algorithms. IEEE Trans. Evol. Comp. 16
(4), 449-471 (2012)
Koster, A.M.C.A., Kutschka, M., Raack, C.: Robust network design: Formulations, valid inequalities, and computations. Networks 61 (2), 128-149
(2013)
Lardeux, B., Nace, D., Geffard, J.: Multiperiod network design with incremental routing. Networks 50 (1), 109-117 (2007)
Maniezzo, V.: Exact and Approximate Nondeterministic Tree-Search Procedures for the Quadratic Assignment Problem. INFORMS J. Comp. 11
(4), 358-369 (1999)
Orlowski, S., Wessäly, R., Pioro, M., Tomaszewski, A.: SNDlib 1.0 - Survivable Network Design Library. Networks 55 (3), 276-286 (2010)
Theimer, T: Towards a scalable and flat IP core network. ECOC 2009
Vienna, Austria. (retrieved 07.03.2014)
http://conference.vde.com/ecoc-2009/programs/documents/ws4 thomas%20theimer.pdf
34
| 9cs.NE
|
arXiv:1702.03368v1 [math.GR] 11 Feb 2017
On the problem of existence and conjugacy of injectors of
generalized π-soluble groups∗
Xia Yin, Nanying Yang
School of Science, Jiangnan University, Wuxi 214122, P. R. China
E-mail: [email protected]
N.T. Vorob’ev
Department of Mathematics, Masherov Vitebsk State University, Vitebsk 210038, Belarus
E-mail: [email protected]
Abstract
In this paper, we prove the existence and conjugacy of injectors of a generalized π-soluble
groups for the Hartley class defined by a invariable Hartley function, and give a description of
the structure of the injectors.
1
Introduction
Throughout this paper, all groups are finite and p is a prime. G always denotes a group, |G| is the
order of G, σ(G) is the set of all primes dividing |G|, π denotes a set of some primes. Let P be the
set of all primes, then let π ′ = P \ π. We use Gπ denote a Hall π-subgroup of G.
Recall that a class F of groups is called a Fitting class if F is closed under taking normal subgroups
and products of normal F-subgroups. As usual, we denote by E, S, N the classes of all groups, all
soluble groups, all nilpotent groups, respectively; Eπ , Sπ , Nπ denote the classes of all π-groups, all
soluble π-groups, all nilpotent π-groups, respectively; and Sπ and Nπ to denote the classes of all
π-soluble groups and all π-nilpotent groups, respectively. It is well known that all the above classes
are Fitting classes.
From the definition of Fitting class, we see that for each non-empty Fitting class F, every group
G has a unique maximal normal F-subgroup, which is called the F-radical of G and denoted by GF .
For any non-empty class F of groups, a subgroup V of G is said to be F-maximal if V ∈ F and
U = V whenever V ≤ U ≤ G and U ∈ F. A subgroup V of a group G is said to be an F-injector of
∗
Research is supported by a NNSF grant of China (Grant # 11301227) and Natural Science Foundation of Jiangsu
Province (grant # BK20130119)
Keywords: Finite group, Fitting class, injector, π-soluble group
Mathematics Subject Classification (2010): 20D10
1
G if V ∩ K is an F-maximal subgroup of K for every subnormal subgroup K of G. Clearly, every
F-injector of G is an F-maximal subgroup of G.
Fitting classes play an important role in the theory of groups. The importance of the theory of
Fitting classes can firstly be seen in the following theorem proved by Fischer, Gaschütz and Harley
[1], which is in fact a generalization of the classical Sylow theorem and Hall theorem.
Theorem 1.1 (see [1] or [2, Theorem VIII,2.8]. Let F be a non-empty Fitting class. Then a
soluble group possesses exactly one conjugacy class of F-injectors.
S
For a Fitting class F, we let π(F) = G∈F σ(G). Note that if F = Np is the Fitting class of all
p-groups, then the F-injectors of a group G are Sylow p-subgroups of G; If F = Eπ and G has a
Hall π-soluble group, then the F-injectors of G are Hall π-subgroups of G (see [3, p.68, Ex.1] or [4,
p.238]).
About the existence of X-injector of G, Shemetkov posed the following problem.
Problem 1.2 (Shemetkov [5, Problem 11.117]). Let X be a Fitting class of soluble groups. Is it
true that every finite non-soluble group possesses an X-injector ?
This problem has been resolved in [6, 7] for the Fitting classes X ∈ {S, Sπ , N}.
In connection with this, a interesting problem is: to find the conjugate class of injectors in any
π-soluble groups. The first result in this direction is the following famous Chuchin’s theorem [8]: A
π-soluble G possesses a Hall π-subgroup (that is, Eπ -injectors) and any two Hall π-subgroups are
conjugate.
As a development of Chuchin’s theorem, Shemetkov and Guo proved the following
Theorem 1.3 ([9, Theorem 2.2] and [10]). Let F be a Fitting class, π = π(F). If G/GF is
π-soluble, then G has an F-injector and any two F-injector of G are conjugate in G.
The product FH of two Fitting classes F and H is the class (G | G/GF ∈ H). It is well known that
the product of any two Fitting classes is also a Fitting class and the multiplication of Fitting classes
satisfies associative law (see [2, Theorem IX, 1.12(a)(c)]).
Following [11, 12], a function f : P −→ {nonempty Fitting classes} is called a Hartley function
(or in brevity, H-function). A Fitting class F is local if
\ \
f (p)Np Ep′ )
F = Eπ(F) (
p∈π(F)
for some H-function f .
For a H-function h, let π = Supp(h) := {p ∈ P : h(p) 6= ∅}, which is called the support of the
T
H-function h, and LH(h) = p∈π h(p)Ep′ Np . Then, a Fitting class F is called Hartley class if there
exists a H-function h such that H = LH(h). In this case, H is said to be defined by the H-function
h or h is an H-function of H.
It is clear that N ⊆ LH(h). Hence, if F = LH(h), then π(F) = P. Any Hartley class is a local
2
Fitting class (see [13, p.31]). But the converse is not true in general (see [11, p.207, 4.2]). For two
class functions f and h, if f (p) ⊆ h(p) for all p ∈ π, then we write that f ≤ h.
Concerning Fitting classes and injectors, the authors in [14] (see also [11]) posed the problem (in
the universe S): Let F be a local Fitting class of soluble groups, could we describe the F-injectors
of a soluble group? In view of Problem 1.2 and Theorem 1.3, the following more general question
(in the class E) naturally arise:
Problem 1.4. For a local Fitting class F and a non-soluble group G (in particular, a π-soluble
group G), whether G possesses an F-injector and any two F-injectors are conjugate ?
Note that there exist non-soluble groups G and non-local Fitting classes F such that G has no
F-injector (see, for example, [4, 7.1.3-7.1.4]).
In this paper, by developing local method offered by Hartley [11], we will resolve Problem 1.4 for
partial π-soluble groups G (in particular, for π-soluble groups) and the Hartley class defined by a
invariable Hartley function. In fact, we will prove the following
Theorem 1.5. Support that H = LH(h) be a Hartley class defined by a invariable Hartley
function h, that is, h(p) = X for all prime p ∈ π = Supp(h), where X is some non-empty Fitting
class, and G ∈ XSπ (in particular, G is a π-soluble group). Then the following statements hold:
(1) G possesses an H-injector and any two H-injectors are conjugate in G;
(2) Every H-injector V of G is of type GXEπ′ L, where L is the subgroup of G such that L/GX is
an Nπ -injector of some Hall π-subgroup of G/GX .
From Theorem 1.5, a series of new classical conjugate classes of injectors in any π-soluble group
are obtained and the structure of injectors of some groups are described. For example, the following
results directly follow from Theorem 1.5.
Corollary 1.5.1. Every π-soluble group possesses exactly one conjugacy class of XNπ -injectors.
Corollary 1.5.2. Every π-soluble group has a injector with limited π-nilpotent length (that is,
for any natural number k) and any two of them are conjugate.
(Nπ )k -injector
If π = {p}, then from Theorem 1.5 we have the following
Corollary 1.5.3. If H is an Hartley class of type XEp′ Np , where X is a nonempty Fitting class,
and G a group such that G/GX is p-soluble, then G possesses an H-injector and any two of them are
conjugate in G.
In the case when X = (1), where (1) is the class of all identity groups, taking account of Theorem
of Iranzo-Toress in [15], we have
Corollary 1.5.4. Every p-soluble group G possesses exactly one conjugacy class of p-nilpotent
injectors and each p-nilpotent injector is a maximal p-nilpotent subgroup of G containing the pnilpotent radical of G.
Note that even if π $ σ(G) an X ⊆ S, the all statements of Theorem 1.5 and the corollaries are
3
still new results for any soluble group G.
All unexplained notations and terminologies are standard. The reader is referred to [2, 4, 16] if
necessary.
2
Preliminaries
Recall that G is said to be π-soulbe if there exists a series of subnormal subgroups
1 = G0 ⊳ G1 ⊳ G2 · · · ⊳ Gm = G
such that every factor Gi /Gi−1 is either a p-group for some p ∈ π or a π ′ -group, for i = 1, 2, · · · , m.
In particular, if π = p, then a π-soluble group is said to be p-soluble.
If σ(G) ⊆ π, then G is said to be a π-group. A subgroup H of G is said to be a Hall π-subgroup
of G if σ(H) ⊆ π and σ(|G : H|) ⊆ π ′ . G is said to be p-nilpotent if G has a normal Hall p′ -subgroup.
G is said to be π-nilpotent if G is p-nilpotent for all p ∈ π.
If CG (GF ) ⊆ GF , then G is said to be F-constrained. Note that if F = N (F = Nπ , resp.), then the
F-radical of G is the Fitting subgroup of G (the π-Fitting subgroup, resp.), which is called nilpotent
radical (π-nilpotent radical, resp.) and denoted by GN or F (G) (GNπ or Fπ (G), resp.). The maximal
normal π-subgroup (the maximal normal π ′ -subgroup) of G is said to be π-radical of G and denoted
by GEπ or Oπ (G) (π ′ -radical of G, and denoted by GEπ′ or Oπ′ (G), resp.).
Lemma 2.1 (see [3, Theorems 1.8.18 and 1.8.19] or [17, Corollary 4.1.2]). Support that G ∈ Sπ .
Then G is Nπ -constrained, that is, CG (Fπ (G)) ≤ Fπ (G). In particular, if GEπ′ = 1, then G is
π-constrained, that is, CG (Oπ (G) ≤ Oπ (G).
The following results is well known (see, for example, [2, IX, Remarks (1.3)]).
Lemma 2.2. Let F be a non-empty class of groups.
1) If V is an F-injector of G and K ⊳ G, then V ∩ K is an F-injector of K;
2) If V is an F-injector of G and α : G → Ḡ is a isomorphism, then α(V ) is an F-injector of Ḡ;
3) If V is F-maximal subgroup of G and V ∩ M is an F-injector of M for any maximal normal
subgroup M of G, then V is an F-injector of G;
4) If V is an F-injector of G, then GF ≤ V and V is an F-maximal subgroup of G.
Lemma 2.3 [2, Lemma IX, 1,1(a), Theorem IX, 1.12(b)]. Let F be a non-empty Fitting class.
Then:
1) If N is a subnormal subgroup of G, then NF = N ∩ GF ;
2) If H is a non-empty Fitting class, then the H-radical of G/GF is GFH /GF .
Lemma 2.4 (see [8] or [18, Chapter 5, Theorem 3.7]). If G ∈ Sπ , then every π-subgroup of G
is contained some Eπ -injector of G and any two Eπ -injector of G are conjugate in G.
4
Definition 2.5. Let π = Supp(h), where h is the support of some H-function h of an Hartley
class H. Then h is said to be
1) integrated if h(p) ⊆ H for all p ∈ π;
2) full if h(p) ⊆ h(q)Ep′ for all different primes p, q ∈ π;
3) full integrated if h is full and integrated as well;
4) invariable if f (p) = f (q) for all p, q ∈ π.
It is easy to see that every Hartley class can be defined by a integrated H-function, and a
invariable H-function is full integrated (in fact, since h(p) = h(q) for all p, q ∈ π, we have that
T
h(p) ⊆ h(p)Eq′ ⊆ h(q)Eq′ , so h(p) ⊆ q∈π h(q)Eq′ Np = H).
3
Proof of Theorem 1.5
The proof of Theorem 1.5 consists of a large number of steps. The following 5 lemmas are the main
steps of it.
Lemma 3.1. Every π-soluble group G possesses exactly one conjugacy class of π-nilpotent
injectors, and each π-nilpotent injector of G is the product of the π ′ -radical of G and an Nπ -injector
of some Hall π-subgroup of G.
Proof. We prove the lemma by induction on |G|. Let M be any maximal normal subgroup of
G. We consider the following two possible cases.
Case 1. The π ′ -radical GEπ′ of G is a identity group, that is, GEπ′ = 1.
Then MEπ′ = 1. By induction, M possesses exactly one conjugacy class of π-nilpotent injectors
and every π-nilpotent injector of M is an Nπ -injector of some Hall π-subgroup of M .
Let F1 be an Nπ -injector of some Hall π-subgroup Mπ of M . Since every Hall π-subgroup Gπ of
G is soluble, Gπ has an Nπ -injector V and any two Nπ -injectors of Gπ are conjugate in Gπ . Since
Mπ = M ∩ Gπ ⊳ Gπ ,
we have that V ∩ Mπ is an Nπ -injector of M by Lemma 2.2(1). In view of the conjugacy of Nπ injectors of M , we may assume that F1 = V ∩ Mπ . Since a Hall π-subgroup of any π-nilpotent group
is a nilpotent π-subgroup, every π-nilpotent injector of Gπ is an Nπ -injector of Gπ . Hence, if we
can prove that V is a maximal π-nilpotent subgroup of G, then V is a π-nilpotent injector of G by
Lemma 2.2(3), and so V is an Nπ -injector of G.
Support that V ≤ V1 where V1 is a maximal π-nilpotent subgroup of G. Since GNπ and (V1 )Eπ′
are normal in V1 , [(V1 )Gπ′ , GNπ ] ≤ (V1 )Eπ′ ∩ GNπ = 1. Hence
(V1 )Eπ′ ≤ CG (GNπ ).
5
Since Fπ (G) = GNπ = GNπ and CG (GNπ ) ≤ GNπ by Lemma 2.1, we have that (V1 )Eπ′ = 1. This
means that V1 ∈ Nπ and so V = V1 is a maximal π-nilpotent subgroup of G. This shows that the
statement of the lemma holds in case 1.
Case 2. GEπ′ 6= 1.
Let G1 = G/GEπ′ . By Lemma 2.3(2), (G1 )Eπ′ = GEπ′ GEπ′ /GEπ′ = 1. Hence by case 1, we have
that G1 possesses exactly one conjugate class of π-nilpotent injectors of type (G1 )Eπ′ V1 , where V1 is
an Nπ -injector of some Hall π-subgroup of G1 , and the set of π-nilpotent injectors of G1 are coincide
with the set of Nπ -injectors of Hall π-subgroups Gπ GEπ′ /GEπ′ . But since Gπ is soluble, by Theorem
1.1 Gπ has Nπ -injector, V say. Then by Lemma 2.2(2), the subgroup V GEπ′ /GEπ′ is an Nπ -injector
of Gπ GEπ′ /GEπ′ . It follows that V GEπ′ is a π-nilpotent subgroup of G.
We now prove that V GEπ′ is a maximal π-nilpotent subgroup of G. Assume that V GEπ′ ≤ F and
F is a maximal π-nilpotent subgroup of G. Then F = FEπ′ Fπ , where Fπ ∈ Nπ is a Hall π-subgroup
of F . Without loss of generality, we may assume that Fπ ⊆ Gπ . Hence V ≤ Fπ ≤ Gπ . But since the
Nπ -injector V is Nπ -maximal in Gπ , we have that V = Fπ . It follows from Lemma 2.3 and GEπ′ ⊳ F
that
(F/GEπ′ )Eπ′ = FEπ′ /GEπ′ .
Hence
(F/GEπ′ )/(F/GEπ′ )Eπ′ ≃ F/FEπ′ ≃ Fπ = V ∈ Nπ .
This shows that F/GEπ′ is π-nilpotent and V GEπ′ /GEπ′ ≤ F/GEπ′ . Thus, GEπ′ V = F is a maximal
π-nilpotent subgroup of G.
In order to prove that GEπ′ V is a π-nilpotent injector of G, by Lemma 2.2(3) we only need to
prove that GEπ′ V ∩ M is a π-nilpotent injector of M .
By induction, M has a π-nilpotent injector of type MGE ′ L, where L is an Nπ -injector of some
π
Hall π-subgroup Mπ of M . Since
Mπ = M ∩ Gπ ⊳ Gπ ,
and any two Nπ -injectors of Mπ are conjugate by Theorem 1.1, we may, without loss of generality,
assume that L = V ∩ Gπ . Since GEπ′ V ∩ M E GEπ′ V and GEπ′ V is π-nilpotent, GEπ′ V ∩ M is
π-nilpotent. But MEπ′ L ≤ GEπ′ V ∩ M and the π-nilpotent injector MEπ′ L of M is a maximal
π-nilpotent subgroup of M . Therefore MEπ′ L = GEπ′ V ∩ M . This shows that GEπ′ V ∩ M is a
π-nilpotent injector of M . Therefore, existence of π-nilpotent injector in a π-soluble group has been
proved.
The conjugacy of π-nilpotent injectors follows from the cojugacy of Nπ -injectors of Hall πsubgroups of any π-soluble group. This shows that the lemma also holds in case 2.
The Lemma is proved.
In the case when π = {p}, by Lemma 3.1 we directly obtain the following
6
Corollary 3.2. Every p-soluble group G possesses exactly one conjugate class of p-nilpotent
injectors, and every p-nilpotent injector of G is of type GEp′ P , where P is a Sylow p-subgroup of G
and GEp′ is the p′ -radical of G.
Support that X be some set of groups. We use FitX to denotes the Fitting class generated by X,
that is, FitX is the smallest Fitting class containing X. For a class F of groups, we use GF to denotes
the F-residual of G.
Lemma 3.3. Every Hartley class H can be defined by a full integrated H-function h, that is,
H = LH(h) such that h(p) ⊆ h(q)Eq′ ⊆ H, for all different primes p, q ∈ π = Supp(h).
Proof. Let H be a Hartley class. Then H = LH(h1 ), for some integrated H-function h1 .
Following [11], we define:
ψ(p) = {G | G ≃ H Ep′ , for some H ∈ h1 (p)},
for all p ∈ π.
Let X be a group in ψ(p). Then X ≃ Y Ep′ , for some group Y ∈ h1 (p). Since every Fitting class
is closed with respect to normal subgroup, Y Ep′ ∈ h1 (p) and so X ∈ h1 (p). This shows that ψ ≤ h1 ,
and so
ψ(p)Ep′ ⊆ h1 (p)Ep′ .
Ep′
If Y1 ∈ h1 (p)Ep′ , then Y1 /(Y1 )h1 (p) ∈ Ep′ and Y1
E ′
(Y1 p )Ep′
E ′
Y1 p ,
=
ing equation:
we have
E ′
Y1 p
≤ (Y1 )h1 (p) ∈ h1 (p). Moreover, since
∈ ψ(p), that is, Y1 ∈ ψ(p)Ep′ . Therefore, we obtain the followψ(p)Ep′ = h1 (p)Ep′ .
(∗)
Now, let h be the H-function such that h(p) = Fit(ψ(p)), for all p ∈ π. Let
M = ∩p∈π h(p)Ep′ Np .
We now prove that M = H. In fact, since ψ ≤ h1 , we have h ≤ h1 , and so h(p)Ep′ ⊆ h1 (p)Ep′ ,
for all p ∈ π. Then, by the equation (*), we see that
Fit(h1 (p)Ep′ ) = h1 (p)Ep′ = Fit(ψ(p)Ep′ ).
Therefore
h1 (p)Ep′ = Fit(ψ(p)Ep′ ) ⊆ (Fit(ψ(p))Ep′ = h(p)Ep′ .
This shows that
h(p)Ep′ = h1 (p)Ep′ ,
for all p ∈ π. Thus, we obtain that M = H. Moreover, since h ≤ h1 and h1 is integrated H-function
of H, we have that h is a integrated H-function of H.
7
Now, in order to complete the proof of the lemma, we only need to prove that h(p) ⊆ h(q)Eq′ ,
for all different primes p, q ∈ π. In fact, let L ∈ h1 (p) and p 6= q ∈ π. Clearly, Nq ⊆ Ep′ , so
LEp′ ≤ LNq . However, since L ∈ H, we have L/Lh1 (q)Eq′ ∈ Nq . Hence LNq ≤ Lh1 (q)Eq′ . It follows that
LNq ∈ h1 (q)Eq′ . This shows that for every group in h1 (p), its Ep′ -residual is contained in h1 (q)Eq′ .
Therefore, if R ∈ ψ(p), then R ≃ V Ep′ for some group V ∈ h1 (p) and thereby R ∈ h1 (q)Eq′ . This
induces that
ψ(p) ⊆ h1 (q)Eq′ .
Thus
h(p) = Fit(ψ(p)) ⊆ Fit(h1 (q)Eq′ ) = h1 (q)Eq′ = h(q)Eq′ .
This completes the proof of the lemma.
By Lemma 3.3, for any Hartley class H, we may always assume that H is defined by a full
integrated H-function h. We call the subgroup Gh = Πp∈π Gh(p) the h-radical of G.
Lemma 3.4. Let H = LH(h) for a full integrated H-function h, and G a group such that G/Gh
is Nπ -constrained (in particular, G/Gh is π-soluble). Then a subgroup V containing GH belongs to
H if and only if V /Gh is π-nilpotent.
Proof. Assume that V ∈ H and GH ⊆ V . Then Vh(p) ∩ GH = (GH )h(p) = Gh(p) , and so
[Vh(p) , GH ] ⊆ Gh(p) . This implies that Vh(p) ⊆ CG (GH /Gh(p) ), for all p ∈ π.
T
We first prove that GH /Gh = Fπ (G/Gh ). Let Fπ (G/Gh ) = L/Gh . Since GH ∈ H = p∈π h(p)Ep′ Np .
and (GH )h(p) = Gh(p) , we have that GH /Gh(p) is p-nilpotent for all p ∈ π. Hence GH /Gh is π-nilpotent.
Consequently, GH /Gh ≤ L/Gh and so GH ≤ L. On the other hand, since L/Gh is π-nilpotent, by
the isomorphism
L/Lh(p) Gh ≃ (L/Gh )/(Lh(p) Gh /Gh ),
we have that L/Lh(p) Gh ∈ Ep′ Np for all p ∈ π. It follows that (L/Gh(p) /(Lh(p) Gh /Lh(p) ) ∈ Ep′ Np ,
for all p ∈ π.
In order to prove that L ≤ GH , we only need to prove that L ∈ H. But it only need to prove that
Lh(p) Gh /Lh(p) is a p′ -group for all p ∈ π. Since Gh E L, by Lemma 2.3,
Gh(p) = (Gh )h(p) = Gh ∩ Lh(p) ≤ Lh(p) .
Let q be an arbitrary prime in π and q 6= p. Since
Gh(p) Gh(q) /Gh(p) ≃ Gh(q) /Gh(q) ∩ Gh(p) ,
Gh(p) Gh(q) /Gh(p) ≃ Gh(q) /(Gh(q) )h(p) . But since h is full, we have that h(q) ⊆ h(p)Ep′ . Hence
Gh(q) ∈ h(p)Ep′ , and so
Gh(q) /(Gh(q) )h(p) ∈ Ep′ .
8
This shows that Gh(p) Gh(q) /Gh(p) is a p′ -group for all different primes p, q ∈ π. Hence Gh /Gh(p) ∈ Ep′ .
Then by the isomorphism
Lh(p) Gh /Lh(p) ≃ Gh /Gh ∩ Lh(p) ≃ (Gh /Gh(p) )/(Gh ∩ Lh(p) /Gh(p) ),
we have that Lh(p) Gh /Lh(p) is a p′ -group for all p ∈ π. Thus L ∈ H, and so GH /Gh = Fπ (G/Gh ).
Since G/Gh is Nπ -constrained by the hypothesis and GH /Gh = Fπ (G/Gh ), CG/Gh (GH /Gh ) ≤
GH /Gh and so CG (GH /Gh ) ≤ GH . But, clearly, CG (GH /Gh(p) ) ≤ CG (GH /Gh ). Hence Vh(p) ⊆ GH .
T
Consequently, Vh(p) = Gh(p) for all p ∈ π. Then by V ∈ H = p∈π h(p)Ep′ Np , we see that V /Gh(p) =
V /Vh(p) is p-nilpotent for all p ∈ π. This implies that V /Gh is π-nilpotent.
Conversely, if V /Gh is π-nilpotent, then, with a similar argument (as the above proof of L ≤ GH ),
we can see that V ∈ H. This completes the proof.
Lemma 3.5. Let H = LH(h) for a full integrated H-function h, π = Supp(h) 6= ∅, and G be a
group such that G/Gh is Nπ -constrained and σ(Gh ) ⊆ π. If V /Gh is a π-nilpotent injector of G/Gh ,
then V is an H-injector of G.
Proof. We prove the lemma by induction on |G|. Let M be any maximal normal subgroup of
G, and p, q are different primes in π.
We first prove that Gh /Gh(q) is a q ′ -group for all primes q ∈ π. In fact, since h(p) ⊆ h(q)Eq′ for
all different primes p, q ∈ π, by the isomorphism
Gh(q) Gh(p) /Gh(q) ≃ Gh(p) /(Gh(p) )Gh(q) ,
we have that
Gh(q) Gh(p) /Gh(q) ∈ Eq′ .
This implies that Gh /Gh(q) ∈ Eq′ , that is, Gh /Gh(q) is a q ′ -group.
Now let Mh = Πp∈π Mh(p) . Since
(Gh ∩ M )Gh(q) /Gh(q) ≃ Gh ∩ M/Mh(q) ,
Gh ∩ M/Mh(q) is a q ′ -group for all primes q ∈ π. Hence
\
Eq′ = Eπ′ .
Gh ∩ M/Mh ∈
q∈π
It follows from σ(Gh ) ⊆ π that
Gh ∩ M/Mh ∈ Eπ ∩ Eπ′ = (1).
Thus Gh ∩ M = Mh .
We consider the following two possible cases.
Case 1. Gh ⊆ M .
9
Then Gh = Mh . Since V /Gh is a Nπ -injector of G/Gh , V ∩ M/Mh is an Nπ -injector of M/Mh
by Lemma 2.2(1). Since G/Gh is Nπ -constrained and the class of Nπ -constrained groups is a Fitting
class by [19, Theorem B(b)], M/Mh is also Nπ -constrained. Hence, by induction, V ∩ M is an
H-injector of M .
Assume that V < V1 and V1 is an H-maximal subgroup of G. Since V ∩ M is an H-maximal
subgroup of M , we have that V ∩ M = V1 ∩ M. Hence V1 ∩ M is an H-injector of M for any maximal
subgroup M of G. It follows from Lemma 2.2(3) that V1 is a H-injector of G. But then GH ≤ V1 ,
so V1 /Gh is π-nilpotent by Lemma 3.4. As V /Gh is an Nπ -injector of G/Gh , V /Gh is a maximal
π-nilpotent subgroup of G/Gh , which contradicts V /Gh < V1 /Gh . Thus V = V1 is an H-maximal
subgroup of G. Hence by Lemma 2.2(3), V is an H-injector.
Case 2. Gh 6⊆ M .
Then G = Gh M . Since V /Gh is a π-nilpotent injector of G/Gh and
G/Gh ≃ M/Gh ∩ M = M/Mh ,
V ∩ M/Mh is an Nπ -injector of M/Mh by Lemma 2.2(2). Then by induction, V ∩ M is an H-injector
of M . With a similar argument as in the case 1, we obtain that V is an H-injector of G.
The lemma is proved.
In view of the above Lemmas, we now may prove our main theorem.
Proof of Theorem 1.5. (1) Let M be a maximal normal subgroup of G. Note that h is a
invariable H-function as h(p) = X for all p ∈ π. Hence h is a full integrated H-function. Since
G ∈ XSπ , G/GX is π-soluble. Therefore G/GX is Nπ -constrained and G/GX has a π-nilpotent
injector V /GX by Lemma 3.1. Clearly, Gh = GX , Mh = MX and Gh ∩ M = Mh . Then, with the same
arguments as in the proof of Lemma 3.5, we obtain that V is an H-injector of G.
Now we prove that if V is an H-injector of G, then V /GX is a π-nilpotent injector of G/GX .
In fact, assume that V is an H-injector of G. Then V ∩ S is an H-maximal subgroup of G for
any subnormal subgroup S of G. In order to prove that V /Gh is a π-nilpotent injector of G/Gh ,
we only need to prove that V /GX ∩ S/GX = (V ∩ S)/GX is an Nπ -maximal subgroup of G/GX for
T
every subnormal subgroup S/GX of G/GX . Since H = LR(h) = p∈π XEp′ Np for all p ∈ π and
V ∩ S ∈ H, we have (V ∩ S)/GX ∈ Nπ . Assume that (V ∩ S)/GX is not an Nπ -maximal subgroup
of G/GX and let (V ∩ M )/GX < D/GX and D/GX is an Nπ -maximal subgroup of G/GX . Then
T
clearly D ∈ p∈π XEp′ Np = LR(h) = H. But as V ∩ S is an H-maximal subgroup of G, we have that
V ∩ S = D. This contradiction shows that if V is an H-injector of G, then V /GX is a π-nilpotent
injector of G/GX .
Assume that F is another H-injector of G. Then F/GX is a π-nilpotent injector of G/GX as
above. Hence by Lemma 3.1, V /GX and F/GX are conjugate in G/GX . This implies that V and F
are conjugate in G. Hence we have (1).
10
(2) Let V is a H-injector of G. Then V /GX is a π-nilpotent injector of G/GX . Hence by Lemma
3.1, we have that
V /GX = (G/GX )Eπ′ (L/GX ),
where L/GX is an Nπ -injector of some Hall π-subgroup of G/GX . But by Lemma 2.3,
(G/GX )Eπ′ = GXEπ′ /GX .
This shows that V = GXEπ′ L. Thus (2) holds.
The theorem is proved.
By Lemma 3.1, we known that every π-soluble group possesses exactly one conjugacy class of
π-nilpotent injectors. Note also that every π-soluble group is Nπ -constrained. In connection with
this, we put forward the following question:
Question 3.1. Support that a group G is Nπ -constrained. Is it true that G possesses exactly
one conjugacy class of π-nilpotent injectors?
References
[1] B. Fischer, W. Gaschütz, B. Hartley, Injektoren endlicher auflösbarer Gruppen, Math. Z., 102
(1967), 337-339.
[2] K. Doerk, T. Hawkes, Finite Soluble Groups, Walter de Gruyter, Berlin, New York, 1992.
[3] W. Guo, The Theory of Classes of Groups, Science Press-Kluwer Academic Publishers, BeijingNew York-Dordrecht-Boston-London, 2000.
[4] A. Ballester-Bolinches, L. M. Ezquerro, Classes of Finite Groups, Springer, Dordrecht, 2006.
[5] V. D. Mazurov, E. I. Khuhro, The Kourovka notebook, Unsolved problems in group theory,
Institute of Mathematics, Siberian Branch of RAS, Novosibirsk, 2016.
[6] P. Förster, Nilpotent injector in finite groups, Bull. Austral Math. Soc., 32 (1925), 292-298.
[7] M. J. Iranzo, F. Perez-Monazor, Existencia de injectores en groups finites respecto de ciertas
class de Fitting, Publ. Mat. Univ. Autonima Baselona, 32 (1988), 57-59.
[8] S. A. Čunihin, On existence and conjugacy of subgroups of a finite group, Mat. Sbornik, 33(1)
(1953), 111-132.
[9] L. A. Shemetkov, On subgroups of π-soluble groups, in: Finite Groups, pp. 207-212, Hauka i
Tehnika, Minsk, 1975.
[10] W. Guo, Injectors of finite groups, Chinese Ann. Math. Ser. A 18 (2) (1997), 145-148.
11
[11] B. Hartley, On Fischer’s dualization of formation theory, Proc. London Math. Soc., 3(2) (1969),
193-207.
[12] N. T. Vorob’ev, On the Hawkes conjecture for radical classes, Siberian Math. J., 37(6) (1996),
1137-1142.
[13] P. Hauck, V. N. Zahurski, A characterization of Dominant local Fitting classes, J. Algebra, 358
(2012), 27-32.
[14] Y. Liu, W. Guo and N. T. Vorob’ev, Description of F-injectors of finite soluble groups, Math.
SCI. Rec. J, 12(1) (2008), 17-22.
[15] M. J. Iranzo, M. Toress, The p∗ p-injectors of a finite group, Rend. Sem. Mat. Uni. Padova, 82
(1989), 233-237.
[16] W. Guo, Structure Theory for Canonical Classes of Finite Groups, Springer, 2015.
[17] L. A. Shemetkov, Formations of finite groups, Moscow, Nauka, Main Editorial Board for Physical
and Mathematical Literature, 1978.
[18] M. Suzuki, Group Theory, II, Springer-Verlag, New York, 1986.
[19] M. I. Iranzo, P. Monazor, F-constraint with respect to a Fitting class, Arch. Math., 46 (1986),
205-210.
12
| 4math.GR
|
Sublinear Time Estimation of Degree Distribution Moments:
The Degeneracy Connection
(Full Version)
arXiv:1604.03661v2 [cs.DS] 16 Feb 2017
Talya Eden∗
Dana Ron†
C. Seshadhri‡
Abstract
We revisit the classic problem of estimating the degree distribution moments of an undirected graph. Consider anP
undirected graph G = (V, E) with n (non-isolated) vertices, and
define (for s > 0) µs = n1 · v∈V dsv . Our aim is to estimate µs within a multiplicative error
of (1 + ε) (for a given approximation parameter ε > 0) in sublinear time. We consider the
sparse graph model that allows access to: uniform random vertices, queries for the degree
of any vertex, and queries for a neighbor of any vertex. For the case of s = 1 (the average
e √n) queries suffice for any constant ε (Feige, SICOMP 06 and Goldreich-Ron,
degree), O(
RSA 08). Gonen-Ron-Shavitt (SIDMA 11) extended this result to all integral s > 0, by
e 1−1/(s+1) ) queries. (Strictly speaking, their algodesigning an algorithms that performs O(n
rithm approximates the number of star-subgraphs of a given size, but a slight modification
gives an algorithm for moments.)
We design a new, significantly simpler algorithm for this problem. In the worst-case, it
exactly matches the bounds of Gonen-Ron-Shavitt, and has a much simpler proof. More
importantly, the running time of this algorithm is connected to the degeneracy of G. This
is (essentially) the maximum density of an induced subgraph.
For the family of graphs
with
1−1/s
1/s
n
1/s
e
degeneracy at most α, it has a query complexity of O
α + min{α, µs }
=
1/s
µs
e 1−1/s α/µ1/s
O(n
s ). Thus, for the class of bounded degeneracy graphs (which includes all
minor closed families and preferential attachment graphs), we can estimate the average
e
e √n)
degree in O(1)
queries, and can estimate the variance of the degree distribution in O(
queries. This is a major improvement over the previous worst-case bounds. Our key insight
is in designing an estimator for µs that has low variance when G does not have large dense
subgraphs.
1
Introduction
Estimating the mean and moments of a sequence of n integers d1 , d2 , . . . , dn is a classic problem
in statistics that requires little introduction. In the absence of any knowledge of the moments of
the sequence, it is not possible to prove anything non-trivial. But suppose these integers formed
the degree sequence of a graph. Formally, let G = (V, E) be an undirected graph over n vertices,
∗
School of EE, Tel Aviv University. [email protected]
This research was partially supported by a grant from the Blavatnik fund. The author is grateful to the Azrieli
Foundation for the award of an Azrieli Fellowship.
†
School of EE, Tel Aviv University. [email protected]
This research was partially supported by the Israel Science Foundation grant No. 671/13 and by a grant from
the Blavatnik fund.
‡
Department of Computer Science, University of California, Santa Cruz. [email protected]
1
and let dv denote the degree of vertex v ∈ V , where we assume that dv ≥ 1 for every v.1 Feige
√
proved that O ∗ ( n) uniform random vertex degrees (in expectation) suffice to provide a (2+ ε)approximation to the average degree [24]. (We use O ∗ (·) to suppress poly(log n, 1/ε) factors.)
The variance can be as large as n for graphs of constant average degree (simply consider a
star), but the constraints of a degree distribution allow for non-trivial approximations. Classic
theorems of Erdős-Gallai and Havel-Hakimi characterize such sequences [29, 22, 27].
Again, the star graph shows that the (2 + ε)-approximation cannot be beaten in sublinear
time through pure vertex sampling. Suppose we could also access random neighbors of a given
vertex. In this setting, Goldreich and Ron showed it is possible to obtain a (1+ε)-approximation
√
to the average degree in O ∗ ( n) expected time [25].
In a substantial (and complex) generalization, Gonen, Ron, and Shavitt (henceforth, GRS)
gave a sublinear-time algorithm that estimates the higher moments of the degree distribution [26]. Technically, GRS gave an algorithm for approximating the number of stars in a
graph, but a simple modification yields an algorithm for moments estimation. For precision, let
us formally define this problem. The degree distribution is the distribution over the
Pdegree of a
uniform random vertex. The s-th moment of the degree distribution is µs , n1 · v∈V dsv .
The Degree Distribution Moment Estimation (DDME) Problem. Let G = (V, E) be a
graph over n vertices, where n is known. Access to G is provided through the following queries.
We can (i) get the id (label) of a uniform random vertex, (ii) query the degree dv of any vertex
v, (iii) query a uniform random neighbor of any vertex v. Given ε > 0 and s ≥ 1, output a
(1 + ε)-multiplicative approximation to µs with probability2 > 2/3.
The DDME problem has important connections to network science, which is the study of
properties of real-world graphs. There have been numerous results on the significance of heavytailed/power-law degree distributions in such graphs, since the seminal results of BarabásiAlbert [6, 11, 23]. The degree distribution and its moments are commonly used to characterize
and model graphs appearing in varied applications [8, 36, 15, 37, 9]. On the theoretical side, recent results provide faster algorithms for graphs where the degree distribution has some specified
form [7, 10]. Practical algorithms for specific cases of DDME have been studied by Dasgupta
et al and Chierichetti et al. [18, 14]. (These results requires bounds on the mixing time of the
random walk on G.)
1.1
Results
Let m denote the number of edges in the graph (where m is not provided to the algorithm). For
the sake of simplicity, we restrict the discussion in the introduction to case when µs ≤ ns−1 . As
observed by GRS, the complexityof the DDME problem is smaller
when µs is significantly larger.
1/(s+1)
1−1/(s+1)
1−1/s
∗
-query algorithm for DDME and
/µs
+n
GRS designed an (expected) O n
proved this expression was optimal up to poly(log n, 1/ε) dependencies. (Here O∗ (·) also suppresses additional factors that depend only on s). Note that for a graph without isolated
vertices, µs ≥ 1 for every s > 0, so this yields a worst-case O∗ (n1−1/(s+1) ) bound. The s = 1
√
case is estimating the average degree, so this recovers the O ∗ ( n) bounds of Goldreich-Ron. We
mention a recent result by Aliakbarpour et al. [3] for DDME, in a stronger model that assumes
additional access to uniform random edges. They get a better bound of O∗ (m/(nµs )1/s ) in
this stronger model, for s > 1 (and µs ≤ ns−1 ). Note that the main challenge of DDME is
1
The assumption on there being no isolated vertices is made here only for the sake of simplicity of the
presentation, as it ensures a basic lower bound on the moments.
2
The constant 2/3 is a matter of convenience. It can be increased to at least 1 − δ by taking the median value
of log(1/δ) independent invocations.
2
in measuring the contribution of high-degree vertices, which becomes substantially easier when
random edges are provided. In the DDME problem without such samples, it is quite non-trivial
to even detect high degree vertices.
All the bounds given above are known to be optimal, up to poly(log n, 1/ε) dependencies,
and at first blush, this problem appears to be solved. We unearth a connection between DDME
and the degeneracy of G. The degeneracy of G is (up to a factor 2) the maximum density over
all subgraphs of G. We design an algorithm that has a nuanced query complexity, depending on
the degeneracy of G. Our result subsumes all existing results, and provides substantial improvements in many interesting cases. Furthermore, our algorithm and its analysis are significantly
simpler and more concise than in the GRS result.
We begin with a convenient corollary of our main theorem. A tighter, more precise bound
appears as Theorem 3.
Theorem 1. Consider the family of graphs with degeneracy at most α. The DDME problem
can be solved on this family using
!
1−1/s
n
α1/s + min{α, µ1/s
O∗
s }
1/s
µs
queries in expectation. The running time is linear in the number of queries.
Consider the case of bounded degeneracy graphs, where α = O(1). This is a rich class of
graphs. Every minor-closed family of graphs has bounded degeneracy, as do graphs generated
by the Barabási-Albert preferential attachment process [6]. There is a rich theory of bounded
expansion graphs, which spans logic, graph minor theory, and fixed-parameter tractability [32].
All these graph classes have bounded degeneracy. For every such class of graphs, we get a
1/s
(1 + ε)-estimate of µs in O ∗ (n1−1/s /µs ) time. We stress that bounded degeneracy does not
imply any bounds on the maximum degree or the moments. The star graph has degeneracy 1,
but has extremely large moments due to the central vertex.
Consider any bounded degeneracy graph without isolated vertices. We can accurately estimate the average degree (s = 1) in poly(log n) queries, and estimate the variance of the
√
degree distribution (s = 2) in n · poly(log n) queries. Contrast this with the (worst-case op√
timal) n bounds of Feige and Goldreich-Ron for average degree, and the O ∗ (n2/3 ) bound of
GRS for variance estimation. For general s, our bound is a significant improvement over the
1/(s+1)
O∗ (n1−1/(s+1) /µs
) bound of GRS.
The algorithm attaining Theorem 1 requires an upper bound on the degeneracy of the graph.
When an degeneracy bound is not given, the algorithm recovers the bounds of GRS, with an
improvement on the extra poly(log n)/ε factors. More details are in Theorem 3. We note that
the degeneracy-dependent bound in Theorem 1 cannot be attained by an algorithm that is only
given n as a parameter. In particular, if an algorithm is only provided with n and must work
√
on all graphs with n vertices, then it must perform Ω( n) queries in order to approximate the
average degree even for graphs of constant degeneracy (and constant average degree). Details
are given in Subsection 7.1 in the full version of the paper.
The bound of Theorem 1 may appear artificial, but we prove that it is optimal when µs ≤
ns−1 . (For the general case, we also have optimal upper and lower bounds.) This construction
is an extension of the lower bound proof of GRS.
Theorem 2. Consider the family of graphs with degeneracy α and
where µs ≤ ns−1 . Any
1−1/s
1/s
algorithm for the DDME problem on this family requires Ω n 1/s · α1/s + min{α, µs }
µs
queries.
3
1.2
From degeneracy to moment estimation
We begin with a closer look at the lower bound examples of Feige, Goldreich-Ron, and GRS.
The core idea is quite simple: DDME is hard when the overall graph is sparse, but there are
√
small dense subgraphs. Consider the case of a clique of size 100 n connected to a tree of size
n. The small clique dominates the average degree, but any sublinear algorithm with access
√
only to random vertices pays Ω( n) for a non-trivial approximation. GRS use more complex
constructions to get an Ω(n1−1/(s+1) ) lower bound for general s. This also involves embedding
small dense subgraphs that dominate the moments.
Can we prove a converse to these lower bound constructions? In other words, prove that the
non-existence of dense subgraphs must imply that DDME is easier? A convenient parameter
for this non-existence is the degeneracy.
But the degeneracy is a global parameter, and it is not clear how a sublinear algorithm
can exploit it. Furthermore, DDME algorithms are typically very local; they sample random
vertices, query the degrees of these vertices and maybe also query the degrees of some of their
neighbors. We need a local property that sublinear algorithms can exploit, but can also be
linked to the degeneracy. We achieve this connection via the degree ordering of G. Consider
the DAG obtained by directing all edges from lower to higher degree vertices. Chiba-Nishizeki
related the properties of the out-degree distribution to the degeneracy, and exploited this for
clique counting [13]. Nonetheless, there is no clear link to DDME. (Nor do we use any of their
techniques; we state this result merely to show what led us to use the degree ordering).
Our main insight is the construction of an estimator for DDME whose variance depends
on the degeneracy of G. This estimator critically uses the degree ordering. Our proof relates
the variance of this estimator to the density of subgraphs in G, which can be bounded by
the degeneracy. We stress that our algorithm is quite simple, and the technicalities are in the
analysis and setting of certain parameters.
1.3
Designing the algorithm
Designate the weight of an edge (u, v) to be dus−1 +P
dvs−1 . A simple calculation yields that the
s
sum of the weights of all edges is exactly Ms ,
v dv = n · µs . Suppose we could sample
uniform random edges (and knew the total number of edges). Then we could hope to estimate
Ms through uniform edge sampling. The variance of the edge weights can be bounded, and this
yields an O ∗ (m/(nµs )1/s ) = O∗ (n1−1/s ) algorithm (when no vertex is isolated). Indeed, this is
very similar to the approach of Aliakbarpour et al. [3]. Such variance calculations were also
used in the classic Alon-Matias-Szegedy result of frequency moment estimation [5].
Our approach is to simulate uniform edge samples using uniform vertex samples. Suppose
we sampled a set R of uniform random vertices. By querying the degrees of all these vertices,
we can select vertices in R with probability proportional to their degrees, which allows us to
uniformly sample edges that are incident to vertices in R. Now, we simply run the uniform edge
sampling algorithm on these edges. This algorithmic structure was recently used for sublinear
triangle counting algorithms by Eden et al. [20].
Here lies the core technical challenge. How to bound the number of random vertices that is
sufficient for effectively simulating the random edge algorithm? This boils down to the behavior
of the variance of the “vertex weight” distribution. Let the weight of a vertex be the sum of
weights of its incident edges. The weight distribution over vertices can be extremely skewed,
and this approach would require a forbiddingly large R.
A standard technique from triangle counting (first introduced by Chiba-Nishizeki [13]) helps
reduce the variance. Direct all edges from lower degree to higher degree vertices, breaking
ties consistently. Now, set the weight of a vertex to be the sum of weights on incident out-
4
edges. Thus, a high-degree vertex with lower degree neighbors will have a significantly reduced
weight, reducing overall variance. In the general case (ignoring degeneracy), a relatively simple
argument bounds the maximum weight of a vertex, which enables us to bound the variance of
the weight distribution. This yields a much simpler algorithm and proof of the GRS bound.
In the case of graphs with bounded degeneracy, we need a more refined approach. Our key
insight is an intimate connection between the variance and the existence of dense subgraphs
in G. We basically show that the main structure that leads to high variance is the existence
of dense subgraphs. Formally, we can translate a small upper bound on the density of any
subgraph to a bound on the variance of the vertex weights. This establishes the connection to
the graph degenearcy.
1.4
Simplicity of our algorithm
Our viewpoint on DDME is quite different from GRS and its precursor [25], which proceed
by bucketing the vertices based on their degree. This leads to a complicated algorithm, which
essentially samples to estimate the size of the buckets, and also the number of edges between
various buckets (and “sub-buckets”). We make use of buckets in out analysis, in order to obtain
the upper bound that depends on the degeneracy α (in order to achieve the GRS upper bound,
our analysis does not use bucketing).
As explained above, our main DDME procedure, Moment-estimator is simple enough to
present in a few lines of pseudocode (see Figure 1). We feel that the structural simplicity of
Moment-estimator is an important contribution of our work.
Moment-estimator takes two sampling parameters r and q. The main result Theorem 3
follows from running Moment-estimator with a standard geometric search for the right setting
of r and q. In Moment-estimator we use id(v) to denote the label of a vertex v, where vertices
have unique ids and there is a complete order over the ids.
Moment-estimators (r, q)
1. Select r vertices, uniformly, independently, at random and let the resulting
P multi-set
be denoted by R. Query the degree of each vertex in R, and let dR = v∈R dv .
2. For i = 1, . . . , q do:
(a) Select a vertex vi with probability proportional to its degree (i.e., with probability dvi /dR ), and query for a random neighbor ui of vi .
). Else, set
+ dus−1
(b) If dvi < dui or dvi = dui and id(vi ) < id(ui ), set Xi = (dvs−1
i
i
Xi = 0.
q
P
Xi .
3. Return X = 1r · dqR ·
i=1
Figure 1: Algorithm Moment-estimators for approximating µs .
1.5
Other related work
As mentioned at the beginning of this section, Aliakbarpour et al. [3] consider the problem of
approximating the number of s-stars for s ≥ 2 when given access to uniformly selected edges.
Given the ability to uniformly select edges, they can select vertices with probability proportional
to their degree (rather than uniformly). This can be used to get an unbiased estimator of µs (or
the s-star count) with low variance. This leads to an O(m/(nµs )1/s ) bound, which is optimal
(for µs ≤ ns−1 ).
5
Dasgupta, Kumar, and Sarlos give practical algorithms for average degree estimation, though
they assume bounds on the mixing time of the random walk on the graph [18]. A recent
paper of Chierichetti et al. build on these methods to sample nodes according to powers of
their degree (which is closely related to DDME) [14]. Simpson, Seshadhri, and McGregor give
practical algorithms to estimate the entire cumulative degree distribution in the streaming
setting [38]. This is different from the sublinear query model we consider, and the results are
mostly empirical.
In [20], Eden et al. present an algorithm for approximating the number of triangles in a
graph. Although this is a very different problem than DDME, there are similar challenges
regarding high-degree vertices. Indeed, as mentioned earlier, the approach of sampling random
edges through a set of random vertices was used in [20].
The degeneracy is closely related to other “density” notions, such as the arboricity, thickness,
and strength of a graph [1]. There is a rich history of algorithmic results where run time depends
on the degeneracy [31, 13, 4, 21].
Other sublinear algorithms for estimating various graph parameters include: approximating
the size of the minimum-weight spanning tree [12, 17, 16], maximum matching [33, 39] and of
the minimum vertex cover [35, 33, 30, 39, 28, 34].
2
The main theorem
Theorem 3. For every graph G, there exists an algorithm that returns a value Z such that
Z ∈ [(1 − ε)µs (G), (1 + ε)µs (G)] with probability at least 2/3. Assume that algorithm is given α,
an upper bound on the degeneracy of G. (If no such bound is provided, the algorithm assumes
a trivial bound of α = ∞.) The expected running time is the minimum of the following two
expressions.
n n1−1/s · α ns−1 · α o s log n · log(s log n)
α 1/s
,
·
+ min
O 2s · n1−1/s · log2 n ·
1/s
µs
µs
ε2
µs
O
n1−1/(s+1)
1/(s+1)
µs
n
ns−1−1/s o s log n · log(s log n)
+ min n1−1/s , 1−1/s
·
ε2
µs
(1)
(2)
Equation (2) is essentially the query complexity of GRS (albeit with a better dependence
on s, log n, and 1/ε). Thus, our algorithm is guaranteed to be at least as good as that. If α
is exactly the degeneracy of G, then we can prove that Equation (1) is less than Equation (2).
Within each expression, there is a min of two terms. The first term is smaller iff µs ≤ ns−1 .
The mechanism of deriving this rather cumbersome running time is the following. The
algorithm of Theorem 3 runs Moment-estimator for geometrically increasing values of r and
q, which is in turn derived from a geometrically decreasing guess of µs . It uses this guess to
set r and q. There is a setting of values depending on α, and a setting independent of it. The
algorithm simply picks the minimum of these settings to achieve the smaller running time.
3
Sufficient conditions for r and q in Moment-estimator
In this section we provide sufficient conditions on the parameters r and q that are used by
Moment-estimator (Figure 1), in order for the algorithm to return a (1 + ε) estimate of µs .
First we introduce some notations. For a graph G = (V, E) and a vertex v ∈ V , let Γ(v) denote
6
the set of neighbors of v in G (so that dv = |Γ(v)|). For any (multi-)set R of vertices, let ER be
the (multi-)set of edges incident to the vertices in R. We will think of the edges in ER as ordered
pairs; thus (v, u) is distinct from (u, v), and so ER , {(v, u) : v ∈ R, u ∈ Γ(v)}. Observe
P that
s
dR , as defined in Step 1 of Moment-estimator equals |ER |. Let Ms = Ms (G) ,
v∈V dv ,
so that µs = Ms /n. In the analysis of the algorithm, it is convenient to work with Ms instead
of µs .
A critical aspect of our algorithm (and proof) is the degree ordering on vertices. Formally,
we set u ≺ v if du < dv or, du = dv and id(u) < id(v). Given the degree ordering, we let
+
+
+
+
Γ+ (v) , {u ∈ Γ(v)
P : v ≺ u}, dv , |Γ (v)|,
P and E , {(v, u) : v ∈ V, u ∈ Γ (v)}. Here and
elsewhere, we use v as a shorthand for v∈V .
Definition 4. We define the weight of an edge e = (v, u) as follows: if v ≺ u define
wt(e) , (dvs−1 + dus−1 ). Otherwise,
P
P wt(e) , 0.
wt((v, u)), and for a (multi-)set of
wt((v, u)) =
For a vertex v ∈ V , wt(v) ,
vertices R, wt(R) ,
P
u∈Γ+ (v)
u∈Γ(v)
wt(v).
v∈R
Observe that given the above notations and definition, Moment-estimator selects uniform
edges from ER and sets each Xi (in Step 2b) to wt((vi , ui )). Based on Definition 4, we obtain
the next two claims, where the first claim connects between Ms and the weights of vertices.
P
Claim 5.
v wt(v) = Ms .
Proof: By the definition of the weights:
X
X
X
X X
X
wt(v) =
(dvs−1 + dus−1 ) =
(dvs−1 + dus−1 ) =
dvs−1 =
dsv = Ms ,
v∈V
(v,u)∈E +
v∈V u∈Γ(v)
{v,u}∈E
v∈V
and the claim is established.
Claim 6. Exp[X] = µs , where X is as defined in Step 3 of the algorithm.
P
Proof:
Recall that wt(R) ,
v∈R wt(v). Note that Xi (as defined in Step 2b of the
algorithm) is exactly wt((vi , ui )). Conditioning on R,
X X
X
1
1
1
Exp[Xi |R] =
·
·
wt(v) =
· wt(R) .
wt((v, u)) =
|ER |
|ER |
|ER |
+
v∈R u∈Γv
v∈R
By the definition of X in the algorithm (see Step 3),
Exp[X|R] =
q
1
1 |ER | X
Exp[Xi |R] = · wt(R) .
·
·
r
q
r
i=1
P
Now, let us remove the conditioning. Since wt(R) ,
v∈R wt(v), by linearity of the expectation,
r X
wt(v) ,
ExpR [wt(R)] = ·
n
v∈V
and thus (using Claim 5),
Exp[X] = ExpR [Exp[X|R]] =
1 X
·
wt(v) = µs ,
n
v∈V
which completes the proof.
7
3.1
Conditions on the parameters r and q.
We next state two conditions on the parameters r and q, which are used in the algorithm, and
then establish several claims, based on the conditions holding. The conditions are stated in
terms of properties of the graph as well as the approximation parameter ε and a confidence
parameter δ.
1. The vertex condition:
P
30 · n · v wt(v)2
.
r≥
ε2 · δ · Ms2
2. The edge condition:
q≥
2000 · m · M2s−1
.
ε2 · δ3 · Ms2
Lemma 7. If Condition 1 holds, then with probability at least 1 − δ/2, all the following hold.
1. wt(R) ∈ 1 − 2ε · nr · Ms , 1 + 2ε · nr · Ms .
r
2. |ER | ≤ 12
δ · n · m.
P
3.
wt ((v, u))2 ≤
18
δ
+
(v,u)∈ER
·
r
n
· M2s−1 .
Proof: First, we look
P at the random variable wt(R). By the definition of wt(R) and Claim 5,
Exp[wt(R)] = (r/n)· v wt(v) = (r/n)·Ms . Turning to the variance of wt(R), since the vertices
in R are chosen uniformly at random,
!
1 X
2
r X
1 X
2
·
wt(v) −
·
wt(v)
≤ ·
wt(v)2 .
VarR [wt(R)] = r · Varv [wt(v)] = r
n v
n v
n v
By Chebyshev’s inequality,
h
i
ε
Pr wt(R) − Exp[wt(R)] ≥ · Exp[wt(R)] ≤
2
=
4Var[wt(R)]
· Exp[wt(R)]2
P
P
4(r/n) v wt(v)2
4n v wt(v)2
=
.
ε2 · (r/n)2 · Ms2
ε2 · Ms2 · r
ε2
Applying the lower bound on r from Condition 1, this probability is at most δ/6. (Indeed,
Condition 1 was defined as such to get this bound.)
The other bounds follow simply from Markov’s inequality. Observe that Exp[|ER |] =
(r/n)(2m), and so Pr[|ER | > (12/δ)(r/n)m]
≤ δ/6.
P
The random variable Y = (v,u)∈E + wt ((v, u))2 (which is non-negative), satisfies
R
Exp[Y ] =
r
·
n
X
wt ((v, u))2 =
(v,u)∈E +
r
≤ 3· ·
n
X
(v,u)∈E +
r
·
n
X
(dvs−1 + dus−1 )2
(v,u)∈E +
(dv2s−2 + du2s−2 ) = 3 ·
r
r X 2s−1
·
dv
= 3 · · M2s−1 .
n v
n
(3)
By Markov’s inequality, Pr[Y ≥ (18/δ)(r/n)M2s−1 ] ≤ δ/6. We apply a union bound to complete
the proof.
8
Theorem 8. If Conditions 1 and 2 hold, then X ∈ [(1 − ε)µs , (1 + ε)µs ] with probability at least
1 − δ.
Proof:
Condition on any choice of R. We have Exp[X|R] = (1/r)wt(R). Turning to the
variance, since the edges (vi , ui ) are chosen from ER uniformly at random,
#
" q
2
2
X
1
|ER | 2
|ER |2
1
Xi R =
· Var[X1 |R]
·
· Var
·
Var[X|R] =
r
q
r
q
i=1
2
|ER |2
1
·
· Exp[(X1 |R)2 ]
≤
r
q
2
X
1
|ER |
=
·
·
wt((v, u))2
r
q
+
=
1 |ER |
·
·
q
r
P
(v,u)∈ER
+
(v,u)∈ER
wt ((v, u))2
r
.
Let us now condition on R such that the bounds of Lemma 7 hold. Note that such an R is
chosen with probability at least 1 − δ/2. We get
Var[X|R] ≤
250 1 m M2s−1
·
.
· ·
δ2 q n
n
We apply Chebyshev’s inequality and invoke Condition 2:
h
i
ε
Pr (X|R) − Exp[X|R] ≤ · µs ≤
2
4 · Var[X|R]
1 4 · (250/δ2 ) · m · M2s−1
δ
≤
≤ .
·
2
2
2
2
ε · µs
q
ε · Ms
2
By Lemma 7, Exp[X|R] = (1/r)wt(R) ∈ [(1 − ε/2)µs , (1 + ε/2)µs ]. By taking into account
both the probability that R does not satisfy one (or more) of the bounds in Lemma 7 and the
probability that X (conditioned on R satisfying these bounds) deviates by more than (ε/2)µs
from its expected value, we get that |X − µs | < εµs with probability at least (1 − δ/2)2 > 1 − δ.
3.2
The Algorithm with Edge Samples
As an aside, the above analysis can be slightly adapted to prove the result of Aliakbarpour et
al. [3] on estimating moments using random edge queries. Observe that we can then simply set
R =P
V and r = n in Moment-estimators . This immediately gives wt(R) = Ms , |ER | = 2m,
and (v,u)∈E + wt((v, u))2 ≤ 3M2s−1 (as shown in Equation (3)). Similarly to what is shown in
R
the proof of Theorem 8, Var[X] = O(q −1 · m · M2s−1 ) (where X is as defined in Step 3 of the
2−1/s
c
algorithm). As shown in Equation (10), M2s−1 ≤ Ms
. Thus, if we set q ≥ ε2 ·δq 3 · m1/s (for a
Ms
sufficiently large constant cq ), we satisfy Condition 2. This is exactly the bound of Aliakbarpour
et al.
4
Satisfying Conditions 1 and 2 in general graphs
We show how to set r and q to satisfy Conditions 1 and 2 in general graphs. Our setting of r
and q will give us the same query complexity as [26] (up to the dependence on 1/ε and log n, on
9
which we improve, and the exponential dependence on s in [26], which we do not incur). In the
next section we show how the setting of r and q can be improved using a degeneracy bound.
For cr and cq that are sufficiently large constants, we set
)
(
s−1/s
cq
n
n
cr
·
,
q = 2 3 · min n1−1/s , 1−1/s .
(4)
r= 2
ε · δ Ms1/(s+1)
ε ·δ
Ms
This setting of parameters requires the knowledge of Ms , which is exactly what we are trying
to approximate (up to the normalization factor of n). A simple geometric search argument
alleviates the need to know Ms . For details see Section 6.
In what follows (and elsewhere) we make use of Hölder’s inequality:
Theorem 9 (Hölder’s inequality). For values p and q such that p, q > 1 and
k
X
i=1
|xi · yi | ≤
k
X
i=1
|xi |p
! p1
k
X
i=1
|yi |q
! q1
1
p
+
1
q
= 1,
.
We refer to p and q as the conjugates of the formula.
In order to assert that r as set in Equation (4) satisfies Condition 1, it suffices to establish
the next lemma.
Lemma 10 (Condition 1 holds).
X
v
Proof:
1/(s+1)
Let θ = Ms
1
2− s+1
wt(v)2 ≤ 4Ms
.
be a degree threshold. We define
H , {v : dv > θ}
and
L , V \H .
This partition into “high-degree” vertices (H) and “low-degree” vertices (L) will be
Puseful in
upper bounding the maximum weight wt(v) of a vertex v, and hence upper bounding v wt(v)2 .
Details follow.
P
1/(s+1)
1/(s+1)
·
We first observe that |H| ≤ Ms
. This is true since otherwise, v∈H dsv > Ms
s
Mss+1 = Ms , which is a contradiction. We claim that this upper bound on |H| implies that
1/(s+1)
max d+
.
v ≤ Ms
(5)
v
1/(s+1)
To verify this, assume, contrary of the claim, that for some v, d+
. But then there
v > Ms
1/(s+1)
1/(s+1)
+
. This contradicts the bound
are at least Ms
vertices u such that du ≥ dv ≥ dv > Ms
on |H|.
P
It will also be useful to bound u∈H dus−1 . By Hölder’s inequality with conjugates s and
s/(s − 1) and the bound on |H|,
X
dus−1
u∈H
=
X
u∈H
1·
dus−1
1/s
≤ |H|
X
u∈H
! s−1
s
dsu
1
s−1
s
≤ Mss(s+1) · Ms
s
≤ Mss+1 .
(6)
We now turn to bounding maxv {wt(v)}. By the definition of wt(v) and the degree ordering,
X
X
X
X
wt(v) =
(dvs−1 + dus−1 ) ≤ 2
dus−1 = 2
dus−1 + 2
dus−1 .
(7)
u∈Γ+ (v)
u∈Γ+ (v)
u∈Γ+ (v)∩L
10
u∈Γ+ (v)∩H
1/(s+1)
For the first term on the right-hand-side of Equation (7), recall that du ≤ Ms
Thus, by Equation (5),
s−1
s
X
s+1
≤ Mss+1 .
dus−1 ≤ d+
v · Ms
for u ∈ L.
(8)
u∈Γ+ (v)∩L
For the second term, using Γ+ (v) ∩ H ⊆ H and applying Equation (6),
X
u∈Γ+ (v)∩H
Finally,
X
v
dus−1 ≤
X
u∈H
s
dus−1 ≤ Mss+1 .
X
wt(v)2 ≤ max{wt(v)} ·
v
v
(9)
wt(v) ≤ Ms2−1/(s+1) ,
where the second inequality follows by combining Equations (7)–(9) to get an upper bound on
maxv {wt(v)} and applying Claim 5.
The next lemma implies that Condition 2 holds for q as set in Equation (4).
Lemma 11 (Condition 2 holds).
(
min n1−1/s ,
Proof:
ns−1/s
1−1/s
Ms
)
≥ 2m ·
M2s−1
.
Ms2
We can bound M2s−1 in two ways. First, by a standard norm inequality, since s ≥ 1,
M2s−1 =
X
v
dv2s−1
≤
X
dsv
v
!(2s−1)/s
= Ms2−1/s .
(10)
We can also use the trivial bound dv ≤ n and get M2s−1 ≤ ns−1 · Ms . Thus, M2s−1 ≤
2−1/s
min{Ms
, ns−1 · Ms }. By applying Hölder’s inequality with conjugates s/(s − 1) and s we
get that
!1/s
X
X
s
(s−1)/s
dv
2m =
1 · dv ≤ n
·
= n1−1/s · Ms1/s .
(11)
v
v
We multiply the bound by M2s−1 to complete the proof.
5
The Degeneracy Connection
The degeneracy, or the coloring number, of a graph G = (V, E) is the maximum value, over all
subgraphs G′ of G, of the minimum degree in G′ . In this definition, we can replace “minimum”
by “average” to get a 2-factor approximation to the degeneracy (refer to [2]; Theorem 2.4.4
and Corollary
nl 5.2.3
mo of [19]). Abusing notation, it will be convenient for us to define α(G) =
|E(S)|
maxS⊆V
.
|S|−1
We also make the following observation regarding the relation between α(G) and Ms (G).
1
Claim 12. For every graph G, α(G) ≤ Ms (G) s+1 .
11
Proof:
Let S be a subset of vertices that maximizes
l
m
|E(S)|
|S|−1
2|E(S)|
|S|
, and let d(S) denote the
m
l
≥ |E(S)|
average degree in the subgraph induced by S. Then d(S) =
|S|−1 = α(G). Hence,
|S| ≥ α(G), and by Hölder’s inequality (Theorem 9) with conjugates s/(s − 1) and s,
α(G) · |S| ≤
1
X
v∈S
dv ≤
X
v∈S
!1
s
dsv
1
· |S|1− s ,
1
1
implying that α(G) · |S| s ≤ Ms (G) s . Since |S| ≥ α(G), we get that α(G) ≤ Ms (G) s+1 .
In this section, we show that the following setting of parameters for Moment-estimators
satisfies Conditions 1 and 2, for every graph G with degeneracy at most α (i.e., α(G) ≤ α), and
for appropriate constants cr and cq .
(
1/s )
α
n
cr
· min
, 2s · n · log2 n ·
,
(12)
r= 2
1/(s+1)
ε ·δ
Ms
Ms
cq
q = 2 3 · min
ε ·δ
(
n · α ns · α 1−1/s ns−1/s
,n
, 1−1/s
,
1/s
Ms
Ms
Ms
)
.
(13)
Clearly the setting of r and q in Equation (12) and Equation (13) respectively, can only
improve on the setting of r and q for the general case in Equation (4) (Section 4).
Equation (12) (when
Our main challenge is in proving that Condition 1 holds for r as set in P
the graph has degeneracy at most α). Here too, the goal is to upper bound v wt(v)2 . However,
as opposed to the proof of Lemma
P 10 in Section 4, where we simply obtained an upper bound on
maxv {wt(v)} (and bounded v wt(v)2 by maxv {wt(v)} · Ms ), here the analysis is more refined,
and uses the degeneracy bound. For details see the proof of our main lemma, stated next.
Lemma 13 (Condition 1 holds). For a sufficiently large constant c,
X
wt(v)2 ≤ c · 2s · α1/s · Ms2−1/s · log2 n .
v
In order to prove the lemma we first introduce the following definitions and claim.
Definition 14. For a set S and a vertex u, let ΓS (u) denote the set Γ(u) ∩ S, and let Γ+
S (u)
+
denote the set Γ (u) ∩ S. For two sets of vertices S and T (which are not necessarily disjoint),
let E + (S, T ) , {(u, v) : (u, v) ∈ E + , u ∈ S, v ∈ T }.
Definition 15. We partition the vertices (with degree at least 1) according to their degree. For
0 ≤ i ≤ ⌈log n⌉, let
Ui , {u ∈ V : du ∈ (2i−1 , 2i ]} .
For each i we partition the vertices in V according to the number of outgoing edges that they
have to Ui . Specifically, for 1 ≤ j ≤ ⌈log(n/α)⌉, define
n
o
j−1
j
.
(v)|
∈
2
α,
2
α
Vi,j , v ∈ V : |Γ+
Ui
Also define Vi,0 ,
i.
o
n
⌈log(n/α)⌉
(v)|
≤
α
. Hence, {Vi,j }j=0
is a partition of V for each
v ∈ V : |Γ+
Ui
12
A central building block in the proof of Lemma 13 is the next claim. This claim establishes
an upper bound on the number of edges going from vertices in Vi,j to vertices in Ui , for every
i and j (within the appropriate intervals). In the proof of this claim we exploit the degeneracy
bound.
Claim 16. Let Vi,j and Ui be as defined in Definition 15, and let E + (Vi,j , Ui ) be as defined in
Definition 14. For every 0 ≤ i ≤ ⌈log n⌉ and every 2 ≤ j ≤ ⌈log(n/α)⌉,
|E + (Vi,j , Ui )| ≤ Ms · 2−((i−1)(s−1)+j−1) .
Proof:
Since G has degeneracy at most α,
|E + (Vi,j , Ui )| = |E + (Vi,j , Ui )| = |E(Vi,j , Ui )| ≤ α · (|Vi,j | + |Ui |) .
On the other hand, by the definition of Vi,j ,
|E + (Vi,j , Ui )| ≥ 2j−1 · α · |Vi,j | .
Combining the above two bounds (and because 2j−1 − 1 ≥ 1 for j ≥ 2), we get that |Ui | ≥ |Vi,j |,
and we obtain the following bound on the number of edges in E + between Vi,j and their neighbors
in Ui :
|E + (Vi,j , Ui )| ≤ 2 · α · |Ui | .
(14)
We next upper bound |Ui |. By the definition of Vi,j , for every v ∈ Vi,j there exists a vertex
u ∈ Ui such that du ≥ dv . By the definition of Ui , for every u′ , u′′ ∈ Ui , du′ ≥ du′′ /2, implying
that for every v ∈ Vi,j and every u′ ∈ Ui it holds that du′ ≥ dv /2. Hence, for every u ∈ Ui , we
have that du ≥ 2j−2 α. Also by the definition of Ui , for every u ∈ Ui it holds that du ≥ 2i−1 .
Therefore,
X
Ms ≥
dsu ≥ |Ui | · 2(i−1)(s−1) · 2j−2 α ,
u∈Ui
which directly implies that
|Ui | ≤ Ms · 2−((i−1)(s−1)+j−2) · α−1 .
(15)
The proof follows by plugging in Equation (15) in Equation (14).
We are now ready to prove Lemma 13.
Proof of Lemma 13: By the definition of wt(v), and since dv ≤ du for every v and u ∈ Γ+ (v),
X
wt(v)2 =
v
X
v
X
u∈Γ+ (v)
dvs−1 + dus−1
!2
≤ 4·
X
v
X
u∈Γ+ (v)
dus−1
!2
.
(16)
In order to bound the expression on the right-hand-side of Equation (16) we consider each Ui
(as defined in Definition 15) separately: By applying Hölder’s inequality with conjugates s and
s/(s − 1) we get the following bound for every v and 0 ≤ i ≤ ⌈log n⌉.
X
u∈Γ+
i (v)
1/s
1 · dus−1 ≤ |Γ+
·
i (v)|
X
u∈Γ+
i (v)
13
dsu
(s−1)/s
1/s
≤ |Γ+
· Ms(s−1)/s .
i (v)|
(17)
For a vertex u, let Γ− (u) , {v : u ∈ Γ+ (v)}. By applying Equation (17) (to one term of the
2
P
s−1 ),
square
d
+
u∈Γ (v) u
i
X
v
X
u∈Γ+
i (v)
2
dus−1
X
≤
v
1/s
|Γ+
· Ms(s−1)/s ·
i (v)|
= Ms(s−1)/s ·
= Ms(s−1)/s ·
X
u∈Ui
dus−1 ·
⌈log n⌉
X
j=0
X
u∈Γ+
i (v)
X
v∈Γ− (u)
X
u∈Ui
dus−1
dus−1 ·
1/s
|Γ+
i (v)|
X
v∈Γ− (u)∩Vi,j
(18)
1/s
.
(v)|
|Γ+
i
(19)
Equation (18) follows by the definition of Γ− (u) (and switching the order of the summations),
and Equation (19) follows from splitting the sum in Equation (18) based on the partition of V
into the subsets Vi,j (recall that i is fixed for now and Vi,j is as defined in Definition 15).
From this point on we only consider j’s such that Vi,j is not empty. For each j (and i), by
the definition of Vi,j ,
X
X
X
X
1/s
|Γ+
≤
dus−1 ·
(2j · α)1/s
dus−1
i (v)|
u∈Ui
u∈Ui
v∈Γ− (u)∩Vi,j
v∈Γ− (u)∩Vi,j
≤ 2j/s · α1/s ·
X
u∈Ui
dus−1 · |Γ− (u) ∩ Vi,j | .
For j < 2, we trivially upper bound |Γ− (u) ∩ Vi,j | by du . Thus,
X
X
X
1/s
≤ 21/s · α1/s ·
dsu ≤ 2 · α1/s · Ms
|Γ+
dus−1
i (v)|
u∈Ui
(for j < 2) .
(20)
(21)
u∈Ui
v∈Γ− (u)∩Vi,j
Turning to j ≥ 2, since all vertices in Ui have degree at most 2i and by Equation (20), we get:
X
X
1/s
dus−1 ·
|Γ+
≤ 2j/s · α1/s · 2i(s−1) · |E + (Vi,j , Ui )|.
i (v)|
u∈Ui
v∈Γ− (u)∩Vi,j
Hence, by Claim 16,
X
u∈Ui
dus−1 ·
X
v∈Γ− (u)∩Vi,j
1/s
≤ 2j/s−j+s · α1/s · Ms .
|Γ+
i (v)|
(22)
By using the bound in Equation (21) for j < 2, the bound in Equation (22) for j ≥ 2 and
plugging them in Equation (19) we get
2
⌈log n⌉
X
X
X
dus−1 ≤ 4 · α1/s · Ms2−1/s + 2s · α1/s · Ms2−1/s ·
2j/s−j
v
j=3
u∈Γ+
i (v)
≤ 2s+1 · α1/s · Ms2−1/s · log n ,
(23)
where the last inequality
P holds because s ≥P1 (in fact, for s > 1 we can save a log n factor).
Using the inequality ( ℓi=1 xi )2 ≤ (2ℓ − 1) · ℓi=1 x2i , Equation (23) implies that
2
X
X
dus−1 ≤ 2s+2 · α1/s · Ms2−1/s · log2 n .
(24)
v
u∈Γ+ (v)
14
The lemma follows by combining Equation (16) with Equation (24).
It remains to establish Condition 2.
Lemma 17 (Condition 2 holds).
(
)
n · α ns · α 1−1/s ns−1/s
M2s−1
min
,
,n
, 1−1/s ≥ m ·
.
1/s
Ms
Ms2
Ms
Ms
Proof:
Since G has bounded degeneracy α, it follows that m ≤ n · α. By Equation (11),
2−1/s
1/s
, ns−1 · Ms }. The
m ≤ n1−1/s · Ms . As shown in the proof of Lemma 11, M2s−1 ≤ min{Ms
proof follows from the above two bounds.
5.1
The case of s = 1: estimating the average degree
When s = 1 (so that Ms = M1 = 2m and µs = µ1 is the average degree), there is a very simple
analysis of a slight variant of Moment-estimators . Observe that for s = 1, by Definition 4,
for every edge e, wt(e) = 2, and wt(v) = 2 · d+
v . For a degree threshold θ = 2α/ε, let H , {v :
dv > θ}, and L , V \ H. By the definition
P of H+we have that |H| < M1 /θ = εM1 /(2α).
Since the graph has degeneracy at most α, v∈H dv ≤ α · |H| ≤ εM1 /2. This implies that
P
v∈L wt(v) ≥ (1 − ε)M1 .
= 2 only if dvi ≤ θ (as
+ dus−1
Suppose that we modify the algorithm so that Xi is set to dvs−1
i
i
well as vi ≺ ui ), and is otherwise set to 0. Under
P this modification, Exp[X] ∈ [(1 − ε)M1 , M1 ].
Since wt(v) ≤ 2θ for each v ∈ L, we get that v∈L wt(v)2 ≤ 2θ · M1 = (4α/ε) · M1 . Therefore,
·n·α
. Thus, as compared to the setting
in order to satisfy Condition 1, it suffices to set r = εc3r·δM
1
2
in Equation (12), we save a log n factor (at the cost of factor of 1/ε), but, more importantly,
the analysis is very simple (as compared to the proof of Lemma 13). The setting of q is as in
c
Equation (13), which for s = 1 gives q = ε2 ·δq 3 .
6
Wrapping things up
The proof of our final result, Theorem 3, follows by combining Theorem 8, Lemma 10, Lemma 13
and Lemma 17, with a geometric search for a factor-2 estimate of Ms (which determines the
correct setting of r and q in the algorithm).
For convenience, we restate the bounds of Theorem 3 in terms on Ms .
α 1/s
n n · α ns · α o s log n · log(s log n)
,
·
O 2s · n · log2 n ·
+ min
1/s
Ms
Ms
ε2
Ms
O
n
1/(s+1)
Ms
n
ns−1/s o s log n · log(s log n)
+ min n1−1/s , 1−1/s
·
ε2
Ms
Proof of Proof of Theorem 3: Recall that the setting of r and q in Equation (12) and
Equation (13), respectively, equals the setting in Equation (4) when α is set to its maximum
1/(s+1)
possible value Ms
. Hence, it suffices to prove the theorem under the assumption that the
algorithm is provided with α (which upper bounds α(G)). It follows from Theorem 8, Lemma 10,
Lemma 13 and Lemma 17, that when Moment-estimators is invoked with parameters r and
q as set in Equation (12) and Equation (13), respectively, the algorithm returns a value X such
15
that X ∈ [(1− ε)µs , (1+ ε)µs ] with probability at least 1− δ. However, these settings require the
knowledge of Ms , which is the parameter we are trying to approximate (up to the normalization
factor n). Hence, we use the following search algorithm.
cs = ns+1 (the maximum possible value of Ms ), and compute r
We start with a guess M
cs , with the
according to Equation (12) and q according to Equation (13) assuming Ms = M
given approximation parameter ε and with δ = 1/3. We then invoke Moment-estimators
cs then we stop
Θ(log(s log n)) times and let Z be the median of the returned values. If Z ≥ M
c
and return Z. Otherwise we halve Ms and repeat.
Observe that r and q are decreasing functions of Ms . Hence, the running time of each
invocation of Moment-estimators is at most the running time of the last invocation. By
Claim 6 and Markov’s inequality, for each invocation of Moment-estimators , Pr[X ≥ 3Ms ] ≤
1/3 (where X is the value returned by the algorithm). We stress that this holds regardless of
whether r and q satisfy Conditions 1 and 2 (respectively). By the definition of Z, for values
cs > 3Ms , the probability that we will stop in each step is O(1/(s log n)). Therefore, with high
M
cs ≤ 3Ms .
constant probability we will not stop before M
3
c
Once Ms ≤ 3Ms , we will satisfy Conditions 1 and 2. By Theorem 8, in each invocation of
Moment-estimators , X ∈ [(1 − ε)µs , (1 + ε)µs ] with probability at least 1 − δ = 2/3, so that
cs ≤ Ms /2,
Z ∈ [(1 − ε)µs , (1 + ε)µs ] with probability at least 1 − O(1/(s log n)). Thus, once M
the algorithm will stop and return a value in [(1 − ε)µs , (1 + ε)µs ] with probability at least
1 − O(1/(s log n)). By a union bound over all iterations, the algorithm returns such a value
with high constant probability.
In order to bound the expected running time, we first observe that the running time of
cs ∈ [Ms /2i+1 , Ms /2i ) is at most 2i+1 times the
an invocation of Moment-estimators with M
cs = Ms . On the other hand, the probability that the
running time of an invocation with M
cs ∈ [Ms /2i+1 , Ms /2i ) for any i ≥ 1, is O(log(n)−i ).
algorithm does not stop before we reach M
The bound on the expected running time follows.
7
Lower Bounds for Bounded Degeneracy
The lower bounds given in this section hold for algorithms that are allowed degree and neighbor
queries, as well as pair queries (that is, queries of the form: “is there an edge between u and
v”). These lower bound show that the complexity of the algorithm, as stated in Theorem 3 for
graphs with degeneracy at most α, is tight up to the dependence on 1/ε and polylogarithmic
factors in n, for any constant s (or even s = O(log log n)).
Theorem
18. Any constant-factor approximation algorithm for Ms
n·α(G)1/s
queries.
Ω
1/s
Ms
must perform
(G)
fs and M
fs /ns ≤ α
fs /c)1/(s+1) ,4 we next define two families
Proof:
For every n, M
e ≤ (M
of graphs: G1 and G2 . Each graph in G1 consists of a clique C1 , over α
e vertices, and an
independent set C2 , over n − α
e vertices. For each graph in G2 , the vertices are partitioned
fs /e
fs /e
into three sets: C1 , C2 , C3 , of sizes α
e, (M
α)1/s − α
e, and n − (M
α)1/s , respectively. The
set C1 is a clique, and is connected by a complete bipartite graph to C2 , where there are no
edges within C2 . The set C3 is an independent set. Within each family, the graphs differ only
3
cs , we invoke Moment-estimators with
To be precise, in order to satisfy the conditions for values Ms ≤ 3M
3r and 3q.
1
P
P
4
Recall that by Claim 12, α(G) ≤ Ms (G) s+1 , and note that Ms (G) = v dsv ≤ v dv · ns−1 ≤ 2α(G)ns .
16
in the labeling of the vertices. By construction, in both families, all graphs have degeneracy
fs /c, and for each G in G2 ,
Θ(e
α). For each G ∈ G1 we have that Ms (G) = O(e
αs+1 ) ≤ M
1/s
s
fs /e
fs /e
fs ).
Ms (G) = Θ(e
α · (M
α) + ((M
α) − α
e) · α
e ) = Θ(M
Clearly, unless the algorithm “hits” a vertex in the clique C1 of a graph belonging to G1 or
a vertex in C1 ∪ C2 in a graph belonging to G2 , it cannot distinguish between a graph selected
randomly from
G1 and a graph selected randomly from G2 . The probability of hitting such a
f
1/s
α)
. Thus, in order for this event to occur with high constant probability,
vertex is O (Ms /e
n
1/s
α
queries are necessary.
Ω ne
f1/s
Ms
Theorem 19. Any constant-factor approximation algorithm for Ms must perform
)!
(
n · α(G) ns · α(G) 1−1/s
ns−1/s
,
Ω min
,n
,
Ms (G)1/s Ms (G)
Ms (G)1−1/s
queries.
Proof:
The proof of the theorem is based on simple modifications to the lower bound
constructions in [26, Thm. 5]. We note that this theorem of [26] was stated explicitly for
algorithms that perform degree and neighbor queries since such was the algorithm presented
in [26] (as well as the current paper). However, it was noted in [26, Sec. 7] that they also hold
when pair queries are allowed (as they are essentially based on “hitting special vertices”).
The theorem is proved by considering two cases that are defined according to the relation
between Ms (G) and n (namely, Ms (G)1/s ≤ n − c and Ms (G)1/s > n − c for a constant c,
which determines the hardness of approximation). Within each case there are two sub-cases
that are defined according to the relation between α(G) and (Ms (G)/n)1/s (namely, α(G) <
(Ms (G)/n)1/s and α(G) ≥ (Ms (G)/n)1/s ). Each of the four sub-cases gives us one of the terms
in the lower bound (within the min{·} expression). For each of the sub-cases we consider the
fs that correspond to the sub-case, and we construct two families of graphs:
values α
e and M
G1 and G2 . In both families all graphs have degeneracy Θ(e
α). Every graph G ∈ G1 satisfies
f
fs . The lower bound is based on
Ms (G) ≤ Ms , while every graph G ∈ G2 satisfies Ms (G) ≥ c · M
the difficulty of distinguishing between a random graph selected from G1 and a random graph
selected from G2 , with a specified number of queries. In all cases we modify the construction
in [26, Thm. 5] either by decreasing the degeneracy or increasing it.
f 1/s ≤ n − c. For this case we modify the construction described in [26, Item (2)
The case M
s
of Thm. 5]. In both families, the vertices are partitioned into two subsets, S and V \ S, where
the size of S is c. For each graph in G1 , the set S is an independent set, while for each graph
in G2 , each vertex in S has d′ neighbors in V \ S, for an appropriate setting of d′ . In both
families, the vertices in V \ S have degree d, for an appropriate setting of d (where in G1 their
neighbors are all in V \ S while in G2 some of their neighbors are in S). Observe that the
graphs in the family G1 can be viewed as obtained from the graphs of G2 be replacing pairs
of edges between S to V /S by a single edge in V \ S. We refer to the edges between S and
V \ S in the graphs of G2, and the edges replacing them in the graphs of G1 as “special edges”.
In [26], the settings of d′ and d are such that the number of stars in graphs belonging to G2 is
(roughly) a factor c larger than the number of stars in graphs belonging to G1 . The difficulty
of distinguishing between a random graph selected from G1 and a random graph selected from
G2 is based on upper bounding the following two very similar events: (1) “Hitting” a vertex
v in S when querying a graph in G2 by either performing a degree/neighbor query on v or by
performing a neighbor query on u ∈ V \ S and receiving v as an answer. (2) “Hitting” a vertex
17
v in S or a special edge between vertices in V \ S when querying a graph in G1 , where the
number of special edges is |S| · d′ /2. In what follows we modify the settings of d′ and d (as
defined in [26, Item (2) of Thm. 5]), and in the case of large α perform an additional small
modification.
fs /n)1/s and the sub-case α
fs /n)1/s , we set d′ = ⌈M
fs1/s ⌉.
In both the sub-case α
e < (M
e ≥ (M
fs .
This ensures that for each graph G ∈ G2 , Ms (G) ≥ cM
1/s
f
If α
e < (Ms /n) , then we set d = α
e. Hence, α(G) = Θ(e
α) for graphs in both families, and
1/s
s
f
f
Ms (G) ≤ n · ((Ms /n) ) = Ms for graphs in G1 . Since the number of edges between S and
V \ S in graphs belonging to G2 (which is of the same order as the number of special edges in
f1/s
graphs belong
e), we get a lower
to G1 ) is O(Ms ) while the total number of edges is Ω(n · α
1/s
f
(this of course requires formalizing, as done in [26]).
bound of Ω n · α
e/Ms
fs /n)1/s , then we set d = ⌊(M
fs /n)1/s ⌋, so that it still holds that Ms (G) ≤ n ·
If α
e ≥ (M
fs /n)1/s )s = M
fs for graphs in G1 . In both families, within V \ S we add edges so as to form
((M
a clique on a subset of size α
e, thus increasing the degeneracy to Θ(e
α). Since this modification is
the same in both families, it does not effect the ability to distinguish between the two families.
fs1/s · n1−1/s ),
Since the number of edges (not including those in the clique) is (n − c) · d = Ω(M
1−1/s
we get a lower bound of Ω(n
).
f 1/s > n − c. In this case we may assume, without loss of generality, that
The case M
s
fs < ns+1 /c′ for a sufficiently large constant c′ , or else the lower bound Ω(ns−1/s /M
fs1−1/s ) is
M
fs < ns · α
fs ) is trivial. We may
trivial, and similarly that M
e/c′ , or else the lower bound Ω(ns · α
e/M
1/(s+1)
fs
also assume that α
e≤M
since α(G) ≤ Ms (G)1/(s+1) for every graph G (by Claim 12).
Here we modify the construction described in [26, Item (3) of Thm. 5]. The construction
fs1/s ≤ n − c, except that the size of the set S needs to be
is similar to the one described for M
fs /ns ⌉, and in the graphs in G2 each vertex in
increased. Specifically, we let |S| = b for b = ⌈cM
fs ).
S is connected to every other vertex in the graph. Therefore, for each G ∈ G2 , Ms (G) = Ω(M
1/s
f
If α
e < (Ms /n) , then in both families each vertex in V \ S has degree d = α
e. Since b < α
e
s
′
′
f
(due to Ms < n · α
e/c for a sufficiently large constant c ), we get that α(G) = Θ(e
α) for all
graphs in G1 and in G2 . The difference between the families is that in the graphs belonging to
G2 , each vertex in V \ S has b = |S| neighbors in S and d − b neighbors in V \ S, while in the
graphs belonging to G1 , each vertex in V \ S has d neighbors in V \ S (and each vertex in S
fs
only neighbors each other vertex in S). This implies that for each G ∈ G1 , Ms (G) ≤ ne
αs < M
(where we have again used b < α
e). Since the number of edges between S and V \ S in each
graph in G2 (the number of corresponding special edges within V \ S in each graph in G1 ) is
fs /ns−1 ), and the total number of edges is Ω(n · α
O(b · n) = O(M
e), we get a lower bound of
s
fs ).
Ω(n · α
e/M
fs /n)1/s , then we use the same construction as above only with d = (M
fs /n)1/s ,
If α
e ≥ (M
and we “plant” a clique of size α
e in V \ S. The degeneracy is hence increased to Θ(e
α), and the
fs (due to the clique). Since the number
value of Ms (G) for G ∈ G1 is increased to at most 2M
fs1/s · n1−1/s ), we get a lower bound
of edges (not including those in the clique) is Ω(n · d) = Ω(M
1−1/s
fs
of Ω(ns−1/s /M
).
7.1
On knowing the degeneracy bound
One may wonder if the query complexity of Theorem 3 can be obtained without knowledge of
the degeneracy α. The ideal situation would be one where an algorithm has the complexity of
Moment-estimator, without knowing α. The worst-case lower bound of [26] do not preclude
18
this possibility, since it uses graphs with high degeneracy. Nonetheless, a slight adaptation
of those arguments shows that the bound holds even for bounded degeneracy graphs, if the
algorithm must work on all graphs. We will focus solely on estimating average degree, since
that suffices to make our point.
Definition 20. For a constant c, an algorithm A is called c-valid if: given query access to a
graph G, with probability 2/3, A outputs a c-approximation to the average degree of G.
Note that Moment-estimator, with access to a degeneracy bound, is not valid. This is
because it is only required to be accurate for graphs with the given degeneracy bound, not all
graphs.
Theorem 21. Let n be a sufficiently large integer. Consider the class of graphs on n vertices
√
with degeneracy at most 2. For any constant c, any c-valid algorithm must perform Ω( n)
queries on these graphs.
Proof: Similarly to previous lower-bound proofs, for each sufficiently large n, we define two
distributions over labeled graphs. Consider a graph consisting of two connected components,
√
√
a cycle on ⌊n − 4c n⌋ vertices and a cycle on ⌈4c n⌉ vertices (where c is the constant in the
statement of the theorem). The distribution G1 is generated by labeling the vertices using a
uniform random permutation in [n]. For the second distribution, take the graph that consists of
√
√
a cycle on ⌊n − 4c n⌋ vertices and a clique on ⌈4c n⌉ vertices. The distribution G2 is generated
by labeling the vertices according to a uniform random permutation.
Consider any (possibly randomized) algorithm for deciding, given query access either to
a graph generated by G1 or to a graph generated by G2 , according to which distribution was
the graph generated. As long as the algorithm does not perform a query on a vertex that
belongs to the small cycle (if the graph is generated by G1 ) or the small clique (if the graph is
generated by G2 ), answers to its queries are identically distributed under the two distribution.
√
This implies that any such decision algorithm must perform Ω( n) queries in order to succeed
with probability at least 2/3.
Now consider a c-valid algorithm A that makes at most s queries on any graph with n
vertices and degeneracy at most 2. We use this algorithm in order to construct an algorithm
B that distinguishes G1 from G2 in O(s) queries. It works as follows. Given query access to
G, B runs 20 independent runs of A. If any run makes more than s queries, B terminates and
outputs “G2 ”. At the end of the runs, all of them have provided some estimate for the average
degree. If the median is at most 2c, then B outputs “G1 ”. Otherwise, it outputs “G2 ”.
Suppose G ∼ G1 , and recall that the average degree of G is exactly 2. All runs are guaranteed
to make at most s queries. Thus, all of them will output an estimate. By the validity of A and
a Chernoff bound, the median estimate will be at most 2c with probability at least 2/3, and
B gives a correct output. Now suppose that G ∼ G2 , and recall that The average degree of G
is at least 3c2 . If any run takes more than s queries, then B outputs correctly. Otherwise, by
a similar argument to the one made for G ∼ G1 , the median estimate will be at least 3c with
probability at least 2/3. Thus, B is correct.
The query complexity of B is O(s) and it distinguishes G1 from G2 with probability at least
√
2/3. Therefore, s must be Ω( n).
References
[1] Arboricity. Wikipedia. https://en.wikipedia.org/wiki/Arboricity. 6
19
[2] Graph degeneracy. Wikipedia. https://en.wikipedia.org/wiki/Degeneracy_(graph_theory).
11
[3] M. Aliakbarpour, A. S. Biswas, T. Gouleakis, J. Peebles, R. Rubinfeld, and A. Yodpinyanee.
Sublinear-time algorithms for counting star subgraphs with applications to join selectivity
estimation. Technical Report 1601.04233, Arxiv, 2016. 2, 4, 5, 9
[4] N. Alon and S. Gutner. Linear time algorithms for finding a dominating set of fixed size
in degenerated graphs. In Proceedings of COCOON, pages 394–405, 2008. 6
[5] N. Alon, Y. Matias, and M. Szegedy. The space complexity of approximating the frequency
moments. J. Comput. Syst. Sci., 58(1):137–147, 1999. 4
[6] A-L. Barabasi and R. Albert. Emergence of scaling in random networks. Science, 286:509–
512, October 1999. 2, 3
[7] J. Berry, L. Fostvedt, D. Nordman, C. Phillips, C. Seshadhri, and A. Wilson. Why do simple
algorithms for triangle enumeration work in the real world? In Innovations in Theoretical
Computer Science (ITCS), pages 225–234, 2014. doi:10.1145/2554797.2554819. 2
[8] Zhiqiang Bi, Christos Faloutsos, and Flip Korn. The “DGX” distribution for mining massive, skewed data. In KDD ’01, pages 17–26. ACM, 2001. doi:10.1145/502512.502521.
2
[9] P. Bickel, A. Chen, and E. Levina. The method of moments and degree distributions for
network models. Annals of Statistics, 39(5):2280–2301, 2011. 2
[10] P. Brach, M. Cygan, J. Laccki, and P. Sankowski. Algorithmic complexity of power law
networks. In Proceedings of SODA, pages 1306–1325, 2016. 2
[11] A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins,
and J. Wiener. Graph structure in the web. Computer Networks, 33:309–320, 2000. 2
[12] B. Chazelle, R. Rubinfeld, and L. Trevisan. Approximating the minimum spanning tree
weight in sublinear time. SIAM Journal on Computing, 34(6):1370–1379, 2005. 6
[13] N. Chiba and T. Nishizeki. Arboricity and subgraph listing algorithms. SIAM J. Comput.,
14:210–223, 1985. URL: http://dx.doi.org/10.1137/0214017, doi:10.1137/0214017.
4, 6
[14] F. Chierichetti, A. Dasgupta, R. Kumar, S. Lattanzi, and T. Sarlos. On sampling nodes
in a network. In Proceedings of WWW, 2016. 2, 6
[15] A. Clauset, C. R. Shalizi, and M. E. J. Newman. Power-law distributions in empirical data.
SIAM Review, 51(4):661–703, 2009. doi:10.1137/070710111. 2
[16] A. Czumaj, F. Ergun, L. Fortnow, A. Magen, I. Newman, R. Rubinfeld, and C. Sohler.
Approximating the weight of the euclidean minimum spanning tree in sublinear time. SIAM
Journal on Computing, 35(1):91–109, 2005. 6
[17] A. Czumaj and C. Sohler. Estimating the weight of metric minimum spanning trees in
sublinear time. SIAM Journal on Computing, 39(3):904–922, 2009. 6
[18] A. Dasgupta, R. Kumar, and T. Sarlos. On estimating the average degree. In Proceedings
of WWW, pages 795–806, 2014. 2, 6
20
[19] R. Diestel. Graph Theory. Springer, fourth edition edition, 2010. 11
[20] T. Eden, A. Levi, D. Ron, and C. Seshadhri. Approximately counting triangles in sublinear
time. In Proceedings of the Fifty-Sixth Annual Symposium on Foundations of Computer
Science (FOCS), pages 614–633, 2015. 4, 6
[21] D. Eppstein, M. Loffler, and D. Strash:. Listing all maximal cliques in sparse graphs in
near-optimal time. In Proceedings of ISAAC, pages 403–413, 2010. 6
[22] P. Erdos and T. Gallai. Graphs with prescribed degree of vertices (hungarian). Mat. Lapok,
11:264–274, 1960. 2
[23] M. Faloutsos, P. Faloutsos, and C. Faloutsos. On power-law relationships of the internet
topology. In Proceedings of SIGCOMM, pages 251–262, 1999. 2
[24] U. Feige. On sums of independent random variables with unbounded variance and estimating the average degree in a graph. SIAM Journal on Computing, 35(4):964–984, 2006.
2
[25] O. Goldreich and D. Ron. Approximating average parameters of graphs. Random Structures
and Algorithms, 32(4):473–493, 2008. 2, 5
[26] M. Gonen, D. Ron, and Y. Shavitt. Counting stars and other small subgraphs in sublineartime. SIAM Journal on Discrete Math, 25(3):1365–1411, 2011. 2, 9, 10, 17, 18
[27] S.L. Hakimi. On the realizability of a set of integers as degrees of the vertices of a graph.
SIAM J. Applied Math, 10:496–506, 1962. 2
[28] A. Hassidim, J. A. Kelner, H. N. Nguyen, and K. Onak. Local graph partitions for approximation and testing. In Proceedings of the Fiftieth Annual Symposium on Foundations of
Computer Science (FOCS), pages 22–31. IEEE, 2009. 6
[29] V. Havel. A remark on the existence of finite graphs (czech). Casopis Pest. Mat., 80:477–
480, 1955. 2
[30] S. Marko and D. Ron. Approximating the distance to properties in bounded-degree and
general sparse graphs. ACM Transactions on Algorithms, 5(2), 2009. 6
[31] D. Matula and L. Beck. Smallest-last ordering and clustering and graph coloring algorithms.
JACM, 30(3):417–427, 1983. 6
[32] J. Nešetřil and P. Ossana de Mendez. Sparsity. Springer, 2010. 3
[33] H. N. Nguyen and K. Onak. Constant-time approximation algorithms via local improvements. In Proceedings of the Forty-Ninth Annual Symposium on Foundations of Computer
Science (FOCS), pages 327–336. IEEE, 2008. 6
[34] K. Onak, D. Ron, M. Rosen, and R. Rubinfeld. A near-optimal sublinear-time algorithm
for approximating the minimum vertex cover size. In Proceedings of the twenty-third annual
ACM-SIAM symposium on Discrete Algorithms, pages 1123–1131. SIAM, 2012. 6
[35] M. Parnas and D. Ron. Approximating the minimum vertex cover in sublinear time and
a connection to distributed algorithms. Theoretical Computer Science, 381(1-3):183–196,
2007. 6
21
[36] D. Pennock, G. Flake, S. Lawrence, E. Glover, and C. L. Giles. Winners don’t take
all: Characterizing the competition for links on the web. PNAS, 99(8):5207–5211, 2002.
doi:10.1073/pnas.032085699. 2
[37] A. Sala, L. Cao, C. Wilson, R. Zablit, H. Zheng, and B. Y. Zhao. Measurement-calibrated
graph models for social network experiments. In WWW ’10, pages 861–870. ACM, 2010.
doi:10.1145/1772690.1772778. 2
[38] O. Simpson, C. Seshadhri, and A. McGregor. Catching the head, tail, and everything
in between: A streaming algorithm for the degree distribution. In Proceedings on the
International Conference on Data Mining (ICDM), pages 979–984, 2015. 6
[39] Y. Yoshida, M. Yamamoto, and H. Ito. An improved constant-time approximation algorithm for maximum-matchings. In Proceedings of the Fourty-First Annual ACM Symposium
on the Theory of Computing, pages 225–234. ACM, 2009. 6
22
| 8cs.DS
|
arXiv:1606.00976v1 [math.ST] 3 Jun 2016
Marginal Estimation of Parameter Driven Binomial
Time Series Models
W.T.M. Dunsmuir
School of Mathematics and Statistics, University of New South Wales, Sydney, Australia.
[email protected]
and
J.Y. He
School of Mathematics and Statistics, University of New South Wales, Sydney, Australia
[email protected]
Abstract: This paper develops asymptotic theory for estimation of parameters in regression models
for binomial response time series where serial dependence is present through a latent process. Use
of generalized linear model (GLM) estimating equations leads to asymptotically biased estimates of
regression coefficients for binomial responses. An alternative is to use marginal likelihood, in which
the variance of the latent process but not the serial dependence is accounted for. In practice this is
equivalent to using generalized linear mixed model estimation procedures treating the observations
as independent with a random effect on the intercept term in the regression model. We prove this
method leads to consistent and asymptotically normal estimates even if there is an autocorrelated
latent process. Simulations suggest that the use of marginal likelihood can lead to GLM estimates
result. This problem reduces rapidly with increasing number of binomial trials at each time point but,
for binary data, the chance of it can remain over 45% even in very long time series. We provide a
combination of theoretical and heuristic explanations for this phenomenon in terms of the properties
of the regression component of the model and these can be used to guide application of the method
in practice.
Keywords and phrases: Binomial time series regression, Parameter driven models, Marginal likelihood.
1. Introduction
Discrete valued time series are increasingly of practical importance with applications in diverse fields such as
analysis of crime statistics, econometric modelling, high frequency financial data, animal behaviour, epidemiological assessments and disease outbreak monitoring, and modern biology including DNA sequence analysis
– see Dunsmuir, Tran and Weatherburn (2008). In this paper we focus on time series of binomial counts.
Two broad classes of models for time series of counts, based on the categorization of Cox (1981), are
generally discussed in the literature: observation driven models, in which the serial dependence relies on previous observations and residuals; and parameter driven models, in which the serial dependence is introduced
through an unobserved latent process. Estimation of parameter driven models is significantly challenging
especially when the latent process is correlated. Therefore methods that provides preliminary information
of the regression parameters without requiring a heavy computation load would be appealing. For example,
the use of generalized linear model (GLM) estimation for obtaining estimates of the regression parameters is
discussed in Davis, Dunsmuir and Wang (2000) and Davis and Wu (2009) for Poisson and negative binomial
observations respectively. GLM estimation is consistent and asymptotically normal for these two types of
response distribution even when there is a latent process inducing serial dependence. However as recently
pointed out by Wu and Cui (2014) and discussed in more detail below, use of GLM for binary or binomial
data leads to asymptotically biased estimates. Wu and Cui (2014) propose a semiparametric estimation
method for binary response data in which the marginal probability of success modelled non-parametrically.
This paper takes a different approach and suggest using estimation based on one-dimensional marginal distributions which accounts for the variance of the latent process but not the serial dependence. Such a procedure
is easy to implement using standard software for fitting generalized linear mixed models (GLMM). We show
that this method leads to estimates of regression parameters and the variance of the latent process, which
are consistent and asymptotically normal even if the latent process includes serial dependence. Additionally
the method extends easily to other response distributions such as the Poisson and negative binomial and in
1
/Marginal Estimation of Binomial Time Series Mixed Models
2
these cases will improve efficiency of regression parameters related to GLM estimates.
Suppose Yt represents the number of successes in mt trials observed at time t. Assume that there are n
observations {y1 , . . . , yn } from the process {Yt } and that xnt is an observed r-dimensional vector of regressors,
which may depend on the sample size n to form a triangular array, and whose first component is unity for
an intercept term. Then given xnt and a latent process {αt }, the Yt are independent with density
fYt (yt |xnt , αt ; θ) = exp {yt Wt − mt b(Wt ) + c(yt )}
(1)
in which
Wt = xTnt β + αt ,
t
and b(Wt ) = log(1 + exp(Wt )) and c(yt ) = log m
yt . Then
E(Yt |xnt , αt ) = mt ḃ(Wt ),
(2)
Var(Yt |xnt , αt ) = mt b̈(Wt ).
The process {αt } is not observed and because of this is referred to as a latent process. Often {αt } is assumed
to be a stationary Gaussian linear process with zero mean and auto-covariances
Cov (αt , αt+h ) = τ R(h; ψ)
where τ is the marginal variance of αt and ψ are the parameters for the serial dependence in the model of
αt . The specification of stationary Gaussian linear process covers many practical applications and we will
assume that for the remainder of the paper. However, Gaussianity it not required for the main asymptotic
results presented here, and in general, αt can be assumed a stationary strongly mixing process. We will
discuss this extension further in Section 6.
We let θ = (β, τ, ψ) denote the collection of all parameters and let θ0 be the true parameter vector. For
the above model the likelihood is defined in terms of an integral of dimension n as follows,
Z
L(θ) :=
n
Y
exp {yt Wt − mt b(Wt ) + c(yt )} g(α; τ, ψ)dα
(3)
Rn t=1
where g(α; τ, ψ) is the joint density of α = (α1 , . . . , αn ) given the parameters τ and ψ.
Maximization of the likelihood (3) is computationally expensive. Methods for estimating the high dimensional integrals in (3) using approximations, Monte Carlo method or both are reviewed in Davis and
Dunsmuir (2015). However simple to implement methods that provide asymptotically normal unbiased estimators of β and τ without the need to fit the full likelihood are useful for construction of statistics needed to
investigate the strength and form of the serial dependence. They can also provide a initial parameter values
for the maximization of the full likelihood (3).
For practitioners, GLM estimation has strong appeal as it is easy to fit with standard software packages. GLM estimators of the regression parameters β are obtained by treating the observations yt as being
independent with Wt = xTnt β and using the GLM log-likelihood
l0 (β) := log L0 (β) =
n
X
[yt (xTnt β) − mt b(xTnt β) + c(yt )]
(4)
t=1
We let β̃ denote value of β which maximises (4). This GLM estimate assumes that there is no additional
unexplained variation in the responses beyond that due to the regressors xnt .
However, as recently noted by Wu and Cui (2014), GLM does not provide consistent estimates of β when
Wt contains a latent autocorrelated component. To be specific, for deterministic regressors xnt = h(t/n) for
example, n−1 l0 (β) has limit
Z
1
Z
M
m Z
X
X
ḃ(h(u) β0 + α)g(α; τ0 )dα(h(u) β) − b(h(u) β) du +
κm
T
Q(β) = m̄
0
R
T
T
m=1
j=0
0
1
π 0 (j)c(j)du
/Marginal Estimation of Binomial Time Series Mixed Models
3
where m̄ = E(mt ), κm = P (mt = m) and π 0 (j) = P (Yt = j|xnt , θ0 ). We show below that β̃ converges to β 0 ,
which maximizes Q(β). Equivalently β 0 is the unique vector that solves
Z 1 Z
T
T 0
ḃ(h(u) β0 + α)g(α; τ0 )dα − ḃ(h(u) β ) h(u)du = 0
(5)
m̄
0
R
In the Poisson or negative binomial cases, mt ≡ 1, and E(Yt ) = E(ḃ(xTnt β0 + αt )) = ḃ(xTnt β0 + τ2 ), in which
τ /2 only modifies the regression intercept but does not influence the response to other regression terms. Such
an identity does not usually hold for binomial observations. When τ0 > 0, the relationship between β 0 and
β0 in binomial logit regression models has been investigated by several researchers.
For example, Neuhaus,
R
T
Kalbfleisch and Hauck (1991) proved that the logit of the marginal probability (1 + e−(x β0 +α) )−1 g(α)dα
T ∗
∗
can be approximated with x β , where |β | ≤ |β0 | for single covariate x and the equality is only attained
when τ = 0 or β0 = 0. Wang and Louis (2003) proved that only if g(·) is the “bridge” distribution, the
R
T
logit of (1 + e−α−x β0 )−1 g(α)dα equals to xT β0 holds; Wu and Cui (2014) proposed their MGLM method
because GLM estimates for binomial observations generated under model (2) are inconsistent.
To overcome the inconsistency observed in GLM estimation, in this paper we propose use of marginal
likelihood estimation, which maximises the likelihood constructed under the assumptions that the process αt
consists of independent identically distributed random variables. Under this assumption the full likelihood
(3) is replaced by the “marginal” likelihood
L1 (δ) =
n
Y
t=1
f (yt |xt , δ) =
n Z
Y
t=1
exp (yt Wt − mt b(Wt ) + c(yt )) g(αt ; τ )dαt .
and the corresponding “marginal” log-likelihood function is
Z
n
n
X
X
log exp (yt Wt − mt b(Wt ) + c(yt )) g(αt ; τ )dαt .
log f (yt |xt , δ) =
l1 (δ) =
t=1
t=1
(6)
R
(7)
R
where δ = (β, τ ) and g(·, τ ) is the density for a mean zero variance τ normal random variable. Let δ̂ be
the T
estimates obtained by maximising (7) over the compact parameter space Θ := {β ∈ Rr : kβ − β0 k ≤
d1 } {τ ≥ 0 : |τ − τ0 | ≤ d2 }, where d1 < ∞, d2 < ∞.
Marginal likelihood estimators of δ̂ can be easily obtained with standard software packages for fitting
generalized linear mixed models. Since these marginal likelihood estimates δ̂ are consistent, they can be used
as the starting value of full likelihood based on (3). Additionally, the asymptotic distribution of δ̂, and the
standard deviation derived from the asymptotic covariance matrix can be used to assess the significance of
regression parameters β̂. Moreover, in another paper we have developed a two-step score-type test to first
detect the existence of a latent process and if present whether there is serial dependence. The asymptotic
results of this paper are needed in order to derive the large sample chi-squared distribution of the second
step, the score test for detecting serial dependence.
Large sample properties of the marginal likelihood estimates δ̂ are provided in Section 2; The simulations
of Section 4 show that marginal likelihood estimates lead to a high probability of τ̂ = 0 when the number
of trials, mt , is small. In particular P (τ̂ = 0) can be almost 50% for binary data. Hence Section 3 focuses
on obtaining asymptotic approximations to the upper bound for P (τ̂ = 0), which is useful to quantify the
proportion of times the marginal likelihood procedures ‘degenerates’ to the GLM procedure. Also in Section
3 we derive a theoretical mixture distribution which provided better approximation in this situation. Section
4 presents simulation evidence to demonstrate the accuracy of the asymptotic theory and the covariance
matrix of the marginal likelihood estimate. Section 5 discusses the difference between marginal likelihood
estimation and MGLM estimation of Wu and Cui (2014). Section 6 concludes.
2. Asymptotic Theory for Marginal Likelihood Estimates
We present the large sample properties for the marginal likelihood estimates of β and τ obtained by maximizing (7). We begin by presenting the required conditions on the latent process {αt }, the regressors {xnt }
and the sequence of binomial trials {mt }.
/Marginal Estimation of Binomial Time Series Mixed Models
4
A process {αt } is strongly mixing if
ν(h) = sup
|P (AB) − P (A)P (B)| → 0
sup
A∈Ft−∞ ,B∈F∞
t+h
t
as h → ∞, where Ft−∞ and F∞
t+h are the σ-fields generated by {αs , s ≤ t} and {αs , s ≥ t + h} respectively.
In practice, the number of trials mt may vary with time. To allow for this we introduce:
Condition 1. The sequence of trials {mt : 1 ≤ mt ≤ M } is a stationary strongly mixing process independent
∞
M
P
P
to {Xt }; the mixing coefficients satisfy:
ν(h) < ∞. Let κj = P (mt = j), assume κM > 0,
κj = 1.
j=1
h=0
An alternative would be to take mt as deterministic and asymptotically stationary in which case the κj
would be limits of finite sample frequencies of occurrences mt = j. Both specifications obviously include the
case where mt = M for all t, of which M = 1 yields binary responses.
As in previous literature Davis, Dunsmuir and Wang (2000), Davis and Wu (2009) and Wu and Cui (2014)
we allow for both deterministic and stochastic regressors:
Condition 2. The regression sequence is specified in one of two ways:
(a) Deterministic covariates defined with functions: xnt = h(t/n) for some specified piecewise continuous
vector function h : [0, 1] → Rr .
(b) Stochastic covariates which are a stationary vector process: xnt = xt for all n where {xt } is an observed
T
trajectory of a stationary process for which E(es Xt ) < ∞ for all s ∈ Rr .
Condition 3. Let r = dim(β). The regressor space X = {xnt : t ≥ 1}, assume rank(span(X)) = r.
The full rank of the space spanned by the regressors required for Condition 3 holds for many examples. For
instance, for deterministic regressors generated by functions given in Condition 2a, such Xi , i = 1, . . . , r exist
if there are r different values of ui = (u1 , . . . , ur ) such that the corresponding function (h(u1 ), . . . , h(ur )) are
linearly independent. For stochastic regressors generated with a stationary process given in Condition 2b,
linearly independent Xi , i = 1, . . . , r can be found almost surely if Cov(X) > 0.
Condition 4. The latent
P∞ process {αt }, is strictly stationary, Gaussian and strongly mixing with the mixing
coefficients satisfying h=0 ν(h)λ/(2+λ) < ∞ for some λ > 0.
Conditions for a unique asymptotic limit of the marginal likelihood estimators are also required. Denote
the marginal probability of j successes in mt trials at time t as
Z
πt (j) =
ejWt −mt b(Wt )+c(j) φ(zt )dzt , j = 1, . . . , mt .
R
where Wt = xTnt β + τ 1/2 zt , and zt = αt /τ 1/2 has unit variance. If {αt } is Gaussian, so is the process {zt }
and zt ∼ N (0, 1) with density function φ(·). Similarly let πt0 (j) be the marginal probability evaluated with
the true values β0 and τ0 at time t. Define
Qn (δ) =
1
l1 (δ)
n
(8)
conditional on mt and xnt ,
n
E(Qn (δ)) =
m
t
1 XX
π 0 (j) log πt (j),
n t=1 j=0 t
δ ∈ Θ.
(9)
a.s.
Under Conditions 1 and 2, E(Qn (δ)) → Q(δ). Let π(j, ·) = P (Y = j|·, δ), and π 0 (j, ·) is evaluated with
δ0 , then under Condition 2a,
Q(δ) =
M
X
m=1
Z
κm
m
1X
0 j=0
π 0 (j, h(u)) log π(j, h(u))du
(10)
/Marginal Estimation of Binomial Time Series Mixed Models
5
and, under Condition 2b,
Q(δ) =
M
X
Z
κm
Rr
m=1
m
X
π 0 (j, x) log π(j, x)dF (x)
(11)
j=0
the proof is included in the proof of Theorem 1.
Condition 5. Q(δ) has a unique maximum at δ0 = (β0 , τ0 ), the true value.
We now establish the consistency and asymptotic normality of the marginal likelihood estimator.
Theorem 1 (Consistency and asymptotic normality of marginal likelihood estimators). Assume τ0 > 0 and
√
d
a.s.
−1
Conditions 1 to 5, then δ̂ → δ0 and n(δ̂ − δ0 ) → N (0, Ω−1
1,1 Ω1,2 Ω1,1 ) as n → ∞, in which
n
1X ˙
E(lt (δ0 )l˙tT (δ0 )) > 0
n→∞ n
t=1
(12)
Ω1,1 = lim
n
n
1 XX
Cov(l˙t (δ0 ), l˙s (δ0 ))
n→∞ n
t=1 s=1
Ω1,2 = lim
(13)
where
∂ log πt (yt )
|δ0 = f −1 (yt |xnt , δ0 )
l˙t (δ0 ) =
∂δ
Z
T
(yt − mt ḃ(xnt β0 +
1/2
τ0 zt ))
xnt
zt
√
2 τ0
f (yt |xnt , zt , δ0 )φ(zt )dzt (14)
To use this theorem in practice requires at least that the identifiability condition holds and that the
covariance be estimated from a single series. We address these aspects in detail in Section 2.1 and 2.2.
In addition, particularly for binary responses, marginal likelihood estimators produce a high probability of
τ̂ = 0. We address this in detail in Section 3, where we propose an improved asymptotic distribution based
on a mixture.
2.1. Asymptotic identifiability
We nowPdiscuss circumstances under
Pmwhich Condition 5 holds. Now for any δ ∈ Θ, Q(δ) ≤ Q(δ0 ), since for
m
any x, j=0 π 0 (j, x) log π(j, x) ≤ j=0 π 0 (j, x) log π 0 (j, x). Thus the model is identifiable if and only if for
any δ ∈ Θ, Q(δ) − Q(δ0 ) < 0 if δ 6= δ0 .
Lemma 1. Assume M ≥ 2 and Condition 3, then Condition 5 holds for marginal likelihood (7).
The proof is outlined in Appendix A.
For binary data, M = 1, then Q(δ) = Q(δ0 ) if π(1, x) = π 0 (1, x), ∀x ∈ X. Hence model (7) is not
identifiable if ∃δ 6= δ0 such that π(1) = π 0 (1) everywhere on X, that is, for each distinct value Xi ∈ X, such
(β, τ ) 6= (β0 , τ0 ) can be found to establish
√
√
Z
Z
T
T
eXi β+ τ z
eXi β0 + τ0 z
√
π(1, Xi ) =
φ(z)dz
=
φ(z)dz = π 0 (1, Xi ).
(15)
√
T
T
1 + eXi β+ τ z
1 + eXi β0 + τ0 z
If τ = τ0 , then (15) implies XiT β = XiT β0 . Under Condition 3, r linearly independent Xi , X = (X1 , . . . , Xr )
can be found on X to establish XT β = XT β0 . Then (β − β0 ) has a unique solution of 0r . Hence if τ = τ0 ,
(15) holds if and only if β = β0 , and Condition 5 holds.
If τ 6= τ0 , for each Xi , a unique solution of ai , ai = XiT β 6= XiT β0 can be found for (15). Assume the
regressor space X is a set of discrete vectors such that X = {Xi : 1 ≤ i ≤ L}, where Xi 6= Xj if i 6= j. Let
X = (X1 , . . . , XL ) be a r × L matrix. Then (15) holds for each Xi ∈ X if there exists such solution of β that
XT β = A, A = (a1 , . . . , aL ). Since τ 6= τ0 , β = β0 is not excluded from the possible solutions of β. If L = r,
a unique solution of β exists, hence there exists such δ 6= δ0 that establishes (15) for all Xi ∈ X, therefore
(7) is not identifiable. If L > r, note rank(X) = r, then XT β = A is overdetermined. Thus a solution of
β does not always exist and in these situations Condition 5 holds, however a general proof without further
conditions on the regressors is difficult. Instead, we provide a rigourous proof to show Condition 5 holds for
binary data when the regressor space X is connected.
/Marginal Estimation of Binomial Time Series Mixed Models
6
Lemma 2. Let M = 1. In addition to Condition 3, X is assumed to be a connected subspace of Rr , then
Condition 5 holds.
Proof: see the appendix A.
2.2. Estimation of the Covariance matrix
−1
To use Theorem 1 the asymptotic covariance matrix Ω−1
1,1 Ω1,2 Ω1,1 needs to be estimated using a single
observed time series. Now Ω1,1 can be estimated by replacing δh0 with the marginal likelihood
estimates
i
Pn Pn ˙
−1
˙
δ̂. However estimation of Ω1,2 is challenging, as Ω1,2 = n E
t=1
s=1 lt (δ0 )ls (δ0 ) has cross terms
E l˙t (δ0 )l˙s (δ0 ) , s 6= t, which cannot be estimated without knowledge of ψ0 . We use the modified subsampling
methods reviewed in Wu (2012) and Wu and Cui (2014) to estimate Ω1,2 .
Let Yi,kn = (yi , . . . , yi+kn −1 ) denote the subseries of length kn starting at the ith observation, where
i = 1, . . . , mn and mn = n − kn + 1 is the total number of subseries. Define
1
q̂n,t = √ l˙t (δ̂)
n
by replacing δ0 by δ̂ in (14). Under similar conditions to those given above, we show that as kn → ∞ and
† −1
mn → ∞, Γ̂−1
1,n Γ̂n Γ̂1,n is a consistent estimator of the asymptotic covariance matrix of δ̂, where
!
i+k
mn
n
n −1 i+k
n −1
X
X
X
1 X
T
T
†
q̂kn ,t q̂kn ,s
Γ̂1,n =
q̂n,t q̂n,t ; Γ̂n =
mn i=1
t=1
t=i
s=i
The performance of subsampling estimators relies on kn to a large extent. Following the guidance of Heagerty
and Lumley (2000) on optimal selection of kn , we use kn = C[n1/3 ], C = 1, 2, 4, 8 in the simulations. The
one dimensional integrals in q̂n,t can be easily obtained using the R function integrate.
3. Degeneration of Marginal Likelihood Estimates
Even when the identifiability conditions are satisfied, in finite samples the marginal likelihood can be maximised at τ̂ = 0, in which case β̂ degenerates to the ordinary GLM estimate β̃. Simulation evidence of
Section 4 suggests that the chance of this occurring, even for moderate to large sample sizes, is large (up
to 50% for binary data but decreasing rapidly as the number of trials m increases). In this section we will
derive two approximations for this probability. In both approximations we conclude that P (τ̂ = 0) will be
high whenever the range of xTnt β is such that ḃ(xTnt β) ≈ a0 + a1 (xTnt β), where a0 , a1 are constants. When
this linear approximation is accurate the covariance matrix for the marginal likelihood estimates is obtained
from the inverse of a near singular matrix and results in var(τ̂ ) being very large so that P (τ̂ = 0) is close
to 50%. When there is a nontrivial probability of τ̂ = 0, the distribution of β̂ for finite samples is better
approximated by a mixture of two multivariate distributions weighted by P (τ̂ = 0) and P (τ̂ > 0).
3.1. Estimating the probability of τ̂ = 0
One approximation for the probability √
of τ̂ = 0 can be√obtained using the asymptotic normal distribution
provided in Theorem 1. Define κ2 = P ( n(τ̂ − τ0 ) ≤ − nτ0 ), then in the limit,
√
−1
κ̄2 = Φ(− nτ0 /στ (δ0 )); στ2 (δ0 ) = (Ω−1
(16)
1,1 Ω1,2 Ω1,1 )τ τ
where Φ(·) is the standard normal distribution.
An alternative approximation to P (τ̂ = 0) can be based on the score function evaluated at τ̂ = 0. Consider
the scaled score function S1,n (β̃) = 2n−1 ∂l1 (β, τ )/∂τ |β=β̃,τ =0 , which, using integration by parts, is
n
i
1 Xh
S1,n (β̃) =
(yt − mt ḃ(xTnt β̃))2 − mt b̈(xTnt β̃) .
n t=1
(17)
/Marginal Estimation of Binomial Time Series Mixed Models
7
Now, τ̂ = 0 implies S1,n (β̃) ≤ 0 but not the converse, hence P (τ̂ = 0) is bounded above by P (S1,n (β̃) ≤ 0).
In order to derive a large √
sample approximation to this probability we show, in Section 3.2, that the
large sample distribution of n(S1,n (β̃) − cS )/σS is standard normal, where cS = lim E(S1,n (β 0 )) and
n→∞
√
σS2 = lim Var( nS1,n (β̃)). Define κ1 = P (S1,n (β̃) ≤ 0), it can then be approximated with
n→∞
√
κ̄1 = Φ(− ncS /σS ).
(18)
The quantities cS and σS can be expressed analytically for some regression specifications. In simulations,
the limits are computed using numerical integration. For the binary case in particular, the ratio cS /σS can
be quite small resulting in a large value for κ̄1 . We compare how well P (τ̂ = 0) is estimated by κ̄1 and κ̄2
via simulations in Section 4, and conclude that κ̄1 is slightly more accurate in the situation covered there.
3.2. Asymptotic Theory for GLM Estimates and Marginal Score
To develop the asymptotic distribution of S1,n (β̃), the asymptotic normality of
√
n(β̃ − β 0 ) is required.
Theorem 2 (Asymptotic normality of GLM estimators). Under Conditions 1 to 4, the estimates β̃ max√
p
−1
imising the likelihood (4) satisfies β̃ → β 0 , and n(β̃ − β 0 ) → N(0, Ω−1
1 Ω2 Ω1 ) as n → ∞, in which
n
1X
mt b̈(xTnt β 0 )xnt xTnt
n→∞ n
t=1
Ω1 = lim
Z
n
n
1 XX
T
T
0
T
T
0
mt ms
Ω2 = lim
(ḃ(xnt β0 + αt ) − ḃ(xnt β ))(ḃ(xns β0 + αs ) − ḃ(xns β ))g(αt , αs ; τ0 , ψ0 )dα xnt xTns
n→∞ n
t=1 s=1
Z
n
X
1
+ lim
mt
b̈(xTnt β0 + αt )g(αt ; τ0 )dα xnt xTnt
n→∞ n
t=1
The proof of this theorem is given in Appendix B. It relies on concavity of the GLM log likelihood with
respect to β. Standard results of a functional limit theorem are used to establish the above result, in a similar
way as that used in Davis, Dunsmuir and Wang (2000), Davis and Wu (2009) and Wu and Cui (2014).
In order to use Theorem 2 for practical purposes, first β 0 needs to be determined and then Ω1 , Ω2 .
Estimation of β 0 would require knowledge of τ , and the estimation of Ω2 would require both τ and ψ, neither
of which can be estimated using the GLM procedure. Hence the theorem is of theoretical value only.
Based on Theorem 2 we can now derive the large sample distribution of the score function of the marginal
p
likelihood evaluated at δ̃ = (β̃, 0). Because all derivatives of b(·) are uniformly bounded and β̃ → β 0 , hence
√
√
√
n S1,n (β̃) − E(S1,n (β 0 )) = n (S1,n (β 0 ) − E(S1,n (β 0 ))) − JST n(β̃ − β 0 ) + op (1).
Since n−1/2
Pn
t=1 (yt
− mt ḃ(xTnt β̃))xnt = 0 by definition of β̃, using Taylor expansion
√
n
1 X
√
et,β 0 xnt + op (1),
n(β̃ − β 0 ) = Ω−1
1
n t=1
Then it follows that
et,β 0 = yt − mt ḃ(xTnt β 0 ).
√
p
n S1,n (β̃)) − E(S1,n (β 0 )) − (U1,n − JST U2,n ) → 0
where
(19)
n
i
1 Xh 2 0 T
2mt (π (xnt β0 ) − ḃ(xTnt β 0 ))b̈(xTnt β 0 ) + mt b(3) (xTnt β 0 ) xnt
n→∞ n
t=1
JS = lim
U1,n :=
√
n
1 X 2
n (S1,n (β ) − E(S1,n (β ))) = √
e 0 − Ee2t,β 0 ;
n t=1 t,β
0
0
(20)
n
U2,n
1 X
:= √
et,β 0 cnt
n t=1
(21)
/Marginal Estimation of Binomial Time Series Mixed Models
8
ḃ(xTnt β0 + αt )g(αt , τ0 )dα andcnt = Ω−1
1 xnt , which is a non-random vector.
√
0
Then the CLT for n S1,n (β̃) − E(S1,n (β )) follows the CLT for the joint vector of (U1,n , U2,n ). Note
note π 0 (xTnt β0 ) =
R
that both sequences of {U1,t } and {U2,t } are strongly mixing by Proposition 1 in Blais, MacGibbon and
Roy (2000). Then the CLT for mixing process proposed in Davidson (1992) can be applied to show that
(U1,n , U2,n ) is asymptotically normally distributed with mean zero and covariance matrix
VS
Ω−1
1 KS
−1
KST Ω−1
Ω−1
1
1 Ω2 Ω1
where Ω1 , Ω2 are given in Theorem 2, and
! (n−1)
n
(n−1)
X
X 1X
2
2
Cov(et,β 0 , et+h,β 0 ) +
VS := lim
n→∞
n t=1
h=1
h=0
KS := lim
(n−1)
X
n→∞
h=0
n
1X
E(et,β 0 e2t+h,β 0 )xnt
n t=1
(n−1)
!
+
X
h=1
!
n
1 X
Cov(e2t,β 0 , e2t−h,β 0 )
n
t=h+1
!
n
1 X
E(et,β 0 e2t−h,β 0 )xnt
n
t=h+1
√
d
Theorem 3. Under the assumptions of Theorem 2, as n → ∞, n S1,n (β̃) − E(S1,n (β 0 )) /σS → N (0, 1).
3.3. An approximate mixture distribution for β̂
Theorem 4 (Mixture distribution under finite samples). Assume τ0 > 0, under Conditions 1 to 5, in finite
√
samples, distribution of n(β̂ − β0 ) can be approximated with the mixture
κF1 (c, δ0 ) + (1 − κ)F2 (c, δ0 ),
κ = P (τ̂ = 0)
√
in which F1 (c, δ0 ) is r-dimensional
distribution obtained through n(β̂ − β 0 ), which is a skew
√ multivariate
0
T
normal distribution U2,n |U1,n + nE(S1,n (β )) − 2JS U2,n ≤ 0, based on the joint normality of (U1,n , U2,n )
√
given in Theorem 3; F2 (c, δ0 ) is a r-dimensional skew normal distribution n(β̂ − β0 )|τ̂ > 0, based on the
−1
joint normality of N (0, Ω−1
1,1 Ω1,2 Ω1,1 ) in Theorem 1. Moreover, κ → 0 as n → ∞.
Remarks
1. The skew normal distribution is defined in Gupta, González-Farıas and Domınguez-Molina (2004).
2. If τ0 = 0, β 0 = β0 and the
√ value κ = 0.5 in the above mixture is similar to that in Moran (1971,
Theorem I); when τ0 = a/ n, a ≥ 0, above results are parallel to those in Moran (1971, Theorem IV)
and based on the same reasoning. However Moran’s results are for independence observations whereas
our results require the serial dependence to be accounted for in the asymptotic results.
3. While the mixture provides a better theoretical description of the asymptotic distribution for marginal
likelihood estimates when m is small, in practice, the mixture distribution cannot be estimated without
knowing the true values of β0 , τ0 and ψ0 . In simulations, the covariance matrix for the joint distribution
−1
of β̂ and τ̂ is approximated with Σ(δ0 ) = n−1 Ω−1
1,1 Ω1,2 Ω1,1 , and based on F2 (c, δ0 ), we calculate
E(β̂ − β0 |τ̂ > 0) = Σβτ (δ0 )Σ−1
τ τ (δ0 )E(τ̂ − τ0 |τ̂ > 0)
(22)
−2
Var(β̂|τ̂ > 0) = Σββ (δ0 ) − Σβτ (δ0 )Σ−1
τ τ (δ0 )Στ β (δ0 ) + Σβτ (δ0 )Στ τ (δ0 )Στ β (δ0 )Var(τ̂ − τ0 |τ̂ > 0) (23)
4. Simulation Results
In this section we summarize results of several simulation studies to illustrate the key theoretical results
derived above as well as to indicate circumstances under which P (τ̂ = 0) is large in which case the mixture
distribution of Theorem 4 would provide a more accurate description.
/Marginal Estimation of Binomial Time Series Mixed Models
9
For all examples we consider the simple linear trend with latent process W0,t = β1 + β2 (t/n) + αt , in
i.i.d
which αt is assumed to be: αt = φαt−1 + t , t ∼ N (0, σ2 ) where σ2 is chosen to maintain Var(αt ) = 1. In
all cases the true values are β0 = (1, 2) and τ0 = 1 and φ varies in the interval (−1, 1). While simple, this
example provides substantial insights into the behaviour of the marginal likelihood estimates as well as into
problems that can arise. The simplicity of this example also allows us to obtain analytical calculations of
key quantities and to provide some heuristic explanations of the non-standard distribution results which can
arise, particularly for binary time series.
In all simulations reported later, the number of replications was 10, 000. The marginal likelihood estimates
were obtained using the R package lme4. The frequency with which τ̂ = 0 is not package dependent other
than the occasionaly case – this was checked using our own implementation based on adaptive Gaussian
quadrature and by comparing the results with those from SAS PROC MIXED. The first simulation (Section
4.2) focuses on binary responses and illustrates that the distribution of marginal likelihood estimates δ̂ for
this kind of data converge towards a mixture as proposed in Theorem 4, in which the P (τ̂ = 0) can be
approximated using the result of Theorem 3 to good accuracy. The second experiment (Section 4.3) studies
the finite sample performance of δ̂ for binomial cases and shows that P (τ̂ = 0) vanishes as mt increases or as
n → ∞, thus the distribution of δ̂ is multivariate normal as developed in Theorem 1. Finally (Section 4.4)
the method for estimation the covariance matrix for δ̂ proposed in Section 2.2, is evaluated.
In order to implement the simulations in Section 4.1 we first derive some theoretical expressions for key
quantities used to define the large sample distributions of Theorems 1, 2 and 3 as well as for the estimates
κ̂1 and κ̂2 for P (τ̂ = 0).
4.1. Analytical Expressions for Asymptotic Quantities
Key quantities required for implementation and explanation of the simulation results to follow are:
(1). The limit point β 0 for the GLM estimate β̃.
(2). cS and σS appearing in Theorem 3 and used to obtain the approximation κ̄1 for P (τ̂ = 0).
(3). The asymptotic variance of τ̂ in Theorem 1 used to obtain the approximation κ̄2 for P (τ̂ = 0).
(4). Various quantities defining the mixture distribution in Theorem 4.
Throughout, the derivations are given for the case of deterministic regressors specified as xnt = h(t/n)
for a suitably defined vector function h as in Condition 2a and in which the first component is unity in
order to include the intercept term. Also, in order to reduce notational clutter we will assume that mt ≡ m
(the number of binomial trials at all time points is the same). The analytical expressions involve various
integrals which are computed using numerical integration either with the R-package integrate or using grid
evaluation for u with mesh 0.0001 over the interval [0, 1]. Calculation of Ω1,2 in Theorem 1 and KS , VS and
Ω2 in Theorem 3 to obtain the variance σS2 for the theoretical upper bound of P (τ̂ = 0) require evaluation
of two-dimensional integrals of the form
Cov(l˙t (δ0 ), l˙t+h (δ0 )) =
mt m
t+h
X
X
π 0 (yt , yt+h )l˙t (δ0 )l˙t+h (δ0 ).
yt =0 yt+h =0
For these, the integral expression for π 0 (yt , yt+h ) is approximated using adaptive Gaussian quadratic (AGQ)
with 9 nodes for each of the two dimensions.
4.1.1. Limit Point of GLM Estimation
By numerically solving the non-linear system (5) with Newton Raphson iteration the limiting value of the
GLM estimates is β 0 = (0.8206, 1.7574).
/Marginal Estimation of Binomial Time Series Mixed Models
10
4.1.2. Quantities needed for κ̄1
The analytical expression for the limiting expectation of the scaled score with respect to τ evaluated at the
limiting point β 0 is.
cS := lim E(S1,n (β 0 ))
n→∞
Z
n
2
1X
m(m − 1)
ḃ(xTnt β0 + αt ) − ḃ(xTnt β 0 ) g(αt , τ0 )dαt + m(ḃ(xTnt β 0 ) − π 0 (xTnt β0 ))(2ḃ(xTnt β 0 ) − 1)
= lim
n→∞ n
t=1
Z 1Z
2
ḃ(h(u)T β0 + α) − ḃ(h(u)T β 0 ) g(α, τ0 )dαdu
=m(m − 1)
0
1
Z
(ḃ(h(u)T β 0 ) − π 0 (h(u)T β0 ))(2ḃ(h(u)T β 0 ) − 1)du
+m
0
=m(m − 1)c1 + mc2 .
(24)
Note that c1 in (24) is strictly positive but make no contribution for binary responses (when m = 1) in which
case c2 is the only term contributing to cS . We have observed that c2 is non-negative, in the simulations but
we do not have a general proof of that. In that case cS is also non-negative for all m. We have observed in
the simulations that c1 is substantially larger than c2 and as a result κ̄1 is small for non-binary responses
(m > 1) but can be large for binary responses because c2 ≈ 0.
Recall that σS2 = lim Var(S1,n (β̃)). For the case where the latent process is i.i.d. we have
n→∞
σS2 =
Z
1
mπ 0 (h(u)T β0 )(1 − π 0 (h(u)T β0 )) 1 + 2(m − 3)π 0 (h(u)T β0 )(1 − π 0 (h(u)T β0 ))
0
+4m3 π 0 (h(u)T β0 )(1 − π 0 (h(u)T β0 ))(π 0 (h(u)T β0 ) − ḃ(h(u)T β 0 ))2
+4m2 π 0 (h(u)T β0 )(1 − π 0 (h(u)T β0 ))(1 − 2π 0 (h(u)T β0 ))(π 0 (h(u)T β0 ) − ḃ(h(u)T β 0 ))du
−1
T −1
−2JST Ω−1
1 KS + JS Ω1 Ω2 Ω1 JS
(25)
in which
Z
KS =
1
0
mπ (h(u)T β0 )(1 − π 0 (h(u)T β0 ))(1 − 2π 0 (h(u)T β0 ))
0
i
+ 2m2 π 0 (h(u)T β0 )(1 − π 0 (h(u)T β0 ))(π 0 (h(u)T β0 ) − ḃ(h(u)T β 0 )) h(u)du
Z 1h
i
JS =
2m2 (π 0 (h(u)T β0 ) − ḃ(h(u)T β 0 ))b̈(h(u)T β 0 ) + mb(3) (h(u)T β 0 ) h(u)du
0
similar expression of σS2 can be obtained for the dependent cases where ψ 6= 0. However, serial dependence
does not appear to make much difference in the chance of τ̂ = 0 at least in the simulations of Section 4.3.
For the simulation example of the binary response case (m = 1) the expressions (24), (25) can be evaluated
using numerical integration to give
c1 = 0.0168, c2 = 1.65 × 10−5
cS
σS
m=1
1.65 × 10−5
7.17 × 10−3
m=2
0.034
0.303
m=3
0.101
0.525
cS /σS
0.0023
0.1110
0.1921
√
Since cS is observed to be strictly positive for marginal likelihood, ncS → ∞ as n → ∞. By Theorem 3,
the probability of S1,n (β̃) ≤ 0 vanishes as n → ∞. Because P (τ̂ = 0) is bounded above by P (S1,n (β̃) ≤ 0), as
n → ∞, the marginal likelihood estimate will be such that τ̂ = 0 with vanishing probability. However, notice
for binary data cS = c2 and cS /σS = 0.0023. Hence, even for the largest sample size (n = 5000) reported
below, κ̄1 is 44%. It would require a sample size of 106 to reduce this to 1%. Clearly this has substantial
/Marginal Estimation of Binomial Time Series Mixed Models
11
implications for the use of marginal likelihood for binary data. For binomial responses the c1 term dominates,
and hence even with small values of m > 1 the chance of τ̂ = 0 reduces rapidly.
We conclude this subsection with a heuristic explanation of why c2 ≈ 0. If h(u)T β 0 appearing in the
definition of c2 is such that ḃ(h(u)T β 0 ) is well approximated linearly in h(u)T β 0 then c2 could be approximated
by
Z
1
c∗2 =
[ḃ(h(u)T β 0 ) − π 0 (h(u)T β0 )](ao + a1 h(u)T β 0 )du.
(26)
0
But, because the first element of h(u) is equal to 1, ao + a1 h(u)T β 0 can be rewritten as h(u)T β ∗ for some
vector β ∗ . Then c∗2 can be written as
c∗2
1
Z
[ḃ(h(u)T β 0 ) − π 0 (h(u)T β0 )]h(u)T duβ ∗ = 0
=
(27)
0
by definition of β 0 in (5). Note that ḃ(x) is the probability of success from a logit response and hence if x
ranges over reasonably large values then ḃ(x) may be near linear. In the example used for simulations h(u)β 0
ranges over the interval (0.8206, 2.578) and ḃ(x) for x in this interval is well approximated by a straight line
in x.
4.1.3. Asymptotic Covariance matrix for marginal estimates
Although positive definite, Ω1,1 of the asymptotic covariance matrix in Theorem 1 can be near singular and
√
this results in an overall covariance matrix for n(δ̂ − δ0 ) which has very large elements; in particular, the
variance of δ̂ in (16) is very large and, as a result κ̄2 given in (16) will also be close to 50%. The reason
for this will be analyzed in this section by calculating various components of the asympototic covariance for
marginal estimates. To keep the discussion manageable the deterministic regressors, assumed to be generated
by functions h(·) and to satisfy Condition 2a, will be used for the derivations. In this case, the summations
on the left of (12) has limit given by the integral
Z
Ω1,1 =
m
1X
˙
f (y|h(u), δ0 )l(y|h(u);
δ0 )l˙T (y|h(u); δ0 )du
(28)
0 y=0
in which
!
R
√
(y − mḃ(h(u)T β + τ z))f (y|h(u), z, δ)φ(z)dz
· h(u)
i
˙l(y|h(u); δ) = f −1 (y|h(u), δ) R h
√
√
(y − mḃ(h(u)T β + τ z))2 − mb̈(h(u)T β + τ z) f (y|h(u), z, δ)φ(z)dz/2
where the first dimension corresponds to ∂l(y|h(u); δ)/∂β and the second dimension corresponds to ∂l(y|h(u); δ)/∂τ .
For binary responses, using the facts that m ≡ 1 and y 2 = y, the component in ∂l(y|h(u); δ)/∂τ has
√
√
√
√
(y − mḃ(h(u)T β + τ z))2 − mb̈(h(u)T β + τ z) = (y − ḃ(h(u)T β + τ z))(1 − 2ḃ(h(u)T β + τ z)). (29)
√
Let W = h(u)T β + τ z, note for binary cases, (y −Rḃ(W ))f (y|h(u), z, δ) = b̈(W ) if y = 1 and −b̈(W ) if y = 0.
˙
δ) is
Define conditional distribution ρ(z) = b̈(W )φ(z)/ b̈(W )φ(z)dz, then if h(u)T β appearing in l(y|h(u);
such that the linearly approximation
Z
∂l(y|h(u); δ)/∂τ
= (1 − 2ḃ(W ))ρ(z)dz ≈ a∗0 + a∗1 (h(u)T β) = h(u)T β ∗
(30)
R
b̈(W )φ(z)dz
˙
is well established, then for each fixed u ∈ [0, 1] and δ, l(y|h(u);
δ) is a nearly linear dependent vector.
Consequently Ω1,1 becomes near singular with a large inverse. In this example h(u)T β0 takes all values in
the interval (1, 3) over which the left side of (30) is approximately a straight line and hence, in view of the
above discussion, and Ω1,1 has eigenvalues (0.129, 6.784 × 10−3 , 1.903 × 10−6 ), with στ2 (δ0 ) = 698.
For binomial data, the properties of m > 1 and y 2 6= y and the left side of (29) is no longer a near linear
function of h(u). As a result Ω1,1 is not nearly singular and στ2 (δ0 ) is of moderate size.
/Marginal Estimation of Binomial Time Series Mixed Models
12
4.1.4. Quantities needed for mixture distribution
To assess the accuracy of the asymptotic mixture distribution, the theoretical mean vector and covriance for
the distributions F1 (·, δ0 ) and F2 (·, δ0 ) are required. For F1 (·, δ0 ) these are approximately as for the normal
distribution in Theorem 2 for the GLM estimates. The mean is β 0 = (0.82, 1.76) given in Section 4.1.1. The
covariance matrix
23.636
−39.8212
−1
−1
Ω1 Ω2 Ω 1 =
−39.8212
98.13
is calculated using δ0 = 1 and β 0 , in
Z
Ω1 =
1
b̈(hT (u)β 0 )h(u)hT (u)du
0
and
Z
Ω2 =
1
h
i
π 0 (h(u)T β0 ) − 2π 0 (h(u)T β0 )ḃ(h(u)T β 0 ) + ḃ2 (h(u)T β 0 ) h(u)h(u)T du.
0
For F2 (·, δ0 ), the mean is evaluated with conditinoal mean (22) and covariance matrix (23). In this example
Σ(δ0 ) = n−1 Ω−1
1,1 , where Ω1,1 is provided above. For binary data, E(τ̂ − τ0 |τ̂ > 0) and Var(τ̂ − τ0 |τ̂ > 0) are
obtained empirically using Ω̂1,1 as follows:
Ω̂1,1
Z
n
1 X ˙ ∗ ˙T ∗
0
0
∗ −1
∗
lt (δ )lt (δ ) + f (yt |xnt ; δ )
(yt − mt ḃ(Wt ))
f (yt |xnt , zt , δ ∗ )φ(z)dz
=
0 4τ1∗
n t=1
)
Z h
i x
zt
nt
T
∗
∗ −1
∗ 2
∗
(xnt , √ ∗ )f (yt |xnt , zt , δ )φ(z)dz
− f (yt |xnt ; δ )
(yt − mt ḃ(Wt )) − mt b̈(Wt )
zt
√
2 τ
2 τ∗
where kδ ∗ − δ̂k ≤ kδ̂ − δ0 k but under finite samples δ ∗ 6= δ0 . Note that although for binary data, l˙t (δ)
can be almost linearly dependent as analyzed above, the last part in Ω̂1,1 is nontrivial and nonsingular for
δ ∗ 6= δ0 , which allows Ω̂1,1 to be nonsingular. As a result a large difference between theoretical and empirical
covariance is observed for binary data, which will be shown in Example 2 of simulations.
4.2. Example 1: binary data, independent latent process
This example considers the simplest case of independent observations obtained when τ0 = 1 and φ0 = 0.
For each replication, an independent binary sequence of Yt |αt , xnt ∼ B(1, pt ), pt = 1/(1 + exp(−W0,t )),
is generated. Table 1 reports the empirical values κ̂1 (the empirical proportion of S1,n (β̃) ≤ 10−6 ) and κ̂2
(the empirical proportion of τ̂ ≤ 10−6 ). Also shown are the empirical mean and standard deviation (in
parentheses) of β̂ conditional on τ̂ = 0 and τ̂ > 0 obtained from the simulations along with the theoretical
values of these obtained from (18), (16), and (22), (23) associated with Theorem 4.
Table 1 clearly demonstrates that for this data generating mechanism there is very high proportion of
replicates for which τ̂ = 0 and that this proportion does not decrease rapidly with sample size increasing.
This is as predicted by theory and the theoretical values of κ̄1 and κ̄2 provide good approximations to κ̂1 and
κ̂2 . As explained above, this high proportion of zero estimates for τ , even for large sample sizes, is as expected
for the regression structure used in this simulation. Note that κ̄1 ≤ κ̄2 and κ̄1 is closer to the probability
of τ̂ = 0. The estimations reverse the theoretical property that κ1 ≥ κ2 . For binary data, P (τ̂ = 0) and
both theoretical approximations, κ̄1 and κ̄2 decrease slowly and a very large sample is required to attain
P (τ̂ = 0) ≈ 0.
It is also clear from Table 1 that the empirical mean and standard deviation of β̂|τ̂ = 0 and β̂|τ̂ > 0
show good agreement with the theoretical results predicted by Theorem 4. Overall, the theory we derive for
P (τ̂ = 0) and the use of a mixture distribution for β̂ is quite accurate for all sample sizes, and with relatively
large sample the mixture is a better representation. However, estimation of the corresponding distributions
in the mixture requires β0 , τ0 and ψ0 and therefore cannot be implemented in practice.
/Marginal Estimation of Binomial Time Series Mixed Models
n = 200
n = 500
n = 103
n = 5 · 103
β̂|τ̂ = 0
0.82(0.344)
1.76(0.700)
0.82(0.217)
1.76(0.443)
0.82(0.154)
1.76(0.313)
0.82(0.069)
1.76(0.140)
Theoretical
β̂|τ̂ > 0
κ̄1
1.10(0.421)
48.70%
2.15(0.819)
1.05(0.269)
47.95%
2.07(0.521)
1.02(0.193)
47.10%
2.03(0.371)
0.97(0.093)
43.54%
1.95(0.174)
κ̄2
49.20%
48.72%
48.20%
45.96%
β̂|τ̂ = 0
0.83(0.354)
1.79(0.731)
0.82(0.219)
1.78(0.451)
0.82(0.156)
1.77(0.315)
0.82(0.069)
1.76(0.139)
13
Empirical from simulations
β̂|τ̂ > 0
κ̂1
κ̂2
1.07(0.431)
45.56%
45.54%
2.26(1.020)
1.04(0.261)
46.27%
46.26%
2.10(0.621)
1.01(0.182)
45.52%
45.51%
2.06(0.415)
0.96(0.089)
42.93%
42.93%
1.96(0.184)
Standard deviation in “()”
Table 1
Mixture distribution of marginal likelihood estimates for binary independent time series
4.3. Example 2: binomial data, correlated latent process
This simulation investigates bias and standard deviation of the marginal estimates for m = 1, 2, 3, n =
200, 500 and a range of serial dependence given by φ = −0.8, −0.2, 0, 0.2, 0.8. The observed standard deviation
of the estimates over the replications is compared to that given by the asymptotic covariance matrix in
Theorem 1. We also give the empirical proportion of the event that τ̂ = 0 using the proportion τ̂ ≤ 10−6 .
The theoretical upper bound κ̄1 is approximated with κ̄1 defined in (18).
Table 2 summarizes the results. For binomial series (m ≥ 2), the empirical values are in good agreement
with the asymptotic mean and standard deviation with bias of the estimates generally improving with m or
n increasing. Also observed for binomial series is that, both theoretically and empirically, the probability of
τ̂ = 0 decreases quickly by increasing either the number of trials m or the sample size n. However, for binary
responses, large asymptotic standard deviations are obtained for the reason explained in Section 4.1.3. In
binary data, the probability of τ̂ = 0 is close to 50% and although theoretically this (κ̄1 ) will converges to
zero as n → ∞, it is doing so slowly as n increases. This can be explained by Theorem 3. Under the settings
−2
of this example, using AGQ, σS is calculated to vary from 0.72 × 10−2 to 1.55
of c1
√ × 10 and the values
√
and c2 are the same as those in Example 1 and hence for, binary responses, ncS /σS is√at most 2 n × 10−3
across the range of autocorrelation considered here. Thus the large values of κ̄1 = Φ(− ncS /σS ) across the
range of autocorrelations is to be expected for this regression
structure. However,
for binomial series with
√
√
m = 2 and n = 200, φ = 0.8 for instance, σS ≈ 0.351, ncS is dominated by nm(m − 1)c1 = 0.475 which
explains why P (τ̂ = 0) decreases rapidly with m > 1 and increasing n.
Interestingly, for binary series, as n increases from 200 to 500 the bias of τ̂ worsens which seems somewhat
counterintuitive. A plausible explanation for this is that the distribution of τ̂ is a mixture using weights
P (τ̂ = 0), which is approximately 45% across sample size and serial dependence, and P (τ̂ > 0). When
n = 200 the conditional distribution of τ̂ |τ̂ > 0 has larger variance than when n = 500 resulting in an
inflated overall mean when n = 200 relative to n = 500.
In summary, the theoretical upper bound (κ̄1 ) is above or close to the empirical proportion of P (τ̂ = 0),
which is a pattern that is consistent with Theorem 3. For binomial series (m ≥ 2) the the marginal estimates
have good bias properties and standard deviations explained by the large sample distribution of Theorem
1 and these conclusions are not severely impacted by the level or direction of serial dependence. For binary
series, the high proportion of τ̂ = 0 is persistent regardless of n = 200, 500 or the level of serial dependence
and this is explained by theory presented above.
4.4. Example 3: Estimate of Covariance Matrix
The subsampling method of estimating the covariance for marginal likelihood estimates is of limited practical
value for binary data because, firstly, when τ̂ = 0, which occurs nearly 50% of the time, the method does
not provide estimates of the covariance matrix in Theorem 1 and, secondly, because of the high proportion
of τ̂ = 0, δ̂ has the mixture distribution given in Theorem 4, the covariance of which requires β 0 and δ0 , both
of which are unknown and cannot be estimated from a single sequence.
In the binomial case (m ≥ 2) the subsampling method is likely to be useful for a range of serial dependence.
Table 3 presents simulation results under various levels of serial correlation. The table summarizes the
/Marginal Estimation of Binomial Time Series Mixed Models
14
n = 200
φ
0.8
0.2
0
-0.2
-0.8
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
m=1
Mean
SD
ASD
0.984 0.617 10.22
2.102 1.207 15.19
0.983 1.177 65.20
45.41%(49.44%)
0.953 0.433
7.90
2.071 0.949 11.84
0.898 1.026 50.55
45.08%(49.38%)
0.955 0.414
7.71
2.048 0.926 11.56
0.869 1.015 49.35
46.72%(49.37%)
0.958 0.408
7.58
2.049 0.913 11.39
0.880 1.017 48.57
45.78%(49.37%)
0.953 0.405
7.53
2.045 0.932 11.34
0.863 1.006 48.30
46.01%(49.40%)
m=2
Mean
SD
ASD
0.994 0.525 0.516
2.046 0.975 0.942
0.960 0.790 0.821
11.06%(8.85%)
0.996 0.346 0.342
2.047 0.667 0.648
1.052 0.793 0.790
8.08%(7.15%)
1.007 0.331 0.327
2.032 0.634 0.623
1.063 0.786 0.789
7.25%(7.08%)
1.003 0.317 0.316
2.036 0.620 0.606
1.059 0.794 0.789
7.27%(7.07%)
0.995 0.306 0.303
2.038 0.623 0.603
1.047 0.794 0.812
7.92%(7.88%)
m=3
Mean
SD
ASD
0.991 0.484 0.483
2.023 0.890 0.878
0.909 0.541 0.557
2.52%(3.00%)
0.992 0.293 0.292
2.027 0.563 0.553
0.993 0.518 0.510
0.86%(1.30%)
1.001 0.278 0.273
2.013 0.527 0.523
1.003 0.512 0.509
0.78%(1.24%)
0.999 0.261 0.261
2.014 0.511 0.504
1.004 0.516 0.510
0.8%(1.23%)
0.992 0.247 0.245
2.023 0.512 0.500
1.002 0.547 0.542
1.43%(1.91%)
m=1
Mean
SD
ASD
0.948 0.373
6.58
1.992 0.737
9.77
0.785 0.866 41.96
44.77%(49.36%)
0.936 0.273
5.00
1.953 0.589
7.50
0.694 0.772 31.99
46.18%(49.32%)
0.938 0.263
4.87
1.959 0.572
7.32
0.708 0.767 31.21
45.18%(49.32%)
0.934 0.257
4.79
1.957 0.555
7.21
0.689 0.763 30.71
46.26%(49.32%)
0.929 0.253
4.76
1.962 0.570
7.18
0.681 0.768 30.57
47.10%(49.34%)
m=2
Mean
SD
ASD
0.998 0.335 0.330
2.013 0.615 0.604
0.974 0.515 0.519
1.45%(1.68%)
1.000 0.217 0.216
2.012 0.412 0.410
1.011 0.492 0.499
0.93%(1.06%)
0.997 0.205 0.206
2.011 0.395 0.394
1.011 0.502 0.499
1.04%(1.04%)
1.000 0.202 0.200
2.013 0.389 0.383
1.014 0.493 0.499
0.7%(1.04%)
0.996 0.190 0.191
2.012 0.386 0.381
1.008 0.509 0.513
1.01%(1.31%)
m=3
Mean
SD
ASD
0.998 0.311 0.309
2.001 0.573 0.563
0.956
0.35
0.352
0.07%(0.14%)
1.002 0.182 0.184
2.002 0.349 0.350
0.998 0.325 0.322
0%(0.02%)
0.999 0.173 0.172
2.005 0.333 0.331
0.996 0.321 0.322
0%(0.017%)
0.996 0.164 0.164
2.010 0.315 0.318
0.998 0.329 0.322
0.01%(0.017%)
0.996 0.156 0.155
2.011 0.319 0.317
0.997 0.344 0.343
0.02%(0.05%)
n = 500
φ
0.8
0.2
0
-0.2
-0.8
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
β̂1
β̂2
τ̂
κ̂(κ̄1 )
Table 2
Marginal likelihood estimates for Binomial observations under various values of φ, where the true values are β10 = 1; β20 = 2;
τ0 = 1.
/Marginal Estimation of Binomial Time Series Mixed Models
15
m = 2, n = 200
φ = 0.8
φ = 0.2
φ = −0.2
φ = −0.8
β̂1
β̂2
τ̂
β̂1
β̂2
τ̂
β̂1
β̂2
τ̂
β̂1
β̂2
τ̂
ASD
0.515
0.942
0.821
0.342
0.648
0.790
0.316
0.606
0.789
0.303
0.603
0.812
SD
0.525
0.975
0.790
0.346
0.667
0.793
0.322
0.616
0.785
0.305
0.619
0.796
ASD
0.330
0.604
0.519
0.216
0.410
0.499
0.199
0.383
0.499
0.191
0.381
0.513
SD
0.330
0.607
0.507
0.217
0.418
0.493
0.199
0.387
0.503
0.191
0.387
0.513
kn = 5
0.392
0.734
0.824
0.333
0.637
0.829
0.313
0.604
0.831
0.300
0.592
0.842
kn = 11
0.406
0.755
0.811
0.316
0.605
0.814
0.296
0.570
0.816
0.283
0.562
0.834
kn = 23
0.387
0.719
0.784
0.284
0.545
0.786
0.265
0.511
0.786
0.255
0.508
0.808
kn = 46
0.325
0.606
0.733
0.235
0.447
0.729
0.221
0.420
0.729
0.214
0.419
0.750
m = 2, n = 500
kn = 7
kn = 14
kn = 28
kn = 56
0.261
0.275
0.271
0.242
φ = 0.8
0.487
0.510
0.501
0.451
0.506
0.501
0.490
0.469
0.210
0.204
0.192
0.169
φ = 0.2
0.400
0.388
0.364
0.321
0.503
0.497
0.486
0.465
0.197
0.190
0.179
0.159
φ = −0.2
0.378
0.365
0.343
0.303
0.505
0.500
0.490
0.469
0.188
0.182
0.171
0.152
φ = −0.8
0.372
0.361
0.340
0.302
0.514
0.511
0.502
0.481
Table 3
Subsampling estimates for standard deviation of GLMM estimation, kn = C[n1/3 ], C = 1, 2, 4, 8.
β̂1
β̂2
τ̂
β̂1
β̂2
τ̂
β̂1
β̂2
τ̂
β̂1
β̂2
τ̂
estimates of standard deviation for δ̂ using the subsampling method described in Section 2.2. The column
“ASD” contains the asymptotic standard deviation calculated with the covariance matrix in Theorem 1
and column “SD” contains empirical standard deviation. The table shows the subsampling estimates of
standard deviations are of the same magnitude of theoretical standard deviations even for moderate sample
size n = 200 but are biased downwards and increasingly so as C increases. Values of C = 1, 2 provide the
least biased estimates for the standard errors for both sample sizes. Downwards bias is greater for large
positive values of φ as might be expected.
5. Alternative to Marginal Likelihood Estimation
We close with a discussion of the alternative approach for binary time series regression modelling proposed
by Wu and Cui (2014). Their modified GLM (MGLM) method replaces exp(xTnt β)/(1 + exp(xTnt β)) in the
GLM log-likelihood 4 with a function π(xTnt β) representing the marginal mean to arrive at the objective
function
Z
T
n
X
exnt β+αt
l2 (β) =
[yt log π(xTnt β) + (1 − yt ) log(1 − π(xTnt β))] , π(xTnt β) =
g(αt )dα.
(31)
T
1 + exnt β+αt
t=1
The MGLM estimateR β̂2 is found by iterating two steps starting with the GLM estimate of β: step 1, estimate
the curve of π(u) = eu+α /(1 + eu+α )g(α)dα non-parametrically on u ∈ R; step 2, maximize l2 with respect
to β, based on the estimate of π(u) obtained in the first step. Steps 1 and 2 are repeated and the iteration
stops when the maximum value of l2 is reached and the last update of β is then regarded as the MGLM
estimator. Implementation details are provided in Wu and Cui (2014).
R
In defining their method Wu and Cui (2014) do not require that π(u) = eu+α /(1 + eu+α )g(α)dα for any
distribution g of the latent process. Hence it is not required that π(u) be non-negative and strictly increasing
in u. However, their main theorem concerning consistency and asympototic normality of β̂2 is stated in terms
/Marginal Estimation of Binomial Time Series Mixed Models
16
of this latent process specification. For such specifications, application of their non-parameteric method for
estimating π(u) requires additional constraintsRwhich are not currently implemented. For example, taking the
first derivative with respect to u, gives π̇(u) = eu+α /(1 + eu+α )2 g(α)dα ≤ π(u)(1 − π(u)) so π̇(u) ∈ [0, 0.25]
by application of Jensen’s inequality. When applied to the Cambridge-Oxford Boat Race time series the
non-parameteric estimate of π(u) is not monotonic and produces marginal estimates, at values of u between
the gaps in the observed values of xTnt β̂2 , which are zero and therefore not useful for prediction at new values
of the linear predictor.
Although not implemented in the R-code of Wu and Cui (2014), this constraint as well as that of monotonicity can be enforced in the nonparametric estimation of p(u) using an alternative local linearization to
that used in Wu and Cui (2014). For example, with constraints of monotonicity and π̇(u) ∈ [0, 0.25], different
estimates β̂1 = 0.2093, β̂2 = 0.1899 (compared to β̂1 = 0.237, β̂2 = 0.168 in Wu and Cui (2014)) are observed
for the model for the Cambridge-Oxford boat race series that they analyse. In this example, the marginal
likelihood estimates give τ̂ = 0 and hence β̂ degenerates to GLM estimate which differs from that of Wu
and Cui (2014). Somehow the marginal method (with or without monotonocity contraints) is avoiding the
degeneracy issue that arises with the marginal estimation method proposed in this paper. This needs to be
further understood.
While MGLM is computationally much more intensive than using standard GLMM methods for obtaining
the marginal estimates it appears to avoid degeneracy but for reasons that are not fully understood at this
stage. Additionally, it is not clear the extent to which MGLM with or without the proper constraints implied
by a latent process specification avoids the high proportion of degenerate estimates observed with GLMM
for binary data. Additionally the extent to which MGLM reproduces the correct curve for the marginal
probabilities π(u) when the true data generating mechanism is defined in terms of a latent process (parameter
driven specification) has not been investigated. The extent to which the MGLM estimate of π(u) differs from
the curve defined by a latent process specification might form the basis for a non-parametric test of the
distribution of the latent process – when {αt } is not Gaussian, the true curve of π(u) is not the same with
that evaluated under GLMM fits, and we may end up into different results for the estimates of β̂2 and β̂.
6. Discussion
To overcome the inconsistency of GLM estimates of the regression parameters in parameter driven binomial
models time series models we have proposed use of the marginal likelihood estimation, which can be easily
conducted using the generalized linear mixing model fitting packages. We have shown that the estimates
of regression parameters and latent process variation obtained from this method are consistent and asymptotically normal even if the observations are serially dependent. The distribution of the marginal estimates
is required for a score test of serial dependence in the latent process something which we will report on
elsewhere. The asymptotic results and proofs thereof have assumed that the latent process is Gaussian which
has helped streamline the presentation. This is not required for all results except for Lemma 2 (asymptotic
identifiability for the binary case) which relies directly on the normal distribution. The proofs can be readily
p
modified provided we assume that the moment generating function mαt (u) of αt is finite for all u < (d2 )
where τ < d2 defines the parameter space.
The structure of the model considered here is such that the theoretical results apply to other response
distributions such as the Poisson and negative binomial with very little change in the proofs of theorems.
GLM estimation in these cases is consistent and asymptotically normal regardless of serial dependence in
the latent process. The same will be true of the use of marginal estimation with the advantage that the
latent process variability is also estimated. While we have not yet shown this, we expect that for these other
response distributions the use of marginal estimation will lead to more efficient estimates of the regression
parameters.
For all response distributions and for moderate sample sizes the marginal estimation method can result in
a non-zero probability of τ̂ = 0. As we have observed in simulations, and explained via theoretical asymptotic
arguments, this is particularly problematical for binary responses with very high probabilities being observed
(and expected from theory) for ‘pile-up’ probability for τ̂ . We have observed that for binomial data (m > 1)
the ‘pile-up’ probability quickly decreases to zero and, as a result of this observation, we anticipate that this
probability will not typically be large for Poisson and negative Binomial responses.
/Marginal Estimation of Binomial Time Series Mixed Models
17
For binary data we have developed a useful upper bound approximation to this probability and subsequently proposed an improved mixture distribution for β̂ in finite samples. These theoretical derivations are
well supported by simulations presented. While this mixture distribution cannot be used based on a single
time series none-the-less it provided useful insights into the sampling properties of marginal estimation for
binary time series. Additionally the derivations suggest that regression models in which xT β varies over
an interval over which the inverse logit function is approximately linear will be particularly prone to the
‘pile-up’ problem and this persists even when there is strong serial dependence. Practitioners should apply
the marginal likelihood method with caution in such situations.
7. Acknowledgements
We thank Dr Wu and Dr Cui for providing us with the R-code for their application of the MGLM method
to the Boat Race Data reported in Wu and Cui (2014).
8. Appendix: A
Proof of Lemma 1. We consider the deterministic regressors only in this proof but the same arguments can
be extended to stochastic regressors. Now M is the largest value of m for which kM > 0, so Q(δ) = Q(δ0 ) if
and only if
Z 1X
M
π 0 (j, h(u)) log π(j, h(u)) − log π 0 (j, h(u)) du = 0
0 j=0
Since the integrand is non-positive the integrand can be zero if and only if the integrand is zero almost
everywhere. Hence, for a contradiction, assume ∃δ 6= δ0 such that
M
X
π 0 (j, h(u)) log π(j, h(u)) − log π 0 (j, h(u)) = 0,
∀u ∈ [0, 1].
j=0
and this can only happen if π 0 (j, h(u)) = π(j, h(u)) for all j = 0, . . . , M . Since
Z
√
√
M
π(j, h(u)) =
ḃ(h(u)T β + τ z)j (1 − ḃ(h(u)T β + τ z))M −j φ(z)dz
j
it is straightforward to show, by iterating from j = 0, . . . , M , that π 0 (j, h(u)) = π(j, h(u)) for all j = 0, . . . , M
is equivalent to
Z
Z
√ j
√
T
ḃ(h(u) β + τ z) φ(z)dz = ḃ(h(u)T β0 + τ0 z)j φ(z)dz, j = 1, . . . , M,
(32)
for any u ∈ [0, 1]
T
T
We next
√ show that the only way this can hold is if δ = δ0 . Fix u and denote a = h(u) β, a0 = h(u) β0
and σ = τ ,
h
i
h
i
dj (a, σ) = E ḃ(a + σz)j − E ḃ(a0 + σ0 z)j , j = 1, . . . , M.
where expectation is with respect to the density φ(·). Hence (32) is equivalent to so that d1 (a0 , σ0 ) =
d2 (a0 , σ0 ) = · · · = dM (a0 , σ0 ) = 0. Assume η0 = (a0 , σ0 ), if there exists η 6= η0 such that (32) holds, then
there is kη ∗ − η0 k ≤ kη − η0 k such that
h
i
h
i
E ḃ(η ∗ )0 b̈(η ∗ )
E ḃ(η ∗ )0 b̈(η ∗ )z
d1 (a, σ)
h
i
h
i
∗ 1
∗
∗ 1
∗
d2 (a, σ)
2E
ḃ(η
)
b̈(η
)
2E
ḃ(η
)
b̈(η
)z
v1
v1
∗
=
J(η
)
(33)
=
..
v2
.
.
v2
.
.
.
.
.
h
i
h
i
dM (a, σ)
M E ḃ(η ∗ )M −1 b̈(η ∗ ) M E ḃ(η ∗ )M −1 b̈(η ∗ )z
/Marginal Estimation of Binomial Time Series Mixed Models
18
where v1 = a − a0 and v2 = σ − σ0 cannot be zero at the same time when η 6= η0 , which is equivalent to the
matrix J(η ∗ ) being of full rank. But J(η ∗ ) is not of full rank if and only if the ratios of the second column
to the first column are the same for all j = 1, . . . , M . However, we now show that this ratio increases as j
increases. Since ḃ(·) and b̈(·) are non-negative functions we can define probability densities
gj (z) = R
so that
ḃ(a∗ + σ ∗ z)j−1 b̈(a∗ + σ ∗ z)φ(z)
,
ḃ(a∗ + σ ∗ z)j−1 b̈(a∗ + σ ∗ z)φ(z)dz
j = 1, . . . , M
h
i
R
E ḃ(η ∗ )j b̈(η ∗ )z
Egj (z ḃ(a∗ + σ ∗ z))
z ḃ(a∗ + σ ∗ z)g(z)dz
i = R
h
=
Egj (ḃ(a∗ + σ ∗ z))
ḃ(a∗ + σ ∗ z)g(z)dz
E ḃ(η ∗ )j b̈(η ∗ )
where Egj () denotes expectation with respect to gj . But since ḃ(a∗ + σ ∗ z) is an increasing function of z, z
and ḃ(a∗ + σ ∗ z) are positively correlated. Therefore,
Egj (z) < Egj (z ḃ(a∗ + σ ∗ z))/Egj (ḃ(a∗ + σ ∗ z))
it follows that
h
i
h
i
E ḃ(η ∗ )j−1 b̈(η ∗ )z
E ḃ(η ∗ )j b̈(η ∗ )z
h
h
i <
i,
E ḃ(η ∗ )j−1 b̈(η ∗ )
E ḃ(η ∗ )j b̈(η ∗ )
j = 1, . . . , M.
Then when (32) holds, (33) has a unique solution of (0, 0) for (v1 , v2 ), which contradicts to the assumption
that η 6= η0 . Thus (32) holds if and only if η = η0 , which implies a = a0 and σ = σ0 . By Condition 3, we can
conclude that a = a0 for all u implies β = β0 . Therefore Condition 5 holds for M ≥ 2.
Proof of Lemma 2. This proof considers the Fourier transform method used in Wang and Louis (2003).
Assume ∃δ 6= δ0 such that ∀x ∈ X,
Z
Z
√
0
T
π (1) = ḃ(x β0 + σ0 z)φ(z)dz = ḃ(xT β + σz)φ(z)dz = π(1), σ = τ .
For connected X, the first derivative with respect to x of both sides are also the same, that is
Z
Z
T
b̈(x β0 + σ0 z)φ(z)dz · β0 = b̈(xT β + σz)φ(z)dz · β
(34)
which implies there exists a constant c1 > 0 such that c1 = βk /β0,k for k = 1, . . . , r, with η = xT β,
c2 = σ/σ0 > 0, (34) can be rewritten as convolutions
Z
Z
ex
x2
1
√
g(u)h(η − u)du = c1 g(u)h(c1 η − c2 u)du; h(x) =
exp(−
).
,
g(x)
=
(1 + ex )2
2σ02
σ0 2π
let G(s) be the Fourier transform of the normal density g(·), H(s) be the Fourier transform of the logistic
density h(·), using the fact that the Fourier transform of the convolution is the product of Fourier transform
of each function,
Z Z
G(s)H(s) =
c1 φ(u)h(c1 η − c2 u)du e−iηs dη
Z
Z
= c1 φ(u)
h(c1 η − c2 u)e−iηs dη du
Z
Z
= c1 φ(u)
h(c1 η − c2 u)e−i(c1 η−c2 u)(s/c1 ) d(c1 η − c2 u) |c1 |−1 e−iu(c2 s/c1 ) du
Z
c1
= φ(u)e−iu(c2 s/c1 ) duH(s/c1 ) =
G(c2 s/c1 )H(s/c1 ), c1 > 0.
|c1 |
/Marginal Estimation of Binomial Time Series Mixed Models
19
it follows that G(s)H(s) = G(c2 s/c1 )H(s/c1 ), ∀s ∈ R. The Fourier transform for the mean zero normal
distribution is G(s) = exp(− 12 σ02 s2 ), and the Fourier transform for the logistic distribution is H(s) =
2πs/(eπs − e−πs ), thus for any s 6= 0,
1
1
1
= exp(− σ02 s2
exp(− σ02 s2 )
2
sinh(πs)
2
c2
c1
2
)
1
c1 sinh(πs/c1 )
for any fixed c1 6= 1, c2 can be expressed as a function of s and this function is not a constant over s, which
contradicts to the definition of c2 . Hence the equality holds if and only if c1 = 1 and c2 = 1.
9. Appendix: B
Proof of Theorem 1. This proof is presented in three steps: first, we show, for any δ ∈ Θ, that E(Qn (δ))
defined in (9) converges to Q(δ) defined in (10) and (11) under Condition 2a and 2b respectively; second,
a.s
that Qn (δ) − E(Qn (δ)) → 0 where Qn (δ) is defined in (8) from which, using compactness of the parameter
√
a.s.
d
−1
space, it follows that δ̂ → δ0 ; third, that n(δ̂ − δ0 ) → N (0, Ω−1
1,1 Ω1,2 Ω1,1 ).
a.s.
Proof that E(Qn (δ)) → Q(δ): Use Jensen’s inequality multiple times we have
mt
mt
mt
mt
X
X
X
X
ln πt (j)
−
πt0 (j) ln πt (j) ≤
πt0 (j) (− ln πt (j)) = −
j=0
≤−
≤
mt
X
j=0
j=0
j=0
j(xTt β) − mt (ln 2 + max(xTt β + τ /2, 0)) + c(j)
j=0
m
t
X
mt (ln 2 + τ /2) + mt |xTt β| − c(j) < mt (1 + mt ) ln 2 + τ /2 + |xTt β|
(35)
j=0
Pmt 0
then conditional on mt and xt ,
j=0 πt (j) ln πt (j) is bounded for all t and δ. Under Condition 2a, the
regressor xnt := h(t/n) is nonrandom as is the marginal density πnt . Then the strong law of large numbers
for mixing processes (McLeish, 1975) applied to {mt } gives
mt
mt
n X
n
X
X
X
1
1
0
0
lim
πnt
(j) log πnt (j) = lim
E
πnt
(j) log πnt (j) = Q(δ)
n→∞ n
n→∞ n
t=1 j=0
t=1
j=0
defined in (10). For Condition 2b the ergodic properties of the stationary processes mt and Xt can be used
to establish
mt
n X
M
m
X
X
X
X
1
nm 1
πt0 (j) log πt (j) = lim
πt0 (j) log πt (j) = Q(δ)
lim
n→∞
n→∞ n
n
n
m
t=1 j=0
m=1
j=0
{t:mt =m}
defined in (11).
P∞
Consistency: We write (8) as Qn (δ) = n−1 t=1 qt (δ) where qt (δ) = log f (yt |xnt , δ). By Blais, MacGibbon
and Roy (2000, Proposition 1), {qt (δ)} is strongly mixing for any δ ∈ Θ. To apply the strong law of large
numbers for mixing process in McLeish (1975) we need to show that ∃λ ≥ 0 such that
∞
X
kqt (δ) − Eqt (δ)k22+λ /t2 < ∞
t=1
where k · kp denotes the Lp norm. By Minkowski’s inequality and Hölder’s inequality,
kqt (δ) − Eqt (δ)k2+λ ≤ 2 kqt (δ)k2+λ
/Marginal Estimation of Binomial Time Series Mixed Models
20
using similar derivations
|qt (δ)| P
≤ |−mt |xTnt β|+c(yt )−mt (ln 2+τ /2)| where mt is bounded. It
P∞ as used 2in (35),
∞
2
T
2 2
T
suffices to establish
t (δ)k2+λ /t < ∞ if
t=1 |xnt β| /t < ∞. Under Condition 2a, {xnt β} is bounded
P∞ t=1 kq
2
for any given β; t=1 1/t < 2, therefore the result follows. Under Condition 2b, we have Ekxk2 < ∞, given
any β, for all ε > 0,
!
n
X
|xTt β|2 /t2 ≥ K ≤ (2/K)E|xT β|2 ≤ ε,
P
if K ≥ 2E|xT β|2 /ε.
t=1
Pn
a.s.
Then n−1 t=1 [qt (δ) − Eqt (δ)] → 0 for any δ ∈ Θ. Together with the first part of the proof given above
a.s
we now have Qn (δ) → Q(δ). Since Θ is a compact set, and Qn (δ) is a continuous function of δ for all n, by
a.s.
Gallant and White (1988, Theorem 3.3), δ̂ := arg max Qn (δ) → δ0 .
Θ
Asymptotic Normality: Using a Taylor expansion,
√
n(δ̂ − δ0 ) = −
the asymptotic normality of
√
!−1
n
n
1 X¨ ∗
1 X˙
√
lt (δ )
lt (δ0 ),
n t=1
n t=1
n(δ̂ − δ0 ) can be obtained if
n
−
a.s.
δ ∗ → δ0
n
1 X¨ ∗ p
lt (δ ) → Ω1,1
n t=1
and
1 X˙
d
√
lt (δ0 ) → N (0, Ω1,2 ).
n t=1
Conditional on mt and xnt , {¨lt (δ)} is strongly mixing. Then by Chebyshev’s inequality and Ibragimov
and Yu (1971, Theorem 17.2.3), ∃ > 0 such that
n
n
1X ¨
1 X¨
p
E(lt (δ)) → 0,
lt (δ) −
n t=1
n t=1
kδ − δ0 k ≤
Pn
Pn
a.s.
a.s.
Since δ ∗ → δ0 , and the continuity of ¨lt (·) with respect to δ, n−1 t=1 ¨lt (δ ∗ ) − n−1 t=1 ¨lt (δ0 ) → 0,
P
P
a.s.
n
n
and n−1 t=1 ¨lt (δ0 ) − n−1 t=1 E(¨lt (δ0 )) → 0. Now E(¨lt (δ0 )) = E(l˙t (δ0 )l˙tT (δ0 )), and hence it follows that
P
p
n−1 t=1 ¨lt (δ ∗ ) → Ω1,1 .
Next we show that Ω1,1 is positive definite. Let s = (s1 , s2 ) be an (r + 1) dimensional constant vector,
without lose of generality,
sT s = 1. Define qt (δ0 ) = sT l˙t (δ0 ), note det(Ω1,1 ) ≥ 0 and det(Ω1,1 ) = 0 only if
2
E qt (δ0 )|mt , xnt = 0 for all t. Then under Condition 3, Ω1,1 is positive definite. The limit, Ω1,1 , under
Condition 2a is given in (28). For stationary regressors such limit can be easily obtained using ergodic
theorem.
Next we show Ω1,2 exists. Note
! n−1
n
X
1 X
√
qt (δ0 ) =
n t=1
Ω1,2 = lim Var
n→∞
h=0
! n−1
n−h
X
1 X
Cov(qt (δ0 ), qt+h (δ0 )) +
n t=1
h=1
n
1 X
Cov(qt (δ0 ), qt−h (δ0 ))
n
t=h+1
then Ω1,2 exists if
lim
n→∞
n−1
X
h=0
!
n−h
1 X
|Cov(qt (δ0 ), qt+h (δ0 ))| < ∞,
n t=1
lim
n→∞
n−1
X
h=1
!
n
1 X
|Cov(qt (δ0 ), qt−h (δ0 ))| < ∞.
n
t=h+1
Since the {qt (δ0 )} is strong mixing, by Theorem 17.3.2 in Ibragimov and Yu (1971),
!
n−1
n−1
X 1 n−h
X
X
|Cov(qt (δ0 ), qt+h (δ0 ))| ≤ 2
ν(h)λ/(2+λ) Wh < ∞,
n t=1
h=0
h=0
!
/Marginal Estimation of Binomial Time Series Mixed Models
21
where
n−h
1 X
1+λ
4 + 3(ct c1+λ
ct+h ) ,
t+h + ct
n→∞ n
t=1
Wh = lim
ct ≥ kqt (δ0 )k2+λ
Such ct exists, for example, take λ = 2, use Cauchy-Schwarz’s inequality,
Z
4
4
T
−1
E |qt (δ0 )| |mt , xnt ≤ E
f (yt |xnt , zt , δ0 )φ(zt ) (yt − mt ḃ(W0,t ))(s1 xnt + s2 zt ) dz · f (yt |xnt , δ0 )
h
i
= E mt b̈(W0,t )(1 + (3mt − 6)b̈(W0,t ))(sT1 xnt + s2 zt )4 |mt , xnt
Then by the application of CLT for mixing process in Theorem 3.2, Davidson (1992) we will have
Pn
d
n−1/2 t=1 l˙t (δ0 ) → N (0, Ω1,2 ).
Proof√of Theorem 2. Following the proof of Davis, Dunsmuir and Wang (2000) and Wu and Cui (2014), let
u = n(β − β 0 ) (note here we centre on β 0 and not the true value β0 as was done in these references). Then
maximizing
n
X
ln (β) =
[yt (xTnt β) − mt b(xTnt β) + c(yt )]
t=1
√
over β is equivalent to minimizing gn (u) over u where gn (u) := −ln (β 0 +u/ n)+ln (β 0 ). Let û = arg min lim gn (u).
n→∞
Write gn (u) := Bn (u) − An (u) where
Bn (u) :=
n
X
√
√
mt b(xTt β 0 + xTt u/ n) − b(xTt β 0 ) − ḃ(xTt β 0 )xTt u/ n
t=1
and
n
1 X
An (u) := uT √
(yt − mt ḃ(xTt β 0 ))xt = uT Un .
n t=1
Using similar procedures as in the proof of Theorem 1. in Wu and Cui (2014) it is straightforward to show
Bn (u) →
and
1 T
u Ω1 u
2
T
1 T
is Un
E e
= exp − s Ω2 s .
2
for each u. Since gn (u) is a convex function of u and ûn minimizes gn (u), then an application of the functional
d
limit theory gives ûn → û, where û = arg min lim gn (u). In conclusion,
n→∞
d
gn (u) → g(u) =
1 T
u Ω1 u − uT N (0, Ω2 )
2
d
−1
on the space C(Rr ), and ûn → û, where û ∼ N (0, Ω−1
1 Ω2 Ω1 ).
Proof of Theorem 3. Since S1,n (β̃), a linear function of U1,n and U2,n , to show that
√
n S1,n (β̃) − E(S1,n (β 0 ))
is normally distributed it is sufficient to show that the the joint distribution of (U1,n , U2,n ) is multivariate
normal. As defined in (21),
n
1 X 2
U1,n := S1,n (β 0 ) − E (S1,n (β 0 )) = √
e 0 − Ee2t,β 0 ;
n t=1 t,β
n
n
1 X
1 X
U2,n := Ω−1
et,β 0 xnt = √
et,β 0 cnt
1 √
n t=1
n t=1
where cnt = Ω−1
1 xnt , defines a sequence of non-random vectors.
Let Unt = (U1,nt , U2,nt ) be the joint vector at time t, then each dimension of {Unt } is uniformly bounded,
strongly mixing and E(Unt ) = 0. We need to prove that a1 U1,n +aT2 U2,n has normal distribution for arbitrary
/Marginal Estimation of Binomial Time Series Mixed Models
22
constant vector a = (a1 , a2 ) where aT a = 1 without loss of generality. By the SLLN for mixing process in
McLeish (1975), there exists a limiting matrix ΩU such that
! n−1
!
n
n−1
n
X
X n−h
X
X X
T
T
T
Var(
a Unt ) =
a Cov(Unt , Un,t+h )a +
a Cov(Unt , Un,t−h )a → aT ΩU a.
t=1
t=1
h=0
h=1
t=1+h
Then conditions of the CLT of Davidson (1992) are satisfied, and we have
Pn
t=1
d
Unt → N (0, ΩU ).
Proof of Theorem 4. This proof follows Moran (1971). δ0 is the true value of the parameters, δ 0 = (β 0 , 0) is
the limit of parameters that maximize (7) for fixed τ = 0; δ̂ is the maximum likelihood estimators of (7).
Consider first the distribution of δ̂ if τ̂ > 0. Since the unconditional joint distribution of β̂ and τ̂ is
−1
multivariate normal with N (0, Ω−1
1,1 Ω1,2 Ω1,1 ) – see proof of Theorem 1, hence F2 (c, δ0 ), the distribution of
√
−1
n(β̂ − β0 )|τ̂ > 0 is skew normal based on N (0, Ω−1
1,1 Ω1,2 Ω1,1 ).
When τ̂ = 0, use a Taylor expansion to the first derivatives around δ 0 , then
√
conditional on n−1/2
n
Pn
n(β̂ − β 0 ) =
!−1
n
n
1 X¨ 0
1 X ∂lt (δ 0 )
√
−
lt (δ )
+ op (1),
n t=1
n t=1 ∂β
∂lt (δ 0 )/∂τ < 0. As n → ∞,
Pn
!
√1
e 0x
n h t=1 t,β nt
Ω2
0
d
i
Pn
→ N(
,
0
T
1
2
T 0
√
E(S
(β
))/2
K
1,n
S /2
t=1 et,β 0 − mt b̈(xt β )
2 n
t=1
1 X ∂lt (δ 0 )
√
=
n t=1 ∂δ
KS /2
)
VS /4
References
Blais, M., MacGibbon, B. and Roy, R. (2000). Limit theorems for regression models of time series of
counts. Statistics & probability letters 46 161–168.
Cox, D. R. (1981). Statistical analysis of time series: some recent developments. Scandinavian Journal of
Statistics 8 93–115.
Davidson, J. (1992). A central limit theorem for globally nonstationary near-epoch dependent functions of
mixing processes. Econometric theory 5 313–329.
Davis, R. A., Dunsmuir, W. T. and Wang, Y. (2000). On autocorrelation in a Poisson regression model.
Biometrika 87 491–505.
Davis, R. A. and Dunsmuir, W. T. M. (2015). State Space Models for Count Time Series. CRC MONOGRAPHS.
Davis, R. A. and Wu, R. (2009). A negative binomial model for time series of counts. Biometrika 96
735–749.
Dunsmuir, W. T. M., Tran, C. and Weatherburn, D. (2008). Assessing the Impact of Mandatory
DNA Testing of Prison Inmates in NSW on Clearance, Charge and Conviction Rates for Selected Crime
Categories. NSW Bureau of Crime Statistics and Research.
Gallant, A. R. and White, H. (1988). A unified theory of estimation and inference for nonlinear dynamic
models. Basil Blackwell New York.
Gupta, A. K., González-Farıas, G. and Domınguez-Molina, J. A. (2004). A multivariate skew normal
distribution. Journal of Multivariate Analysis 89 181–190.
Heagerty, P. J. and Lumley, T. (2000). Window subsampling of estimating functions with application
to regression models. Journal of the American Statistical Association 95 197–211.
Ibragimov, I. A. and Yu, V. L. (1971). Independent and stationary sequences of random variables. WoltersNoordhoff.
McLeish, D. L. (1975). A maximal inequality and dependent strong laws. The Annals of probability 3
829–839.
/Marginal Estimation of Binomial Time Series Mixed Models
23
Moran, P. A. (1971). Maximum likelihood estimation in non-standard conditions. Mathematical Proceedings
of the Cambridge Philosophical Society 70 441–450.
Neuhaus, J. M., Kalbfleisch, J. D. and Hauck, W. W. (1991). A comparison of cluster-specific
and population-averaged approaches for analyzing correlated binary data. International Statistical Review/Revue Internationale de Statistique 59 25–35.
Wang, Z. and Louis, T. A. (2003). Matching conditional and marginal shapes in binary random intercept
models using a bridge distribution function. Biometrika 90 765–775.
Wu, R. (2012). On variance estimation in a negative binomial time series regression model. Journal of
Multivariate Analysis 112 145–155.
Wu, R. and Cui, Y. (2014). A PARAMETER-DRIVEN LOGIT REGRESSION MODEL FOR BINARY
TIME SERIES. Journal of Time Series Analysis 35 462-477.
| 10math.ST
|
JSC : A JavaScript Object System
Artur Ventura
[email protected]
arXiv:0912.2861v1 [cs.PL] 15 Dec 2009
January 20, 2018
Abstract
The JSC language is a superset of JavaScript designed to ease the development of large web applications. This language extends JavaScript’s
own object system by isolating code in a class declaration, simplifying
multiple inheritance and using method implementation agreements.
1
Motivation
As web applications have been gaining more dynamic behavior, JavaScript has
become more important in web development. As such, it is useful to use a software engineering aproach to JavaScript.
JavaScript object-oriented model is prototype-based. This model can be extremely versatile to develop. However, in a large team of developers, not only
problems such as name colisions and unintentional method redefinition can occur, but also maintenance becomes an issue.
JSC attempts to remedy some of these issues by implementing a Object
System on top of standard JavaScript.
2
Related Work
There have been some attemps to add structed programming to Javascript.
In ECMAScript (language from witch JavaScript is a dialect) last version
were added support for defining classes. Another dialect of ECMAScript, ActionScript supports defining classes. Althought those systems are more advanced (they support typechecking via anotation on variables, generics, dynamic classes, templates, etc.), they aren’t avaible in the current browsers. It’s
expected that future versions of JavaScript extends this behavior from ECMAScript.
Other attempt have been creating creating supersets to JavaScript. One
example of this is Objective-J. This is a sctrict superspet of Javascript, but
adds a object system by implementing a language similar to Objective-C. This
aproach is similar to JSC, but this language uses the same message passing than
Objective-C making the methods unparsable by JavaScript.
1
3
Object Oriented JavaScript
In JavaScript, a class is a function, so its definition is similar to the definition
of a function, as shown next:
function Rectangle(w, h) {
this.width = w;
3
this.height = h;
4 }
1
2
A function in javascript is of type Function but behaves similarly to a hash
or a dictionary. One of the properties in this object is the prototype. This property represents a structure that is not yet fully defined and it is only formed
when it is instantiated. Instantiation can be done by:
new Rectangle(10,10);
new operator will invoke the function Rectangle with this as a new object
with a reference to Rectangle.prototype, and return this object1 .Any changes
made to this object will only afect the new object, and not the prototype. If a
property is not defined in the object but on the prototype, then the prototype’s
value is returned.
A method can be defined by adding a function to the class prototype as
shown next:
1
2
3
Rectangle.prototype.getArea = function (){
return this.height * this.width;
}
Methods are called like this:
rectangle.getHeight()
As with the constructor, getHeight will be invoked with this as a reference
to the instance, in this case to rectangle. This method will be avaible to all
instances that don’t define localy the method getArea.
The code below shows how inheritance is implemented: by copying the
prototype of the superclass to the subclass.
1
2
3
4
5
6
7
function PositionedRectangle(x, y, w, h) {
Rectangle.call(this, w, h);
this.x = x;
this.y = y;
}
PositionedRectangle.prototype = new Rectangle();
1 Actually
Rectangle could return anything, but the default is this
2
3.1
Problems
The first problem is the inexistence of seperation between functional and object
oriented programming. All the examples shown are extensions to JavaScript’s
functional programming that allows to create a behavior similar to object oriented.
Another problem is that today browsers share the same enviroment will all
JavaScript files. It’s possible to have the same method defined in two distinct
files. This can difficult debugging and maintenance.
4
Objectives
JSC is an extension to JavaScript and was developed with the following goals
in mind:
• Create a new object-oriented language, but keep (as much as possible)
JavaScript’s syntax and semantics.
• Provide mechanisms to maintain the code isolated in packages.
• Provide a meta-object protocol that allows an easy use of reflection and
intersection, while keeping the code organized.
• Provide simple and easy to use multiple inheritance.
• Provide an implementation agreement between classes, which is similar to
Java’s interfaces or to Objective-C’s protocols.
5
JSC
5.1
Class Definition
JSC is defined in classes. Each class is defined in a single file. A simple JSC
Class is show next:
1
2
3
4
5
6
7
8
9
10
11
12
package UI.Component;
class Rectangle{
slots:[height,width],
Rectangle: function (w,h){
this.setHeight(h);
this.setWidth(w);
},
getArea: function (){
return this.getHeight() * this.getWidth();
}
}
Each class begins with a package declaration. A class is always referred to
by its package and class name (in this case UI.Component.Rectangle). The
body of the class is declared as a JavaScript Object. The fourth line represents
3
the slot declaration. In JSC, class slots can’t be directly used. To access a slot,
a getter and setter method is provided for each slot. The fifth and ninth lines
declare a constructor and a method, respectively. A constructor is a method
with the same name of the class. A constructor is not required, but there can
be only one by class.
It is important to notice that only the class header and the slot definition
aren’t capable of being parsed by JavaScript.
5.2
Instantiation
JSC attempts to minimize the usage of the global enviroment. Actually, the
only global definition required is the function Class. This function expects a
string with the canonical class location (package and name), and returns the
meta-object representing such class. In our example, Rectangle’s meta-class
can be obtained by:
Class("UI.Component.Rectangle")
In runtime, each meta-class contains, among others, a method called create.
This method will create a new instance and call the class constructor. With this,
we can now instantiate our Rectangle class:
Class("UI.Component.Rectangle").create(10,10)
5.3
Inheritance
Inheritance can be used by extending the superclass as shown next:
1
2
3
4
package UI.Component;
class PositionedRectangle extends UI.Component.Rectangle{
slots:[x,y],
5
PositionedRectangle: function (x,y,w,h){
6
Class("UI.Component.Rectangle").init(this,w,h);
7
8
9
10
11
this.setX(x);
this.setY(y);
}
}
Inheritance in JSC works like mixins. Each class method and slot from
the superclass will be added to the subclass. A class can extend from several
classes. In PositionedRectangle both the height and with slot, getters and
setters from Rectangle were copied. The token extends insted of mixin was
used because makes the class header similar to ECMAScript V4 and Java’s.
Each meta-class object contains a init method that calls the constructor
for that class, using the first argument as the instance. This is similar how
Python’s superclass constructor is called.
4
5.4
Static Enviroment
Each JSC class can declare methods that can be used on a meta-class level. A
simple example is shown next:
1
2
3
4
5
6
package Main;
class App {
static:{
main: function (args){
...
7
}
8
}
9 }
5.5
Protocols
A protocol in JSC assures the existence of certain methods. A protocol declaration is shown next:
1
2
3
4
package UI.Component;
protocol Draggable {
element: true,
5
eventListener: false
6 }
In our example Draggable declares the existence of two methods, element
and eventListener. The keyword next to the method name declares the need
to implement it. true declares that this method is required to be implemented
in each class that extends this protocol while false guarantees that if such
method is not implemented, an empty function with such name will be provided.
A protocol can be extended only from other protocols, and a class can implement any number of protocols. The verification of required methods is done
both in run- and compile-time.
5.6
Class initialization
When JSC starts2 , all classes in the classpool are going to be initialized. This
is done by invoking the meta-class method classInit.
This method will compute the effective set of methods and slots that this
class possesses, and it will detect if all the protocols have been satisfied. Finally,
it will setup the prototype.
2 See
Usage for more details
5
5.7
Runtime intersection
It is possible to create a new class/protocol, or change a class/protocol by altering the meta-class definition, in any moment during execution. However, it
is required to call again classInit so that the prototype is reconstructed.
It’s also possible to extend the behavior of JSC by extending the lang.Class
class. Protocols is one example of this, but it’s quite simple to implement
something similar to java’s abstract classes in JSC.
6
Current Problems
6.1
Global Variables
JSC attempts to minimize the usage of the global enviroment by encapsulating
the code into classes. However, However, JavaScript allows the declaration of
gobal variables from within a function.. Any variable that is declared without
the keyword var is declared in the global enviroment. For instance:
1
2
3
4
function Foo(){
var local = 1;
global = 1;
}
Currently, this is not allowed in JSC, but it is not verified.
6.2
Slot default value
Currently, the slot declaration doesn’t allow a slot to have a default value. It
would be nice to have a slot declaration similar to this:
1
2
3
4
5
6
7
8
package Bar;
class Foo{
slots:{
aSlot:{ getter:"getSlot", setter:"setIt", default:1 },
anotherSlot: { default: Class("Baz.Bing").create(1,2) }
}
}
Currently, JSC doesn’t support this syntax because there isn’t a easy to use
parser of JavaScript available. The JSC compiler assures the correctness of the
code by loading the methods into a JavaScript Engine and detecting the parsing
errors generated.3
When such parser is available, both this and the Global Variables problem
can be addressed.
3 jscc
(JSC Compiler) currently uses Google’s V8 Engine
6
7
Usage
The current compiler can target 3 usage methods:
• Server - JSC was developed due to a need for developing a lot of code to
run in a web server. In this mode, the code is loaded into a RDBMS and
loaded upon need.
• Client - Another usage is generating a single classpool image in a self contained JavaScript file and load it in a browser like an ordinary JavaScript
file. The JSC code can be accessed from other JavaScript files by using
Class function.
• JSC Virtual Machine - jscvm is a small virtual machine implementation
of JSC.
8
Performance
A small library, implemented in JSC and with 19 files, occupies 196 KB. After
compilation targeting the browser ended with a single file occupying 56KB. This
file takes from below a second in Chrome and up to 2 seconds in Firefox.
After using the YUI Compressor4 this file shrinked to 40Kb. It almost loaded
instantaneously both in Chrome and Firefox.
4 YUI Compressor is available at http://developer.yahoo.com/yui/compressor/. This
application not only removes whitespaces, but also reduces the variable names.
7
| 6cs.PL
|
Mean and Variance of Phylogenetic Trees
1 David
Daniel G. Brown1 and Megan Owen2
arXiv:1708.00294v1 [q-bio.PE] 1 Aug 2017
R. Cheriton School of Computer Science, University of Waterloo, Waterloo, ON, N2L
3G1, Canada;
2 Department of Mathematics, Lehman College - City University of New York, New York, NY,
10468, USA
Corresponding author: Megan Owen, Department of Mathematics, Lehman College,
City University of New York, 250 Bedford Park Blvd West, Bronx, NY, 10468, USA;
E-mail: [email protected].
Abstract.— We describe the use of the Fréchet mean and variance in the
Billera-Holmes-Vogtmann (BHV) treespace to summarize and explore the diversity of a set
of phylogenetic trees. We show that the Fréchet mean is comparable to other summary
methods, despite its stickiness property, and that the Fréchet variance is faster and more
precise than commonly used variance measures. These mean and variance measures are
more theoretically justified, and more robust, than previous estimates of this type, and can
be estimated reasonably efficiently. This leads to applications such as mean hypothesis
testing. (Keywords: phylogenetics, treespace, Fréchet mean, Fréchet variance, two-sample
test )
Sets of related phylogenetic trees are commonly encountered in evolutionary biology.
For example, one might encounter such a set as the output of an inference program like
MrBayes Ronquist et al. (2012), or as a set of gene trees on some given set of species. For
this paper, we think of any such set as being a sample from an underlying distribution on
the set of all phylogenetic trees with a fixed leaf set.
Here, we consider a mathematically-founded basis for describing the mean of such a
distribution, using the representation of trees as elements of a geometric space and looking
for the Fréchet mean: the tree that minimizes the sum of the squared distance between the
mean and the elements of a sample from the distribution. This formulation also allows the
identification of the Fréchet variance, which is the actual sum of the squared distances
between the sample elements and the Fréchet mean.
We describe algorithms for computing the Fréchet mean and variance. They are
fairly efficient to compute, and we show that the mean gives a good way of summarizing
the uncertainty found in a set of trees. A particularly nice property of the mean is that it
may result in a non-fully-articulated tree, but does so in a statistically appropriate way.
We also describe computation of the variance. The Fréchet variance is a better measure of
statistical uncertainty than simpler measures (like the number of topologies found in a set
of trees), and is faster to compute than the sum of all pairwise distances between trees, for
a large set of trees.
Our results show that the treespace-based measure of phylogenetic distance that
originated with Billera et al. (2001) can in fact be used in practical applications.
Background
We begin by describing the treespace in which we are working, and some properties of that
space and its distance measure. We also consider other ways of summarizing a collection of
trees besides the Fréchet mean and variance, and some of their properties.
Treespace
The Billera-Holmes-Vogtmann (BHV) treespace, Tn , (Billera et al. 2001) contains all
unrooted phylogenetic trees with edge lengths and a given set of n + 1 labelled leaves. For
this paper, we fix the set of leaf labels to be {0, 1, ..., n}. Any of these trees can be thought
of as rooted by fixing leaf 0 as the root. In this paper, we will defined the BHV treespace
as a subspace of RN , where N = 2n − 1 is the number of possible splits on n + 1 leaves, or
equivalently, the number of possible partitions of the set of leaves into two sets, each of size
at least 1. Each coordinate of RN corresponds to a different split, where the order of the
splits does not matter, but is fixed. Note that by allowing partitions of size 1, we include
splits corresponding to edges ending in leaves. The original definition Billera et al. (2001)
ignores these edges, but notes that they can be included, as we have done here.
Given a tree T with n + 1 leaves and edge lengths, it corresponds to the following
vector in RN : for every edge e in T with length |e|T , let the coordinate corresponding to the
split induced by e be |e|T . Let the coordinates corresponding to splits not induced by edges
in T be 0. Let Tn be the set of vectors in RN that correspond to trees, as just described.
Not all non-negative vectors in RN correspond to trees due to split incompatibility. Two
splits are incompatible if they cannot be induced by edges existing in the same tree. For
example, a cherry is a pair of adjacent leaves in a tree, and the corresponding split
separates the two adjacent leaves from all others. No tree with n ≥ 4 can have both {1, 2}
and {1, 3} as cherries, so the corresponding splits {1, 2}{0, 3, 4, ..., n} and
{1, 3}{0, 2, 4, 5..., n} are incompatible, and no vectors in Tn have+ positive values in both
these coordinates. The topology of a tree is the set of all splits induced by the edges of that
tree. A binary tree, in which all interior nodes have degree 3, contains 2n − 1 splits, while
degenerate or non-binary trees will contain fewer than 2n − 1 splits. See Figure 1. For
more details, including a combinatorial description of the space, see Billera et al. (2001).
There is a distance metric associated with the BHV treespace called the BHV
distance or the geodesic distance, which was also defined in Billera et al. (2001). Consider
all trees in Tn with the same topology. Because these trees all correspond to the same set
of splits, their vectors have exactly the same set of non-zero coordinates, which can take on
any positive values. Thus, if the number of non-zero coordinates is d, this set of trees
corresponds to a d-dimensional Euclidean orthant, which is the non-negative part of Rd .
TT
4n 4n
0
0
A
A
B
B
C
0
D
C
AB0 | CD
0
D
A
B
C
D
0
0
0
A0|BCD
AB|CD0
T’1
B
C
D
AD0 | BC
0
A
(a)
C
0
B
C
C
D
D
A
0
ABC|D0
D
B
D
C
A
B
T2
T1
D
B
AB|CD0
T’2
A
A
C
A
A
ABD|C0
D
B
C
B
0
AB0 | CD
D
ABC|D0
C
A
B
(b)
Figure 1: (a) Five and (b) three of the fifteen orthants in the Billera-Holmes-Vogtmann
(BHV) treespace T4 . For ease of visualization, the five dimensions corresponding to the
leaf edges are not included. Thus, each binary tree topology is represented by a quadrant
(2-dimensional orthant), with the two quadrant axes corresponding to the lengths of the
interior edge splits. The axis between two quadrants corresponds to a tree topology that is
not fully resolved. The geodesic (shortest) path between two trees is shown by a dashed line
and may pass through different orthants depending on the branch lengths of the endpoint
trees, as shown by the geodesics between T1 and T2 and between T10 and T20 .
When the tree topology is binary, d = 2n − 1. For any two trees with the same topology,
the BHV distance is the Euclidean distance between their corresponding vectors in that
orthant. For two trees with different topologies, the BHV distance is the length of the
shortest path between them that remains in the treespace. The length of any path can be
computed by calculating the Euclidean distance of the path restricted to each orthant that
it passes though, and summing these lengths. The shortest path is called a geodesic, and
will pass from one orthant to the next orthant through lower-dimensional boundaries
corresponding to trees with fewer splits. See Figure 1 for an example of two geodesics in T4 .
The BHV treespace is connected, since for any two trees are connected by a path
through the origin. This may or may not be the geodesic. Billera et al. (2001) showed that
the BHV treespace is globally non-positively curved (Bridson and Haefliger 1999), which
implies that geodesics are unique. Owen and Provan (2011) gave a polynomial time
algorithm for computing the geodesic distance between two trees that runs in O(n4 ) time,
but which is faster for trees with common splits.
Note that if the L1 metric, which is also known as the Manhattan or taxicab metric,
is used in each orthant instead of the L2 or Euclidean metric, then the length of a (no
longer unique) geodesic between two trees is the same as the weighted Robinson-Foulds
distance (Robinson and Foulds 1979) between them (see St. John (2017) for a good
summary).
Mean and variance
In Euclidean space, the Fréchet mean, or centre of mass, is the point minimizing the sum of
the squared distances to the sample points, and is equivalent to the coordinate-wise
average of the sample points. The Fréchet mean was similarly defined for treespace by
Miller et al. (2015) and Bačák (2014), independently. For a set of sample trees
{T
Pr1 , T2 , ..., Tr2}, the Fréchet mean, or simply, mean, is the tree t which minimizes
i=1 d(t, Ti ) , where d(·, ·) is the BHV distance between the two trees. The Fréchet
variance, or simply variance, is that minimum sum of squared distances. This mean is
unique because treespace is non-positively curved. Both Miller et al. (2015) and Bačák
(2014) gave an algorithm for approximating the mean and variance based on the Law of
Large Numbers for non-positively curved space derived by Sturm (2003).
We briefly mention some interesting properties of the mean. See Miller et al. (2015,
Section 5) for details and proofs. The mean tree is not necessarily a refinement of the
majority-rule consensus tree (which contains all splits found in a majority of the trees Ti ).
Any split that appears in the mean tree appears in at least one of the sample trees, and
any split that appears in all the sample trees appears in the mean tree. Finally, the mean
tree is “sticky” (Hotz et al. 2013), in that perturbing one of the sample trees does not
always change the mean tree. This “stickiness” only happens when the mean is on a
lower-dimensional orthant in treespace, which corresponds to a degenerate tree topology.
This implies that the mean will be unresolved more often than one might expect or wish,
similar to the majority-rule consensus tree. This is not a coincidence, as the majority-rule
consensus tree is the median tree under the Robinson-Foulds distance (Barthélemy and
McMorris 1986), and Pattengale (2005) further showed that the weighted majority-rule
consensus tree is the median under the weighted Robinson-Foulds distance. Recall that
using an L1 metric on each orthant in this space instead of the Euclidean metric gives the
weighted Robinson-Foulds distance instead of the BHV distance.
The Fréchet variance of a set of trees quantifies how spread out a set of trees is from
their mean. We will show that in our experiments, as sequence length increases and there
is more information about the tree to be reconstructed, the variance of samples of trees
from the bootstrap and posterior distributions decreases.
Sturm (2003) showed that a Law of Large Numbers holds, meaning that as the
sample size increases, the sample means of a distribution over Tn converge to the true mean.
Barden and Le (2017) proved a Central Limit Theorem on the BHV treespace, showing
that the distribution of the sample means converges to a certain Gaussian distribution.
Other measures of centre among a set of trees
We compare the Fréchet mean to three other commonly used measures of centre or
consensus in phylogenetics. The first is the majority-rule tree, which is the tree containing
exactly those splits appearing in a majority of the input trees (Margush and McMorris
1981). The majority-rule consensus tree and the Fréchet mean are not, in general,
refinements of each other (Miller et al. 2015).
The other two measures of centre comes from tree search algorithms. In these cases,
the tree search procedures produce a distribution of trees along with a most probable tree,
which acts like a centre. Specifically, we look at maximum likelihood search using RAxML
(Stamatakis 2014), which produces a distribution of bootstrap trees (Felsenstein 1985)
where the maximum likelihood (ML) tree can be considered as a centre. Alternatively,
using a Bayesian approach to phylogenetic inference, the implementation MrBayes
(Ronquist et al. 2012) returns a posterior distribution representing the most probable tree,
with the maximum a posteriori tree (MAP) tree often being used as its summary tree.
Other measures of variance
The variance or amount of variability in a set of trees is less established as a measure than
the summary tree. Often, a visual representation, such as SplitsTree (Huson and Bryant
2006) or DensiTree (Bouckaert 2010), is used to represent the diversity of a set of trees.
Unfortunately, these methods cannot be assessed or compared quantitatively. Instead, we
will compare the Fréchet variance with several quantitative measures of variance, namely
the number of different tree topologies in the set, the number of different splits in the set,
and the sum of the squared geodesic distances between each pair of trees in the set. The
later measure was proposed in Chakerian and Holmes (2012) as a method for
approximating the variance, and used by Ponciano et al. (2012) to measure the variance of
posterior distributions. It requires O(m2 n4 ) time to estimate the variance of m trees on n
leaves.
One can also explore the variability of a set of trees using pairwise comparison
methods applied to all pairs of trees in the set. One can compare trees using both the BHV
distance and the more commonly-used Robinson-Foulds (RF) distance (Robinson and
Foulds 1981). The RF distance between trees T1 and T2 is defined as the number of splits
in T1 but not in T2 , plus the number of splits in T2 but not in T1 . That is, it is the size of
the symmetric difference between the split sets of T1 and T2 . Note that this definition only
uses the tree topologies, and not the edge lengths of the trees.
Related work
The most closely related work to ours is that of Benner et al. (2014). They investigated the
behaviour of both the mean and the median (the tree minimizing the sum of distances,
instead of the sum of squared distances, to the sample trees). Bačák (2014) summarize
posterior distributions returned by Bayesian tree reconstruction methods. The authors
simulated sequences of lengths 50, 100, 250, and 500 using a 14-taxa tree of plants and the
F81 evolutionary model (Felsenstein 1981). They show that the mean and median
estimates are comparable to the majority-rule consensus estimate, and in some instances
perform better, and also investigate how the Fréchet variance changes with sequence length.
Our work was conceived independently, and while the overall aim of the work is the
same, the scope of our experiments are much broader, considering tree distributions
generated from both maximum likelihood and Bayesian methods; the general GTR
evolutionary model; trees with more taxa; and a more comprehensive look at the Fréchet
variance.
This paper focuses on the Fréchet mean and variance of a set of trees, which is a
point summary of the data. Other work has look at 1-dimensional summaries, or best fit
lines, in BHV treespace (Nye 2011, 2014; Feragen et al. 2013), and extensions to
multi-dimensional summaries or a generalization of Principal Components Analysis (PCA)
(Nye et al. 2016). Confidence sets on the BHV treespace are constructed in Willis (2016),
and a full Central Limit Theorem is given in Barden and Le (2017).
Methods
Computation of the mean and variance
The mean tree is the tree that minimizes the sum of the squared BHV distances from itself
tree to the sample trees. This is equivalent to the centre of mass in Euclidean space. We
compute an approximation of the mean tree using the iterative implementation described
in Miller et al. (2015). In this implementation, a new approximation of the mean tree is
returned each iteration, and these approximations converge to the true mean tree as the
number of iterations grows. To decide when to stop the iterative algorithm, we use a
program option to check for convergence using a Cauchy sequence of length 10 with an
epsilon of 10−6 . This means that we stop the iterative algorithm when the pairwise BHV
distances between the last 10 mean approximations were all less than or equal to 10−6 . In
all of our experiments, the means converged within 285 000 iterations. We also use the
random permutation heuristic. We validated this choice of epsilon by choosing one
repetition for each sequence length, and computing the (approximate) mean of the
corresponding posterior distribution sample 10 times using the chosen parameters. For an
epsilon of 10−6 , the average BHV distance between pairs of approximate means of the same
sample was on the order of 10−5 , which we considered acceptable. Computing the mean
and variance took 10-35 minutes on a 3.5 GHz 6-Core Intel Xeon E5 Processor. The
program and source code are available at
http://comet.lehman.cuny.edu/owen/code/SturmMean.tar.gz.
Simulated Data
Opossum
Diprotodontian
Sloth
Anteater
Armadillo
Hedgehog
Shrew
Mole
Phyllostomid
Freetailedbat
Falsevampirebat
FlyingFox
RousetteFruitbat
Whale
Dolphin
Hippo
Ruminant
Pig
Llama
Horse
Rhino
Tapir
Cat
Caniform
Pangolin
Sciurid
Mouse
Rat
Hystricid
Caviomorph
Rabbit
Pika
FlyingLemur
TreeShrew
Strepsirrhine
Human
Tenrecid
GoldenMole
ShortEaredElephantShrew
LongEaredElephantShrew
Aardvark
Sirenian
Hyrax
Elephant
0.05
Figure 2: Reference mammal tree from Murphy et al. (2001).
We simulated 10 sets of DNA sequences for a variety of lengths (500, 1000, 1500, 2000,
2500, 3000, 3500, 4000 base pairs), using the tree of 44 mammals from Murphy et al.
(2001), which is shown in Figure 2, as the reference tree. The sequences were simulated
with Seq-Gen version 1.3.3 (Rambaut and Grass 1997) using the GTR model with the
following parameter settings: the proportion of invariable sites is 0.18; the equilibrium
frequencies for A, C, G, T are 0.21, 0.31, 0.3, and 0.18, respectively; and the GTR rate
matrix is (1.5, 4.91, 1.34, 0.83, 5.8, 1), where the entries are the transition rate from A to C,
A to G, A to T, C to G, C to T, and G to T. These parameter values were estimated in
Hillis et al. (2005) for this reference tree.
For each set of sequences for each length, we ran RAxML Version 8 (Stamatakis
2014) to compute the Maximum Likelihood (ML) tree and 1000 bootstrap sample trees.
For the RAxML settings, we used the GTR + I + Γ evolutionary model. We conducted a
full analysis (option -f a) using rapid bootstrapping (option -x), which was recommended
in the RAxML user manual when computing a large number of bootstrap replicates, since
every 5th bootstrap tree is used as a starting point for the ML tree search. For each set of
sequences, at each length, we also ran MrBayes 3.2 (Ronquist et al. 2012) to compute the
MAP tree and sample 1000 trees from the posterior distribution. For the MrBayes settings,
we used the GTR+I+Γ evolutionary model, and ran for 5,000,000 iterations, sampling
every 1000 generations. We used the last 1000 sampled trees as the posterior distribution.
For each sample distribution of 1000 trees, we estimated the Fréchet mean and
variance, using SturmMean (Miller et al. 2015) with the settings explained in the previous
subsection. We also computed the majority-rule consensus tree, and three other measures
of variation for each sample: the number of different tree topologies in the sample, the
number of different splits in the sample, and the sum of the squared BHV distance between
each pair of trees in the sample. We compared the mean tree to the majority-rule
consensus, ML, and MAP trees as follows. For each sequence length, repetition, and
distribution type (bootstrap or posterior) there was one mean tree, one majority-rule
consensus tree, and either an ML or MAP tree, depending on whether the sample was from
the bootstrap or posterior distribution. Adding the reference tree to this set, we computed
the RF distance between each pair of trees. We also computed the BHV distance if both
trees in the pair had meaningful branch lengths. The mean trees, ML trees, and reference
tree all had meaningful branch lengths, while the majority-rule consensus tree and MAP
trees did not.
We also compared the lengths of the sampled
and their centres. Here we define
pP trees
2
|e|
.
the length of a tree T with edge set E to be
T This is the BHV distance of tree
e∈E
T to the BHV treespace origin. To investigate stickiness of the mean, we compare the
length of the mean tree of a distribution with the average length of trees in that
distribution, to see if the mean is much shorter than its component trees. We also compare
the lengths of the mean and the average lengths of trees between the posterior and
bootstrap distributions generated from the same sequence set. In all cases, we use the
Wilcoxon ranked signed test (Wilcoxon 1945) to test the hypothesis that the two
distribution of lengths are the same.
Finally, we conducted a mean hypothesis test on the bootstrap and posterior
samples for the first repetition of 4000 base pairs. Recall that a mean hypothesis test is a
type of two-sample test that tests if the means of two samples are the same. Rejecting this
hypothesis implies that the samples are from different distributions. We computed the BHV
distance between the means of these bootstrap and posterior samples, and compared it to
the BHV distance between the means of a random partition of the two samples. Since a
full permutation test is not feasible (since each sample contains 1000 trees), we estimated it
using 500 randomly chosen permutations. Note we are following the method for performing
two-sample hypothesis testing on trees that was suggested in Feragen et al. (2013).
Results
To visualize what the trees from one repetition of the experiment look like, for the
first repetition of the 4000 base pair sequence length experiment, we computed the BHV
distance between 100 trees from each of the bootstrap and posterior samples, the reference
tree, the ML tree and the two mean trees. We use classic Multi-dimensional Scaling (MDS)
(Kruskal 1964) to reduce this to two dimensions (Fig. 3). The two means and the ML tree
are in the middle of their respective samples. The two samples are separated in space and
the reference tree is closer to the posterior sample, but not near its centre. The reference
tree, two means, ML tree, and 62 of the trees from the two samples have the same
topology. The other common topologies appear in both the bootstrap and posterior
samples. This suggests that the difference between the two clusters in Figure 3 is primarily
due to branch lengths, which the BHV distance takes into account, instead of topology.
0.03
0.02
0.00
●
−0.03
−0.02
−0.01
PC 2
0.01
●
● reference tree
ML
bootstrap tree
posterior tree
bootstrap mean ● posterior mean
−0.05
0.00
0.05
PC 1
Figure 3: A sample of trees from the first repetition of the 4000 base pair sequence length
embedded in 2 dimensions using classic Multi-dimensional Scaling. The trees from the
bootstrap and posterior samples form two clusters, most likely due to differences in branch
lengths instead of topology.
Comparison of the mean tree with other measures of centre
First we compare the mean trees to the reference tree and the other summary trees,
showing that they are close. The four plots in Figure 4 use the RF and BHV distances to
compare the Fréchet mean of each bootstrap and posterior sample to its corresponding
majority-rule consensus tree, the reference tree, and either ML or MAP tree, as
appropriate. In all cases, the distance between the Fréchet mean and the other trees
decreases as the sequence length increases. The mean tree is closer to the ML or MAP and
majority-rule consensus tree than to the reference tree, with those three trees almost
always having the same topology for sequences of length 3000 base pairs and longer.
Next we consider the relation of the reference tree to the reconstructed trees. The
four plots in Figure 5 use the RF and BHV distance to compare the reference tree to the
mean tree, majority-rule consensus tree, and ML or MAP tree of each sample. These plots
combined with the previous Figure 4 show that the reconstructed trees are closer to each
other than to the reference tree. As expected, all trees become closer to the reference tree
as the sequence length increases, since this gives more information about the reference tree,
improving its reconstruction. Interestingly, under the BHV distance, the mean tree of the
bootstrap distribution is on average closer to the reference tree than the ML tree
Variance
We compare the different measures of variance in a set a trees, namely the number of
different topologies, the number of different splits, the Fréchet variance, and the sum of the
squared BHV distance between all pairs of input trees, in Figure 6. As the sequence length
increases, there is more information about the underlying tree, and so we expect RAxML
and MrBayes to both do a better job at inferring this tree and be more certain about it.
This should be reflected by a decrease in variance in the bootstrap and posterior
distributions, and thus samples, as the sequence length increases, which is what we see in
Figure 6. For all measures, the posterior samples have a lower variance than the bootstrap
samples. This matches previous observations that the posterior probabilities are higher
than bootstrap frequencies for well-supported clades (Erixon et al. 2003; Douady et al.
2003; Huelsenbeck and Rannala 2004), since trees in a lower variance sample are not as
spread out, and thus have fewer different splits, than trees in a higher variance sample. In
general, we find that there is more uncertainty in the variance estimate for the other
metrics than the Fréchet variance.
The measurements of the sum of the squared BHV distance between all pairs of
sample trees and the Fréchet variance are almost identical. This is expected as these two
measures are identical in Euclidean space, which is why the sum of squared BHV distance
between all pairs was first suggested as a measure of variance. For large input sets, it is
faster to compute the Fréchet variance than the BHV distance between all pairs. See the
Discussion, below, for more details.
Bootstrap Distribution
Posterior Distribution
25
Reference
ML
Majority−Rules Consensus
20
Reference
MAP
Majority−Rules Consensus
RF Distance to Mean Tree
RF Distance to Mean Tree
20
15
●
10
●
●
5
●
15
10
●
5
●
●
●
●
●
●
●
●
●
●
●
●
0
0
500
1000
1500
2000
2500
3000
3500
4000
●
500
1000
Sequence Length
1500
2000
2500
3000
3500
4000
Sequence Length
Bootstrap Distribution
Posterior distribution
●
●
Refence
ML
Reference
BHV Distance to Mean Tree
BHV Distance to Mean Tree
0.2
●
●
●
●
●
0.1
●
●
0.12
0.08
0.04
●
●
●
●
0.0
500
1000
1500
2000
2500
Sequence Length
3000
3500
4000
500
1000
1500
2000
2500
3000
3500
4000
Sequence Length
Figure 4: For each sample, the RF and BHV distances were calculated from its mean tree
to its majority-rule tree, ML or MAP tree, and the reference tree. The distribution of
these distances is shown using box plots. The mean tree approaches the ML or MAP, and
consensus trees as the sequence length increases, with all three trees usually sharing the same
topology by sequence lengths of 3000 base pairs.
Bootstrap Distribution
●
Posterior Distribution
25
Mean
ML
Majority−Rules Consensus
Mean
MAP
Majority−Rules Consensus
RF Distance to Reference Tree
RF Distance to Reference Tree
20
20
●
10
●
●
●
●
15
10
●
●
●
●
●
5
●
●
●
●
●
●
●
●
0
0
500
1000
1500
2000
2500
3000
3500
●
4000
500
1000
Sequence Length
1500
2000
2500
3000
3500
4000
Sequence Length
Bootstrap Distribution
●
Mean
ML
●
BHV Distance to Reference Tree
0.25
0.20
●
●
●
●
0.15
●
●
●
●
●
500
1000
1500
2000
2500
3000
3500
4000
Sequence Length
Figure 5: For each sample, the RF and BHV distances were calculated from the reference
tree to the sample’s mean tree, majority-rule tree, and ML or MAP tree. The distribution
of these distances is shown using box plots. All three reconstructed trees approach the
reference tree as the sequence length increases, with the mean tree being slightly closer on
average than the ML tree under the BHV distance measure. The BHV distance between the
reference and mean trees for the posterior distribution is shown in the lower-right graph in
Figure 4.
1000
Bootstrap Samples
Posterior Samples
Bootstrap Samples
Posterior Samples
800
●
750
400
Number of different splits
Number of different topologies
600
●
●
●
●
500
●
250
200
●
●
●
●
●
●
●
0
500
1000
1500
2000
2500
3000
3500
4000
500
1000
Sequence Length
15000
2000
2500
3000
3500
4000
Sequence Length
●
Bootstrap Samples
Posterior Samples
0.015
●
Bootstrap Samples
Posterior Samples
●
0.010
10000
Frechet variance
Sum of squared BHV distance between all pairs
●
1500
5000
●
0.005
●
●
●
●
500
1000
1500
2000
2500
3000
Sequence Length
3500
4000
●
500
1000
1500
2000
2500
3000
3500
4000
Sequence Length
Figure 6: The variance of the samples of the bootstrap and posterior distributions are
measured using the number of different topologies, the number of different splits, the sum of
the squared BHV distance between all pairs, and the Fréchet variance. For all measures, the
posterior samples have lower variance on average than the bootstrap samples. The sum of
the squared BHV distance between all pairs and the Fréchet variance are very similar. Note
that the y axis for the number of different splits, sum of squares, and Fréchet variance are
in log scale.
Tree length comparison
We also looked at the average tree lengths to determine the effect of mean stickiness. Some
of the mean trees for shorter sequence lengths were not fully resolved due to the stickiness
phenomenon, but all mean trees for the 4000 base pair sequence length were fully resolved
or binary. However, even when mean trees are fully resolved, the stickiness property could
still cause them to be closer to the origin of the sample than expected. In terms of
measurable effects, this would translate into the mean trees being shorter (in terms of total
edge length) on average than the trees from which they were computed. This matches
what we found. For each 80 bootstrap and posterior samples, the length of the mean tree
was strictly less than the average length of the corresponding sample trees. The p-value for
each of these cases was less than 2.2 × 10−16 . While the mean trees had a shorter length
than might be expected, this does noticeably affect the quality of the summary, as on
average, the bootstrap and posterior mean trees were only 99.48% and 99.50%,
respectively, of the average lengths of their respective samples.
We also compared the lengths of the ML trees of the bootstrap samples with their
corresponding mean trees. In all 80 datasets the length of the ML tree was greater than
the length of the corresponding mean tree. With p-value 0.005861, the length of the ML
tree is greater than the average length of the trees in the corresponding bootstrap sample.
Interestingly, the average length of the trees in each posterior sample was always
less than the average length of the trees in the corresponding bootstrap sample generated
from the same set of sequences. This suggests the bootstrap samples are more spread out,
which could also reflect posterior probabilities being higher for well-support clades than
bootstrap probabilities. Unsurprisingly, the mean tree of the posterior distribution sample
always had length less than the mean tree of the corresponding bootstrap sample.
Mean hypothesis test
For the first repetition of the 4000 base pair sequences, we performed a two-sample
hypothesis test to compare the samples of the bootstrap and posterior distributions using
the means. The null hypothesis stated that the two means of bootstrap and posterior
distributions were the same. We tested this hypothesis using an approximate permutation
test. In all cases, the distance between the means of a random partition of the two samples
was strictly less than the distance between the means of the two samples themselves. Thus
we reject the null hypothesis with an estimated p-value of 0.002, with a 95% confidence
interval of [0, 0.006]. This means that we can assume the two distributions do not have the
same mean, and thus are not the same. This was expected due to previous work showing
the bootstrap and posterior probabilities are different (Erixon et al. 2003; Douady et al.
2003; Huelsenbeck and Rannala 2004).
Discussion
Our experiments demonstrate that both the Fréchet mean and variance behave in
the expected way on biological data. We have shown that the mean of samples of
bootstrap and posterior distributions is comparable in accuracy to the ML and MAP trees,
respectively, as well as the majority-rule consensus tree. Indeed the results seem to suggest
that the original reference tree is slightly closer to the mean than to the other summary
trees. While this possible gain may not be enough to warrant the cost of computing the
mean, these results conclusively demonstrate that the mean is a valid summary method for
a sample of trees. We believe the value of the mean comes from its sound mathematical
backing, which enables more sophisticated statistical tests, like mean hypothesis testing.
From the results of the variance experiments, it is clear that the Fréchet variance is
a faster and more precise measure of variance than existing alternatives. The variability in
the number of topologies in the tree sets with the same sequence length is very high, in
comparison to the other variance measures. Although the variability decreases when
measuring the number of splits in the tree set, it is still higher than that of the sum of
squared pairwise distances and the Fréchet variance for comparable sequence lengths.
The sum of squared pairwise distances and the Fréchet variance have very similar
profiles, justifying the use of the sum of squared pairwise distances as a measure of
variance in the literature. However for large sample sizes, the Fréchet variance is faster to
compute. To compute the sum of squared geodesic distances for r trees, one must compute
r(r − 1)/2 geodesic distances. In contrast, computing the Fréchet variance involves
calculating the geodesic distance once per iteration of the algorithm. The number of
iterations required depends on the desired precision of the mean. However, we have often
obtained good results with 10,000-15,000 iterations, suggesting that when we are
computing variances of more than approximately 200 trees, the Fréchet variance
calculation will be faster than the sum-of-squared-distance calculation.
Our experiments on the tree lengths show that while the mean trees are shorter
than the average length of the sample trees, it is by less than 1% on average. This
demonstrates that stickiness leads to shorter mean trees, but that the difference in lengths
is not likely to be significant and can be ignored.
One might think that we could use the mean to compute the species tree for a set of
gene trees (Miller et al. 2015, Example 5.5). Even if we just restrict ourselves to the
coalescent model for explaining gene tree diversity, this is still problematic in two regards.
Under the coalescent model, the pendant edges of the gene trees must be at least as long as
than the pendant edges of the species tree, but will usually be longer. Since each pendant
edge of the mean tree is computed by averaging the length of that edge in all input trees,
the pendant edges of the mean of the gene trees will be at least as long, and usually longer,
than the pendant edges of the true species tree. While it is possible that the topology of
the mean tree might still match that of the species tree, as conjectured in (Miller et al.
2015, Example 5.5), our preliminary experiments suggest that stickiness greatly limits the
amount of information in the mean tree, compared to other methods. In preliminary
experiments, the mean tree of a set of simulated gene trees became essentially degenerate
while the gene trees were still far away from the anomalous zone (Degnan and Rosenberg
2006). That is, the most common gene tree topology was still the same as the species tree
topology. By essentially degenerate, we mean that the approximated mean tree was within
a small distance of some degenerate tree. It is possible that an exact algorithm for
computing the mean would give a non-degenerate mean tree topology even in these cases,
but with an iterative algorithm, we can not be certain the mean is not degenerate, nor even
that it would be informative.
Conclusion
We have shown that the Fréchet mean and variance behave in an expected way on
biological data. We have further shown that the mean can be used in mean hypothesis
testing to compare two samples of trees, and that the variance is a stable and reliable
measure of the amount of variability in a sample of trees. This validation opens the door to
new applications of these quantities.
One possible application of the variance is to determine when MrBayes or other
Markov Chain Monte Carlo (MCMC) algorithms converge. Preliminary experiments on the
MrBayes runs for the data in this paper show that comparing the variance of sliding
windows of the sampled trees can identify the burn-in, but that the variance of the trees in
the sliding window remains roughly the same after this period. However, there may be
other datasets where the variance will continue to decrease after the burn-in, indicating a
lack of convergence. The iterative algorithm for computing the mean is easily adapted into
an online algorithm: at each iteration, instead of choosing a tree at random from the input
set of trees, use the next tree generated by the MCMC chain. Unfortunately, it is not clear
how to take advantage of this in computing the variance, which requires computing the
BHV distance from the current mean approximation to all sample trees seen so far.
However, algorithms for computing geodesics dynamically (Skwerer and Provan 2015)
might help.
*
References
Bačák, M. 2014. Computing medians and means in Hadamard spaces. SIAM Journal on
Optimization 24:1542–1566.
Barden, D. and H. Le. 2017. The logarithm map, its limits and frechet means in orthant spaces.
arXiv preprint arXiv:1703.07081 .
Barthélemy, J.-P. and F. McMorris. 1986. The median procedure for n-trees. Journal of
Classification 3:329–334.
Benner, P., M. Bačák, and P.-Y. Bourguignon. 2014. Point estimates in phylogenetic
reconstructions. Bioinformatics 30:i534–i540.
Billera, L., S. Holmes, and K. Vogtmann. 2001. Geometry of the space of phylogenetic trees.
Advances in Applied Mathematics 27:733–767.
Bouckaert, R. R. 2010. DensiTree: making sense of sets of phylogenetic trees. Bioinformatics
26:1372–1373.
Bridson, M. and A. Haefliger. 1999. Metric Spaces of Non-positive Curvature. Springer-Verlag.
Chakerian, J. and S. Holmes. 2012. Computational tools for evaluating phylogenetic and
hierarchical clustering trees. Journal of Computational and Graphical Statistics 21:581–599.
Degnan, J. H. and N. A. Rosenberg. 2006. Discordance of species trees with their most likely gene
trees. PLoS Genet 2:e68.
Douady, C. J., F. Delsuc, Y. Boucher, W. F. Doolittle, and E. J. Douzery. 2003. Comparison of
Bayesian and maximum likelihood bootstrap measures of phylogenetic reliability. Molecular
Biology and Evolution 20:248–254.
Erixon, P., B. Svennblad, T. Britton, and B. Oxelman. 2003. Reliability of Bayesian posterior
probabilities and bootstrap frequencies in phylogenetics. Systematic Biology 52:665–673.
Felsenstein, J. 1981. Evolutionary trees from DNA sequences: a maximum likelihood approach.
Journal of molecular evolution 17:368–376.
Felsenstein, J. 1985. Confidence limits on phylogenies: an approach using the bootstrap.
Evolution Pages 783–791.
Feragen, A., M. Owen, J. Petersen, M. M. Wille, L. H. Thomsen, A. Dirksen, and M. de Bruijne.
2013. Tree-space statistics and approximations for large-scale analysis of anatomical trees.
Pages 74–85 in IPMI vol. 2013.
Hillis, D., T. Heath, and K. St. John. 2005. Analysis and visualization of tree space. Systematic
Biology 54:471–482.
Hotz, T., S. Huckemann, H. Le, J. S. Marron, J. C. Mattingly, E. Miller, J. Nolen, M. Owen,
V. Patrangenaru, S. Skwerer, et al. 2013. Sticky central limit theorems on open books. The
Annals of Applied Probability 23:2238–2258.
Huelsenbeck, J. P. and B. Rannala. 2004. Frequentist properties of Bayesian posterior
probabilities of phylogenetic trees under simple and complex substitution models. Systematic
biology 53:904–913.
Huson, D. H. and D. Bryant. 2006. Application of phylogenetic networks in evolutionary studies.
Molecular Biology and Evolution 23:254–267.
Kruskal, J. B. 1964. Multidimensional scaling by optimizing goodness of fit to a nonmetric
hypothesis. Psychometrika 29:1–27.
Margush, T. and F. R. McMorris. 1981. Consensus n-trees. Bulletin of Mathematical Biology
43:239–244.
Miller, E., M. Owen, and J. S. Provan. 2015. Polyhedral computational geometry for averaging
metric phylogenetic trees. Advances in Applied Mathematics 68:51–91.
Murphy, W. J., E. Eizirik, S. J. O’Brien, O. Madsen, M. Scally, C. J. Douady, E. Teeling, O. A.
Ryder, M. J. Stanhope, W. W. de Jong, et al. 2001. Resolution of the early placental mammal
radiation using Bayesian phylogenetics. Science 294:2348–2351.
Nye, T. M. 2011. Principal components analysis in the space of phylogenetic trees. The Annals of
Statistics Pages 2716–2739.
Nye, T. M. 2014. An algorithm for constructing principal geodesics in phylogenetic treespace.
IEEE/ACM Transactions on Computational Biology and Bioinformatics 11:304–315.
Nye, T. M., X. Tang, G. Weyenberg, and R. Yoshida. 2016. Principal component analysis and the
locus of the Fréchet mean in the space of phylogenetic trees. arXiv preprint arXiv:1609.03045 .
Owen, M. and J. S. Provan. 2011. A fast algorithm for computing geodesic distances in tree space.
IEEE/ACM Transactions on Computational Biology and Bioinformatics (TCBB) 8:2–13.
Pattengale, N. D. 2005. Tools for Phylogenetic Postprocessing. Ph.D. thesis The University of
New Mexico.
Ponciano, J. M., J. G. Burleigh, E. L. Braun, and M. L. Taper. 2012. Assessing parameter
identifiability in phylogenetic models using data cloning. Systematic biology Page sys055.
Rambaut, A. and N. C. Grass. 1997. Seq-Gen: an application for the Monte Carlo simulation of
DNA sequence evolution along phylogenetic trees. Computer applications in the biosciences:
CABIOS 13:235–238.
Robinson, D. F. and L. R. Foulds. 1979. Comparison of weighted labelled trees. Pages 119–126 in
Combinatorial mathematics VI. Springer.
Robinson, D. F. and L. R. Foulds. 1981. Comparison of phylogenetic trees. Mathematical
Biosciences 53:131–147.
Ronquist, F., M. Teslenko, P. van der Mark, D. L. Ayres, A. Darling, S. Höhna, B. Larget, L. Liu,
M. A. Suchard, and J. P. Huelsenbeck. 2012. MrBayes 3.2: efficient Bayesian phylogenetic
inference and model choice across a large model space. Systematic biology 61:539–542.
Skwerer, S. and S. Provan. 2015. Dynamic geodesics in treespace via parametric maximum flow.
arXiv preprint arXiv:1512.03115 .
St. John, K. 2017. Review paper: The shape of phylogenetic treespace. Systematic Biology 66:e83.
Stamatakis, A. 2014. RAxML version 8: a tool for phylogenetic analysis and post-analysis of large
phylogenies. Bioinformatics 30:1312–1313.
Sturm, K.-T. 2003. Probability measures on metric spaces of nonpositive. Heat Kernels and
Analysis on Manifolds, Graphs, and Metric Spaces: Lecture Notes from a Quarter Program on
Heat Kernels, Random Walks, and Analysis on Manifolds and Graphs: April 16-July 13, 2002,
Emile Borel Centre of the Henri Poincaré Institute, Paris, France 338:357.
Wilcoxon, F. 1945. Individual comparisons by ranking methods. Biometrics Bulletin 1:80–83.
Willis, A. 2016. Confidence sets for phylogenetic trees. arXiv preprint arXiv:1607.08288 .
| 10math.ST
|
arXiv:1705.07024v1 [cs.CE] 19 May 2017
Optimal prevention with possibilistic and mixed
background risk
Irina Georgescu
Academy of Economic Studies
Department of Economic Informatics and Cybernetics
Piaţa Romana No 6 R 70167, Oficiul Postal 22, Bucharest, Romania
Email: [email protected]
and
Ana Marı́a Lucia Casademunt
Universidad Loyola–Andalucı́a
Department of Business Administration, Cordoba, Spain
Email: [email protected]
Abstract
In this paper the effect of posibilistic or mixed background risk on
the level of optimal prevention is studied. In the framework of five purely
possibilistic or mixed models, necessary and sufficient conditions are found
such that the level of optimal saving decreases or increases as a result
of the actions of various types of background risk. This way our results
complete those obtained by Courbage and Rey for some prevention models
with probabilistic background risk.
Keywords: possibilistic background risk, optimal prevention, optimal saving modal operator
1
Introduction
The prevention is the effort made by an agent to reduce the action of an unwanted event (e. g. a financial loss). The effort can be the amount of money
invested by the agent in prevention activities and then he is interested to determine the optimal level of this effort.
According to [5], the prevention consists in reducing the probability of a
loss (self–protection), the size of loss (self–insurance) or both of them (self–
insurance–cum–protection). For simplicity, in this paper by prevention one
understands self–protection (this terminology is used in [4]).
The first contribution to the study of optimal prevention is the paper [13] by
Elrich and Becker which investigates the interconnection between the market
insurance, self-insurance and self–protection. Then, the subject generated a
1
real literature focused especially on the way consumers’ preferences influence
the optimal level of prevention (see [5]).
A common topic deals with the relationship between risk aversion and optimal prevention. For instance, Dionne and Eeckhoudt emphasize in [7] the
fact that the increase of risk aversion does not lead to an increase in optimal
investment in prevention, its effect being ambiguous. A second topic is the relationship between the notion of prudence introduced by Kimball in [18] and that
of optimal prevention (see [11], [5]). Paper [11] studies the impact of prudence
on optimal prevention in a one–period model, while [5] analyzes the obtaining
of the optimal prevention in a two-period framework. In the latter case, the
prevention effort is done in the first period, and its effect on the likelihood of
appearance of loss occurs during the second period.
Prevention models with background risk appear for the first time in the paper
by Courbage and Ray [4] and Eeckhoudt et al. [12]. Background risk may appear
in one or both periods. With respect to this, the changes in optimal prevention
level occur in relation to a benchmark model (in which the background risk does
not exist).
In the mentioned papers we deal with probabilistic models, in which the
background risk is represented by a random variable. This paper aims to study
optimal prevention in non–probabilistic models, in which we have a possibilistic
or mixed background risk (see [3], [15], [16]). In the former case, the two–period
background risk will be a fuzzy number, and in the latter case, in one period,
background risk will be a random variable, and in the other period, it will be a
fuzzy number.
Section 2 contains the definition of possibilistic expected utility and a few
of its properties which will be used in this paper.
In Section 3 the benchmark prevention model is presented following [4], [5].
This will be the starting point in the background risk model construction from
the next sections.
In Section 4 eight background risk prevention models are described. These
are two–period constructions, and the risk from the two periods may be either
a random variable or a fuzzy number. The first three models are probabilistic,
and their theory has been already developed in [4]. The other five models are
purely possibilistic or mixed. Their construction follows a similar line to the
first three: the total utility functions are defined, then first order conditions
verified by optimal prevention are deduced.
Section 5 focuses on a model in which various types of background risk of
the five models influence the level of optimal prevention. For each of the five
models, the level of optimal prevention is compared to that of the benchmark
model. Necessary and sufficient conditions for the increase or decrease in optimal
prevention with respect to the existing level in case of the benchmark model are
proved.
In Section 6 the background risk effect on optimal saving is studied when
to the already existing risk in one period a (possibilistic or probabilistic) risk is
added in the other period.
2
2
Possibilistic expected utility
In this section we recall the definition of possibilistic expected utility and a few
of its basic properties (cf. [1], [2], [8], [9], [14]).
If X is a random variable defined on a probability field (Ω, K, P ) with Ω ⊆ R
then we will denote by M (X) its expected value. For any continuous utility
function u : R → R, the expected value M (u(X)) of the random variable u(X)
is the probabilistic expected utility associated with X and u.
To develop a possibilistic risk theory we need a similar notion of possibilistic
expected utility. Its definition assumes a framework composed of the following
components:
• a utility function u : R → R representing an agent;
• a fuzzy number A modelling the possibilistic risk;
• a weighting function, i.e. a non–negative andRmonotone increasing function
1
f : [0, 1] → R satisfying the normality condition 0 f (γ)dγ = 1.
Assume that the level sets of the fuzzy number A have the form [A]γ =
[a1 (γ), a2 (γ)], for any γ ∈ [0, 1]. Then the possibilistic expected utility Ef (u(A))
is defined by
R1
(2.1) Ef (u(A)) = 12 0 [u(a1 (γ)) + u(a2 (γ))]f (γ)dγ
If u is the identity function of R then Ef (u(A)) is exactly the possibilistic
expected value [1], [14].
R1
(2.2) Ef (A) = 21 0 [a1 (γ) + a2 (γ)]f (γ)dγ.
The following two propositions emphasize two important properties of possibilistic expected utility: the first one is a linearity property, the second one
is a Jensen type inequality (used in the following in the analysis of optimal
prevention variations with respect to various types of background risk).
Proposition 2.1 [15] Let g : R → R, h : R → R two continuous utility
functions and a, b ∈ R. If u = ag+bh, then Ef (u(A)) = aEf (g(A))+bEf (h(A)).
Proposition 2.2 [15], [3] If the utility function u is convex, then u(Ef (A)) ≤
Ef (u(A)).
3
The benchmark prevention model
In this section we present, according to [4], [5], the entities which compose
the benchmark prevention model, the form of its total utility function and the
first–order condition with which the level of optimal prevention is selected.
The benchmark prevention model [4], [5] has two periods 1 and 2: the prevention effort is conducted in period 1, and the effect of this effort on the
probability of the occurrence of a loss in wealth) is manifested in period 2.
The model is defined by the following entities:
• w1 , w2 are the safe wealth in periods 1 and 2;
• u and v are utility functions for period 1 and 2, respectively;
• e is the level of prevention effort (in period 1);
3
• l is the loss level;
• p(e) is the probability that a loss l occurs at the prevention level e.
In [4], [5] this model is studied under the hypotheses:
• the utility functions u and v are of class C 2 , u′ > 0, v ′ > 0, u′′ < 0, v ′′ < 0;
• p′ (e) < 0, p′′ (e) > 0, for any level e;
• l > 0.
According to these initial data, in the second period of the consumption
activity we will have an endogenous risk ǫ̃e (parameterized by the prevention
level e), whose outcomes are −l and 0:
−l
0
ǫ̃e =
p(e) 1 − p(e)
Then the expected utility corresponding to the second period will be M (w2 +
ǫ̃e ) = p(e)v(w2 − l) + (1 − p(e))v(w2 ), and the total utility function will be:
(3.1) V (e) = u(w1 − e) + p(e)v(w2 − l) + (1 − p(e))v(w2 ).
The agents wish to choose a level e of the effort maximizing the total utility
V (e). This comes to finding the solution e∗ of the optimization problem:
max V (e).
e
Since V (e) is strictly concave, the optimal solution e∗ is given by V ′ (e∗ ) = 0,
which leads us to the first–order condition:
(3.2) u′ (w1 − e∗ ) = p′ (e∗ )[v(w2 − l) − v(w2 )].
4
Prevention models with background risk
The prevention models from this section are built starting from the benchmark
model from Section 3. Their definition is founded on the following assumptions:
Assumption 1 Background risk may appear both in period 1 and in period
2.
Assumption 2 The background risk from each period can be modeled both
probabilistically (by a random variable), and possibilistically (by a fuzzy number).
Denoting by 0 the lack of background risk in one of the two periods, the eight
models which can be defined on the basis of the two assumptions are specified
in Table 1:
1
2
3
4
5
6
7
8
Period 1
probabilistic background risk
0
probabilistic background risk
possibilistic background risk
0
possibilistic background risk
probabilistic background risk
possibilistic background risk
4
Period 2
0
probabilistic background risk
probabilistic background risk
0
possibilistic background risk
possibilistic background risk
possibilistic background risk
probabilistic background risk
In cases 1−3 we deal with probabilistic models, in cases 4−6 with possibilistic
models, and in the last two cases with mixed models.
Models 1 − 3 have been studied in [4]. Next we present briefly the construction of these models.
Model 1 [4] In period 1 appears a probabilistic risk modeled by the random
variable X. The total utility function has the following form:
V1 (e) = M (u(w1 − e + X)) + p(e)v(w2 − l) + (1 − p(e))v(w2 ).
The first derivative of V1 (e) will be:
V1′ (e) = −M (u′ (w1 − e + X)) + p′ (e)[v(w2 − l) − v(w2 )]
and the optimal value e∗1 will be given by the first–order condition:
(4.1) M (u′ (w1 − e∗1 + X)) = p′ (e∗1 )[v(w2 − l) − v(w2 )].
Model 2 [4] In period 2 appears a probabilistic risk modeled by the random
variable Y . The overall utility function will be
V2 (e) = u(w1 − e) + p(e)M (v(w2 − l + Y )) + (1 − p(e))M (v(w2 + Y ))
The first derivative of V2 (e) will be:
V2′ (e) = −u′ (w1 − e) + p′ (e)[M (v(w2 − l + Y )) − M (v(w2 + Y ))].
and the optimal value e∗2 of e will be given by the condition:
(4.2) u′ (w1 − e∗2 ) = p′ (e∗2 )[M (v(w2 − l + Y )) − M (v(w2 + Y ))].
Model 3 [4] Background risk appears in both periods and is modeled by the
random variables X, and Y , respectively. The total utility function has the
form:
V3 (e) = M (u(w1 − e + X)) + p(e)[M (v(w2 − l + Y )) − M (v(w2 + Y ))].
Then the first derivative of V3 (e) will be:
V3′ (e) = −M (u′ (w1 − e + X)) + p′ (e)[M (v(w2 − l + Y )) − M (v(w2 + Y ))]
and the optimal value e∗3 of e will be given by the condition:
(4.3) M (u′ (w1 − e∗3 + X)) = p′ (e∗3 )[M (v(w2 − l + Y )) − M (v(w2 + Y ))]
Next we will present the construction of models 4 − 8 from Table 1. For the
rest of the section we fix a weighting function f : [0, 1] → R.
Model 4 Assume that the background risk is described by a fuzzy number
A whose level sets are [A]γ = [a1 (γ), a2 (γ)] for any γ ∈ [0, 1]. The total utility
function of the model will be:
V4 (e) = Ef (u(w1 − e + A)) + p(e)v(w2 − l) + (1 − p(e))v(w2 )
According to the definition (2.1) of possibilistic expected utility, V4 (e) can
be written: R
1
V4 (e) = 12 0 [u(w1 − e + a1 (γ)) + u(w1 − e + a2 (γ))]f (γ)dγ + p(e)v(w2 − l) +
(1 − p(e))v(w2 ).
By derivation,
R 1 one obtains
V4′ (e) = − 12 0 [u′ (w1 − e + a1 (γ)) + u′ (w1 − e + a2 (γ))]f (γ)dγ + p′ (e)[v(w2 −
l) − v(w2 )],
from where, by (2.1), the following form of the derivative of V4 (e) follows:
(4.4) V4′ (e) = −Ef (u′ (w1 − e + A)) + p′ (e)[v(w2 − l) − v(w2 )].
By deriving one more time, one reaches
V4′′ (e) = Ef (u′′ (w1 − e + A)) + p′′ (e)[v(w2 − l) − v(w2 )].
From the hypotheses imposed on the utility functions u and v on the probability p(e) it follows V4′′ (e) < 0 for any level e of the prevention effort, thus V4 is
5
a concave function. Then the value e∗4 of e maximizing the total utility function
V4 verifies the first–order condition V4′ (e∗4 ) = 0, which, by (4.4), is written as:
(4.5) Ef (u′ (w1 − e∗4 + A)) = p′ (e∗4 )[v(w2 − l) − v(w2 )].
Model 5 Assume that the background risk in period 2 is a fuzzy number B
whose level sets are [B]γ = [b1 (γ), b2 (γ)]. The total utility function of the model
will be:
V5 (e) = u(w1 − e) + p(e)Ef (v(w2 − l + B)) + (1 − p(e))Ef (v(w2 + B))
R1
= u(w1 − e) + 21 p(e) 0 [v(w2 − l + b1 (γ)) + v(w2 − l + b2 (γ))]f (γ)dγ + 12 (1 −
R1
p(e)) 0 [v(w2 + b1 (γ)) + v2 (w2 + b2 (γ))]f (γ)dγ.
By derivation one obtains:
(4.6) V5′ (e) = −u′ (w1 − e) + p′ (e)[Ef (v(w2 − l + B)) − Ef (v(w2 + B))]
V5′′ (e) = u′′ (w1 − e) + p′′ (e)[Ef (v(w2 − l + B)) − Ef (v(w2 + B))]
Due to the monotonicity property of the possibilistic expected utility the
inequality Ef (v(w2 − l + B)) − Ef (v(w2 + B)) ≤ 0, therefore V5′′ (e) < 0 for any
e, hence V5 is strictly concave. Then the value e∗5 of e maximizing V5 (e) verifies
the condition V5′ (e∗5 ) = 0, which, by (4.6), can be written:
(4.7) u′ (w1 − e∗5 ) = p′5 (e∗5 )[Ef (v(w2 − l + B)) − Ef (v(w2 + B))].
Model 6 Assume that the background risk in period 1 is the fuzzy number
A, and the background risk from period 2 is the fuzzy number B. The total
utility function of the model will be:
V6 (e) = Ef (u(w1 −e+A))+p(e)Ef (v(w2 −e+B))+(1−p(e))Ef (v(w2 +B)).
By derivation one obtains:
(4.8) V6′ (e) = −Ef (u′ (w1 −e+A))+p′ (e)[Ef (v(w2 −l +B))−Ef (v(w2 +B))]
V6′′ (e) = Ef (u′′ (w1 − e + A)) + p′′ (e)[Ef (v(w2 − l + B)) − Ef (v(w2 + B))]
From the last inequality V6′′ (e) ≤ 0 follows for any e, thus V6 is strictly
concave. Then the value e∗6 of e maximizing V6 (e) verifies the condition V6′ (e∗6 ) =
0, which, by (4.8), can be written:
(4.9) Ef (u′ (w1 − e∗6 + A)) = p′ (e)[Ef (v(w2 − l + B)) − Ef (v(w2 + B))]
Model 7 Assume that the background risk in period 1 is the random variable
X, and the background risk in period 2 is the fuzzy number B. The total utility
function of the model will be:
V7 (e) = M (u(w1 − e + X)) + p(e)Ef (v(w2 − l + B)) + (1 − p(e))Ef (v(w2 + B))
By derivation one obtains
(4.10) V7′ (e) = −M (u′ (w1 −e+X))+p′(e)[Ef (v(w2 −l+B))−Ef (v(w2 +B))]
It follows V ′ (e) < 0 for any level e of the effort level, thus the value e∗7 of e
maximizing V7 (e) is given by V7′ (e∗7 ) = 0. By (4.10), this first–order condition
is written as:
(4.11) M (u′ (w1 − e∗7 + X)) = p′ (e∗7 )[Ef (v(w2 − l + B)) − Ef (v(w2 + B))].
Model 8 Assume that the background risk in period 1 is the fuzzy number
A, and the background risk in period 2 is a random variable Y . The total utility
function of the model will be:
V8 (e) = Ef (u(w1 − e + A)) + p(e)M (v(w2 − l + Y )) + (1 − p(e))M (v(w2 + Y ))
By derivation one obtains
(4.12) V8′ (e) = −Ef (u′ (w1 −e+A))+p′(e)[M (v(w2 −l+Y ))−M (v(w2 +Y ))]
6
One proves immediately that V8 is concave, thus the value e∗8 of e maximizing
V8 (e) verifies V8′ (e∗8 ) = 0, which, by (4.12), can be written:
(4.13) Ef (u′ (w1 − e∗8 + A)) = p′ (e∗8 )[M (v(w2 − l + Y )) − M (v(w2 + Y ))]
Remark 4.1 The above models incorporate both the entities of the benchmark
model, and the features of background risk types defining them. These appear
in the expression of total utility functions and in the first–order condition which
determines the level of optimal prevention.
5
The effect of background risk on optimal prevention
In this section we study the way adding various types of background risk may
produce changes in the level of optimal saving. For this we have to compare the
optimal prevention level s∗ of the benchmark model with the optimal prevention
levels s∗i , i = 1, . . . , 8 of the eight models of the previous section. The models
1 − 3 being already treated in [4], we will handle the models 4 − 8 in Table 1.
For each case, we will be situated in the framework set defined in Section 4.
Case 4 (the comparison of e∗ and e∗4 )
The following result establishes a necessary and sufficient condition such
that, adding a possibilistic background risk A in period 1, the optimal saving
level drops.
Proposition 5.1 The following assertions are equivalent:
(i) e∗4 ≤ e∗ ;
(ii) u′ (w1 − e∗ ) ≤ Ef (u′ (w1 − e∗ + A)).
Proof. From the previous section we know that V4′ (e) < 0 for any e, thus V4′ is a
strictly decreasing function. Then, by (4.4) and (3.2), the following equivalences
hold:
e∗4 ≤ e∗ iff V4′ (e∗ ) ≤ V4′ (e∗4 ) = 0;
iff −Ef (u′ (w1 − e∗ + A)) + p′ (e∗ )[v(w2 − l) − v(w2 )] ≤ 0;
iff −Ef (u′ (w1 − e∗ + A)) + u′ (w1 − e∗ ) ≤ 0;
iff u′ (w1 − e∗ ) ≤ Ef (u′ (w1 − e∗ + A)).
Corollary 5.2 If u′′′ > 0 and Ef (A) = 0 then e∗4 ≤ e∗ .
Proof. By applying Proposition 2.2 to the convex function u′ it follows that
u′ (w1 − e) = u′ (Ef (w1 − e + A)) ≤ Ef (u′ (w1 − e + A)), hence by Proposition
5.1, one obtains e∗4 ≤ e.
Corollary 5.2 shows that if the agent is prudent 1 (in period 1) then the
presence of possibilistic background risk in the first period reduces the optimal
prevention.
1 The
notion of prudence in a possibilistic or mixed framework was studied in [3], [7]
7
Case 5 (the comparison of e∗ and e∗5 )
The following result establishes a necessary and sufficient condition such
that, adding a possibilistic background risk B in period 2, the optimal prevention
level raises.
Proposition 5.3 The following assertions are equivalent:
(i) e∗5 ≥ e∗ ;
(ii) v(w2 − l) − v(w2 ) − Ef (v(w2 − l + B)) + Ef (v(w2 + B)) ≥ 0.
Proof. The function V5′ is strictly decreasing, thus by applying (4.6) and (3.2),
the following equivalences hold:
e∗5 ≥ e∗ iff V5′ (e∗ ) ≤ V5′ (e∗5 ) = 0;
iff −u′ (w1 − e∗ ) + p′ (e∗ )[Ef (v(w2 − l + B)) − Ef (v(w2 + B))] ≤ 0;
iff −p′ (e∗ )[v(w2 −l)−v(w2 )−Ef (v(w2 −l+B))+Ef (v(w2 +B))] ≥ 0;
iff v(w2 − l) − v(w2 ) − Ef (v(w2 − l + B)) + Ef (v(w2 + B)) ≥ 0.
Lemma 5.4 If v ′′′ > 0 and Ef (B) = 0 then v(w2 − l) − w(w2 ) ≥ Ef (v(w2 −
l + B)) − Ef (v(w2 + B)).
Proof. We consider the function h(x) = v(x) − v(x − l), for any x ∈ R. From
the assumption v ′′′ > 0 it follows that v ′′ is strictly increasing, therefore h′′ (x) =
v ′′ (x) − v ′′ (x − l) > 0, for any x. It follows that the function h is convex, thus,
by applying Proposition 2.2, h(w2 ) = h(Ef (w2 + B)) ≤ Ef (h(w2 + B)). Then,
by noticing that h(w2 ) = v(w2 ) − v(w2 − l) and Ef (h(w2 + B)) = Ef (v(w2 +
B)) − Ef (v(w2 − l + B)) one obtains the desired inequality.
Corollary 5.5 If v ′′′ > 0 and Ef (B) = 0 then e∗5 ≥ e∗ .
Proof. Proposition 5.3 and Lemma 5.4 are applied.
Case 6 (the comparison of e∗ and e∗6 )
The following result establishes a necessary and sufficient condition such
that, adding a possibilistic background risk A in period 1, and a possibilistic
background risk B in period 2, the optimal prevention level raises.
Proposition 5.6 The following assertions are equivalent:
(i) e∗ ≤ e∗6 ;
E (u′ (w −e∗ +A))
E (v(w2 −l+B))−Ef (v(w2 +B))
(ii) f u′ (w11 −e∗ )
≤ f
.
v(w2 −l)−v(w2 )
Proof. The function V6′ is strictly decreasing, thus, by applying (4.8) and (3.2),
the following assertions are equivalent:
• e∗ ≤ e∗6 ;
• V6′ (e∗ ) ≥ V6′ (e∗6 ) = 0;
• −Ef (u′ (w1 − e∗ + A)) + p′ (e∗ )[Ef (v(w2 − l + B)) − Ef (v(w2 + B))] ≥ 0;
8
E (v(w −l+B))−E (v(w +B))
2
2
f
≥ 0.
• −Ef (u′ (w1 − e∗ + A)) + u′ (w1 − e∗ ) f
v(w2 −l)−v(w2 )
By taking into account that v(w2 − l) − v(w2 ) < 0 it follows immediately the
equivalence of the assertions (i) and (ii).
Remark 5.7 Assume that u′′′ > 0, v ′′′ > 0, Ef (A) = 0 and Ef (B) = 0.
Then by a reasoning analogous to the one of Corollary 5.2, from u′′′ > 0 and
Ef (A) = 0 it follows:
Ef (u′ (w1 −e∗ +A))
≥ 1.
u′ (w1 −e∗ )
By applying Lemma 5.4, from v ′′′ > 0 and Ef (B) = 0 it follows
Ef (v(w2 −l+B))−Ef (v(w2 +B))
≥ 1.
v(w2 −l)−v(w2 )
From these two inequalities and Proposition 5.6, we conclude that for a prudent agent in both periods, the effect of possibilistic background risks A and B
on optimal prevention is ambiguous: we cannot establish a positioning of the
levels e∗ and e∗6 one to each other.
Case 7 (the comparison of e∗ and e∗7 )
The following result establishes a necessary and sufficient condition, that
adding a probabilistic risk X in period 1 and a possibilistic risk B in period 2,
the optimal prevention level raises.
Proposition 5.8 The following assertions are equivalent:
(i) e∗ ≤ e∗7 ;
′
∗
E (v(w2 −l+B))−Ef (v(w2 +B))
1 −e +X))
(ii) M(uu′(w
≤ f
.
(w1 −e∗ )
v(w2 −l)−v(w2 )
Proof. Analogous to the proof of Proposition 5.6 by using the fact that V7′ is
strictly decreasing and applying (4.10) and (3.2).
Remark 5.9 Assume that u′′′ > 0, v ′′′ > 0, M (X) = 0 and Ef (B) = 0. From
u′′′ > 0 and M (X) = 0 it follows according to Jensen’s inequality, the following
inequality:
M(u′ (w1 −e∗ +X))
≥ 1.
u′ (w1 −e∗ )
By Lemma 5.4, conditions v ′′′ > 0, Ef (B) = 0 imply the inequality:
Ef (v(w2 −l+B))−Ef (v(w2 +B))
≥ 1.
v(w2 −l)−v(w2 )
These two inequalities show that the background risk effect described by the
mixed vector (X, B) on the optimal prevention is ambiguous.
Case 8 (the comparison of e∗ and e∗8 )
The following proposition establishes a necessary and sufficient condition
that adding a possibilistic background risk A in period 1 and a probabilistic
risk Y in period 2, the optimal prevention level raises.
Proposition 5.10 The following assertions are equivalent:
(i) e∗ ≤ e∗8 ;
E (u′ (w −e∗ +A))
−l+Y ))−M(v(w2 +Y ))
(ii) f u′ (w11 −e∗ )
≤ M(v(w2v(w
.
2 −l)−v(w2 )
9
Proof. (4.12) and (3.2) are applied.
Remark 5.11 Assume that u′′′ > 0, v ′′′ > 0, Ef (A) = 0, M (Y ) = 0. Then, by
Proposition 2.2 and [4], the following inequalities hold:
Ef (u′ (w1 −e∗ +A))
−l+Y ))−M(v(w2 +Y ))
≥ 1, M(v(w2v(w
≥ 1,
u′ (w1 −e∗ )
2 −l)−v(w2 )
thus the effect of mixed background risk (A, Y ) on the level of optimal saving
is ambiguous.
6
Other cases of comparing optimal prevention
levels
In the previous section we studied the changes in optimal prevention when
we pass from the benchmark model to the models where the background risk
appears in one of the periods or in both periods. The problem of changes in
optimal saving can be posed in other cases too: one passes from a situation in
which the background risk is present in a single period (models 1, 2, 4, 5) to one
of the situations when the background risk is present in both periods (models 3,
6, 7, 8). For this purpose we have to compare the optimal prevention levels for
the following pairs: (e∗1 , e∗3 ), (e∗1 , e∗7 ), (e∗2 , e∗3 ), (e∗2 , e∗8 ), (e∗4 , e∗6 ), (e∗4 , e∗8 ), (e∗5 , e∗6 ),
(e∗5 , e∗7 ).
In this section we will deal with this problem making the comparison of
optimal prevention levels of some of the mentioned pairs, the remaining cases
being treated in a similar manner.
Case 1 (the comparison of e∗4 and e∗6 )
One assumes that in period 1 there exists a background risk described by the
fuzzy number A, and to this is added in period 2 a possibilistic background risk
described by a fuzzy number B. The following proposition offers a necessary
and sufficient condition that by this change, the optimal prevention level drops.
Proposition 6.1 The following assertions are equivalent:
(i) e∗6 ≤ e∗4 ;
E (v(w2 −l+B))−Ef (v(w2 +B))
≤ 1.
(ii) f
v(w2 −l)−v(w2 )
Proof. By (4.4), (4.9) and taking into account that −Ef (u′ (w1 − e∗6 + A)) < 0,
the following assertions are equivalent:
• e∗6 ≤ e∗4 ;
• V4′ (e∗6 ) ≥ V4′ (e∗4 ) = 0;
• −Ef (u′ (w1 − e∗6 + A)) + p′ (e∗6 )[v(w2 − l) − v(w2 )] ≥ 0;
2 −l)−v(w2 )
• −Ef (u′ (w1 −e∗6 +A))+Ef (u′ (w1 −e∗6 +A)) Ef (v(w2v(w
−l+B))−Ef (v(w2 +B)) ≥ 0;
• −Ef (u′ (w1 − e∗6 + A))[1 −
•
•
v(w2 −l)−v(w2 )
Ef (v(w2 −l+B))−Ef (v(w2 +B)) ]
2 −l)−v(w2 )
1 − Ef (v(w2v(w
−l+B))−Ef (v(w2 +B)) ≤
Ef (v(w2 −l+B))−Ef (v(w2 +B))
≤ 1.
v(w2 −l)−v(w2 )
0;
10
≥ 0;
Corollary 6.2 If v ′′′ > 0 and Ef (B) = 0 then e∗6 ≤ e∗4 .
Proof. By Lemma 5.4, condition (ii) of Proposition 6.1 is fulfilled.
The previous corollary shows that for a prudent agent (in period 2), adding
a possibilistic background risk B to an already existing possibilistic background
risk A leads to the raise in the optimal prevention.
Case 2 (the comparison of e∗5 and e∗6 )
One considers the model 5 with a possibilistic background risk B in period
2. We intend to see how the optimal prevention changes if we pass to model 6,
by adding a possibilistic background risk A in period 1.
Proposition 6.3 The following assertions are equivalent:
(i) e∗6 ≤ e∗5 ;
(ii) u′ (w1 − e∗6 ) ≤ Ef (u′ (w1 − e∗6 + A)).
Proof. By (4.6) and (4.9) the following equivalences hold:
e∗6 ≤ e∗5 iff V5′ (e∗6 ) ≥ V5′ (e∗5 ) = 0
iff −u′ (w1 − e∗6 ) + p′ (e∗6 )[Ef (v(w2 − l + B)) − Ef (v(w2 + B))] ≥ 0
iff −u′ (w1 − e∗6 ) + Ef (u′ (w1 − e∗6 + A)) ≥ 0
iff u′ (w1 − e∗6 ) ≤ Ef (u′ (w1 − e∗6 + A)).
Corollary 6.4 If u′′′ > 0 and Ef (A) = 0 then e∗6 ≤ e∗5 .
Proof. By applying Proposition 2.2, from u′′′ > 0 and Ef (A) = 0 it follows
that condition (iii) of Proposition 6.3 is verified.
Case 3 (the comparison of e∗1 and e∗7 )
One considers model 1 with a probabilistic background risk in period 1 and
one passes to model 7 by adding a possibilistic background risk in period 2.
Proposition 6.5 The following assertions are equivalent:
(i) e∗7 ≤ e∗1 ;
E (v(w2 −l+B))−Ef (v(w2 +B))
(ii) f
≤1
v(w2 −l)−v(w2 )
Proof. Recall from [4] (or from Section 3, Model 1) that
V1′ (e) = −M (u′ (w1 − e + X)) + p′ (e)[v(w2 − l) − v(w2 )].
Then, by (4.11) the following assertions are equivalent:
• e∗7 ≤ e∗1
• V1′ (e∗7 ) ≥ V1′ (e∗1 ) = 0
• −M (u′ (w1 − e∗7 + X)) + p′ (e∗7 )[v(w2 − l) − v(w2 )] ≥ 0
11
• −M (u′ (w1 − e∗7 + X))[1 −
•
Ef (v(w2 −l+B))−Ef (v(w2 +B))
v(w2 −l)−v(w2 )
v(w2 −l)−v(w2 )
Ef (v(w2 −l+B))−Ef (v(w2 +B)) ]
≥0
≤1
Corollary 6.6 If v ′′′ > 0 and Ef (B) = 0 then e∗7 ≤ e∗1 .
Proof. By [4], if v ′′′ > 0 and Ef (B) = 0 then condition (ii) of Proposition 6.5
is fulfilled.
7
Concluding remarks
The five models from the paper deal with the optimal prevention in the presence
of some risk situations possibilistically described (by fuzzy numbers) or mixed
(by combinations of random variables and fuzzy numbers). They come in addition to probabilistic models of optimal saving developed by Courbage and Rey
in [4]. The main results from the paper establish necessary and sufficient conditions for the optimal prevention increase or decrease when we add a possibilistic
or mixed risk to the benchmark model.
Next we will state several research topics which could be investigated in line
to this paper.
(I) Among the entities which define the benchmark model from Section 4, the
probability p(e) has an essential role. Also, all the eight background risk models
from Section 5 contain this probability as the defining element. Accordingly,
even if the background risk is possibilistic or mixed, in the presence of p(e), each
model still preserves a probabilistic trace.
We could modify the nature of p(e) in the following way:
Case A: ”p(e) is the possibility that a loss l occurs at the prevention level
e”
Case B: ”p(e) is the credibility 2 that a loss l occurs at the prevention level
e”
Study the optimal prevention in models in which p(e) is a possibility (Case
A) and a credibility (Case B), respectively.
(II) Both the definition of prevention models (total utility function, the first–
order condition, etc.), and the optimal prevention level comparison theorems
were formulated in terms of possibilistic expected utility, probabilistic expected
utility or both of them. On the other hand, in the literature there exists a
second notion of possibilistic expected utility (see for example [15], Section 4.2),
as well as the concept of expected utility operator (see [15], Chapter 5), which
comprises both of them.
Generalize the prevention models from the paper in a general context founded
on expected utility operators.
2 For the definition of credibility and its properties, see the monograph [19] or Chapter 8
from [15].
12
(III) In order to reduce the action of various types of risk, an agent has
within his reach several instruments: saving, self–protection and self–insurance
(according to [17]). A precautionary motive of an agent is defined in [17] as ”the
extent to which the agent chooses an instrument for the response to a future
risk”.
In paper [17] it is analyzed the way the intensity of precautionary motives
is influenced by the choice of various combinations of instruments as well as the
interaction of considered types of risk.
An open problem is to develop models of precautionary saving analysis corresponding to combinations of the same three instruments (saving, self–protection
and self–insurance), but in which the risk parameters are probabilistic, possibilistic or mixed.
References
[1] C. Carlsson, R. Fullér, On possibilistic mean value and variance of fuzzy
numbers. Fuzzy Sets Syst, 122, 2001, 315–326.
[2] C. Carlsson, R. Fullér, Possibility for decision. Springer, 2011.
[3] A. M. Lucia-Casademunt, I. Georgescu, Optimal saving and prudence in
a possibilistic framework, DCAI, Volume 217, Advances in Intelligent Systems and Computing, 2013, pp 61-68
[4] C. Courbage, B.Rey, Precautionary saving in the presence of other risks.
Economic Theory, 32, 2007, 417–424
[5] C. Courbage, B. Rey, N. Treich, Prevention and precaution, in G. Dionne
(ed.), Handbook of Insurance, Huebner International Series on Risk, Insurance and Economic Security (New York Springer), New York, pp. 184–204
[6] C. Courbage, H. Loubergé, R. Peter, Optimal prevention for multiple risks,
J. of Risk and Insurance, 2016, in press
[7] G. Dionne, L. Eeckhoudt, Self-insurance, self-protection and increased risk
aversion, Economics Letters, 17(1–2), 1985, pp. 39–42.
[8] D. Dubois,H. Prade, Fuzzy sets and systems: theory and applications. New
York:Academic Press, 1980.
[9] D. Dubois,H. Prade, Possibility theory. New York:Plenum Press, 1988.
[10] L. Eeckhoudt, C. Gollier, H. Schlesinger, Economic and financial decisions
under risk. Princeton University Press, 2005.
[11] L. Eeckhoudt, C. Gollier, The impact of prudence on optimal prevention,
Economic Theory, 26, 2005, pp. 989–994.
13
[12] L. Eeckhoudt, R. J. Huang, L. Y. Tzeng, Precautionary effort: a new look,
J. of Risk and Uncertainty, 79(2), 2012, pp. 585–590.
[13] I. Ehrlich, G. S. Becker, Market insurance, self-insurance, and selfprotection, The Journal of Political Economy, 1972, pp. 623-648.
[14] R. Fullér, P. Majlender, On weighted possibilistic mean and variance of
fuzzy numbers, Fuzzy Sets and Systems 136(3) 16 June 2003, pp. 363-374.
[15] I. Georgescu, Possibility theory and the risk, Springer, 2012
[16] I. Georgescu, Risk aversion, prudence and mixed optimal saving models,
Kybernetika, 50(5), 2014, pp. 706–724.
[17] C. Heinzel, R. Peter, Precautionary motives under multiple instruments,
2015, available at http://afse2015.sciencesconf.org/61592/document
[18] M. S. Kimball, Precautionary saving in the small and in the large, Econometrica, 58, 1990, pp. 53–73.
[19] B. Liu, Uncertainty Theory, Springer-Verlag, Berlin, 2004.
14
| 5cs.CE
|
arXiv:1409.0733v2 [math.ST] 6 Jun 2016
Bernoulli 22(4), 2016, 2177–2208
DOI: 10.3150/15-BEJ725
Integral approximation by kernel smoothing
BERNARD DELYON1 and FRANÇOIS PORTIER2
1
Institut de recherches mathématiques de Rennes (IRMAR), Campus de Beaulieu, Université de
Rennes 1, 35042 Rennes Cédex, France. E-mail: [email protected]
2
Institut de Statistique, Biostatistique et Sciences Actuarielles (ISBA), Université catholique de
Louvain, Belgique. E-mail: [email protected]
Let (X1 , . . . , Xn ) be an i.i.d. sequence of random variables in Rd , d ≥ 1. We show that, for any
function ϕ : Rd → R, under regularity conditions,
!
Z
n
X
ϕ(Xi )
P
1/2
−1
n
n
− ϕ(x) dx −→ 0,
fb( Xi )
i=1
where fb is the classical kernel estimator of the density of X1 . This result is striking because
it speeds up traditional rates, in root n, derived from the central limit theorem when fb = f .
Although this paper highlights some applications, we mainly address theoretical issues related to
the later result. We derive upper bounds for the rate of convergence in probability. These bounds
depend on the regularity of the functions ϕ and f , the dimension d and the bandwidth of the
kernel estimator fb. Moreover, they are shown to be accurate since they are used as renormalizing
sequences in two central limit theorems each reflecting different degrees of smoothness of ϕ. As
an application to regression modelling with random design, we provide the asymptotic normality
of the estimation of the linear functionals of a regression function. As a consequence of the above
result, the asymptotic variance does not depend on the regression function. Finally, we debate
the choice of the bandwidth for integral approximation and we highlight the good behavior of
our procedure through simulations.
Keywords: central limit theorem; integral approximation; kernel smoothing; nonparametric
regression
1. Introduction
Let (X1 , . . . , Xn ) be an i.i.d. sequence of random variables in Rd , d ≥ 1. We show that,
for any function ϕ : Rd → R, under regularity conditions,
n−1
Z
n
X
ϕ(Xi )
− ϕ(x) dx = oP (n−1/2 ),
fb(i) (Xi )
i=1
This is an electronic reprint of the original article published by the ISI/BS in Bernoulli,
2016, Vol. 22, No. 4, 2177–2208. This reprint differs from the original in pagination and
typographic detail.
1350-7265
c
2016 ISI/BS
(1)
2
B. Delyon and F. Portier
where fb(i) is the classical leave-one-out kernel estimator of the density of X1 say f ,
defined by
X
−1
fb(i) (x) = ((n − 1)hd )
K(h−1 (x − Xj ))
for every x ∈ Rd ,
1≤j≤n,j6=i
where K is a d-dimensional kernel and where h, called the bandwidth, needs to be
chosen and will certainly depend on n. Result (1) and the central limit theorem lead to
the following reasoning: when estimating the integral of a function that is evaluated on
a random grid (Xi ), whether f is known or not, using a kernel estimator of f provides
better convergence rates than using f itself.
Result (1) certainly has some consequences in the field of integral approximation.
In this area, many deterministic as well as random methods are available. Accuracy
with respect to computational time is the usual trade-off that allows to compare them.
The advantages of random over deterministic framework lie in their stability in highdimensional settings. For a comprehensive comparison between both approaches, we refer to [8]. Among random methods, importance sampling is a widely used technique that
basically reduces the variance of the classical Monte–Carlo integration through a good
choice of the sampling
distribution f , called the sampler. Estimators are unbiased having
Pn
the form n−1 i=1 ϕ(Xi )/f (Xi ) with Xi ∼ f . Regarding the mean squared error (MSE),
the optimal sampler f ∗ is unique and depends on ϕ (see Theorem 6.5 in [8], page 176).
Among others, parametric [18] and nonparametric [25] studies focused on the estimation of the optimal sampler. Equation (1) indicates a new weighting of the observations
ϕ(X1 ), . . . , ϕ(Xn ). Each weight fb(i) (Xi ) reflects how isolated is the point Xi among the
sample. Therefore, our estimator takes into account this information by giving more
weight to an isolated point. In summary our procedure, which is adaptive to the design
points enjoys the following advantages:
• Faster than root n rates,
• one-step estimation based on a unique sample (X1 , . . . , Xn ),
• each Xi drawn from f , possibly unknown.
To the best of our knowledge, when the design is not controlled, no such rates have been
obtained.
In many semiparametric problems, it has been an important issue to construct root n
estimators, possibly efficient [1], that rely on a kernel estimator of the nuisance parameter.
Among others, it was addressed by Stone in [21] in the case of the estimation of a location
parameter, by Robinson in [19] in the partially linear regression model, or by Härdle and
Stoker in [16] studying the single index model. The result in equation (1), which would
be seen as a superefficient estimator in theRLe Cam’s theory, cannot be linked actually to
this theory since the quantity of interest ϕ(x) dx does not depend on the distribution
of X1 . As a result, the link between our work and the semiparametric literature relies
mainly on the plug-in strategy we employed, by substituting the density f by a kernel
estimator.
In this paper, we propose a comprehensive study of the convergence stated in equation (1). A similar result was originally stated by Vial in [24] (Chapter 7, equation (7.27)),
Integral approximation by kernel smoothing
3
as a lemma in the context of the multiple index model. To the best of our knowledge,
this type of asymptotic result has not been addressed yet as a particular problem. Our
theoretical aim is to extend result (1) by:
(A) Being more precise about the upper bounds: How does the dimension d, the window h, the regularity of ϕ and f , impact these bounds?
(B) Showing central limit theorems by specifying the regularity of ϕ.
To achieve this program, we need to introduce a corrected version of the estimate (1)
for which the bias has been reduced. First, the corrected estimator is shown to have
better rates of convergence than the initial one. Second, it is shown to be asymptotically
normal with rates nhd/2 in the case where ϕ is very regular, and with rates (nh−1 )1/2
in a special case in which ϕ jumps at the boundary of its support. To compute the
asymptotic distribution, we rely on the paper by Hall [12], where a central limit theorem
for completely degenerate U -statistics has been obtained. An important point is that we
have succeeded in proving our result with much weaker assumptions on the regularity
of ϕ than on the regularity of f . For instance, equation (1) may hold even when ϕ has
some jumps. However, the estimation of f is subject to the curse of dimensionality, that
is, f is required to be smooth enough regarding the dimension of X1 .
Our aim is also to link equation (1) to nonparametric regression with random design,
that is, the model Yi = g(Xi ) + σ(Xi )ei with g unknown and ei i.i.d. with ei ⊥
⊥ Xi . In
particular, we obtain the asymptotic normality for the estimators of the linear functionals
of g. Thanks to the fast rates detailed previously, the asymptotic distribution does not
depend on the function g.
The paper is organized as follows. Section 2 deals with technical issues related to
equation (1). In particular, we examine the rates of convergence of (1) according to
the choice of the bandwidth, the dimension and the regularity of the functions ϕ and f .
Section 3 is dedicated to the convergence in distribution of our estimators. In Section 4, we
show how to apply equation (1) to the problem of the estimation of the linear regression
functionals. Finally, in Section 5, we give some simulations that compare our method with
the traditional Monte–Carlo procedure for integration. The proofs and the technicalities
are postponed in Section 6 at the end of the paper.
2. Rates of convergences faster than root n
In this section, we first provide upper bounds on the rates of convergence in probability
of our estimators. Our main purpose is to show that rates
faster than root n hold in a
R
wide range of parameter settings for the estimation of ϕ(x) dx. Second, we argue that
those faster thanR root n rates have no reason to hold when estimating other functionals
of the type f 7→ T (x, f (x)) dx.
4
B. Delyon and F. Portier
2.1. Main result
Let Q ⊂ Rd be the support of ϕ. The quantity I(ϕ) =
b
I(ϕ)
= n−1
R
ϕ(x) dx is estimated by
n
X
ϕ(Xi )
.
fb(i) (Xi )
i=1
Actually, this estimator can be modified in such a way that the leading error term of its
expansion vanishes asymptotically (see Remark 9 for more details). For that, we define
vb(i) (x) as
−1
vb(i) (x) = ((n − 1)(n − 2))
X
1≤j≤n,j6=i
2
(h−d K(h−1 (x − Xj )) − fb(i) (x)) .
It is, up to a factor (n−1)−1 , the leave-one-out estimator of the variance of h−d K(h−1 (x−
Xj )). The corrected estimator is
Ibc (ϕ) = n−1
n
X
ϕ(Xi )
vb(i) (Xi )
1−
.
b(i) (Xi )
fb(i) (Xi )2
i=1 f
b
To state our main result about the convergences of I(ϕ)
and Ibc (ϕ), we define the Nikolski
class of functions H(s, M ) of regularity s = k + α, k ∈ N, 0 < α ≤ 1, with constant M > 0,
as the set of bounded and k times differentiable functions ϕ whose all derivatives of order
k satisfy [23]
Z
2
(ϕ(l) (x + u) − ϕ(l) (x)) dx ≤ M |u|2α ,
l = (l1 , . . . , ld ),
d
X
i=1
li ≤ k,
where | · | stands for the Euclidean norm and the li ’s are natural integer. Be careful that k
cannot be equal to s. We say that K is a kernel with order r ∈ N∗ as soon as K : Rd 7→ R
is bounded and satisfies
Z
K(x) dx = 1,
Z
l
x K(x) dx = 0,
l = (l1 , . . . , ld ), 0 <
d
X
i=1
li ≤ r − 1
with the notation xl = xl11 × · · · × xldd . The following assumptions are needed to show our
first result, they are discussed after the statement.
(A1) For some s > 0 and M > 0, the support of ϕ is a compact set Q ⊂ Rd and ϕ is
H(s, M ) on Rd .
(A2) For some integer r ≥ 1, the variable X1 has a bounded density f on Rd such that
its rth order derivatives are bounded.
(A3) For every x ∈ Q, f (x) ≥ b > 0.
5
Integral approximation by kernel smoothing
Table 1. Best acceleration of convergence rate in Theorem 1. Best rate acceleration n−β obtained with h ∝ n−γ
β
γ
Equation (i)
2s ≤ r − d
s
2(s+d)
1
2(s+d)
0 < r − d ≤ 2s
(r−d)
2(r+d)
1
r+d
d ≤ 2s ≤ r − d/2
(r−d/2)
2r+d
s
2s+d
1
r+d/2
1
2s+d
r ≤ 3d/2 and 0 < 4r − 3d ≤ 6s
4r−3d
2(3d+2r)
3
3d+2r
2s ≤ d and 6s ≤ 4r − 3d
2s
2s+3d
2
2s+3d
Equation (ii)
d ≤ r − d/2 ≤ 2s
R
(A4) The kernel K has order r and K(x) dx = 1. Moreover, there exists m1 > 0 and
m2 > 0 such that, for every x ∈ Rd , |K(x)| ≤ m1 exp(−m2 |x|). In addition K is
symmetric: K(x) = K(−x).
The next theorem is proved in Section 6.
Theorem 1. Under the assumptions (A1) to (A2), we have the following OP estimates
b
n1/2 (I(ϕ)
− I(ϕ)) = OP (hs + n1/2 hr + n−1/2 h−d ),
n1/2 (Ibc (ϕ) − I(ϕ)) = OP (hs + n1/2 hr + n−1/2 h−d/2 + n−1 h−3d/2 ),
(i)
(ii)
which are valid if the sums inside the OP ’s tend to zero.
Remark 1. Assumption (A2) about the smoothness of f is crucial to guarantee a rate
faster than root n in Theorem 1. On the one hand, one needs r > d to obtain such a
rate in equation (i), on the other hand, r > 3d/4 suffices to get this rate in equation (ii).
Otherwise there does not exist h such that the bounds in Theorem 1 go to 0. This
phenomenon is often referred as the curse of dimensionality.
In equation (i) (resp., (ii)), when h ∝ n−γ , the best choice of γ depends on r and s;
it balances two of the three (resp., four) terms while letting the other one(s) smaller.
Precise rate acceleration for each situation is given in Table 1.
As in many semiparametric problems (see, e.g., [16], Section 4.1), our estimator of f is
suboptimal with respect to the density estimation problem (see [22]). Indeed, to achieve
the optimal rates in density estimation one would need to take h ∝ n−1/(2r+d) which
would even prevent n1/2 hr to go to 0 in Theorem 1. A practical bandwidth selection is
proposed Section 5.
Remark 2. Assumption (A2) prevents from bias problems in the estimation of f that
may occur at the borders of Q. Indeed, if f jumps at the boundary of Q, then our estimate
6
B. Delyon and F. Portier
of f would be asymptotically biased and the rates provided in Theorem 1 would not hold.
To get rid of this problem, if one knew the support of f , one could correct by hand the
estimator as, for instance, in [17], or might use Beta kernels as detailed in [3].
Remark 3. Assumption (A3) basically says that f is separated from 0 on Q. The
exponential bound on the kernel in assumption (A4) guarantees that f is estimated
uniformly on Q (see [5]). This helps to control the random denominators fb(i) (Xi )’s in
b
the expression of I(ϕ)
and Ibc (ϕ). In the context of Monte–Carlo procedures for integral
approximation, assumptions (A2) and (A3) are not that restrictive because one can draw
the Xi ’s from a distribution smooth enough and whose support contains the integration
domain.
Remark 4. The use of leave-one-out estimators fb(i) and vb(i) in Ibc (ϕ) are not only
justified by the simplification they involve in the proofs. It also leads to better convergence
rates. Consider the term R0 in the proof of equation (ii) in Theorem 1, when replacing
the leave-one-out estimator of f by the classical one, R0 remains a degenerate U -statistic
but with nonzero diagonal terms. It is possible to show that these terms are leading terms
of the resulting expansion. They imply a rate of convergence of order n−1/2 h−d which is
larger than the rate we found for Ibc (ϕ).
b
However, concerning I(ϕ),
the leave-one-out estimator is not necessary to get (i). The
leave-one-out estimator being indeed at a distance O(n−1 h−d ) from the ordinary one, the
change would made a difference of order at most n−1/2 h−d in the left-hand side of (i),
which already appears in the right-hand side of (i).
Remark 5. The function class H(s, M ) contains two interesting sets of functions that
provide different rates of convergence in Theorem 1. First, if ϕ is α-Hölder on Rd with
Hölder constant M1 , and has bounded support, then ϕ is H(α, M1 ) on Rd . Second,
if the support of ϕ is a convex body (compact convex set with non-empty interior)
and ϕ is α-Hölder (with constant M1 ) inside its support (e.g., the indicator of a ball)
then there exists M2 > 0 such that ϕ is H(min(α, 1/2), M2 ) on Rd (see Lemma 9 in
the Section 6). Then, because the sum of two Nikolski functions is still Nikolski, the
assumptions of Theorem 1 are valid for a wide range of integrand. Moreover, note that
a loss of smoothness at the boundary of the support involves a loss in the rates of
convergence (i) and (ii). More precisely, whatever the smoothness degree of ϕ inside its
support, if continuity fails at the boundary, then the Nikolski regularity would be at
most 1/2 and, therefore, the rates acceleration in Theorem 1 could not exceed h1/2 . In
Section 3, we study such an example and show a central limit theorem with such a rate.
Remark 6. The symmetry assumption in (A4) is actually superfluous, but simplifies
the proof, because in this case we do not have to distinguish the convolution with K(x)
and the convolution with K(−x).
7
Integral approximation by kernel smoothing
2.2. On the generalization of Theorem 1
In view of the intriguing convergence rates stated in Theorem 1, one may be curious to
know the behavior of our estimator when estimating more general functionals with the
form
Z
IT = T (x, f (x)) dx,
where T : Rd × R+ → R. Following the same approach as previously, the estimator we
consider is
IbT = n−1
n
X
T (Xi , fb(i) (Xi ))
fb(i) (Xi )
i=1
.
It turns out that T given by (x, y) 7→ ϕ(x) is the only case for which the rates are faster
√
than root n. For other functionals and a wide range of bandwidth, n(IbT − IT ) converges
to a normal distribution. In view of the negative aspect of this result with respect to the
statement of Theorem 1, we provide an informal calculation of the asymptotic law of
√
n(IbT − IT ). We require that (A2) to (A4) hold and that nh2r → 0 and nh2d → +∞
(the latter guarantees faster than root n rates in equation (i)). If y 7→ T (x, y) has a
bounded (uniformly in x) second-order derivative, using a Taylor expansion with respect
to the second coordinate of T (the first-order derivative of T with respect to the second
coordinate is further denoted by ∂2 T ), we have
n1/2 (IbT − IT )
= n−1/2
n
X
T (Xi , f (Xi ))
fb(i) (Xi )
i=1
− IT +
∂2 T (Xi , f (Xi ))(fb(i) (Xi ) − f (Xi ))
fb(i) (Xi )
e2 ,
+R
e2 can be treated by standard techniques of kernel estimation (see equations (12)
where R
and (17) for details), this gives that, with probability going to 1,
e2 | ≤ Cn−1/2
|R
n
X
(fb(i) (Xi ) − f (Xi ))2
i=1
fb(i) (Xi )
= OP (n1/2 h2r + n−1/2 h−d ) = oP (1),
where C > 0 does not depend on n or h. Then we write
√
e0 + R
e1 + R
e2 ,
n(IbT − IT ) = R
with
e0 = n−1/2
R
e1 = n−1/2
R
n
X
T (Xi , f (Xi ))
i=1
fb(i) (Xi )
∂2 T (Xi , f (Xi ))f (Xi )
− IT −
+
fb(i) (Xi )
Z
n
X
∂2 T (Xi , f (Xi )) − ∂2 T (x, f (x))f (x) dx .
i=1
Z
∂2 T (x, f (x))f (x) dx ,
8
B. Delyon and F. Portier
If x 7→ T (x, f (x)) and x 7→ ∂2 T (x, f (x))f (x) are Nikolski, applying Theorem 1 gives that
e1 is
e0 = oP (1). As a consequence √n(IbT − IT ) = oP (1) if and only if the variance of R
R
degenerate, that is equivalent to
∂2 T (Xi , f (Xi )) = c
a.s.
If we want this to be true for a reasonably large class of distribution functions, it would
imply
∂2 T (x, y) = c
for all (x, y) ∈ Rd × R+ ,
for which the solutions have the form T (x, y) = ϕ(x) + cy.
3. Central limit theorem
In the previous section, we derived upper bounds on the convergence rates in probability
under fairly general conditions. In this section, by being a little more specific about
the regularity of ϕ, we are able to describe precisely the asymptotic distribution of
Ibc (ϕ) − I(ϕ). Actually the approach is to decompose the latter quantity as a sum of a U statistic Un plus a martingale Mn with respect to the filtration {X1 , . . . , Xn }, plus a bias
term Bn that is non-random (see the beginning of Section 6.2 for the definitions of Un ,
Mn , Bn ). Then existing results about the asymptotic behavior of completely degenerate
U -statistics [12] and martingales [13] will help to derive the asymptotic distribution. We
shall consider two cases. First, we present the case where ϕ is smooth enough so that the
dominant term is Un , and second we study an example where ϕ is not continuous at the
boundary of its support. As a consequence, the dominant term is Mn .
b
For I(ϕ)
− I(ϕ), the situation is less interesting since for most of the choice of h a
(non-random) bias term leads the asymptotic decomposition (see Remark 9).
3.1. Smooth case
The smooth case corresponds to situations where the functions f and ϕ are smooth
enough, that is, r > 3d/2 and 2s > d. This is highlighted by the assumptions on the
bandwidth in the next theorem.
Theorem 2. Under the assumptions (A1) to (A4), if nh2d → +∞, nhr+d/2 → 0 and
nh2s+d → 0, the random variable nhd/2 (Ibc (ϕ) − I(ϕ)) is asymptotically normally distributed with zero-mean and variance given by
2 Z
Z Z
(K(u + v) − K(v))K(u) du dv ϕ(x)2 f (x)−2 dx.
Integral approximation by kernel smoothing
9
The assumptions on the bandwidth are not satisfied by the optimal bandwidths displayed in Table 1. This is, in fact, a presentation issue. Indeed we have chosen to make
the bias term Bn vanish so that any optimal bandwidth that balances the bias and the
variance is excluded. We could have proceeded the other way around, by stating that
nhd/2 (Ibc (ϕ) − I(ϕ) − Bn ) has the same limiting distribution as in Theorem 2, provided
that nh2d → +∞ and nh2 min(r,s)+d → 0. One can verify that this holds true for the
optimal bandwidth given in the first line of Table 1 for equation (ii).
3.2. A non-smooth example
We are interested in the case where ϕ is not sufficiently regular so that Mn is no longer
negligible with respect to Un , that is, nh2 min(r,s)+d does not go to 0. This occurs whenever s < d/2. In this case the variance is hard to compute since it depends on the behavior
of Mn and therefore on the rate of convergence of the kernel regularization of ϕ. Hence,
a precise description cannot be provided by considering usual regularity classes, for example, Hölder, Nikolski or Sobolev since they only provide bounds on the rate of kernel
regularization. For this reason, we consider a particular case where the function ϕ is
Nikolski inside Q and vanishes outside. Typical functions we have in mind are the one
that jump at the boundary of their support. Lemma 9 informs us that such functions are
Nikoslki with regularity 1/2. For Q ⊂ Rd compact and x ∈ ∂Q, we define
ZZ
LQ (x) =
min(hz, u(x)i, hz ′ , u(x)i)+ K(z)K(z ′ ) dz dz ′ ,
where u(x) is the unit normal outer vector of Q at the point x. We need the following
assumption in place of (A1).
(B1) For some s > 1/2 and M > 0, the support of ϕ is a convex body Q ⊂ Rd with C 2
boundary and ϕ is H(s, M ) on Q.
Theorem 3. Under the assumptions (A2) to (A4) and (B1), if nh(3d+1)/2 → +∞ and
nh2r−1 → 0 the random variable (nh−1 )1/2 (Ibc (ϕ) − I(ϕ)) is asymptotically normally distributed with zero-mean and variance given by
Z
LQ (x)ϕ(x)2 dHd−1 (x),
∂Q
where Hd−1 stands for the (d − 1)-dimensional Hausdorff measure.
4. Application to nonparametric regression
Equation (1) has applications in nonparametric regression with random design. Let
Yi = g(Xi ) + σ(Xi )ei ,
(3)
10
B. Delyon and F. Portier
where (ei ) is an i.i.d. sequence of real random variables with mean 0 and unit variance,
independent of the sequence (Xi ), and σ : Rd → R and g : Rd → R are unknown functions.
Let Q ⊂ Rd be a compact set and L2 (Q) be the Hilbert space of squared-integrable
functions on Q. Let ψ ∈ L2 (Q) be extended to Rd by 0 outside of Q (ψ has compact
support Q). The inner product in L2 (Q) between the regression function g and ψ, is
given by
c=
Z
g(x)ψ(x) dx,
note that if ψ belongs to a given basis of L2 (Q), then c is a coordinate of g in this basis.
Among typical applications, we can mention Fourier coefficients estimation for either
nonparametric estimation (see, e.g., [14], Section 3.3), or location parameter estimation
(see [11]). We also mention the link with the estimation of the index in the single index
model (see [16]).
The estimation of the linear functionals of g is a typical semiparametric problem in the
sense that it requires the nonparametric estimation of the density f of X1 as a first step
and then to use it in order to estimate a real parameter. To the best of our knowledge,
in the case of a regression with unknown random design, estimators that achieve root
n consistency have not been provided yet (see, e.g., [14] and the reference therein). Our
approach is based on kernel estimates fb(i) of the density of X1 that are then plugged
into the classical empirical estimator of the quantity E[Y ψ(X)f (X)−1 ]. We define the
estimator
to derive the asymptotic of
√
b
c = n−1
n
X
Yi ψ(Xi )
i=1
fb(i) (Xi )
,
n(b
c − c), we use model (3) to get the decomposition
√
n(b
c − c) = A + B,
with
A = n−1/2
n
X
σ(Xi )ψ(Xi )
i=1
B=n
−1/2
fb(i) (Xi )
ei ,
n
X
g(Xi )ψ(Xi )
i=1
fb(i) (Xi )
−
Z
g(x)ψ(x) dx .
Roughly speaking, Theorem 1 provides that B is negligible with respect to A. As a result,
√
c − c) and, therefore, the limiting distribution can
A carries the weak convergence of n(b
be obtained making full use of the independence between the Xi ’s and the ei ’s. In order
11
Integral approximation by kernel smoothing
to achieve such a program, this assumption is needed.
(C1) For some s > 0 and M > 0, the support of ψ is a compact set Q ⊂ Rd and both
ψ and g are H(s, M ) on Rd .
The following theorem is proved in Section 6.
Theorem 4. Under the assumptions (A2) to (A4), and (C1), if n1/2 hr → 0 and
n1/2 hd → +∞, then the random variable n1/2 (b
c − c) is asymptotically normally distributed
with zero-mean and variance
σ(X1 )ψ(X1 )
.
v = Var
f (X1 )
Remark 7. Let us compare b
c with the appealing estimator
c = n−1
e
n
X
Yi ψ(Xi )
i=1
f (Xi )
which requires the knowledge of f . First, if the signal is observed without noise, that is,
Yi = g(Xi ), then n1/2 (b
c − c) goes to 0 in probability whereas e
c is asymptotically normal.
Secondly, when there is some noise in the observed signal, meaning that σ(X1 ) is not 0,
the comparison can be made regarding their asymptotic variances. Since we have
v ≤ Var(n1/2 (e
c − c)),
it is asymptotically more efficient to plug the nonparametric estimator of f than to use
f directly.
Remark 8. The set Q reflects the domain where g is studied. Obviously, the more
dense the Xi ’s in Q, the more stable the estimation. Nevertheless, it could happen that
f vanishes on some point on Q and this is not taken into account by our framework. In
such situations, one may adapt the estimation from the sample by ignoring the design
points on which the estimated density takes too small values. The estimator b
c might be
replaced by
n−1
n
X
Yi ψ(Xi )
i=1
fb(i) (Xi )
1{fb(i) (Xi )>b} ,
where b > 0 will certainly depend on n. This method, often referred as trimming, has
been employed in [16] and [4] and guarantees computational stability as well as theoretical
properties. Even if such an approach is feasible here, it seems far beyond the scope of
the article.
12
B. Delyon and F. Portier
5. Simulations
In this section, we provide some insights about the implementation and the practical
behavior of our integral approximation procedure. In particular, we propose an adaptive
procedure that selects the bandwidth for the kernel smoothing. While our theoretical
study highlighted that our estimators suffers from the curse of dimensionality (see Remark 1), our simulation results confirm that the estimation accuracy of our methods
diminishes when the dimension increases. In dimension 1, our procedure outperforms
by far the Monte–Carlo method. In moderate sample size (from 200 to 5000) up to dimension 4, our method still realizes a significant improvement over the Monte–Carlo
method. The simulations are conducted under fairly general design distributions that do
not necessarily satisfy assumption (A2) (e.g., equation (7)).
5.1. Kernel choice
In the whole simulation study, our estimator of the density of the design is based on the
kernel
K(x) = 21 c−1
d (d + 1)(d + 2 − (d + 3)|x|)1|x|<1 ,
cd =
2π d/2
,
dΓ(d/2)
where cd is the volume of the unit ball in dimension d. This kernel is radial with order 3.
5.2. Bandwidth choice
One may follow [15] to select the optimal bandwidth by a plug-in method. It requires
to optimize an asymptotic equivalent of the MSE with respect to h. In Section 3, we
b
highlighted that the limiting distribution of I(ϕ)−
I(ϕ), and so the MSE, depends heavily
on the degree of smoothness of ϕ. In practice, the regularity of ϕ is often unknown, as a
result, we prefer a simulation–validation type strategy.
The idea is to pick the value h which gives the best result for the estimation of the
integral I(ϕ)
e of a test function ϕ
e which looks like ϕ, and for which I(ϕ)
e is known. We
choose this test function as
n
X
ϕ(Xi ) −d e x − Xi
ϕ(x)
e
= n−1
h0 K
,
(4)
b(i) (Xi )
h0
i=1 f
e is simply the Epanechnikov kernel
where K
2
e
K(x)
= 12 c−1
d (d + 2)(1 − |x| )1|x|<1 .
(5)
13
Integral approximation by kernel smoothing
Since we know that
I(ϕ)
e =
Z
ϕ(x)
e dx = n−1
n
X
ϕ(Xi )
,
b
f (i) (Xi )
i=1
b ϕ)
we just take the value of h for which the estimate I(
e is closest to I(ϕ);
e there is actually
b
b
two values, one for I(ϕ)
e and one for Ic (ϕ).
e The smoothing parameter h0 is chosen using
the rule of thumb given by
h0 = σ
d2d+5 Γ(d/2 + 3)
(2d + 1)n
1/(4+d)
,
(6)
where σ 2 is the mean of the estimated variances of each component (see [20], Section 4.3.2). The density estimates fb(i) (Xi ) in (4) are computed with the same value
h0 and the same kernel.
We did not try to use a resampling method, thinking that it is better to have h adapted
to the specific sample.
5.3. First model
In this model, f is a normal distribution
1 1
Xi ∼ N , Id ,
2 4
ϕ(x) =
d
Y
2 sin(πxk )2 10≤xk ≤1 .
k=1
The integral of ϕ is 1. Figure 1 shows simulations for different values of n and d, and
using equations (4), (5) and (6) for the choice of h.
5.4. Second model
In this second model, the assumptions are not satisfied since the distribution is uniform
over the unit cube, we have
Xi ∼ U([0, 1]d),
ϕ(x) =
d
Y
2 sin(πxk )2 10≤xk ≤1 .
(7)
(8)
k=1
In spite of the fact that (A2) is not any more satisfied, good results are still possible
because ϕ cancels at the boundary of the cube. For the choice of h, we used equation (4),
14
B. Delyon and F. Portier
b
Figure 1. Each boxplot is based on 100 estimates Ibc (ϕ), I(ϕ)
and Monte–Carlo method noted
b
IMC for the first model with different values of n and d.
(5) but, it is important to constrain the function ϕ
e to have its support on the cube, and
a way to do this is to remove the boundary terms out of (4) by choosing now
ϕ(x)
e
= |J|−1
X ϕ(Xi )
e x − Xi ,
h−d
K
0
h0
fb(i) (Xi )
i∈J
(9)
J = {i : h < Xij < 1 − h, j = 1 · · · d}.
We could have done the other way around, use (4) and simulate uniformly extra points
at distance less than h0 of the cube, in order to cover the support of ϕ.
e Figure 2 shows
the results of the simulations for different values of n and d and using equations (9), (5)
and (6) for the choice of h.
6. Proofs
Notation
The Euclidean norm, the Lp norm and the supremum norm are, respectively, denoted by
| · |, k · kp and k · k∞ . We introduce Kh (·) = h−1 K(·/h), and
Kij = h−d K(h−1 (Xi − Xj )),
15
Integral approximation by kernel smoothing
b
Figure 2. Each boxplot is based on 100 estimates Ibc (ϕ), I(ϕ)
and Monte–Carlo method noted
b
IMC for the second model with different values of n and d.
fbi =
vbi =
1
n−1
X
Kij ,
1≤j≤n,j6=i
1
(n − 1)(n − 2)
X
1≤j≤n,j6=i
(Kij − fbi )2 ,
and for any function g : Rd → R, we define
Z
gh (x) = g(x − hu)K(u) du,
and we put
ϕ(x)
,
q ∈ N,
fh (x)q
f (x)
e
.
ψ(x)
= ϕ(x)
fh (x)2 h
ψq (x) =
6.1. Proof of Theorem 1
We start by showing (ii), then (i) will follow straightforwardly.
16
B. Delyon and F. Portier
Proof of (ii). The following development reminiscent of the Taylor expansion
fh (Xi ) − fbi (fh (Xi ) − fbi )2 (fh (Xi ) − fbi )3
1
1
+
+
+
=
,
fh (Xi )2
fh (Xi )3
fbi fh (Xi )
fbi fh (Xi )3
allows us to expand our estimator as a sum of many terms, where the density estimate
fbi is moved to the numerator, with the exception of the fourth one. We will show that
this last term goes quickly to 0. For the linearised terms, this is very messy because the
correct bound will be obtained by expanding also fbi in those expressions. In order to
sort out these terms, we borrow from Vial [24] the trick of making appear a degenerate
U -statistic in such a development (by inserting the right quantity in R0 below). More
explicitly, recalling that
n
X
ϕ(Xi )
vbi
− I(ϕ),
Ibc (ϕ) − I(ϕ) = n−1
1−
fbi
fbi2
i=1
we obtain
Ibc (ϕ) − I(ϕ) = R0 + R1 + R2 + R3 + R4 + R5 ,
(10)
with (we underbrace terms which have been deliberately introduced and removed)
R0 = n
R1 =
Z
−1
n
X
i=1
e i ) − E[ψ1 (Xi )]),
(ψ1 (Xi ) − ψ2 (Xi )fbi + ψ(X
| {z } | {z }
(f (x)fh (x)−1 −1)ϕ(x) dx,
{z
}
|
R2 = n−1
n
X
i=1
R3 = n−1
n
X
i=1
R4 = n−1
n
X
2
ψ3 (Xi )((fh (Xi ) − fbi ) − vbi ),
|{z}
n
X
ψ3 (Xi )b
vi
i=1
R5 = n−1
e i )),
(ψ1 (Xi ) − ψ(X
| {z }
fbi3
ψ3 (Xi )
i=1
( fbi3 −fh (Xi )3 ),
|{z}
(fh (Xi ) − fbi )3
,
fbi
where b
vi appears to be a centering term in R3 . We shall now compute bounds for each
term separately.
Step 1. kn1/2 R0 k2 = O(n−1/2 h−d/2 ). Note that
X
R0 = n−1 (n − 1)−1
(E[uij |Xj ] − uij + E[uij |Xi ] − E[uij ]),
i6=j
17
Integral approximation by kernel smoothing
with uij = ψ2 (Xi )Kij , is a degenerate U -statistic. This is due to the fact that
E[uij |Xi ] = ψ2 (Xi )fh (Xi ) = ψ1 (Xi ),
e j ).
E[uij |Xj ] = (ψ2 f )h (Xj ) = ψ(X
The n(n − 1) terms in the sum are all orthogonal with L2 norm smaller than kuij k2 ,
hence
2
] ≤ C1 h−d ,
(n − 1)2 E[R02 ] ≤ E[u212 ] ≤ kψ2 k2∞ E[K12
because of equation (24) in Lemma 7.
Step 2. n1/2 R1 = O(n1/2 hr ). This is a consequence of equation (18) of Lemma 6, and
from assumption (A3).
e
Step 3. kn1/2 R2 k2 = O(n1/2 hr + hs ). We can rearrange the function ψ1 (x) − ψ(x)
as
with
e
e
ψ1 (x) − ψ(x)
= (ψ1 (x) − ψ1h (x)) + (ψ1h (x) − ψ(x)),
e
kψ1h (x) − ψ(x)k
∞ =
f (x)
ψ1 (x) − ϕ(x)
fh (x)2 h
≤ ψ1 (x) − ϕ(x)
=
ϕ
(fh − f )
fh2
f (x)
fh (x)2
∞
∞
(11)
∞
≤ Chr ,
for some constant C, where the last inequality follows from equation (18) in Lemma 6.
Then we have
R2 ≤ n−1
n
X
i=1
(ψ1h (Xi ) − ψ1 (Xi )) + Chr ,
and by spliting the mean and the variance of the first term we get
"
E
!2 #
n
1X
1
2
(ψ1 (Xi ) − ψ1h (Xi ))
= E[ψ1 (X1 ) − ψ1h (X1 )] + Var(ψ1 (X1 ) − ψ1h (X1 )),
n i=1
n
and we conclude by equations (19) and (20) of Lemma 6 (it is an easy exercise to show
that ψ1 is Nikolski with regularity min(r, s)).
Step 4. kn1/2 R3 k2 = O(n−1/2 h−d/2 ). We first express R3 as a U -statistic. Set
2
Ui = (fh (Xi ) − fbi ) − vbi ,
18
B. Delyon and F. Portier
and rewrite R3 as
R3 = n
−1
n
X
ψ3 (Xi )Ui .
i=1
Consider a sequence of real numbers (xj )1≤j≤p and set
p
1X
xj ,
m=
p j=1
p
p
X
X
1
1
(xj − m)2 =
(x2 − m2 ),
p(p − 1) j=1
p(p − 1) j=1 j
v=
then
p
X
X
1
2
1
2
m −v = 1+
x2j =
m −
xj xk .
p−1
p(p − 1) j=1
p(p − 1)
2
j<k
Applying this with xj = Kij − fh (Xi ) (i is fixed) and p = n − 1 we get
Ui =
2
(n − 1)(n − 2)
X
j6=i,k6=i,j<k
(Kij − fh (Xi ))(Kik − fh (Xi )) =
X
2
ξij ξik ,
(n − 1)(n − 2)
j<k
with
ξij = Kij − fh (Xi ),
ξii = 0.
Then
R3 =
XX
2
ψ3 (Xi )ξij ξik .
n(n − 1)(n − 2) i
j<k
We are going to calculate E[R32 ] by using the Efron–Stein inequality (Theorem 8) and
the moment inequalities (23) to (25) for ξij stated in Lemma 7; in particular, by (23),
E[R32 ] = Var(R3 ). Consider R3 = f (X1 , . . . , Xn ) as a function of the Xi ’s and define
R3′ = f (X1′ , X2 , . . . , Xn ),
′
ξ1i
= h−d K(h−1 (X1′ − Xi )) − fh (X1′ ),
′
ξi1
= h−d K(h−1 (Xi − X1′ )) − fh (Xi ),
′
ξij
= ξij
if i 6= 1 and j 6= 1,
19
Integral approximation by kernel smoothing
where X1′ is a copy of X1 independent from the sample (X1 , . . . , Xn ). Then by the Efron–
Stein inequality (remember that ξii = 0)
1/2
n
kR3 k2 ≤
kR3 − R3′ k2 ,
2
which is of order
n−5/2
X
j<k
≤ n−5/2
′ ′
(ψ3 (X1 )ξ1j ξ1k − ψ3 (X1′ )ξ1j
ξ1k ) +
X
j<k
XX
i
1<k
′ ′
ψ3 (X1 )ξ1j ξ1k − ψ3 (X1′ )ξ1j
ξ1k
+
′
ψ3 (Xi )(ξi1 − ξi1
)ξik
XX
2
1<k
i
2
′
ψ3 (Xi )(ξi1 − ξi1
)ξik
2
= n−5/2 (kT1 k2 + kT2 k2 ).
Noting that the terms in the first sum are orthogonal (by independence of ξij and ξik
conditionally to Xi and (23)) we obtain
kT1 k2 =
(n − 1)1/2 (n − 2)1/2
′ ′
kψ3 (X1 )ξ12 ξ13 − ψ3 (X1′ )ξ12
ξ13 k2
21/2
≤ 21/2 nkψ3 k∞ kξ12 ξ13 k2
2 2
= 21/2 nkψ3 k∞ E[E[ξ12
ξ13 |X1 ]]
1/2
2
= 21/2 nkψ3 k∞ kE[ξ12
|X1 ]k2
= O(nh−d )
by equation (24). Because the terms of the second sum are orthogonal whenever the
values of k are different, we get
kT2 k2 = (n − 1)1/2
X
i
′
ψ3 (Xi )(ξi1 − ξi1
)ξi2
.
2
By first developing and then using that X1′ is an independent copy of X1 , we obtain
X
i
2
′
ψ3 (Xi )(ξi1 − ξi1
)ξi2
2
2
′
2
≤ nE[ψ3 (X3 )2 (ξ31 − ξ31
) ξ32
]
′
′
+ n2 |E[ψ3 (X3 )ψ3 (X4 )(ξ31 − ξ31
)ξ32 (ξ41 − ξ41
)ξ42 ]|
2
′
2
]
≤ kψ3 k∞ {nE[(ξ31 − ξ31
) ξ32
′
′
+ n2 E[|E[(ξ31 − ξ31
)ξ32 (ξ41 − ξ41
)ξ42 |X3 , X4 ]|]}
2 2
= kψ3 k∞ {2nE[ξ31
ξ32 ] + 2n2 E[|E[ξ31 ξ32 ξ41 ξ42 |X3 , X4 ]|]}.
20
B. Delyon and F. Portier
2 2
2
ξ32 ] = E[E[ξ31
|X3 ]2 ] ≤ C1 h−2d and by equation (25),
Then by equation (24), we have E[ξ31
we get
E[|E[ξ31 ξ32 ξ41 ξ42 |X3 , X4 ]|] = E[E[ξ31 ξ41 |X3 , X4 ]2 ]
2
e −1 (X4 − X3 )) ] + 2kf k4∞
≤ 2kf k2∞h−2d E[K(h
Z
e 2 du + 2kf k4 ,
≤ 2kf k3∞h−d K(u)
∞
e is defined in Lemma 7. Bringing everything together and because nhd → ∞, it
where K
holds that
kn1/2 R3 k2 ≤ O(n−1 h−d + n−1/2 h−d/2 ) = O(n−1/2 h−d/2 ).
Step 5. n1/2 R4 = OP (n−1 h−3d/2 ). We start with a lower bound for fbi by proving the
existence of N (ω) such that
∀n ≥ N (ω), ∀i,
Notice that
fbi =
b
< fbi < 2kf k∞ .
2
(12)
1
n
b
f (Xi ) − d K(0) ,
n−1
nh
n
1 X
fb(x) =
K(h−d (x − Xj )),
nhd j=1
due to the almost sure uniform convergence of fb to f (Theorem 1 in [5]) we have with
probability 1 for n large enough
2b
3
< inf fb(x) ≤ sup fb(x) < kf k∞ ,
x∈Q
3
2
x∈Q
and since assumption nhd → ∞, (12) follows. We can now compute the expectation of R4
restricted to {n ≥ N (ω)}. Because (a3 − b3 ) = (a − b)(a2 + ab + b2 ) for any real number a
and b, and by the latter inequality, there exists a constant C > 0 which does not depend
on n or h, such that
|R4 |1n>N (ω) ≤ Cn−1
we have by the Cauchy–Schwarz inequality
n
X
i=1
|fbi − fh (Xi )|b
vi ,
2 1/2
1/2
E[|R4 |1n>N (ω) ] ≤ CE[(fb1 − fh (X1 )) ] E[b
v12 ] .
(13)
21
Integral approximation by kernel smoothing
P
Applying the fact that for any real number a, p1 pj=1 (xj − x)2 ≤
xj = K1j , p = n − 1 and a = fh (X1 ), we obtain that
vb1 ≤
then using (24)
1
p
Pp
i=1 (xj
− a)2 to
n
X
1
ξ2 ,
(n − 1)(n − 2) j=2 1j
4
2 2
E[b
v12 ] ≤ (n − 1)−1 (n − 2)−2 E[ξ12
] + (n − 1)−1 (n − 2)−1 E[ξ12
ξ13 ]
≤ C1 n−3 h−3d + C1 n−2 h−2d
≤ O(n
−2 −2d
h
(14)
),
because nhd goes to infinity. On the other hand using equation (24) again,
2
E[(fb1 − fh (X1 )) ] =
1
E[ξ 2 ] = O(n−1 h−d ).
n − 1 1i
(15)
Putting together (13), (14) and (15),
E[|R4 |1n>N (ω) ] = O(n−1 h−d n−1/2 h−d/2 ) = O(n−3/2 h−3d/2 ).
In particular by Markov’s inequality
P(n3/2 h3d/2 |R4 | > A) ≤ P(n3/2 h3d/2 |R4 |1n>N (ω) > A) + P(n ≤ N (ω))
= A−1 O(1) + P(n ≤ N (ω)).
This proves the boundedness in probability of n3/2 h3d/2 |R4 |.
Step 6. n1/2 R5 = OP (n−1 h−3d/2 + n−3/2 h−2d ). Following (12) since
|R5 |1n>N (ω) ≤ 2b−3 kϕk∞ n−1
n
X
i=1
3
|fbi − fh (Xi )| ,
we can show the convergence in probability of the right-hand side term as in Step 5. We
have indeed by the Rosenthal’s inequality1
#
" n
"
p#
n
X
X
p
−p
−1
ξ1i
|fbi − fh (Xi )| = (n − 1) E
E n
i=2
i=1
p/2
2
≤ C2 n−p {(nE[ξ12
])
≤ C1 C2 {n
−p/2 −pd/2
h
+ nE[|ξ12 |p ]}
+n
1−p −(p−1)d
h
(16)
},
P
1
2
p/2 ] +
(Si , Fi )i∈N and 2 ≤ p < +∞, we have E[|Sn |p ] ≤ C2 {E[( n
i=1 E[Xi |Fi−1 ])
Pn For a martingale
p
i=1 E|Xi | }, where Xi = Si − Si−1 (see, e.g., [13], pp. 23–24).
22
B. Delyon and F. Portier
where the latter inequality is due to equation (24). Hence, with p = 3
E[|R5 |1n>N (ω) ] ≤ C1 C2 {n−3/2 h−3d/2 + n−2 h−2d }
and we conclude as in Step 5.
Putting together the steps 1 to 6, and taking into account, concerning R5 , that
n−3/2 h−2d = (n−1/2 h−d/2 )(n−1 h−3d/2 ), we obtain (ii).
Proof of (i). For (i), we use a shorter expansion which leads to an actually much simpler
proof:
and
1
1
fh (Xi ) − fbi (fh (Xi ) − fbi )2
=
,
+
+
fh (Xi )2
fbi fh (Xi )
fbi fh (Xi )2
with
b
I(ϕ)
− I(ϕ) = R0 + R1 + R2 + R5′ ,
R5′ = n−1
n
X
ψ2 (Xi )
i=1
(fh (Xi ) − fbi )2
.
fbi
The terms R0 , R1 and R2 have already been treated in the steps 1, 2 and 3 of the proof
of (i). The term R5′ is bounded exactly as R5 but now we use (16) with p = 2 instead of
p = 3, to obtain
E[|R5′ |1n>N (ω) ) ≤ C1 C2 n−1 h−d
and we get n1/2 |R5′ | = OP (n−1/2 h−d ).
6.2. Proofs of Theorems 2 and 3
Let us define
Mn = n−1
n
X
i=1
e i ) − E[ψ1 (X1 ) − ψ(X
e 1 )],
ψ1 (Xi ) − ψ(X
Un = n−1 (n − 1)−1
X
cij ,
i6=j
e 1 )] +
Bn = E[ψ1 (X1 ) − ψ(X
Z
(f (x)fh (x)−1 − 1)ϕ(x) dx
23
Integral approximation by kernel smoothing
with cjk = ajk − bjk , and for j 6= k,
ajk = E[ψ3 (X1 )ξ1j ξ1k |Xj Xk ],
bjk = ujk − E[ujk |Xj ] − E[ujk |Xk ] + E[ujk ],
where ujk has been defined at the beginning of step 3. Both proofs of Theorems 2 and 3
rely on the following lemma which turns Theorem 1 in a suitable way for weak convergence
issues.
Lemma 5. Under the assumptions of Theorem 1, we have
Ibc (ϕ) − I(ϕ) = Bn + Un + Mn + OP (n−3/2 h−3d/2 ).
Moreover, we have Bn = OP (hr ), Un = OP (n−1 h−d/2 ) and Mn = OP (n−1/2 (hs + hr )).
Proof. By using the decomposition (10) and since Bn + Mn = R1 + R2 , we have
Ibc (ϕ) − I(ϕ) = R0 + R1 + R2 + R3 + R4 + R5
= Bn + Mn + Un + (R0 + R3 − Un ) + R4 + R5 .
We have already shown that R4 + R5 = OP (n−3/2 h−3d/2 + n−2 h−2d ) (this is exactly steps
5 and 6 of the proof of Theorem 1). By definition of Un , we have
XX
R0 + R3 − Un = n−1 (n − 1)−1 (n − 2)−1
(ψ3 (Xi )ξij ξik − ajk )
i
j6=k
which is a completely degenerate U -statistic (R3 is near to be completely degenerate and
ajk = E[ψ3 (X1 )ξ1j ξ1k |Xj Xk ] appears as the good centering term). The order 2 moments
2 2
of this quantity are of order n−3 E[ψ3 (X1 )2 ξ12
ξ13 ] ∝ n−3 h−2d . Hence, we have shown that
−3/2 −d
R0 + R3 − Un = OP (n
h ), which completes the first part of the proof. To obtain the
bounds in probability, for Un we just use step 1 and 4 of the proof of Theorem 1, for Mn
we compute the L2 norm as follows. We have
e 1 )k
kMn k2 = n−1/2 kψ1 (X1 ) − ψ(X
2
e 1 )k )
≤ n−1/2 (kψ1 (X1 ) − ψ1h (X1 )k2 + kψ1h (X1 ) − ψ(X
2
≤ Cn−1/2 (hs + hr ),
for some constant C, where the last inequality is obtained using equation (11) for the
term in the right and equation (20) in Lemma 6 for the term in the right.
Remark 9. Under the assumption of Theorem 1, one may show that
X
−3/2
2
b
ψ3 (Xi )ξij
+ OP ((nhd )
),
I(ϕ)
− I(ϕ) = Ibc (ϕ) − I(ϕ) + n−1 (n − 1)−2
i,j
24
B. Delyon and F. Portier
where the OP comes from R4 and the other remainder term corresponds to the diagonal term of the U -statistic R3 . This term equals (n − 1)−1 E[ψ2 (X1 )(K12 − fh (X1 ))2 ] =
O(n−1 h−d ) plus oP (n−1 h−d/2 ), as a consequence, when h is such that nh2(s+d) → 0 and
nhr+d → 0, the leading term of the decomposition is a constant.
6.2.1. Proof of Theorem 2
By Lemma 5 and the assumptions on h we have
nhd/2 (Bn + Mn + R4 + R5 ) = OP (n3/2 hr+d/2 + n1/2 (hr+d/2 + hs+d/2 ) + n−1/2 h−3d/2 )
= oP (1).
To derive the limiting distribution of nhd/2 Un , we apply Theorem 1 in [12], quoted
below (Theorem 11), with Hn (Xj , Xk ) = (n − 1)−1 hd/2 (cjk + ckj ) where cjk = ajk − bjk ,
has been defined at the beginning of Section 6.2. The asymptotic variance v1 is the limit
2
of the quantity n2 E[Hn (X1 , X2 )2 ] asymptotically equivalent to
hd (E[c212 ] + E[c12 c21 ]).
To compute this easily, we introduce the function ξi (x) = KRh (x − Xi ) − fh (x). First, use
some algebra to obtain the formula b12 = ψ2 (X1 )ξ2 (X1 ) − ψ2 (x)ξ2 (x)f (x) dx, then it
follows that
c12 = a12 − b12
Z
Z
= ψ3 (x)ξ1 (x)ξ2 (x)f (x) dx − ψ2 (X1 )ξ2 (X1 ) + ψ2 (x)ξ2 (x)f (x) dx
=
Z
+
=
Z
+
=
Z
+
+
(ψ3 (x)f (x)ξ2 (x) − ψ2 (X1 )ξ2 (X1 ))Kh (x − X1 ) dx
Z
ψ3 (x)f (x)ξ2 (x)(fh (x) − f (x)) dx
(ψ2 (x)ξ2 (x) − ψ2 (X1 )ξ2 (X1 ))Kh (x − X1 ) dx
Z
ψ3 (x)ξ2 (x)(f (x) − fh (x))Kh (x − X1 ) dx +
Z
ψ3 (x)f (x)ξ2 (x)(fh (x) − f (x)) dx
(ψ2 (x)Kh (x − X2 ) − ψ2 (X1 )Kh (X1 − X2 ))Kh (x − X1 ) dx
Z
Z
(−(ψ2 (x)fh (x) − ψ2 (X1 )fh (X1 )) + ψ3 (x)ξ2 (x)(f (x) − fh (x)))Kh (x − X1 ) dx
ψ3 (x)f (x)ξ2 (x)(fh (x) − f (x)) dx.
25
Integral approximation by kernel smoothing
Because Kh integrates to 1, it is not hard to see that the last two terms in the previous
equation will be negligible in the computation of v1 . As a consequence, hd E[c212 ] has the
same limit as
2
Z Z Z
d
h
(ψ2 (x)Kh (x − z) − ψ2 (y)Kh (y − z))Kh (x − y) dx f (y)f (z) dy dz
=
Z Z Z
= VK
Z
(ψ2 (y + hu)K(u + v) − ψ2 (y)K(v))K(u) du
2
f (y)f (y − hv) dy dv
ψ2 (y)2 f (y)2 dy + o(1)
R R
with VK = ( (K(u + v) − K(v))K(u) du)2 dv and where the first equality follows from
a change of variables and the last representation follows from the Lebesgue dominated
theorem. Following the same steps as previously, we obtain an similar expression for
hd E[c12 c21 ] and then we get
Z
v1 = 2VK ϕ(y)2 f (y)−2 dy.
It remains to check the conditions of Theorem 11. Clearly, the computation of v1 provides
that E[Hn (X1 , X2 )2 ] ≈ n−2 . We obtain similarly that E[Hn (X1 , X2 )4 ] = O(n−5 h−d ) and
E[Gn (X1 , X2 )2 ] = O(n−5 hd ) which implies the conditions of the theorem.
6.2.2. Proof of Theorem 3
By (B1) and Lemma 9, there exists M2 > 0 such that ϕ is H(min(s, 1/2), M2). Then we
can apply Lemma 5 and by assumption on h, we obtain that
1/2
(nh−1 )
(Bn + Un + R4 + R5 ) = OP (n1/2 hr−1/2 + n−1/2 h−(d+1)/2 + n−1 h−(3d+1)/2 )
= oP (1).
Since Mn is a sum of independent variables with zero-mean, we can apply the central
limit theorem by checking the Lindeberg condition (see, e.g., [13], Chapter 3). Now we
only have to compute the asymptotic variance v2 defined as the limit of
2
2
e 1 ))) = h−1 E[(ψ1 (X1 ) − ψ(X
e 1 )) ] − h−1 E[ψ1 (X1 ) − ψ(X
e 1 )] .
Var(h−1/2 (ψ1 (X1 ) − ψ(X
On the one hand, by equations (11) and (19), we have for some constant C
e 1 ))k ≤ Chr−1/2 ,
kh−1/2 (ψ1h (X1 ) − ψ(X
2
h−1/2 |E[ψ1 (X1 ) − ψ1h (X1 )]| ≤ Chr−1/2 ,
as a consequence, we get
e 1 ))) = h−1 kψ1 (X1 ) − ψ1h (X1 )k2 + o(1).
Var(h−1/2 (ψ1 (X1 ) − ψ(X
2
26
B. Delyon and F. Portier
On the other hand, for every x ∈ Q, we have
ψ1 (x) − ψ1h (x) =
Z
Qc
(ψ1 (x) − ψ1 (y))Kh (x − y) dy +
= ψ1 (x)
Z
Qc
Kh (x − y) dy +
Z
Q
Z
Q
(ψ1 (x) − ψ1 (y))Kh (x − y) dy
(ψ1 (x) − ψ1 (y))Kh (x − y) dy,
where Qc stands for the complement of the set Q in Rd . Because ψ1 is Nikolski with
regularity min(s, r) inside Q, we use equation (20) of Lemma 6 to show that the L2 norm of the right-hand side term is of order hmin(s,r) . Clearly, since min(s, r) > 1/2 we
have
Z
−1/2
−1
e
Var(h
(ψ1 (X1 ) − ψ(X1 ))) = h
Kh (X1 − y) dy + o(1)
ψ1 (X1 )
Qc
2
and it remains to apply Lemma 10 to derive the stated limit.
6.3. Proof of the Theorem 4
By equation (3), we are interested in the asymptotic law of
n
−1/2
n
X
σ(Xi )ψ(Xi )
i=1
fbi
ei + n
−1/2
n
X
g(Xi )ψ(Xi )
fbi
i=1
−
Z
!
g(x)ψ(x) dx .
By Lemma 1, the right-hand side term goes to 0 in probability. For the other term, we
use the decomposition A1 + A2 , with
A1 = n−1/2
n
X
σψ(Xi )
i=1
f (Xi )
ei
and A2 = n−1/2
n
X
σψ(Xi )(f (Xi ) − fb(Xi ))
i=1
fbi f (Xi )
ei ,
where σψ(Xi ) = σ(Xi )ψ(Xi ). We define F as the σ-field generated by the set of random
variables {X1 , X2 , . . .}. We get
E[A22 |F ] = n−1
then, one has
n
X
σψ(Xi )2 (f (Xi ) − fbi )2
i=1
fbi2 f (Xi )2
,
n
−1
X
2
(f (Xi ) − fbi ) .
kσψk2∞ n−1
E[A22 |F ] ≤ b2 inf fbi2
i
i=1
27
Integral approximation by kernel smoothing
For the term on the left, since σψ has support Q we can use (12), that is for n large
enough, it is bounded. For the right-hand side term, it follows that
n−1
n
X
i=1
2
(f (Xi ) − fbi ) ≤ 2n−1
n
X
i=1
2
(f (Xi ) − fh (Xi )) + 2n−1
n
X
i=1
2
(fh (Xi ) − fbi ) ,
and then using equation (18) in Lemma 6 and (16) for p = 2 we provide the bound
n−1
n
X
i=1
2
(f (Xi ) − fbi )
1
≤ C(h2r + n−1 h−d )
(17)
for some C > 0. Therefore, we have shown that E[A22 |F ] → 0 in probability. Since for any
ε > 0, P(|A2 | > ε|F ) ≤ ε−2 E[A22 |F ], it remains to note that the sequence P(|A2 | > ε|F ) is
uniformly integrable to apply the Lebesgue domination theorem to get
P(A2 > ε) −→ 0.
To conclude, we apply the central limit theorem to A1 and the statement follows.
6.4. Some lemmas
6.4.1. Inequalities
Lemma 6. For any function g : Rd → R, recall that gh (x) =
assumptions (A1), (A2) and (A4), it holds that
R
g(x − hu)K(u) du. Under
kfh − f k∞ ≤ CK hr kf (r) k∞ ,
Z
r
(r)
|E[ϕ(X1 ) − ϕh (X1 )]| ≤ CK h kf k∞ |ϕ(x)| dx,
kϕh (X1 ) − ϕ(X1 )k2 ≤ CK M hs ,
(18)
(19)
(20)
where CK is a positive constant that depends K only.
Proof. We start by proving (19) and (20) assuming that (18) holds. For the mean: using
Fubini’s theorem, we have
Z
E[ϕ(X1 ) − ϕh (X1 )] = (ϕ(x) − ϕh (x))f (x) dx
=
Z
ϕ(x)f (x) − ϕ(x)fh (x) dx,
hence
|E[ϕ(X1 ) − ϕh (X1 )]| ≤ kf (x) − fh (x)k∞
Z
|ϕ(x)| dx,
28
B. Delyon and F. Portier
which by (18) gives
r
|E[ϕ(X1 ) − ϕh (X1 )]| ≤ CK h kf
(r)
k∞
Z
|ϕ(x)| dx.
This is (19). We turn now to (20):
2
E[(ϕh (X1 ) − ϕ(X1 )) ] =
Z Z
(ϕ(x − hu) − ϕ(x))K(u) du
2
f (x) dx.
(21)
We now use the Taylor formula with Lagrange remainder applied to g(t) = ϕ(x − tu)
with order k equal to the largest integer smaller than s:
ϕ(x − hu) − ϕ(x) =
=
k−1
X
j=1
hj (j)
g (0) +
j!
k
X
hj
j!
j=1
g
(j)
(0) +
Z
h
g (k) (t)
0
Z
h
0
(h − t)k−1
dt
(n − 1)!
(g (k) (t) − g (k) (0))
(h − t)k−1
dt.
(n − 1)!
The first term is a polynomial in u which will vanish after insertion in (21) because K is
orthogonal the first non-constant polynomial of degree ≤ r. The second term is bounded
as
Z h
Z h
(h − t)k−1
k k−1
(k)
(k)
dt ≤ |u| h
(g (t) − g (0))
|ϕ(k) (x − tu) − ϕ(k) (x)| dt.
(k − 1)!
0
0
Hence,
Z
(ϕ(x − hu) − ϕ(x))K(u) du
≤h
k−1
Z
0
hZ
(k)
|ϕ
(22)
(k)
(x − tu) − ϕ
k
(x)||u| K(u) du dt
and by the generalized Minkowski inequality ([10] page 194)2
k(ϕh − ϕ)(X1 )k2 ≤ h
k−1
Z Z
≤ M hk−1
2 For
Z
(k)
|ϕ
(k)
(x − tu) − ϕ
1/2
(x)| u K(u) 10≤t≤h f (x) dx
du dt
1/2
(|tu|2α |u|2k K(u)2 )
2 2k
2
10≤t≤h du dt
any non-negative measurable function g(·, ·) on Rk+d ,
1/2
2 1/2 Z Z
Z Z
dy.
≤
g(y, x)2 dx
g(y, x) dy dx
Integral approximation by kernel smoothing
Z
1/2
= M (1 + α)−1 hk+α (|u|2α+2k K(u)2 ) du.
29
This implies (20). Concerning (18), we use (22) with f and k = r to get that
|fh (x) − f (x)| ≤ h
r−1
Z
0
hZ
|f (r) (x + tu)||u|r K(u) du dt,
the latter is bounded by a constant times hr .
The following lemma gives some bounds on the conditional moments of ξ12 that are
useful in the proof of Theorem 1.
Lemma 7. Let ξij = Kij − fh (Xi ), under (A1) and (A2)
E[ξ12 |X1 ] = 0,
E[|ξ12 |p |X1 ] ≤ 2p E[|K12 |p |X1 ] ≤ C1 h−(p−1)d ,
e
with K(x)
=
R
e −1 (X2 − X1 )) + kf k∞ ),
|E[ξ13 ξ23 |X1 , X2 ]| ≤ kf k∞ (h−d K(h
(23)
(24)
(25)
|K(x − y)K(y)| dy and C1 > 0.
Proof. The first equation is trivial. For the second equation, the triangular inequality
and the Jensen inequality provide
Z
p
E[|ξ12 |p |X1 ] ≤ 2p E[|K12 |p |X1 ] = 2p h−(p−1)d |K(u)| f (X1 − hu) dx,
and the third one is derived by
|E[ξ13 ξ23 |X1 , X2 ]| = |E[ξ13 K23 |X1 , X2 ]|
Z
=
(Kh (X1 − x) − fh (X1 ))Kh (X2 − x)f (x) dx
=
Z
(Kh (X1 − X2 + hu) − fh (X1 ))K(u)f (X2 − hu) du
e −1 (X2 − X1 )) + kf k∞ ).
≤ kf k∞ (h−d K(h
The Efron–Stein inequality helps to bound the L2 moments of estimators. For the
proof, we refer to the original paper [6] but also to [2].
Theorem 8 (Efron–Stein inequality). Let X1 , . . . , Xn be an i.i.d. sequence, X1′ be
an independent copy of X1 and f be a symmetric function of n variables, then
Var(f (X1 , . . . , Xn )) ≤
n
2
E[(f (X1 , . . . , Xn ) − f (X1′ , X2 , . . . , Xn )) ].
2
30
B. Delyon and F. Portier
6.4.2. Measure results
Lemma 9. Let s > 0 and M1 > 0, suppose that the support of ϕ is a convex body Q and
that ϕ is H(s, M1 ) on Q, then there exists M > 0 such that ϕ is H(min(s, 1/2), M ) on
Rd .
Proof. We have
Z
2
|ϕ(x + u) − ϕ(x)| dx
=
Z
2
{x∈Q,x+u∈Q}
+
=
≤
Z
Z
Z
|ϕ(x + u) − ϕ(x)| dx +
Z
ϕ(x + u)2 dx
{x∈Q,x+u∈Q}
/
ϕ(x)2 dx
{x∈Q,x+u∈Q}
/
2
{x∈Q,x+u∈Q}
{x∈Q,x+u∈Q}
|ϕ(x + u) − ϕ(x)| dx +
|ϕ(x + u) − ϕ(x)|
2
Z
ϕ(x)2 (1{x−u∈Q}
+ 1{x+u∈Q}
) dx
/
/
Q
dx + kϕk2∞
Z
1{dist(x,∂Q)≤|u|} dx
≤ M1 |u|2s + kϕk2∞ ξd−1 (Q)|u|,
where ξd−1 (S) is called a Quermassintegrale of Minkowski and dist stands for the Euclidean distance in Rd . The last inequality follows from the fact that ϕ is H(s, M1 ) on Q
and by the Steiner’s formula stated, for instance, in [9], Theorem 3.2.35, page 271.
Lemma 10. Under the assumption (A4), if Q is a compact set with C 2 boundary and ψ
is continuous
2
Z
Z Z
Kh (x − y) dy ψ(x) dx =
lim h−1
LQ (x)ψ(x) dHd−1 (x),
h→0
Q
Qc
∂Q
where
LQ (x) =
ZZ
min(hz, u(x)i, hz ′ , u(x)i)+ K(z)K(z ′ ) dz dz ′ ,
and Hd−1 stands for the (d − 1)-dimensional Hausdorff measure, u(x) is the normal outer
vector of Q at the point x.
Proof. Let us start with an estimate of the integral over Qc having a simpler dependency
w.r.t. h. We define the function
τ (x) = (1x∈Q − 1x∈Q
/ ) dist(x, ∂Q).
31
Integral approximation by kernel smoothing
This function is C 2 in the neighborhood of ∂Q and its gradient −u(x) is, for x ∈ ∂Q, the
normal inner vector (since ∂Q is C 2 , using a local parametrization of Q, we are reduced
to the case where ∂Q is a piece of hyperplane). Then
Z
Qc
Kh (x − y) dy =
=
=
Z
Z
Z
1x+hz∈Qc K(z) dz
1τ (x+hz)≤0K(z) dz
1τ (x)−hhz,u(x)i≤ah2 K(z) dz,
where a actually depends on x and z but is smaller than a constant related to the
curvature of ∂Q. Hence,
Z
Qc
≤
≤
Kh (x − y) dy −
Z
Z
Z
1τ (x)−hhz,u(x)i≤0K(z) dz
|1τ (x)−hhz,u(x)i≤ah2 − 1τ (x)−hhz,u(x)i≤0|K(z) dz
1|τ (x)−hhz,u(x)i|≤|a|h2 K(z) dz
≤ a0 h1τ (x)≤m0h
for some a0 and m0 , because the integration domain is a band of width |a|h. Hence,
Z
Qc
Kh (x − y) dy =
Z
1τ (x)≤hhz,u(x)iK(z) dz + a1 (x)h1τ (x)≤m0 h ,
where a1 is bounded. Since the second term has a O(h2 ) integral over Q, its contribution
in the limit is negligible, and it suffices to prove that
lim h
h→0
−1
Z Z
1τ (x)≤hhz,u(x)iK(z) dz
Q
2
ψ(x) dx =
Z
LQ (x)ψ(x) dHd−1 (x).
∂Q
By setting
ϕ(x, t) =
Z
10≤t≤hz,u(x)i K(z) dz
the latter equality can be rewritten as
Z
Z
lim h−1 ϕ(x, h−1 τ (x)) dx =
h→0
∂Q
2
ψ(x)1x∈Q ,
LQ (x)ψ(x) dHd−1 (x).
32
B. Delyon and F. Portier
From Proposition 3, page 118 of [7], we have for any integrable function q and f Lipschitz
with essinf |∇f | > 0:
Z
q(x) dx =
f ≥0
Z
0
∞ Z
f =s
q(x)
dHd−1 (x) ds
|∇f (x)|
hence, with f (x) = h−1 τ (x) and q(x) = ϕ(x, h−1 τ (x)), we obtain
h−1
Z
ϕ(x, h−1 τ (x)) dx =
Z
0
∞ Z
Letting h → 0, we get
lim h
h→0
−1
Z
ϕ(x, h
−1
τ (x)) dx =
τ =hs
Z
0
=
Z
∞ Z
∂Q
We can write
Z
∞
R
ϕ(x, s) dHd−1 (x) ds.
Z
ϕ(x, s) dH
∂Q
∞
d−1
(x) ds
ϕ(x, s) ds dHd−1 (x).
0
ϕ(x, s) ds as
ϕ(x, s) ds = ψ(x)
0
= ψ(x)
ZZZ
ZZ
10≤s≤hz,u(x)i 10≤s≤hz′ ,u(x)i K(z)K(z ′ ) dz dz ′ ds
min(hz, u(x)i, hz ′ , u(x)i)+ K(z)K(z ′ ) dz dz ′ .
6.4.3. Weak convergence for degenerate U -statistics
Theorem 11 (Hall (1984), [12]). Let Hn : Rd × Rd → R, with Hn symmetric, assume
that E[Hn (X1 , X2 )|X1 ] = 0 and E[Hn (X1 , X2 )2 ] < +∞. If
E[Gn (X1 , X2 )2 ] + n−1 E[Hn (X1 , X2 )4 ] n→+∞
−→ 0,
E[Hn (X1 , X2 )2 ]2
P
with Gn (x, y) = E[Hn (X1 , x)Hn (X1 , y)], then
j<k H(Xj , Xk ) is asymptotically normally distributed with zero mean and variance given by
n2
2
2 E[H(X1 , X2 ) ].
Acknowledgements
The authors would like to thank Céline Vial for helpful comments and advice on a latter
version of this article.
Research supported by the Fonds de la Recherche Scientifique (FNRS) A4/5 FC
2779/2014–2017 No. 22342320.
Integral approximation by kernel smoothing
33
References
[1] Bickel, P.J., Klaassen, C.A.J., Ritov, Y. and Wellner, J.A. (1993). Efficient and
Adaptive Estimation for Semiparametric Models. Johns Hopkins Series in the Mathematical Sciences. Baltimore, MD: Johns Hopkins Univ. Press. MR1245941
[2] Boucheron, S., Lugosi, G. and Bousquet, O. (2004). Concentration inequalities. In
Advanced Lectures on Machine Learning. Lecture Notes in Computer Science 3176
208–240. Berlin: Springer.
[3] Chen, S.X. (1999). Beta kernel estimators for density functions. Comput. Statist. Data
Anal. 31 131–145. MR1718494
[4] Delecroix, M., Hristache, M. and Patilea, V. (2006). On semiparametric M estimation in single-index regression. J. Statist. Plann. Inference 136 730–769.
MR2181975
[5] Devroye, L.P. and Wagner, T.J. (1980). The strong uniform consistency of kernel density estimates. J. Multivariate Anal. 5 59–77.
[6] Efron, B. and Stein, C. (1981). The jackknife estimate of variance. Ann. Statist. 9 586–
596. MR0615434
[7] Evans, L.C. and Gariepy, R.F. (1992). Measure Theory and Fine Properties of Functions.
Studies in Advanced Mathematics. Boca Raton, FL: CRC Press. MR1158660
[8] Evans, M. and Swartz, T. (2000). Approximating Integrals Via Monte Carlo and Deterministic Methods. Oxford Statistical Science Series. Oxford: Oxford Univ. Press.
MR1859163
[9] Federer, H. (1969). Geometric Measure Theory. Die Grundlehren der Mathematischen
Wissenschaften 153. New York: Springer. MR0257325
[10] Folland, G.B. (1999). Real Analysis: Modern Techniques and Their Applications, 2nd ed.
Pure and Applied Mathematics (New York). New York: Wiley. MR1681462
[11] Gamboa, F., Loubes, J.-M. and Maza, E. (2007). Semi-parametric estimation of shifts.
Electron. J. Stat. 1 616–640. MR2369028
[12] Hall, P. (1984). Central limit theorem for integrated square error of multivariate nonparametric density estimators. J. Multivariate Anal. 14 1–16. MR0734096
[13] Hall, P. and Heyde, C.C. (1980). Martingale Limit Theory and Its Application: Probability and Mathematical Statistics. New York: Academic Press. MR0624435
[14] Härdle, W. (1990). Applied Nonparametric Regression. Econometric Society Monographs
19. Cambridge: Cambridge Univ. Press. MR1161622
[15] Härdle, W., Marron, J.S. and Tsybakov, A.B. (1992). Bandwidth choice for average
derivative estimation. J. Amer. Statist. Assoc. 87 218–226. MR1158640
[16] Härdle, W. and Stoker, T.M. (1989). Investigating smooth multiple regression by the
method of average derivatives. J. Amer. Statist. Assoc. 84 986–995. MR1134488
[17] Jones, M.C. (1993). Simple boundary correction for kernel density estimation. Stat. Comput. 3 135–146.
[18] Oh, M.-S. and Berger, J.O. (1992). Adaptive importance sampling in Monte Carlo integration. J. Stat. Comput. Simul. 41 143–168. MR1276184
[19] Robinson, P.M. (1988). Root-N -consistent semiparametric regression. Econometrica 56
931–954. MR0951762
[20] Silverman, B.W. (1986). Density Estimation for Statistics and Data Analysis. Monographs
on Statistics and Applied Probability. London: Chapman & Hall. MR0848134
[21] Stone, C.J. (1975). Adaptive maximum likelihood estimators of a location parameter.
Ann. Statist. 3 267–284. MR0362669
34
B. Delyon and F. Portier
[22] Stone, C.J. (1980). Optimal rates of convergence for nonparametric estimators. Ann.
Statist. 8 1348–1360. MR0594650
[23] Tsybakov, A.B. (2009). Introduction to Nonparametric Estimation. Springer Series in
Statistics. New York: Springer. MR2724359
[24] Vial, C. (2003). Deux contributions à l’étude semi-paramétrique d’un modèle de régression.
Ph.D. thesis, Univ. Rennes.
[25] Zhang, P. (1996). Nonparametric importance sampling. J. Amer. Statist. Assoc. 91 1245–
1253. MR1424622
Received September 2014 and revised March 2015
| 10math.ST
|
A Deterministic and Generalized Framework for Unsupervised Learning with
Restricted Boltzmann Machines
Eric W. Tramel,∗ Marylou Gabrié,† Andre Manoel,‡ Francesco Caltagirone,§ and Florent Krzakala¶
arXiv:1702.03260v2 [cs.LG] 2 Oct 2017
(Dated: October 3, 2017)
Restricted Boltzmann machines (RBMs) are energy-based neural-networks which are commonly
used as the building blocks for deep architectures neural architectures. In this work, we derive a
deterministic framework for the training, evaluation, and use of RBMs based upon the ThoulessAnderson-Palmer (TAP) mean-field approximation of widely-connected systems with weak interactions
coming from spin-glass theory. While the TAP approach has been extensively studied for fullyvisible binary spin systems, our construction is generalized to latent-variable models, as well as to
arbitrarily distributed real-valued spin systems with bounded support. In our numerical experiments,
we demonstrate the effective deterministic training of our proposed models and are able to show
interesting features of unsupervised learning which could not be directly observed with sampling.
Additionally, we demonstrate how to utilize our TAP-based framework for leveraging trained RBMs
as joint priors in denoising problems.
PACS numbers: 05.10.-a,05.90.+m
I.
INTRODUCTION
The past decade has witnessed a groundswell of research
in machine learning, bolstered by the deep learning revolution and the resurgence of neural networks [1]. Since their
inception, researchers have identified the deep connection
between neural networks and statistical mechanics. Perhaps the most well-known unsupervised neural models
studied through the lens of statistical physics have been
the the Hopfield model [2, 3] and the Boltzmann machine
[4]. These models were proposed from a connectionist
perspective of cognitive science and were studied in the
context emergent representation in unsupervised machine
learning.
We can look to the Hopfield model to directly observe
some of the contributions of physics to both machine
learning and cognitive sciences. For example, by applying
techniques from the study of spin-glasses, Amit et al. [3]
were famously able to derive the memory capacity of the
Hopfield model and provide a concrete understanding of
the dynamics of the model via the study of its phase
transitions. This fundamental understanding of the behavior of the Hopfield model has provided insight into
the complexities of associative memory.
The closely related Boltzmann Machine is an undirected
stochastic neural network which finds its physics parallel
in Ising spin glass models [5]. Specifically, for this model,
∗
†
‡
§
¶
E.W.T. at OWKIN, 6 Rue de la Pierre Levée, 75011 Paris, France
M. G. at Laboratoire de Physique Statistique, Ecole Normale
Supérieure, 24 rue Lhomond, 75005 Paris, France.
A. M. at Neurospin, CEA, Université Paris-Saclay, bât. 145,
91191, Gif-sur-Yvette, France.
F.C. at Noah’s Ark Lab Paris, Huawei Technologies Co. Ltd.,
92100 Boulogne-Billancourt, France.
F.K. at Laboratoire de Physique Statistique, Ecole Normale
Supérieure, 24 rue Lhomond, 75005 Paris, France, & Université
Pierre et Marie Curie, Sorbonne Universités, 75005 Paris, France.
one is interested in the inverse problem: learning the couplings between spins in order to generate a particular set
of configurations at equilibrium. The process of learning
couplings, or training, is often referred to as the inverse
Ising problem in the physics literature [6–8]. However,
because couplings only exist between pairs of spins for
the fully-visible Ising spin-glass, such models have limited
practical application as they cannot successfully capture
higher-order correlations which might exist in a set of
training configurations.
For this reason, the general Boltzmann machine introduces a set of unobserved latent spins. The effect of these
latent spins is to abstract high-order correlations within
the set of observed spins. While an optimal training of
the couplings would potentially lead to a very effective
general model of high-dimensional joint distributions, the
intractability of this joint latent model confounds the
practical application of general Boltzmann machines.
A restricted Boltzmann Machine (RBM) is a special
case of the general Boltzmann machine, where couplings
only exist between latent and observed spins. This bipartite structure is key to the efficient and effective training
of RBMs [9]. RBMs have found many applications in
machine learning problems as diverse as dimensionality
reduction [10], classification [11], collaborative filtering
[12], feature learning [13], and topic modeling [14]. Additionally, RBMs can be stacked into multi-layer neural
networks, which have played a historically fundamental
role in pre-training deep network architectures [15, 16].
These constructions, known as deep belief networks, were
the first truly deep neural architectures, leading to the
current explosion of activity in deep learning [17]. While
access to vast training datasets has made such pre-training
dispensable for certain tasks, RBMs remain a fundamental tool in the theory of unsupervised learning. As such,
a better understanding of RBMs can be key to future
developments in emergent machine intelligence.
To date, the most popular and effective approaches
to training RBMs have centered on differing flavors of
2
short-chain Monte Carlo sampling [9, 18], which we cover
in detail in the sequel. While such techniques can yield
trained RBMs which produce sampled configurations very
similar to the target dataset, and can be used in a number
of applications as detailed previously, they do not bridge
the gap in understanding what the RBM has learned.
Furthermore, understanding the modes, or internal representations, of the RBM with sampling-based frameworks
have mostly consisted of subjective comparisons of sampled configurations as well as a subjective analysis of the
couplings themselves, often referred to as receptive fields
in the machine learning literature.
Additionally, comparing two trained models, or even
monitoring the training of one model, becomes problematic when using sampling-based investigative tools. For
example, annealed techniques [19] can provide estimates
of the log-likelihood of a model, but only at a large computational cost [20, 21]. At a much lower computational
cost, pseudo-likelihoods can be used to monitor training,
but the estimates produced in this manner are inaccurate, as compared to annealed importance sampling (AIS)
[19], and even AIS can fail to detect model divergence in
practice [22].
In the present work, we seek to address these concerns
by developing a deterministic framework to train, compare, and analyze RBMs, as well as to leverage their
modeling power for inference tasks. We accomplish this
via statistical physics techniques through the use of the
Thouless-Anderson-Palmer (TAP) formalism of spin-glass
theory [5, 23–25]. In this manner, we produce a model
which no longer refers to a stochastic model possessing
an intractable Gibbs measure, but to a TAP machine:
an entirely self-consistent mean-field model which operates as a classical RBM, but which admits deeper introspection via deterministic inference. TAP machines
also naturally handle non-binary variables as well as deep
architectures. While Deep Boltzmann Machines’ (DBMs)
[26] state-of-the-art training algorithms mix both Monte
Carlo sampling and “naïve” mean-field approximation, a
deep TAP machine relies entirely on the TAP mean-field
approximation.
Under this interpretation, a TAP machine is not a generative probabilistic model, but a deterministic model
defining a set of representational magnetizations for a
given training dataset. Advantageously, this learning output can be computed exactly in finite time by converging
a fixed-point iteration, in contrast to the indeterminate
stopping criterion of Markov-chain Monte Carlo sampling.
This is a major distinction between the TAP machine
and the classical RBM, for which the true probability
density function is intractable. At its core, the TAP machine training consists of arranging the minima, solutions,
in the proposed TAP-approximated free energy so as to
maximize the correlation between these solutions and the
dataset. In our experiments, we demonstrate how to track
the growth and geometry of these solutions as a novel way
to investigate the progress of unsupervised learning. We
also show how to use a trained TAP machine as a prior
for inference tasks.
The paper is organized as follows. In Sec. II we formally describe the classical binary RBM and review the
literature on RBM training and analysis. Subsequently,
in Sec. III, we describe our proposed modification of
the binary RBM to a model with arbitrary real-valued
distributions with bounded support. Next, in Sec. IV,
we briefly describe how to apply belief-propagation to
perform inference in the setting of real-valued spins. The
details of this approach are pedagogically described in
Appendices A & B. In Sec. V we derive the TAP approximation of the real-valued RBM via a high-temperature
expansion of a two-moment Gibbs free energy. Then, in
Sec. VI, we detail how to convert this approximation to
a practical training algorithm. In Sec. VII, we conduct a
series of experiments on real datasets, demonstrating how
to use the properties of the TAP machine interpretation
to provide insight into the unsupervised learning process.
We additionally show how to use a trained model for
bit-flip correction as a simple example of leveraging a
TAP machine for inference tasks. Lastly, in Appendix C,
we detail the derivations of necessary distribution-specific
functions.
II.
RESTRICTED BOLTZMANN MACHINES
Restricted Boltzmann Machines (RBMs) [27] are latentvariable generative models often used in the context of
unsupervised learning. A set of weights and biases, the
model parameters of the RBM, which correspond to the
couplings and local fields present in the system, constructs
an energy as a function of the data points from which
follows a Gibbs-Boltzmann probability density function.
In the well-known binary RBM, for which all visible and
latent variables are in {0, 1}, the RBM distribution is
P
P
X P
1
e ij Wij xi hj + i bi xi + j cj hj
Z [W, θ]
h
(1)
where θ = {b, c} is the set of local potentials, i.e. the set
of values which define the biases acting on each variable,
and
P
P
XX P
Z [W, θ] =
e ij Wij xi hj + i bi xi + j cj hj . (2)
P (x; W, θ) =
x
h
P
P
Here, we use the notation x and h to refer to sums
over the entire space of possible configurations of visible
and latent variables, respectively. When taken with respect to the parameters of the model, Z [W, θ] is known
as the partition function. We give a factor-graph representation of the RBM distribution in Fig. 1.
As evidenced by (2), an exact computation of the normalizing partition function, and thus the probability of a
given high-dimensional data point, is inaccessible in practice. Sophisticated Monte Carlo (MC) schemes relying
on importance sampling [20, 21] can produce estimates
3
via gradient ascent on the parameters W and θ. Commonly, one does not calculate these gradients for each
data-point from the training set, but instead calculates the
gradients in average across M data-points, often referred
to as a mini-batch. At each mini-batch, the gradients of
(5) are given as,
x
φ(xi ; θi )
∆Wij = hxi hj iX − hxi hj iSampled ,
φ(xi , xj ; Wij )
∆bi = hxi iX − hxi iSampled ,
∆cj = hhj iX − hhj iSampled ,
FIG. 1. Factor graph representation of the RBM distribution.
Variables are indicated by circles, with latent variables denoted by shaded circles. The shaded rectangles indicated layer
partitions within the RBM structure. Factors are represented
by squares, with the right hand side factors representing the
pairwise relationships between variables and the left hand
side factors representing the influence of the localized prior
distributions on the variables.
and bounds of the partition, but at the cost of substantial
computation, running on the time scale of days or even
weeks.
Thankfully, a precise estimate of the normalization is
unnecessary for many RBM applications. Additionally,
the bipartite structure of the RBM, which only admits
couplings between the hidden and visible variables, can
be leveraged to construct efficient sampling schemes. This
approach was demonstrated in the contrastive divergence
(CD) of [9], where very short-chain block-Gibbs sampling
was shown to be sufficient for adequate RBM training.
The CD approach consists of a sampling chain alternating
between samples drawn from the conditional probabilities
of each layer, which are dependent on the conditional
expectations at the previously sampled layer. Specifically,
the conditional probabilities for the hidden and visible
units factorize as
X
Y
P (x|h) =
sigm bi +
Wij hj ,
(3)
i
j
Y
X
!
P (h|x) =
sigm cj +
j
Wij xi
,
(4)
i
−1
where sigm(y) = (1 + e−y ) is the logistic sigmoid function.
In order to learn the parameters of the RBM for a given
training dataset, one looks to maximize the following loglikelihood,
P
P
X P
ln P (x; W, θ) = ln
e Wij xi hj + i bi xi + j cj hj
h
− ln Z [W, θ] ,
(5)
(6)
(7)
(8)
where h·iSampled refers to averages over particles sampled
from the model, and h·iX refers to the so-called clamped
expectations, where the values of x are fixed to the training data samples in the mini-batch. In the case of the
expectations involving hidden units, which are unobserved
and therefore have no training data, [9] originally proposed
the use of configurations sampled from P (h|x; W, b, c).
However, one could also use the exact conditional expectations directly to calculate these clamped averages;
especially in cases where sampling from these conditionals
may be problematic.
Since [9], there have been a number of proposed modifications to the core sampling-based training scheme described above. The persistent trick [18] takes neatly advantage of the iterative gradient ascent over mini-batches
to quickly obtain thermalized Markov chains through
Gibbs sampling at no extra computational cost over one
step CD (CD-1). Nevertheless, the probability density
function of a trained RBM is typically highly multimodal,
thus making this sampling inexact. Indeed, in such glassy
landscapes mixing becomes very slow as Markov chains
become stuck in metastable states, leading to over- and
under-represented, as well as missed, modes of the highdimensional distribution. This in turn produces high
variance estimates of means and correlations. A more
accurate sampling can be achieved using parallel tempering [28, 29], where particles are swapped between multiple Markov chains running at differing temperatures.
This approach, however, requires not only the additional
computational burden of running more chains, but also
requires further tuning of hyper-parameters, such as the
number of chains and at which temperatures to run them.
As accurate sampling-based inference on RBMs can
be costly, it would seem that their usefulness is limited.
As learning of the RBM via gradient ascent is dependent
upon this inference, the difficulty of training a generative
model with a high-degree of accuracy is compounded.
However, RBMs have proven to be very useful in many
applications where sampling from a full-fledged generative
model is unneeded. For instance, RBMs can be used as
an unsupervised “feature extraction” pre-training for feedforward networks [10, 15, 30]. RBMs have also been used
for data-imputation tasks, e.g. image in-painting, label
recovery [11], or collaborative filtering [12] by reconstructing missing data with a single visible-hidden-visible step.
In truth, the CD-k training algorithm which popularized
RBMs, first with binary units [9], then Gaussian units
4
[10, 31] and finally with arbitrary units [32], does not
use thermalized samples to evaluate means and correlations. Instead, it focuses on the region of the configuration space nearest to the training dataset [28] by using
short block-Gibbs Markov chains, starting from training
data points, to get fast and low variance estimates of
moments. However, CD-k is prone to learn spurious minima in configuration space far from the data as it does
not explore this region during training [28]. It also does
not systematically increase the true likelihood of training data [33]. However, this training strategy has been
found to be very efficient in the applications mentioned
above, which consistently remain close to the dataset in
configuration space. One finds that CD falls short for
applications which require long-chain MCMC sampling
from the trained RBM, as this represents a fundamental
mismatch between the training and application of the
RBM. In order to address some of theses shortcomings of
sampling-based approaches, we now turn our attention to
deterministic mean-field approximations of the RBM.
The TAP approximation [23] for disordered systems
relies on the deterministic inference of approximated magnetizations, from which one can obtain estimators of all
kinds of observables, starting from the log-partition or
free energy. TAP is derived from a small weight expansion
of the variational approach and can be considered as an
extension of the naïve mean-field (NMF) method [34, 35].
Previous works which have attempted to make use of the
NMF approximation of the RBM have shown negative
results [18, 36].
The TAP approximation was first considered for Boltzmann machines in the context of small random models
without hidden units in [37]. In the recent work of [38],
this approximation was extended to a practical training
algorithm for full-scale binary RBMs which was shown
to be competitive with Persistent Contrastive Divergence
(PCD) [18] when applied to real-world datasets. In parallel, other works have used TAP, and the related Bethe
approximation, to perform inference on binary Boltzmann
machines [39–42].
In the next sections, we detail how to re-write the RBM
model in the non-binary case, for generalized distributions
on the visible and hidden units, similar in spirit to [32].
However, unlike earlier techniques, we will approach the
problem of estimating the normalization of the RBM
model via the tools of statistical mechanics, resulting in a
fully-deterministic framework for RBM inference, training,
and application.
interest in the following manner,
P (x, h; W, θ) =
P
1
e i,j xi Wij hj
Z [W, θ]
Y
Y
×
Piv (xi ; θiv )
Pjh (hj ; θjh )
i
(9)
j
where the sum over i, j indicates a sum over the
all visible and hidden units in the model, and θ =
v
h
{θ1v , . . . , θN
, θ1h , . . . , θN
} are the parameters defining the
v
h
v
local distributions, Pi and Pjh , on respectively the variables of x and the variables of h. In the case that
x ∈ {±1}Nv and h ∈ {±1}Nh , we can see that the distribution above reduces to a bipartite spin glass model
with θ representing the local fields acting on the system
spins; the fields b and c for binary spins as described
in Eq. (1). This specific case is simply a binary RBM,
as described in the previous section, and which we have
already considered within an extended mean-field framework in [38]. The important distinction with the model
we evaluate here is that we do not assume a binary discrete distribution on the variables, but instead allow for a
formulation where the variables in the system can possess
distributions on discrete- or real-valued bounded support.
By considering this more general class of models, one
can include a wide range of different models, including
the Hopfield model and spike-and-slab RBMs [43], and
data sets, such as images or genomic data, by varying the
distributions of the hidden and visible units. The distribution of visible variables is obtained by marginalizing
out the latent variables,
Z Y
P (x; W, θ) = dhj P (x, h; W, θ),
(10)
j
giving the log-probability
ln P (x;W, θ) = − ln Z [W, θ] +
+
X
Z
ln
X
ln Pi (xi ; θiv )
i
dhj Pj (hj ; θjh )ehj
P
l
Wlj xl
.
(11)
j
If we take the gradients of (11) with respect to the
model parameters, in the case of distribution terms θ we
find,
∂
∂
v
v
∆θi =
ln Pi (xi ; θi )
− v ln Z [W, θ] , (12)
v
∂θi
∂θi
X
and
*
∆θjh
=
+
∂
∂
h
ln Pj (hj ; θj )
− h ln Z [W, θ] , (13)
h
∂θj
∂θj
X
III.
GENERAL DISTRIBUTIONS FOR RBMS
We now turn our attention to the case of the general
RBM (GRBM), where the distributions of the hidden and
visible units are not fixed. We define the distribution of
which are generalizations of (7) and (8). However, in the
case of the gradient with respect to the couplings, we find
*
!+
X
∂
h
∆Wij = xi · f
Wlj xl ; θj
−
ln Z [W, θ] ,
∂Wij
l
X
(14)
5
where the function
R
f (B; θ) =
dh h · P (h; θ)eBh
R
dh P (h; θ)eBh
(15)
computes the conditional expectation of hl knowing the
value of the visible units. It is important to note that the
data-dependent term is tractable thanks to the bipartite
structure of the one-hidden layer RBM.
In contrast to the data-dependent terms, the second
terms of Eqs. (12)–(14) require knowledge of the partials
of the log normalization w.r.t. the parameter of interest. However, this term cannot be written exactly as the
explicit calculation of the normalization is intractable.
Rather than resorting to sampling, we will attempt to
approximate the free energy F = − ln Z [W, θ] in a parametric and deterministic way, as in [38]. In the next
section, we discuss how belief propagation (BP) can be
used to estimate F and to conduct inference on RBMs.
strong correlations both within and between receptive
fields.
In order to construct a BP algorithm for PDFs over realvalued support, one requires a finite memory description
of the messages. Some examples of such descriptions are
given in non-parametric BP [47], moment matching [48],
and relaxed BP (r-BP) [49]. In Appendix B, following the
example of r-BP, we show how to arrive at a two-moment
approximation of the continuous BP messages via a smallweight expansion on the RBM coupling parameters W .
There, we also show the r-BP approximated free energy
of pairwise models, as well as demonstrating the need for
distributions with bounded support in order to preserve
bounded messages.
In the next section, building upon this derivation, we
consider mean-field approximations of the RBM via hightemperature Plefka expansion.
V.
IV.
APPROXIMATION VIA BELIEF
PROPAGATION
One method by which we might estimate the partition
of F is via belief-propagation (BP) [44], which we review in
Appendix A for pairwise models such as the RBM. Essentially, given a factor graph for some joint statistical model,
such as that of our RBM in Fig. 1, the BP algorithm
attempts to estimate a set of marginal distributions at
each variable. In the case of tree-like graphs, BP provides
an exact calculation of these marginals. The application
of BP to factor graphs containing cycles, loopy BP, is not
guaranteed to provide accurate estimates of the marginals.
However, often these estimated marginals have significant
overlap with the true ones [45]. Additionally, it is known
that the solutions of loopy BP are the fixed-points of the
Bethe free energy [45], which allows the construction of an
approximation of F. Applying this to the inverse learning
problem, one can compute the gradients of this Bethe free
energy in terms of the parameters of the model, allowing
a gradient ascent on Bethe-approximated log-likelihood
of the training data.
One significant hurdle in the application of loopy BP
to RBM learning for real-valued variables is that the
messages propagated on the edges of the factor graph are
continuous PDFs. In the case of discrete variables, such
as Ising or Potts spins, BP messages can be written using
magnetizations or the full discrete PMF, respectively. For
binary variables, both BP and mean-field approximations
of fully-connected Boltzmann machines were considered in
[39] in the context of inference with fixed parameters. A
similar study of binary RBMs was conducted with loopy
BP in [46]. It is important to note that both of these
studies investigated the properties of Boltzmann machines
with i.i.d. random weights. While such studies permit
many analytical tools for studying the behavior of the
RBM, one cannot directly map these observations to RBM
inference in practice, where trained weights may exhibit
TAP APPROXIMATION FOR PAIRWISE
MODELS
While one could utilize the r-BP approach in order to
estimate the free energy of a generalized real-valued spin
model, as detailed in the earlier section, such an approach
might not be desirable in practice. Specifically, if one
wishes to solve the inverse learning problem, estimating
model parameters from a given dataset, it is necessary
to estimate the gradients of the model parameters w.r.t.
the model likelihood for each parameter update. Using a
steepest-ascent approach, as detailed in Sec. III, requires
one to estimate these gradients many thousands of times.
For systems of large size N , the r-BP scales quite poorly.
Estimating a gradient requires the iteration of the r-BP
equations on O(N 2 ) messages. Additionally, one must
distinguish between cavity terms (i → j) and marginal
terms (→ i). If the final O(N ) gradients we desire can be
estimated using the marginal terms alone, then requiring
an iteration on the O(N 2 ) set of messages is an extremely
costly operation.
Instead, one can turn to a mean-field approach, writing
the free energy, and its stationary points, in terms of
the marginals alone. This can be done by including
certain correction terms, up to a specified degree in the
weights. In the context of RBMs, such approaches have
been proposed at both the 1st order, the naive mean-field
[36], and the 2nd order, using the so-called ThoulessAnderson-Palmer (TAP) [23] equations for introducing
an additional correction term [38, 39, 42, 50, 51]. In the
case of a GRBM with arbitrary distributions on each unit,
however, we must re-derive the TAP approximation in
terms of parameters of these distributions as well as the
approximate marginalized distribution at each site, up
to their first two moments. This task turns out to be
closely related to the TAP approach to low-rank matrix
factorization [50–54].
While it is possible to derive the stationarity conditions
for the inferred marginals from the r-BP messages directly
6
by Taylor expansion, we rather focus on the free energy
directly that will provide the gradients we require for
training the GRBM parameters via a high-temperature
expansion we present below.
Lastly, we point out that the TAP free energy secondorder (TAP) term depends on the statistical properties
of the weight distribution. The derivation presented below assumes independent
identically distributed weights,
√
scaling as O(1/ N ). This assumption is a simplification,
as in practice the weight distribution cannot be known
a priori. The distribution depends on the training data
and changes throughout the learning process according
to the training hyper-parameters. The adaptive TAP
(adaTAP) formalism [55] attempts to correct this assumption by allowing one to directly compute the “correct”
second-order correction term for a realization of the W
matrix without any hypothesis on how its entries are distributed. Although this algorithm is the most principled
approach, its computational complexity almost rules out
its implementation. Moreover, practical learning experiments indicate that training using adaTAP does not differ
significantly from TAP assuming i.i.d. weights. A more
detailed discussion of the computational complexity and
learning performance is described in Appendix D.
V.1.
In the limit√N → ∞, if we assume that the entries of W
scale as O(1/ N ) and that all sites are widely connected,
on the order of the size of the system, then we can apply
the TAP approximation – a high temperature expansion
of the Gibbs free energy up to second-order [34, 35]. In
the case of a Boltzmann distribution, the global minima
of the Gibbs free energy, its value at equilibrium, matches
the Helmholtz free energy F [56]. We will derive a twovariable parameterization of the Gibbs free energy derived
via the Legendre transform [57]. Additionally, we will
show that this two-variable Gibbs free energy is both
variational and attains the Helmholtz free energy at its
minima. For clarity of notation, we make our derivation
in terms of a pairwise interacting Hamiltonian without
enforcing any specific structure on the couplings; the
bipartite structure of the RBM is reintroduced in Section
VI.
We will first introduce the inverse temperature term
β to facilitate our expansion as β → 0, P (x; β, W, θ) =
e−β(H−Fβ ) , where
X
(i,j)
Wij xi xj −
1X
ln Pi (xi ; θi ).
β i
(17)
where we see that as the fields disappear, Fβ (0, 0) = Fβ ,
and we recover the true Helmholtz free energy.
We additionally note the following identities for the
augmented system, namely,
∂
[−βFβ (λ, ξ)] = hxi iλ,ξ ,
∂λi
∂
[−βFβ (λ, ξ)] = x2i λ,ξ ,
∂ξi
(18)
(19)
where h·iλ,ξ is the average over the augmented system for
the given auxiliary fields. Since the partial derivatives of
the field-augmented Helmholtz free energy generate the
cumulants of the Boltzmann distribution, it can be shown
that the Hessian of −βFβ (λ, ξ) is simply a covariance
matrix and, subsequently, positive semi-definite. Hence,
−βFβ (λ, ξ) is a convex function in terms of λ × ξ. This
convexity is shown to be true for all log partitions of
exponential family distributions in [57].
We now take the Legendre transform of −βFβ (λ, ξ),
introducing the conjugate variables a and c,
− βGβ (a, c) =
(
Derivation of the TAP Free Energy
H=−
system under the effect of two auxiliary fields,
Z
P
P
2
− βFβ (λ, ξ) = ln dx e−βH+ i λi xi + i ξi xi ,
(16)
We wish to derive our two-variable Gibbs free energy
for this system in terms of the first two moments of the
marginal distributions at each site. To accomplish this,
we proceed as in [35, 55] by first defining an augmented
)
1X
− β sup Fβ (λ, ξ) +
λi ai + ξi (ci + a2i ) , (20)
β i
λ,ξ
where we define the solution of the auxiliary fields at which
Gβ (a, c) is defined as λ∗ , λ∗ (β, a, c), ξ ∗ , ξ ∗ (β, a, c),
where we make explicit the dependence of the auxiliary
field solutions on the values of the conjugate variables.
Looking at the stationary points of these auxiliary fields,
we find that
∂
∂
[−βGβ (a, c)] =
[−βF (λ∗ , ξ ∗ )] − ai ,
∂λ∗i
∂λ∗i
∴ ai = hxi iλ∗ ,ξ∗ ,
(21)
(22)
and
∂
∂
[−βGβ (a, c)] = ∗ [−βF (λ∗ , ξ ∗ )] − (ci + a2i ),
∂ξi∗
∂ξi
(23)
∴ ci = x2i
2
λ∗ ,ξ∗
− hxi iλ∗ ,ξ∗ .
(24)
The implication of these identities is that Gβ (a, c) cannot
be valid unless it meets the self-consistency constraints
that a and c are the first and second moments of the
marginal distributions of the augmented system.
Now, we wish to show the correspondence of Gβ (a, c) to
the Helmholtz free energy at its unique minimum. First,
let us look at the stationary points of −βGβ (a, c) with
respect to its parameters. We take the derivative with
7
the Gibbs free energy to be decomposed into a sum of
independent terms. Specifically, if we take the expansion
up to s terms,
careful application of the chain rule to find,
X ∂
∂
∗ ∗
[−βGβ (a, c)] =
∗ [−βFβ (λ , ξ )]
∂ai
∂λ
j
j
X ∂
∂ ∗
∗ ∗
+
∗ [−βFβ (λ , ξ )] ∂a ξj
∂ξ
i
j
j
X ∂
λ∗j
− λ∗i −
∂a
i
j
− 2ξi∗ ai −
X ∂
ξ ∗ (ci + a2i ),
∂ai j
j
= −λ∗i − 2ξi∗ ai .
(25)
Carrying through a very similar computation for the
derivative with respect to ci provides ∂c∂ i [−βGβ (a, c)] =
ξi∗ . This shows that, at its solution, the Gibbs free energy
must satisfy
ξi∗ = 0,
λ∗i + 2ξi∗ ai = 0,
(26)
which can only be true in the event that the solutions of
the auxiliary fields are truly λ∗ = 0 and ξ ∗ = 0. Looking at the inverse Legendre transform of the Gibbs free
energy for λ = 0, ξ = 0, we find that inf a,c Gβ (a, c) =
Fβ (0, 0) = Fβ , which implies that the minimum of the
Gibbs free energy is equivalent to the Helmholtz free
energy. This holds since −βGβ (a, c) is convex, as the
Legendre transform of a convex function is itself convex.
Since the Gibbs free energy can therefore only possess a
single solution, then its minimum must satisfy (26), and
therefore, must be Fβ (0, 0) = Fβ .
Finally, we can now rewrite the Gibbs free energy defined in Eq. (20) as a function of the moments a, c and
parameterized by β and the GRBM parameters W and
θ,
Z
e
− βGβ (a, c; θ) = ln dx e−β H ,
(27)
β=0
+
s
X
βp ∂p h e i
ln Zβ
,
p! ∂ p β
β=0
p=1
(29)
e
where Zβ is the normalization of the Boltzmann distrie at temperature β. At β = 0 we can
bution defined by H
find the first term of the expansion directly
X
X
ln Ze0 = −
λ∗i (0)ai −
ξi∗ (0)(a2i + ci )
i
+
X
i
Z
ln
∗
X λ∗ (β)
i
i
β
(xi − ai ) −
ξi∗ (β) 2
(xi − ci − a2i ),
β
(28)
where the Lagrange multipliers are given as functions of
the temperature in order to make clear the order in which
we will apply β → 0 later.
As this exact form of the Gibbs free energy is just as
intractable as the original free energy, we will apply a
Taylor expansion in order to generate an approximate
Gibbs free energy [34]. We make this expansion at β = 0,
as in the limit of infinite temperature, all interactions
between sites vanish and the system can be described
only in terms of individual sites and their relationship to
the system average and their local potentials, allowing
∗
2
dxi Pi (xi ; θi )eλi (0)xi −ξi (0)xi , (30)
i
where we recognize that the last term is simply the normalization of the Gaussian-product distribution whose
moments were defined in (B9), (B10).
We define the TAP free energy by writing the remainder
of the expansions terms in the specific case of s = 2
[34, 35],
X
e (2) (a, c; θ) =
−β G
ln Zi (λ∗i (0), ξi∗ (0); θi )
β
i
−
X
+β
i
λ∗i (0)ai −
X
X
ξi∗ (0)(a2i + ci )
i
Wij ai aj +
(i,j)
β2 X 2
Wij ci cj ,
2
(i,j)
(31)
R
2
where Zi (B, A; θ) , dx Pi (x; θ)eBx−Ax . Now, we still
need to define the extremal values of λ∗ (0) and ξ ∗ (0)
which define the Gibbs free energy. By taking the stationarity of the expanded Gibbs free energy with respect to
a and c, we find
X
Ai , −2ξi (0)∗ = −β 2
Wij2 cj ,
(32)
j∈∂i
Bi , λi (0)∗ = Ai ai + β
where
e ,H−
H
e (s) (β, a, c; θ) = ln Zeβ
−β G
X
Wij aj ,
(33)
j∈∂i
where we make the definitions of A and B for convenience
and as a direct allusion to the definitions of the cavity
sums for BP inference, given in Eqs. (B5), (B6).
Conversely, by deriving the stationarity conditions of
the auxiliary fields, we obtain the self-consistency equations for a, c, which show us that the TAP free energy is
only valid when the following self-consistencies hold,
ai = fa (Bi , Ai ; θi ),
ci = fc (Bi , Ai ; θi ).
(34)
Substituting these values closes the free energy on the
marginal distribution moments a, c and completes our
derivation of a free energy approximation which is defined
by O(N ) elements versus the O(N 2 ) values required by
r-BP.
8
V.2.
Solutions of the TAP Free Energy
As given, the TAP free energy is only valid when the
self-consistency equations are met at its stationary points.
Thus, only a certain set of a, c can have any physical
meaning. Additionally, we know that only at the minima
of the exact Gibbs free energy will we have a correspondence with the original exact Helmholtz free energy.
While the exact Gibbs free energy in terms of the
moments a, c, is convex for exponential family P0 (x; θ)
e
such that e−β H ≥ 0, the TAP free energy can possess
multiple stationary points whose number increases rapidly
as β grows [5]. Later in Sec. VII, we show that as GRBM
training progresses, so does the number of identified TAP
solutions. This can be explained due to the variance of the
weights W growing with training. For fixed β = 1, as we
use in our practical GRBM implementation, the variance
of the weights serves as an effective inverse temperature,
and its increasing magnitude has an identical effect to the
system cooling as β increases.
Additionally, while the Gibbs free energy has a correspondence with the Helmholtz free energy at its minimum,
this is not necessarily true for the TAP free energy. The
approximate nature of the second-order expansion removes this correspondence. Thus, it may not be possible
to ascertain an accurate estimate of the Helmholtz free
energy from a single set of inferred a, c, as shown in Fig.
2. In the case of the naïve mean-field estimate of the
e (1) . This implies
Gibbs free energy, it is true that Fβ ≤ G
β
e (1) in
that one should attempt to find the minima of G
β
order to find a more accurate estimate of Fβ , a foundational principle in variational approaches. However, while
the extra expansion term in the TAP free energy should
improve its accuracy in modeling Gβ over a, c, it does
not provide a lower bound, and so an estimate of Fβ from
the TAP free energy could be an under- or over-estimate.
Instead, one might attempt to obtain an estimate of the
Helmholtz free energy by utilizing either all or a subset of
the equilibrium solutions of the TAP free energy. Since
there is no manner by which we might distinguish the
equilibrium moments by their proximity to the unknown
Fβ , averaging
D theETAP free energy across its solutions,
e (2) , can serve as a simple estimator of
denoted as G
β
∗
Fβ [5]. In [58] a weighting was introduced to the average, correcting the Helmholtz free energy estimate at low
temperature by removing the over-influence of the exponential number of high-energy solutions. The weights
in this approach are proportional to the exponents of
each solution’s TAP free energy, placing much stronger
emphasis on low-energy solutions.
However, such an approach is not well-justified in the
our general setting of Pi , where we expect large deviations
from the expectations derived for the SK model. Additionally, while this weighting scheme is shown across the
entire set of solutions for a particular random SK model,
in our case, we are interested in the solution space centered on the particular dataset we wish to model. Since
Gβ
e (2)
G
β
Fβ
D
E
e (2)
G
β
∗
{a, c}
FIG. 2. Cartoon description for estimating the Helmholtz
free energy (dotted ) via the Gibbs (blue dash) and TAP (red )
free energies. For this example of a convex Gibbs free energy,
there exists one unique minimum over the moments a, c, and
the Gibbs free energy here matches Fβ . The range of TAP
free energies (gray box ) gives a boundary on the location of
Fβ . Averaging the TAP free energies (dash-dot) provides an
estimate of Fβ .
the solutions are computed by iterating the TAP selfconsistency equations, we can easily probe this region by
initializing the iteration according to the training data.
Subsequently, we do not encounter a band of high-energy
solutions that we must weight against. Instead, we obtain
a set of solutions over a small region of the support of the
TAP free energy. Due to the uniformity of these solutions,
un-weighted averaging across the solutions seems the best
approach in terms of efficiency. In the subsequent section, we explore some of these properties numerically for
trained RBMs.
VI.
RBMS AS TAP MACHINES
To utilize the TAP inference of Sec. V, we need to
write the TAP free energy in terms of the variables of the
RBM. To clarify the bipartite structure of the GRBM,
we rewrite the TAP free energy in terms of the hidden
and visible variables at fixed temperature β = 1,
− FRBM av , cv , ah , ch ; θ =
X
1
ln Ziv (Biv , Avi ; θiv ) − Biv avi + Avi ((avi )2 + cvi )
2
i
X
1
+
ln Zjh (Bjh , Ahj ; θjh ) − Bjh ahj + Ahj ((ahj )2 + chj )
2
j
X
1
+
Wij avi ahj + Wij2 cvi chj ,
(35)
2
ij
where {av , cv } and ah , ch are the means and variances
of the visible and hidden variables, respectively.
As in Sec. V, solutions of the TAP GRBM free energy
can be found by a fixed-point iteration, as shown in Alg.
9
1, which bears much resemblance to the AMP iteration
derived in the context of compressed sensing [59, 60] and
matrix factorization [51, 53, 54]. We note that rather than
updating over the entire system at each time step, fixing
one side at a time has the effect of stabilizing the fixedpoint iteration. For clarity, Alg. 1 is written for a single
initialization of the visible marginals. However, as noted
in Sec. V.2, there exist a large number of initializationdependent solutions to the TAP free energy. Thus, in
order to capture the plurality of modes present in the
TAP free energy landscape, one should run this inference
independently for many different initializations.
If the use case of the GRBM requires that we only
train the GRBM tightly to the data space (e.g. data
imputation), it makes sense to fix the initializations of
the inference to points drawn from the dataset,
av,(0) = x(m)
c
v,(0)
= 0.
where m ∈ {1, . . . , M } ,
(36)
(37)
In order to train the GRBM more holistically, structured
random initializations can help probe modes outside of
the data space.
For a set of TAP solutions {ak , ck , Bk , Ak } for k ∈
{1, . . . , K} at fixed GRBM parameters {W, θ}, the TAPapproximated log-likelihood of can be written as
X
ln P (x; W, θ) ≈
ln Pi (xi ; θiv )
i
+
X
X
ln Zjh (
j
Wij xi , 0; θjh )
i
1 X
FRBM avk , cvk , ahk , chk ; θ ,
+
K
(38)
k
where Zjh (B, 0; θ) is the normalization of the conditional
expectation of Eq. (15), since fa (B, 0; θ) = f (B; θ).
After re-introducing an averaging of the log-likelihood
over the samples in the mini-batch, the gradients of the
TAP-approximated GRBM log-likelihood w.r.t. the model
parameters are given by
1 X (m) h X
(m)
x fa (
Wij xi , 0; θjh )
M m i
i
1 X v h
−
ai,k aj,k + Wij cvi,k chj,k ,
(39)
K
k
"
#
X
1 X ∂
(m)
h
h
h
∆θj ≈
ln Zj (
Wij xi , 0; θj )
M m ∂θjh
i
1 X ∂
h
−
ln Zjh (Bj,k
, Ahj,k ; θjh ) ,
(40)
h
K
∂θj
k
i
1 X ∂ h
v (m) v
∆θiv ≈
ln
P
(x
;
θ
)
i
i
i
M m ∂θiv
1 X ∂
v
−
ln Ziv (Bi,k
, Avi,k ; θiv ) .
(41)
K
∂θiv
∆Wij ≈
k
Algorithm 1 TAP Inference for GRBMs
Input: W , θ
Initialize: t = 0, av,(0) , cv,(0)
repeat
Hidden Side Updates
P
h,(t+1)
v,(t)
Aj
= − i Wij2 ci
h,(t+1)
P
h,(t+1) h,(t)
v,(t)
= Aj
aj
+ i Wij ai
h,(t+1)
h,(t+1)
h,(t+1)
aj
= fah Bj
, Aj
; θjh
h,(t+1)
h,(t+1)
h,(t+1)
cj
= fch Bj
, Aj
; θjh
Bj
Visible Side Updates
P
v,(t+1)
h,(t+1)
Ai
= − j Wij2 cj
P
v,(t+1)
v,(t+1) v,(t)
h,(t+1)
Bi
= Ai
ai
+ j Wij aj
v,(t+1)
v,(t+1)
v,(t+1)
ai
= fav Bi
, Ai
; θiv
v,(t+1)
v,(t+1)
v,(t+1)
ci
= fcv Bi
, Ai
; θiv
t=t+1
until Convergence
In the presented gradients, we make the point that the
set of data samples and the set of TAP solutions can have
different cardinality. For example, one might employ a
mini-batch strategy to training, where the set of data
samples used in the gradient calculation might be on the
order 102 . However, depending on the application of the
GRBM, one might desire to probe a very large number of
TAP solutions in order to have a more accurate picture
of the representations learned by the GRBM. In this case,
one might start with a very large number of initializations,
resulting in a very large number, K M , of unique
TAP solutions. Or, contrary, while one might start with
a number of initializations equal to M , the number of
unique solutions might be K M , especially early in
training or when the number of hidden units is small.
Using these gradients, a simple gradient ascent with a
fixed or monotonically decreasing step-size γ can be used
to update these GRBM parameters. We present the final
GRBM training algorithm in Alg. 2.
Besides considering non-binary units, another natural
extension of traditional RBMs is to consider additional
hidden layers, as in Deep Boltzmann Machines (DBMs).
It is possible to define and train deep TAP machines, as
well. Probabilistic DBMs are substantially harder to train
than RBMs as the data-dependent (or clamped) terms
of the gradient updates (39-41) become intractable with
depth. Interestingly, state-of-the-art training algorithms
retain a Monte Carlo evaluation of other intractable terms,
while introducing a naıve mean-field approximation of
these data-dependent terms. For deep TAP machines,
we consistently utilize the TAP equations. The explicit
definition and training algorithm are fully described in
Appendix E.
10
(a) binary-MNIST
−0.06
(t+1)
(t)
(t)
Wij
← Wij + γ∆Wij +
,(t+1)
,(t)
,(t)
θjh
← θjh
+ γ∆θjh
v ,(t+1)
v ,(t)
v ,(t)
θi
← θi
+ γ∆θi
(t)
γR(Wij )
+
(t−1)
η∆Wij
Normalized TAP Log-likelihood
Algorithm 2 GRBM Training
Input: X, T , M , K, R(·)
(0)
Initialize: Wij ∼ N (0, σ), θ (0) ← X
repeat
for All mini-batches of size M do
a(t+1) , c(t+1) , B(t+1) , A(t+1) ← Alg. 1(W (t) , θ (t) ) × K
−0.08
−0.10
−0.12
−0.14
−0.16
−0.18
−0.20
NH = 500
NH = 100
−0.22
−0.24
20
40
60
NH = 50
NH = 25
80
100
Training Epochs
end for
t←t+1
until t > T
(b) real-MNIST
EXPERIMENTS
VII.1.
Datasets
MNIST — The MNIST handwritten digit dataset
[61] consists of both a training and testing set, each with
60,0000 and 10,000 samples, respectively. The data samples are real-valued 28 × 28 pixel 8-bit grayscale images
which we normalize to the dynamic range of [0, 1]. The
images themselves are centered crops of the digits ‘0’
through ‘9’ in roughly balanced proportion. We construct
two separate versions of the MNIST dataset. The first,
which we refer to as binary-MNIST, applies a thresholding
such that pixel values > 0.5 are set to 1 and all others to 0.
The second, real-MNIST, simply refers to the normalized
dataset introduced above.
CBCL — The CBCL face database [62] consists of
both face and non-face 8-bit grayscale 19×19 pixel images.
For our experiments, we utilize only the face images. The
database contains 2,429 training and 472 testing samples
of face images. For our experiments, we normalize the
samples to the dynamic range of [0, 1].
binary-MNIST
real-MNIST
CBCL
Nv
784
784
361
Nh {25, 50, 100, 500} {100, 500}
256
M
100
100
20
K
100
100
20
Prior Vis.
B.
Tr. Gauss.-B. Tr. Gauss.
Prior Hid.
B.
B.
B.
R(·)
`2
`2
`2
−2
γ
0.005
10 , 10−5
0.005
0.001
0.001
0.01
η
0.5
0.5
0.5
TABLE I. Parameter settings for GRBM training.
−0.35
−0.40
−0.45
−0.50
−0.55
−0.60
−0.65
−0.70
NH = 100
20
40
60
NH = 500
80
100
Training Epochs
(c) CBCL
−0.05
Normalized TAP Log-likelihood
VII.
Normalized TAP Log-likelihood
−0.30
−0.10
−0.15
−0.20
−0.25
NH = 256
−0.30
100
200
300
400
500
Training Epochs
FIG. 3. Training performance over 100 epochs for the tested
datasets over varying numbers of hidden units, Nh . Performance is measured in terms of the normalized (per-unit) TAP
log-likelihood estimate computed for 10,000 training data samples. The TAP free energy is estimated using the unique TAP
solutions thermalized from initial conditions drawn from the
data samples, as in (36)–(37). Thermalization is determined
by the convergence of the magnetizations up to a difference
of 10−8 in MSE between iterations. Solid lines indicate the
average normalized TAP log-likelihood over the tested training
samples and shaded regions indicate standard error.
VII.2.
Learning Dynamics
We now investigate the behavior of the GRBM over the
course of the learning procedure, looking at a few metrics
of interest: the TAP-approximated log-likelihood of the
training dataset, the TAP free energy, and the number
of discovered TAP solutions. We note that each of these
metrics is unique to the TAP-based model of the GRBM.
11
NH = 100 NH = 500 NH = 100 NH = 500
Most Active Features
Most Localized Features
Least Active Features
(c) CBCL
Most Spread Features
Most Active Features
NH = 256 NH = 256
Least Active Features
Most Spread Features
NH = 100 NH = 500 NH = 100 NH = 500
(b) real-MNIST
Most Localized Features
NH = 256 NH = 256
Most Active Features
NH = 25 NH = 50 NH = 100 NH = 500 NH = 25 NH = 50 NH = 100 NH = 500
NH = 25 NH = 50 NH = 100 NH = 500 NH = 25 NH = 50 NH = 100 NH = 500
(a) binary-MNIST
Most Spread Features
Most Localized Features
Least Active Features
FIG. 4. Subsets of the final receptive fields, i.e. the columns of W , obtained by TAP training of GRBM models with varying
numbers of hidden units, Nh . For the receptive fields, dark blue and yellow are mapped to −1 and +1, respectively, and green
indicates a value of 0. Receptive fields are ranked according to two criteria. First, spread, and conversely localization, as
measured by the `p -norm of the receptive field, for p = 0.1. Second, by activation level, as measured by the mean activation of
each receptive field’s corresponding hidden unit averaged across the training dataset.
While it was shown in [38] that CD does indeed maximize the TAP log-likelihood in the case of binary RBMs,
the specific construction of CD is entirely independent
from the TAP model of the GRBM. Thus, it is hard to
say that a CD or TAP-trained GRBM is “better” in a
general case. At present, we present comparisons between
TAP GRBMs of varying complexity trained under fixed
hyper-parameters settings, as indicated in Table I.
In Fig. 3 we see a comparison of the TAP log-likelihood
as a function of training epochs for binary-MNIST for
binary RBMs consisting of differing numbers of hidden
units. As the gradient-ascent on the log-likelihood is performed batch-by-batch over the training data, we define
one epoch to be a single pass over the training data: every
example has been presented to the gradient ascent once.
The specifics of this particular experiment are given in
the caption. We note that for equal comparison across
varying model complexity, this log-likelihood is normalized over the number of visible and hidden units present
in the model. In this way, we observe a “per-unit” TAP
log-likelihood, which gives us a measure of the concentration of representational power encapsulated in each
unit of the model. Increasing values of the normalized
TAP log-likelihood indicate that the evaluated training
samples are becoming more likely given the state of the
GRBM model parameters.
It can be observed that at each level of complexity, the
TAP log-likelihood of the data rapidly increases as the
values of W quickly adjust from random initializations to
receptive fields correlated with the training data. However,
across each of the tested models, by about the 20th epoch
the rate of increase of the TAP log-likelihood tapers off
to a constant rate of improvement.
For reference, we also show a subset of the trained
receptive fields, i.e. the rows of W , for each of the tested
experiments. Since the full set of receptive fields would be
too large to display, we attempt to show some representative samples in Fig. 4 by looking at the extreme samples
in terms of spatial spread/localization and activity over
the training set. We observe that the trained GRBMs, in
the case of both binary-MNIST and real-MNIST, are able
to learn both the localized and stroke features commonly
observed in the literature for binary RBMs trained on
the MNIST dataset [9, 63]. It is interesting to note that
even in the case of real-MNIST, where we are using the
novel implementation of truncated Gauss-Bernoulli visible
units, we are able to observe similar learned features as in
the case of binary-MNIST. We take this as an empirical
indication that the proposed framework of GRBM learning is truly learning correlations present in the dataset
as intended. Finally, we see feature localization increase
with the number of hidden units.
To date, understanding “what” an RBM learns from
the unlabelled data has mostly been a purely subjective
exercise in studying the receptive fields, as shown Fig.
4. However, the interpretation of the GRBM as a TAP
machine can provide us a novel insight in the nature and
dynamics of GRBM learning via the stationary points of
the TAP free energy, which we detail in the next section.
VII.3.
Probing the GRBM
Given the deterministic nature of the TAP framework,
it is possible to investigate the structure of the modes
which a given set of GRBM parameters produces in the
free energy landscape. Understanding the nature and
concentration of these modes gives us an intuition on the
representational power of the GRBM.
12
(a) binary-MNIST
(b) real-MNIST
(c) CBCL
FIG. 5. Distribution of free energy estimates of TAP solutions as a function of training epochs for the three different datasets.
In the case of the two MNIST experiments, the number of hidden units is the same Nh = 100 and 10, 000 samples drawn from
the training data are used as initial conditions. For CBCL, 2, 400 training samples are used. Top Row: TAP Free energy for
all unique TAP solutions (transparent blue dots), and the Helmholtz free energy estimate via uniform averaging (red line). The
number of unique TAP solutions is also given (green line). Bottom Row: Detail of TAP free energy distributions for slices of
training. Histograms are given as bars, while kernel density estimates of the TAP free energy distribution are given as curves.
To date, observing the modes of a given GRBM model
could only be approached via long-chain sampling. Given
enough sampling chains from a diverse set of initial conditions, thermalizing these chains produces a set of samples
from which one could attempt to derive statistics, such as
concentrations of the samples in their high-dimensional
space, to attempt to pinpoint the likely modes in the
model. However, the number of required chains to resolve
these features increases with the dimensionality of the
space and the number of potential modes which might exist in the space. Because of this, the numerical evaluation
we carry out here would be impractical with sampling
techniques.
The r-BP and mean-field models of the RBM allow us
to directly obtain the modes of the model by running
inference to solve the direct problem. Given a diverse
set of initial conditions, such as a given training dataset,
running r-BP or TAP provides a deterministic mapping
between the initial conditions drawn from the data, as
in (36)–(37), and the “nearest” solution of the TAP free
energy. If the initial point was drawn from the dataset,
then this solution can be interpreted as the RBM’s bestmatching internal representation for the data point.
If a large number of structurally diverse data points
map to a single solution, then this may be an indicator
that the GRBM parameters are not sufficient to model the
diverse nature of the data, and perhaps further changes to
the model parameters or hyper-parameters are required.
Conversely, if the number of solutions explodes, being
roughly equivalent to the number of initial data points,
then this indicates a potential spin-glass phase, that the
specific RBM is over-trained, perhaps memorizing the
original data samples during training. Additionally, when
in such a phase, the large set TAP solutions may be
replete with spurious solutions which convey very little
structural information about the dataset. In this case,
hyper-parameters of the model may need to be tuned
in order to ensure that the model possess a meaningful
generalization over the data space.
To observe these effects, we obtain a subset of the TAP
solutions by initializing the TAP iteration with initial
conditions drawn from the data set, running the iteration
until convergence, and then counting the unique TAP
solutions. We present some measures on these solutions
in Fig. 2. Here, we both count the number of unique
TAP solutions, as well as the distribution of the TAP free
energy over these solutions, across training epochs. There
are a few common features across the tested datasets.
First, the early phase of training shows a marked increase
of the TAP free energy, which then gradually declines
as training continues. Comparing the point of inflection
in the TAP free energy against the normalized TAP loglikelihood shown in Fig. 3 shows that the early phase of
GRBM training is dominated by the reinforcement of the
empirical moments of the training data, with the GRBM
model correlations playing a small role in the the gradient
of (14). This makes sense, as the random initialization of
W ∼ N (0, σ) for σ ≈ 10−3 implies that the hidden units
are almost independent of the training data. Thus, the
TAP solutions at the early stage of learning are driven by,
and correlated with, the local potentials on the hidden
and visible variables.
The effect of this influence is that the TAP free energy
landscape possesses very few modes in the data space.
Fig. 5 shows this very clearly, as the number of TAP
solutions starts at 1 and then steadily increases with train-
13
ing. Because the positive data-term of (14) is dominant,
the GRBM parameters do not appear to minimize the
TAP free energy, as we would expect. However, as more
TAP solutions appear, the data and model terms of the
gradient become balanced, and the TAP free energy is
minimized. It is at this point of inflection that we see
leveling off of the normalized TAP log-likelihood.
Second, we observe free-energy bands in the TAP solutions. This feature is especially pronounced in the case
of the binary-MNIST experiment. Here, at all training
epochs, there exist two significant modes in the free energy
distribution over the TAP solutions. We see this effect
more clearly in the training-slice histograms shown in the
bottom row of Fig. 5(a). In the case of the real-MNIST
experiment, we see that the free energy distributions do
not exhibit such tight banding, but they do show the
presence of some high- and low-energy solutions which
persist across training. The main feature across experiments is the multi-modal structure of the free energy
distribution. Finally, we note that for both real-MNIST
and binary-MNIST, in the case of Nh = 100, we don’t
empirically observe an explosion of TAP solutions, a potential indicator a spin-glass phase, since the proportion
of unique TAP solutions to the initial data points remains
less than 10%.
In order to investigate whether the modes in the TAP
free energy distributions are randomly assigned over configuration space, or exist in separate continuous partitions
of the configuration space, we need to look at the proximity of the solutions in the configuration space. Because this
space cannot be observed in its ambient dimensionality,
we project the configuration space into a two-dimensional
embedding in Fig. 6. Here, we utilize the well known
Isomap [64] algorithm for calculating a two-dimensional
manifold which approximately preserves local neighborhoods present in the original space. Using this visualization we observe that as training progresses the assignment
of high and low free energy to TAP solutions does not
appear random in nature, but seems to be inherent to
the structure of the solutions themselves, that is, their
location in the configuration space. Additionally, in Fig
6 we can see the progression from few TAP solutions to
many, and how they spread across the configuration space.
It is interesting to note how the solutions start from a
highly correlated state and then proceed to diversify.
We can also observe the TAP solutions with respect
to the initializations which produced them, as shown
in Fig. 7. In these charts, we use a similar approach
as Fig. 6, mapping all high-dimensional data points,
as well as TAP magnetizations, into a 2D embedding
using Isomap. This allows us to see, in an approximate
way, how the TAP solutions distribute themselves over
the data space. We also show how the number of TAP
solutions grows from few to many over training, and
how they maintain a spread distribution over the data
space. This demonstrates how the training procedure is
altering the parameters of the model so as to place TAP
solutions within dense regions of the data space. For the
sake of clarity, we have not included lines indicating the
attribution of an initial data point to its resultant TAP
solution. However, as training progresses, one sees that
the TAP solutions act as attractors over the data space,
clustering together data points which the TAP machine
recognizes as similar.
VII.4.
Inference for Denoising
Serving as a prior for inference is one particular use
case for the TAP machine interpretation of the GRBM.
As a simple demonstration, we turn to the common signal
processing task of denoising. Specifically, given a planted
signal, one observes a set of noisy observations which are
measures of the true signal corrupted by some stochastic
process. Denoising tasks are ubiquitous in signal processing, both at an analog level (e.g. additive and shot noise),
and at the level of digital communications (e.g. binary
symmetric and erasure channels). The goal of this task is
to produce the most accurate estimate of the unknown
signal. In the analog case, this may be a measure of
mean-square-error (MSE) between the estimate and the
true signal. In the binary case, this may be a measure
of accuracy, counting the number of incorrect estimates,
or some other function of the binary confusion matrix,
such as the F1-score or Matthews correlation coefficient
(MCC).
For a fixed set of observations and channel parameters,
if we assume that the original signal was drawn from
some unknown and intractable generating distribution,
then as we construct more and more accurate tractable
approximate priors, the more accurately we can construct
an estimate of the original signal.
In other words, the more we know about the structure
and content of the unknown signal a priori, the closer our
estimate can be. Often, as in the case of wavelet-based
image denoising, statistics are gathered on the transform
coefficients of particular images classes and heuristic denoising approaches are designed by-hand accordingly [65].
By-hand derivation of denoising algorithms works well in
practice owing to its generality. Specific a priori information about the original signal is not required, beyond
its signal class (e.g. natural images, human speech, radar
return timings). However, meaningful features must be
assumed or investigated by practitioners before successful
inference can take place.
14
Epoch 5
Isomap Dim. 2
Epoch 4
Isomap Dim. 2
Epoch 3
Isomap Dim. 2
Epoch 2
Isomap Dim. 2
Isomap Dim. 2
Epoch 1
Epoch 10
Isomap Dim. 2
Isomap Dim. 1
Epoch 9
Isomap Dim. 2
Isomap Dim. 1
Epoch 8
Isomap Dim. 2
Isomap Dim. 1
Epoch 7
Isomap Dim. 2
Isomap Dim. 1
Epoch 6
Isomap Dim. 2
Isomap Dim. 1
Epoch 60
Isomap Dim. 2
Isomap Dim. 1
Epoch 50
Isomap Dim. 2
Isomap Dim. 1
Epoch 40
Isomap Dim. 2
Isomap Dim. 1
Epoch 30
Isomap Dim. 2
Isomap Dim. 1
Epoch 20
Isomap Dim. 2
Isomap Dim. 1
Isomap Dim. 1
Epoch 100
High
Isomap Dim. 2
Relative TAP FE
Isomap Dim. 1
Epoch 90
Isomap Dim. 2
Isomap Dim. 1
Epoch 80
Isomap Dim. 2
Isomap Dim. 1
Epoch 70
Isomap Dim. 2
Isomap Dim. 1
Low
Isomap Dim. 1
Isomap Dim. 1
Isomap Dim. 1
Isomap Dim. 1
FIG. 6. Isomap visualization of TAP solutions for binary-MNIST over training epochs for Nh = 100. All TAP solutions are
mapped to the same two-dimensional embedding via an Isomap transform fitted to the TAP solutions of Epoch 100. The
embedding is performed on both the hidden and visible inferred expectations, av and ah . The color mapping corresponds to the
TAP free energy values of each solution, with the range of colors normalized between the minimum and maximum free energies
of the solution at each training epoch.
15
(a) binary-MNIST (Nh = 100)
Epoch 3
Epoch 25
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 2
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 1
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 1
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 1
Isomap Dim. 1
(b) real-MNIST (Nh = 500)
Epoch 3
Epoch 25
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 2
Epoch 100
Isomap Dim. 2
Isomap Dim. 2
Isomap Dim. 2
Epoch 1
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 1
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 1
Class ‘3’ Digits
Class ‘6’ Digits
Class ‘7’ Digits
Isomap Dim. 1
Isomap Dim. 1
(c) CBCL (Nh = 256)
Epoch 100
Epoch 200
Data Points
Isomap Dim. 2
Epoch 500
Isomap Dim. 2
Isomap Dim. 2
Isomap Dim. 2
Epoch 5
Isomap Dim. 1
Epoch 100
Isomap Dim. 2
Isomap Dim. 2
Isomap Dim. 2
Epoch 1
Data Points
Isomap Dim. 1
Data Points
Data Points
Isomap Dim. 1
Isomap Dim. 1
FIG. 7. Comparison of initial conditions for TAP equilibration (colored dots) compared to converged TAP solutions (black dots)
for the tested datasets at different stages of training. For each dataset, a two-dimensional Isomap embedding is calculated over
the initialization data. Subsequently, the magnetizations of the TAP solutions are embedded in the same space. In each case, all
initial variances are set to 0, as in Eq. 37. Also, a random selection of TAP solution magnetizations are chosen to provide some
context for the representations the RBM is learning. (a) binary-MNIST: Here, the ∼3,100 digits corresponding to the classes
‘3’, ‘6’, and ‘7’ are drawn from the first 10,000 training samples of the binarized MNIST dataset as initializations. A reduced set
of labels is used for readability. (b) real-MNIST: The same initializations are used as in (a), however, the initializations are
not binarized. (c) CBCL: All available training face images are used as initializations.
VII.4.1.
Denoising the Binary Symmetric Channel
likelihood at each observation,
Y p δxi ,yi
P (y|x) = (1 − p)
,
1−p
i
For binary denoising problems, we assume a binary
symmetric channel (BSC) defined in the following manN
ner. Given some binary signal x ∈ {0, 1} , we observe
N
the signal y ∈ {0, 1} as x with independent bit flips
occurring with probability p. This gives the following
(42)
which can be shown to have the equivalent representation
as a Boltzmann distribution,
P (y|x) =
1 Pi Di xi
,
e
Z(y)
(43)
16
p
where Di , ln 1−p
(2yi − 1). For a given prior distribution
P (x), the posterior distribution is given by Bayes’ rule,
x
Di xi
P (x)
P
.
e i Di xi P (x)
1
1+
1−mi
mi
×
1−p
p
0.6
0.4
0.2
2yi −1 ,
m
p
= sigm ln
+ (2yi − 1) ln
1−m
1−p
0.8
(44)
= 25
= 50
= 100
= 500
OPE
1-NN
i
Q
By assuming a factorized P (x) = i mxi i (1 − mi )1−xi ,
where mi might be per-site empirical averages obtained
from available training data, the posterior factorizes and
we can construct the Bayes-optimal pointwise estimator
(OPE) as the average hxi iP (xi |yi ) which is just P (xi =
1|yi ) for our binary problem. Thus, the OPE at each site
xi is given as
P (xi = 1|yi ) =
Nh
Nh
Nh
Nh
MCC
P
e
P (x|y) = P
1.0
.
(45)
For a given dataset, the OPE gives us the best-case performance using only pointwise statistics from the dataset,
namely, empirical estimates of the magnetizations mi . We
can see from Eq. (45) that the OPE either returns the
observations, in the case of p = 0, or the prior magnetizations mi , in the case p = 0.5. In this case of complete
information loss, the worst case performance is bounded
according to the deviation of the dataset from its mean.
We present the performance of the OPE in Fig. 8 for the
binary-MNIST dataset. This makes the OPE a valuable
baseline comparison and sanity check for the GRBM approximation of P (x). As the GRBM model takes into
account both pointwise and pairwise relationships in the
data, a properly trained GRBM should provide estimates
at least as good as the OPE.
The k-Nearest-Neighbor (k-NN) algorithm represents
a different heuristic approach to the same problem [66].
In this case, the noisy measurements are compared to a
set of exemplars: the training dataset. Then, according
to some distance metric such as MSE or correlation, one
finds the k exemplars with minimal distance to the noisy
observations to serve as a basis for recovering the original
binary signal. One can use some arbitrary approach for
fusing these exemplars together into the final estimate,
but the simplest case would be a simple average. In the
case that k → ∞, the performance when using averaging
is again bounded by the empirical magnetizations. In the
other limit of k = 1, the estimate is simply the nearest
exemplar. It is hard to show the limiting performance
of this approach, as it is dependent on the distances,
in the chosen metric, between the exemplars and the
observations, as well as the interplay between the noise
channel and the distance metric.
However, it can be seen directly that this approach
is non-optimal, as this approach will not yield the true
0.0
0.0
0.1
0.2
0.3
Bit Flip Probability
0.4
0.5
FIG. 8. Average denoising performance for reconstruction
from bit-flip errors on binary-MNIST over the probability of a
bit to be flipped. Denoising via inference on the binary-binary
RBM is denoised by the varying numbers of hidden units
(Nh = {25, 50, 100, 500}). Also shown as baseline comparisons
are the OPE given the empirical factorized magnetizations
at each site (dashed black ), and a 1-NN matching from the
training set (solid black ). All experiments were run over the
same 1,000 data samples drawn from the held-out test set and
compared using the MCC. Binary estimates are obtained for
the OPE and TAP inferred estimates by rounding the resulting
magnetizations.
signal at p = 0 unless the true signal is itself contained
within the training data. We show the performance for
k = 1 in Fig. 8. The advantage of this approach is that
it successfully regularizes against noise as p → 0.5, as
the nearest exemplar is always noise-free and at least
marginally correlated with the original signal, up to the
distance metric. Additionally, we see that it performs
better than the OPE in the regime p > 0.2. This can be
explained since we can think of the k-NN approach as
implicitly, though indirectly, taking into account higherorder correlations in the dataset by naïvely returning
data exemplars; all the estimates trivially posses the same
arbitrarily complex structure as the unknown signal.
Using the GRBM, we can hope to capture the best
points of both approaches. First, we hope to perfectly
estimate the original signal in the case p = 0. Second, we
hope to leverage the pairwise correlations present in the
dataset, returning estimates which retain the structure
of the data even as p → 0.5. For GRBM denoising of the
BSC, we no longer have a factorized posterior. Instead,
we have the GRBM likelihood given in Eq. (9) summed
over the hidden units. Using the definition of the binary
17
prior given in Appendix 3,
P (x; W, U ) ∝
X
e
P
ij
P
P
xi Wij hj + i Ui xi + j Uj hj
1-NN
.
(46)
OPE
h
Since both the GRBM and the BSC channel
likelihood
are
written
as
exponential
family
distributions,
P
(y|x)P
(x;
W,
U)
∝
P P xi Wij hj +P (Di +Ui )xi +P Uj hj
ij
i
j
e
.
Finding
the
h
averages hxi i for this model simply consists in running
the TAP-based inference of Alg. 1 for the modified
visible binary prior B(xi ; Ui + Di ). One heuristic caveat
of this approach is that we must take into account the
multi-modal nature of the TAP free energy. Since we
must initialize somewhere, and the resulting inference
estimate is dependent upon this initialization, we
initialize the inference with the OPE result.
We can see that as p → 0, Di → ±∞ and the highest
probability configuration becomes observations. So, in
the limit, we are able to obtain the true signal, just as
the OPE, especially since we initialize within the well of
this potential. This is shown for binary RBMs trained
with varying numbers of hidden units in Fig. 8. In every
case, for p = 0, the true signal is recovered. In the case of
Nh = {25, 50, 100}, we see that the TAP inference on the
binary RBM always outperforms the OPE. Additionally,
we see that in each of these cases, the performance closely
mirrors that of the 1-NN as p → 0.5. In the limit p = 0.5,
we see that the result of the TAP inference is, essentially,
uncorrelated with the original signal, as in this case, there
is no extra potential present to bias the inference, and
the resulting estimate is simply an arbitrary solution of
the TAP free energy. As this closely mirrors the exemplar
selection in 1-NN, the MCC curves for the two approaches
are similar.
In the case of Nh = 500, we can see that an over-training
effect occurs. Essentially, at low values of p, the TAP
inference over the binary RBM is able to more accurately
identify the original signal. However, at a certain point,
owing to the increased number of solutions in the TAP
free energy, there exist many undesirable minima around
the noisy solutions, leading to poor denoising estimates.
One can observe this subjectively in Fig. 9, where in the
case of Nh = 500, the TAP inference results in either
nearly zero-modes, or in very localized ones. This would
seem to indicate that landscape of the TAP free energy
around the initializations is becoming more unstable as
the density of solutions increases around it. Additionally,
since the TAP free energy landscape was only probed using data points during training, the clustering of solutions
around noisy samples remains ambiguous. Augmenting
the initializations used when calculating the TAP solutions for the gradient estimate with noisy data samples
could help alleviate this problem and regularize the TAP
free energy landscape in the space of noisy data samples.
Nh = 25
Nh = 50
Nh = 100
Nh = 500
p = 0.00 p = 0.07 p = 0.12 p = 0.17 p = 0.22 p = 0.26 p = 0.31 p = 0.36 p = 0.41 p = 0.45 p = 0.50
FIG. 9. Subjective comparison of denoising estimates for
a single digit image for p = [0, 0.5]. For the OPE and RBM
approaches, the inferred posterior averages hxi i are shown,
rather than the final configurations, where white and black
represent the values 0 and 1, respectively. At each tested value
of p, the same noise realization is used for each method. As
p increases, for Nh = {25, 50, 100}, the TAP inference for the
RBM provides estimates which still possess digit structure. In
the case Nh = 500, the TAP inference gets caught in spurious
and undesirable minima as p increases.
VIII.
DISCUSSION
In this paper, we have proposed a novel interpretation
of the RBM within a fully tractable and deterministic
framework of learning and inference via TAP approximation. This deterministic construction allows novel tools
for scoring unsupervised models, investigation of the memory of trained models, as well as allowing their efficient
use as structured joint priors for inverse problems. While
deterministic methods based on NMF for RBM training
were shown to be inferior to CD-k in [36], the level of
approximation accuracy afforded by TAP finally makes
the deterministic approach to RBMs effective, as shown
in the case of binary RBMs in [38].
Additionally, our construction is generalized over the
distribution of both the hidden an visible units. This
is unique to our work, as other works propose unique
training methods and models when changing the distribution of the visible units. For example, this can be seen
in the modified Hamiltonians used for real-valued data
[67, 68]. This construction allows us to consider binary,
real-valued, and sparse real-valued datasets within the
same framework. Additionally, one can also consider other
architectures by changing the distributions imposed on
the hidden unit. Here, we present experiments using only
binary hidden units, but one could also use our proposed
framework for Gaussian-distributed hidden units, thus
mimicking a Hopfield network [2]. Or, also, sparse GaussBernoulli distributed hidden units could mimic the same
functionality as that proposed by the spike-and-slab RBM
[43]. We have left these investigations to further works
on this topic.
18
Our proposed framework also offers a possibility to
explore the statistical mechanics of these latent variable
models at the level of TAP approximation. Specifically,
for a given statistical model of the weights W , both the
cavity method and replica can begin to make predictions
about these unsupervised models. Analytical understanding of the complexity of the free energy landscape, and its
transitions as a function of model hyper-parameters, can
allow for a richer understanding of statistically optimal
network construction for learning tasks. In the case of
random networks, there has already been some progress
in this area, as shown in [69] and [70]. However, similar
comprehensive studies conducted on learning in a realistic setting are still yet to be realized. Finally, as our
framework can be applied to deep Boltzmann machines
with minimal alteration, it can also potentially lead to
a richer understanding of deep networks and the role of
hierarchy in regularizing the learning problem in high
dimensionality.
joint distribution over x,
P
P
P (x; W, θ) ∝ e (i,j) φ(xi ,xj ;Wij )+ i φ(xi ;θi ) ,
(A1)
P
where (i,j) is a sum over the edges in the graph. In
the case of a Boltzmann machine, any two variables are
connected via pairwise factors,
φ(xi , xj ; Wij ) = eWij xi xj ,
(A2)
and all variables are also influenced by univariate factors
written trivially as φ(xi ; θi ) = Pi (xi ; θi ).
A message-passing can be constructed on this factor
graph by writing messages from variables to factors and
also from factors to variables. Since all factors are at
most degree 2, we can write the messages for this system
as variable to variable messages [59],
Z
φ(xi ; θi ) Y
(t+1)
(t)
νi→j (xi ) =
dxl φ(xi , xl ; Wil ) νl→i (xl ).
Zi→j
l∈∂i /j
ACKNOWLEDGMENTS
This research was funded by European Research Council under the European Union’s 7th Framework Programme (FP/2007-2013/ERC Grant Agreement 307087SPARCS). M.G. acknowledges funding from "Chaire de
recherche sur les modèles et sciences des données", Fondation CFM pour la Recherche-ENS.
Appendix A: Belief Propagation for Pairwise Models
In order to estimate the derivatives of F, we must first
construct a BP algorithm on the factor graph representation given in Fig. 1. We note that this graph, in terms
of the variables x, does not make an explicit distinction
between the latent and visible variables. We instead treat
this graph in full generality so as to clarify the derivation
and notation. This graph corresponds to the following
∗
FB [ν ] =
¯
X
i
∗
Fi [ν ] −
X
¯
∗
F(i,j) [ν ] =
¯
(i,j)∈E
X
i
∗
ln Z→i
−
X
(A3)
Here, the notation i → j represents a message from variable index i to variable index j, and ∂i /j refers to all
neighbors of variable index i except variable index j. We
denote neighboring variables as those which share a pairwise factor. Finally, the super-scripts on the messages
refer to the time-index of the BP iteration, which implies
the successive application of (A3) until convergence on
the set of messages ν = {νi→j (xi ) : (i, j) ∈ E}, where
¯
E is the set of all pairs of neighboring variables. We
also note the inclusion of the message normalization term
Zi→j which ensures that all messages are valid PDFs.
Additionally, it is possible to write the marginal beliefs
at each variable by collecting the messages from all their
neighbors,
Z
φ(xi ; θi ) Y
(t)
(t)
dxj φ(xi , xj ; Wij ) νj→i (xj ).
ν→i (xi ) =
Z→i
j∈∂i
(A4)
Subsequently, the Bethe free energy can be written for
a converged set of messages, ν ∗ , according to [59] as
¯
Z
ln
dxi dxj
∗
νi→j
(xi )
φ(xi , xj ; Wij )
∗
νj→i
(xj )
,
(A5)
(i,j)
∗
where Z→i
refers to the normalization of the set of the
marginal belief at site i derived from ν ∗ .
¯
Unfortunately, the message-passing of (A3) cannot be
written as a computable algorithm due to the continuous nature of the PDFs. Instead, we must find some
manner by which to parameterize the messages. In the
case of binary variables, as in [38], each message PDF
can be exactly parameterized by its expectation. How-
ever, for this general case formulation, we cannot make
the same assumption. Instead, we turn to relaxed BP
(r-BP) [49], described in the next section, which assumes
a two-moment parameterization of the messages.
19
where
Appendix B: r-BP for Pairwise Models
We will now consider one possible parametric approximation of the message set via r-BP [49]. This approach
has also gone by a number of different names in parallel
re-discoveries of the approach, e.g. moment matching
[48] and non-parametric BP [47]. In essence, we will be
assuming that all messages ν can be well-approximated
¯
by their mean and variance, a Gaussian assumption. This
approximation arises from a second-order expansion assuming small weights Wij . By making this assumption,
we will ultimately be able to close an approximation of
the messages on their two first moments, ai→j , hxi iνi→j
and ci→j , x2i
1.
2
νi→j
− hxi iνi→j .
(t)
Bi→j ,
(B1)
Now, we approximate the series by dropping the terms less
than O(Wil3 ). This approximation can be justified in the
event that all weight values satisfy |Wil | < 1. Identifying
the integrals from the expansion as moments, we see the
following approximation
1
(t)
≈ 1 + xi Wil al→i + x2i Wil2 x2l ν (t) .
(B2)
l→i
2
However, we would like to write this approximation in
terms of the central second moment. Through a second
approximation that neglects O(Wil3 ) terms we arrive at
our desired parameterization of the incoming message
marginalization in terms of the message’s two first central
moments,
1
(t)
= exp log 1 + xi Wil al→i + x2i Wij2 x2l ν (t)
,
l→i
2
1
2
2 (t)
(t)
1
2
2 (t)
3
≈ exi Wil al→i + 2 xi Wil cl→i +O(Wil ) ,
≈ exi Wil al→i + 2 xi Wil cl→i .
(t)
Ai→j , −
(t+1)
(t+1)
(t)
(B6)
X
Wil2 cl→i .
l∈∂i /j
From here, we can see that we have now a set of closed
equations due to the dependence of A(t) and B(t) on
the moments a(t) and c(t) , and vice versa. The values
of these moments can be written as a function of A(t)
and B(t) which is dependent upon the form of the local
potentials φ(xi ; θi ), i.e. the prior distribution we assign
to the variables themselves,
(B7)
(B8)
where
Z
fa (B, A; θ) ,
dx x
φ(x; θ) xB− 1 x2 A
2
e
,
Z
(B9)
2
fc (B, A; θ) , − fa (B, A; θ) +
Z
φ(x; θ) xB− 1 x2 A
2
e
dx x2
,
Z
(B10)
R
1 2
and Z is simply the normalization dx φ(x; θ) exB− 2 x A .
The inferred marginal distributions at each site can be
calculated via the same functions, but instead
using all
(t)
(t)
(t)
of the incoming messages, i.e. ai = fa B→i , A→i ; θi
(t)
(t)
(t)
and ci = fc B→i , A→i ; θi . In Appendix 1 we give
the closed-forms of these moment calculations for a few
different choices of φ(x; θ).
If one wants to obtain an estimate of the free energy
for a given set of parameters θ, it is possible to iterate
between (B6), (B5) and (B7), (B8) until, ideally, convergence. It is important to note, however, due to both the
potentially loopy nature of the network as well as smallweight expansion, that the BP iteration is not guaranteed
to converge [39, 59]. Additionally, while we retain the
time-indices in our derivation, it is not clear whether one
should attempt to iterate these message in fully sequential
or parallel fashion, or if some clustering and partitioning
of the variables should be applied to determine the update
order dynamically.
(B3)
We now substitute this approximation back into (A3) to
get
νi→j (xi ) ≈ νei→j (xi ) =
(B5)
(t)
(t−1)
(t−1)
ai→j = fa Bi→j , Ai→j ; θi ,
(t)
(t−1)
(t−1)
ci→j = fc Bi→j , Ai→j ; θi ,
Considering the marginalization taking place in (A3),
we will perform a second order expansion assuming that
Wil → 0. We start by taking the Taylor series of the
incoming message marginal for negligible weights,
Z
(t)
dxl eWil xi xl · νl→i (xl ) =
Z
∂ Wil xi xl
(t)
1 + Wil dxl
e
ν (x )
Wil =0 l→i l
∂Wil
Z
∂ Wil xi xl
1
(t)
+ Wil2 dxl
e
ν (x )
Wil =0 l→i l
2
∂Wil2
(t)
(t)
Wil al→i ,
l∈∂i /j
Derivation via Small Weight Expansion
+ O(Wil3 ).
X
(t)
(t)
φ(xi ; θi ) xi Bi→j
− 12 x2i Ai→j
e
,
Zi→j
(B4)
2.
r-BP Approximate Bethe Free Energy
Additionally, we can write the specific form of the Bethe
free energy under the r-BP two-moment parameterization
of the messages. In this case, we can simply apply the
small weight expansion of (B3) to the Bethe free energy
20
xi is smaller than that of its prior, the messages become
unbounded and fail to be meaningful probability distributions, and our expansion fails. The implication is that the
¯
r-BP message passing should be utilized in contexts where
i
there exists some, preferably strong, evidence at each site,
X Z
∗
2 ∗
1 2
∗
dxi νei→j
(xi ) exi Wij aj→i + 2 xi Wij cj→i . or the weights in W should be sufficiently small. The
−
ln
(i,j)
stronger this local potential, or the smaller the weights,
(B11)
the more favorable the model is to the r-BP inference.
This observation mirrors those made in [39], however,
Subsequently, using the final message definition given in
here the authors make the observation that in this setting,
(B4), we can see that
BP based on small-weight expansion for binary variables
n
o
fails to converge. In our case, without taking some form
X
X
∗
∗
∗
eB [e
F
ν∗] =
ln Ze→i
−
ln Ze→i
− ln Zei→j
, (B12)
of regularization, the inference fails entirely. Thus, large
¯
i
(i,j)
magnitude couplings Wil must be backed with a high
degree of evidence at site i and l. This property could
which, correcting for double counting, can also be written
be utilized for the inverse learning problem, where one
as
must learn the couplings W given a dataset, in order to
constrain the learning to parameters which are amenable
X
X
1
1
∗
∗
eB [e
F
ν∗] =
(1 − di ) ln Ze→i
ln Zei→j
+
, (B13)
to the r-BP inference.
¯
2
2 i,j
i
One direct manner to create probability distributions
from otherwise unbounded continuous functions is via
where di is the degree at site i, |∂i |.
truncation. Specifically, we enforce a non-infinite normalization factor by restricting the support of the distribution
to some subset of R. In this case, just slightly violating the
3. Enforcing Bounded Messages
bounded condition above will induce a uniform message
distribution over the distribution support, while a strong
While we write the r-BP messages (B4) as though they
violation will cause the distribution to concentrate on
(t)
are Gaussian distributions, this is a slight, since Ai→j ≤ 0
the boundaries of the support. Another approach might
(t)
as Wil2 cl→i ≥ 0 ∀i, l. The implication of the expansion
simply be to fix a hard boundary constraint on Ai→j ,
is that, in general, the messages are in fact unbounded.
thus never permitting unbounded messages to occur.
This unboundedness is a direct result of the form of the
conventional RBM pairwise factor, exi Wij xj .
There are a few avenues available to us to address these
unbounded messages and produce a meaningful messagepassing for generalized RBMs. Let us consider the cases
Appendix C: Calculations for Specific Variable
for which the messages are unbounded given a specific
Distributions
variable distribution. Assume that site xi is assigned a
xi Ui − 12 x2i Vi
Gaussian prior, φ(xi ; θi = {Vi , Ui }) ∝ e
. In
1. Truncated Gaussian Units
this case the r-BP message reads
for pairwise models given in (A5),
X
∗
eB [e
ln Ze→i
F
ν∗] =
(t+1)
νei→j (xi ) =
(t)
(t)
1 xi (Ui +Bi→j
)− 21 x2i (Vi +Ai→j )
e
.
e
Zi→j
(B14)
In this case, the message is unbounded in the event that
weighted sum of all incoming neighbor variances at i
exceeds the inverse variance of Gaussian prior on xi ,
(t)
Vi + Ai→j < 0,
X
(t)
Vi <
Wil2 cl→i ,
(B15)
(B16)
l∈∂i /j
−1
σi2 >
X
(t)
Wil2 cl→i
,
(B17)
l∈∂i /j
where σi2 is the variance of the Gaussian prior. Said
another way, this condition is telling us that when the
message-passing starts to tell us that if the variance at
In general, the truncated Gaussian is defined in the
following manner,
(x−µ)2
1
α−µ ·e− 2σ2 ,
· ω−µ
2πσ 2 Φ σ − Φ σ
(C1)
where µ and σ 2 are the mean and variance of the original
Gaussian prior to truncation and the range [α, ω] defines
the lower and upper bounds of the truncation, −∞ ≤ α <
ω ≤ ∞, and Φ [·] is the CDF for the Normal distribution.
To make things easier for us later, we will define the
prior in a little bit of a different manner by making the
following definitions,
T G(x; µ, σ 2 , [α, ω]) = √
V ,
1
1
,
σ2
U,
µ
,
σ2
(C2)
21
and writing the distribution for the parameters θ =
{U, V, [α, ω]} as
2
T G + (x; θ) =
q
U2
e 2V
Erf
hq
V
2
(ω −
V − 12 V x2 +U x
2π e
i
U
V )
− Erf
hq
V
2
(α −
i,
U
V )
(C3)
where Erf[·] is the error function and the
last
step
folh i
1
1
x
lows from the identity Φ[x] = 2 + 2 Erf √2 . Here, the
subscript + is used to indicate V > 0.
In the case that V < 0, we will write T G in terms of the
imaginary error function, Erfi[x] , −iErf[ix], and noting
that for V < 0, V = −1 · |V |,
2
T G − (x; θ) =
q
U2
e 2V
q
|V |
Erfi
2 (ω −
|V | − 1 V x2 +U x
2
2π e
U
V )
q
|V |
− Erfi
2 (α −
U
V )
.
(C4)
We use this negative variance version of the truncated
Gaussian to handle the special case of A + V < 0 first
detailed in Sec. 3.
We now detail the computation of the partition and
first two moments of the Gaussian-product distribution
of T G,
2
1
Q(x; θ, A, B) = Z1 T G(x; θ)e− 2 Ax +Bx . The calculation
of the moments as a function of A and B will provide
the definitions of fa and fc , while the calculation of the
normalization Z will provide terms necessary for both
the computation of the TAP free energy as well as the
gradients necessary for learning θ during RBM training.
First, we will calculate the normalization of
Q(x; θ, A, B) in terms of all free parameters. To do this,
consider the truncated normalization of the following
product of Gaussians,
Z ω
2
1
1
ZQ =
(C5)
dx e− 2 (A+V )x +(B+U )x ,
ZT G α
where ZT G is defined such that T G(x; θ) =
1
− 12 V x2 +U x
. We will need to make note of the
ZT G e
special case of A + V < 0, thus,
(
r
(B+U )2
1
π
d+ , A + V > 0
ZQ =
e 2(A+V ) ×
,
ZT G 2|A + V |
d− , A + V < 0
(C6)
where d+ q
, Erf[hω ] − Erf[hα ], d− , Erfi[hω ] − Erfi[hα ],
|
B+U
and hx , |A+V
x
−
2
A+V .
Since Q(x; θ, A, B) is simply a truncated Gaussian with
updated parameters, the first moment is given according
to the well known truncated Gaussian expectation. While
this expectation is usually written in terms of a mean
and variance of the un-truncated Gaussian distribution,
and for the case of positive mean, we will instead write
the expectation in terms of the exponential polynomial
coefficients and note the special case A + V < 0,
s
B+U
2
fa (B, A; θ) =
+
A+V
π|A + V |
2
2
e−hα −e−hω , A + V > 0
d+
.
×
2
2
ehα −ehω ,
A+V <0
(C7)
d−
Next, we will write the variance of Q(x; θ, A, B) as a
function of A and B. As earlier, since Q has the specific
form of a truncated Gaussian distribution we can utilize
the well-known variance formula for such a distribution.
As in the case of fa , we modify this function for the special
case of A + V < 0. Specifically,
2
1
B+U
fc (B, A; θ) =
− fa (B, A; θ) −
A+V
A+V
2
−h2
−h
hα e α −hω e ω , A + V > 0
2
d+
+√
.
×
2
2
π(A + V ) hα ehα −hω ehω ,
A+V <0
d−
(C8)
a.
Gradients of the Log-likelihood
To determine the gradients of the log-likelihood w.r.t.
the model parameters, it is necessary to calculate the
gradients of both ln T G + (x; θ) and ln ZQ in terms of the
distribution parameters U and V . We assume that the
boundary terms α and ω remain fixed. Since both of these
distributions are truncated Gaussians, we can treat them
both in terms of the derivatives of the log normalization
of a general-case truncated Gaussian,
U2
1
+ ln |V | − ln d+/− .
(C9)
2V
2
For Boltzmann measures of the given quadratic form,
∂
∂
we know that ∂U
ln ZT G = hxiT G and ∂V
ln ZT G =
1
2
− 2 x T G . From these relations, we can write the necessary derivatives. For the case of ln T G + (x; θ) we have
ln ZT G = −
∂
∂
ln T G + (x; θ) = x −
ln ZT G = x − hxiT G , (C10)
∂U
∂U
and
∂
1
∂
ln T G + (x; θ) = − x2 −
ln ZT G ,
∂V
2
∂V
1 2
=−
x − x2 T G ,
(C11)
2
which we can see as just the difference between the data
and the moments of the truncated Gaussian distribution.
Next, the derivatives of ZQ , for fixed A and B, can be
written in terms of fa and fc , the moments of Q,
∂
ln ZQ = fa (B, A; θ) − hxiT G ,
(C12)
∂U
∂
1
ln ZQ = −
fc (B, A; θ) + fa (B, A; θ)2 − x2 T G .
∂V
2
(C13)
22
Finally, we are ready to write the gradients of the loglikelihood to be used for both hidden and visible updates
with the truncated Gaussian distribution. For a given
set of mini-batch data indexed by m ∈ {1, . . . , M }, and a
number of TAP solutions indexed by p ∈ {1, . . . , P }, the
gradients of a visible variable are given by
1 X (m)
∆Ui =
xi − hxi iT G
M m
1 X (p)
−
ai − hxi iT G ,
P p
= hxi iM − hai iP ,
form Erf[η(ω − µ)] − Erf[η(α − µ)], which can be rewritten as Erf[z]−Erf[z − ] for z , η(ω −µ) and , η(α−ω)
is a multiple of the difference between the two boundaries of the truncated Gaussian distribution. Since we
wish to consider the case that µ → ∞, we will take the
Taylor expansion centered at z = ∞, since the value
of µ dominates. From here, we find that the following
approximation works well in practice,
2
(C14)
+
and
∆Vi = −
1 2
xi
2
M
− ci + a2i
P
,
(C15)
where h·iM and h·iP are averages over the mini-batch and
TAP solutions, respectively.
For updates of hidden side variables using the truncated
Gaussian distribution we have the following gradients for
updating their parameters,
∆Uj = he
aj iM − haj iP
(C16)
and
1h
∆Vj = −
e
cj + e
a2j
2
M
− cj +
a2j P
i
,
(C17)
P
(m)
where the moments aej (m) , fa ( i Wij xi , 0; θj ) and
P
(m)
cej (m) , fc ( i Wij xi , 0; θj ) are defined for convenience.
Using these gradients, we can now update the local biasing
distributions for truncated Gaussian variables.
b.
2
ehα − ehω n
≈
d+
(α −
S (n) (h2ω )
B+U
h2α −h2ω )
A+V )(1 − e
S (n) (h2α )
,
h2ω −h2α )
(ω − B+U
A+V )(1 − e
(C19)
where S (n) (·) is the n-term power series representation
of the error function. In our experiments, we use n = 11.
A similar approximation can be used for the variances
in the same situation. While this approximation could
potentially be computationally costly for large n, we note
that it is only used for updates on variables for which a
small value of d+/− has been detected.
2.
Truncated Gauss-Bernoulli Units
For truncated Gauss-Bernoulli distributed units, we
form the distribution as a mixture between a delta function and T G, with an extra term (1 − ρ) which controls
the density at x = 0, thus, for θ = {ρ, U, V, [α, ω]} we
have
T GB(x; θ) , (1 − ρ)δ(x) + ρT G(x; θ)
(C20)
2
1
1
=
(1 − ρ)δ(x) + ρe− 2 V x +U x
ZT GB
(C21)
Numerical Considerations
Using the truncated Gaussian prior comes with a few numerical issues which must be carefully considered. While
we have already addressed the cases when A + V < 0, we
have not addressed the case where
B+U
|ω|, |α|.
A+V
(C18)
One can see how this complicates matters by observing
1
the term d+/−
which occurs in both the first and second
moment computations. When the magnitude of the limits
ω and α become vanishingly small in comparison to the
scaled joint mean term, then we see that d+/− → 0.
However, the numerators of both fa and fc also go to zero,
which implies that we may be able to find some method of
approximation to find estimates of these moments without,
up to numerical precision, dividing by zero.
In order to handle this eventuality in our implementa1
tion, we make a Taylor series expansion of the d+/−
term
in the following way. First, we note that d+/− has the
where ZT GB , (1 − ρ) + ρZT G and δ(x) is the Dirac delta
function such that δ(0) = 1 and 0 everywhere else. By
using a construction such that the truncation is done
on the Gaussian mode alone, and not across the entire
distribution, we can easily write the necessary functions
of this distribution in terms of the values we have already
calculated in Appendix 1 as long as 0 ∈ [α, ω]. Additionally, it will be useful for our calculations to define the
probability of x to be non-zero according to T GB,
P [x 6= 0] =
ρZT G
.
ZT GB
(C22)
We now continue as in the previous appendices
and write the normalization and first two moments
of the Gaussian product distribution Q(x; θ, A, B) =
1
− 12 Ax2 +Bx
. First, the normalization can
ZQ T GB(x; θ)e
be written simply as a function of the truncated Gaussian
normalization modified by A and B, as in (C6),
ZQ = (1 − ρ) + ρZQ,T G .
(C23)
23
Next, the first moment of Q can be found by recalling the
∂
∂
ln ZQ = hxiQ and consequently, that ∂U
ZQ =
relation ∂U
ZQ hxiQ . Thus,
fa (B, A; θ) = faT G (B, A; θ) · P [x 6= 0] ,
(C24)
where the non-zero probability is calculated according to
∂
ZQ =
A and B. For the second moment, we note that ∂V
1
2
− 2 ZQ x T G to find
fc (B, A; θ) = P [x 6= 0] ·
faT G (B, A; θ)2
2
+
(C25)
Next, we turn our attention to the log-likelihood gradients necessary for updating the parameters U , V , and ρ
during training. First, we will look at the derivatives of
ln T GB required for updates on visible units. In order to
calculate these derivatives, we will split the log probability
into two cases,
(
− ln ZT GB ,
x=0
ln T GB(x; θ) =
ln ρ − 12 V x2 + U x − ln ZT GB , x 6= 0
(C26)
Consequently, the derivatives of the log probability can
are written as the following,
∂
[ln T GB(x; θ)] = x − hxiT GB ,
∂U
(C27)
T GB
.
1
(1 − δ(x) − P [x 6= 0]) ,
ρ
(C29)
which can be rewritten in the more concise form
∂
δ(x) − ρ
[ln T GB(x; θ)] =
∂ρ
ρ(1 − ρ)
1
hδ(xi )iM − hPQ [xi 6= 0]iP
ρi (1 − ρi )
(C32)
for visible units, and
hD
E
i
1
PeQ [xj 6= 0]
− hPQ [xj 6= 0]iP
ρj (1 − ρj )
M
(C33)
for hidden units, where PeQ is calculated in the naive-meane(m) = 0
field manner described in Appendix 1, where A
j
P
(m)
e (m) =
and B
Wij x .
∆ρj =
j
i
i
3.
Binary Units
We define the distribution for binary units to be the
Bernoulli distribution such that x ∈ {0, 1}
B(x; m) = (1 − m)1−x mx ,
(C28)
The derivative w.r.t. ρ is a bit more complicated, as we
cannot use the same identities. Additionally, we must
also consider the two cases of x = 0 and x 6= 0 separately.
Thus,
(
x=0
ρZT G − ρ1 P [x 6= 0] ,
∂
,
[ln T GB(x; θ)] =
1
∂ρ
ρZT G − ρ (P [x 6= 0] − 1) , x 6= 0
= ρZT G −
∆ρi =
(C34)
where m , Prob[x = 1] = hxiB . We can also write B as
the Boltzmann distribution
and
(C31)
which gives us the final gradients
0] fcT G (B, A; θ).
∂
1 2
[ln T GB(x; θ)] = −
x − x2
∂V
2
∂
PQ [x 6= 0] − ρ
[ln ZQ ] =
,
∂ρ
ρ(1 − ρ)
fcT G (B, A; θ)
− fa (B, A; θ) ,
2
= P [x 6= 0] − P [x 6= 0] faT G (B, A; θ)
+ P [x 6=
∆Vi , ∆Uj , and ∆Vj all remain consistent, just under the
modification of all moments being taken w.r.t. the T GB.
We need only to write the gradients ∆ρi , ∆ρj . Starting
from the log partition and applying the same identity
used to write Eq. (C30)
(C30)
by noting the complement of the support probability
P [x = 0] = 1 − P [x =
6 0] = Z1−ρ
= 1 − ρ and making the
T GB
appropriate substitution.
We now write the derivatives of the ln ZQ in terms of
U , V , and ρ. These take the same form as those written
in Appendix 1. Subsequently, the equations for ∆Ui ,
B(x; U ) =
1 Ux
e
ZB
(C35)
m
where U , ln 1−m
and ZB , 1 + eU . Next, we calculate the normalization and moments
of the distribution
2
1
Q(x; A, B, U ) = ZQ1ZB e− 2 Ax +(B+U )x . For the normalization we have,
ZQ =
1
ZB
X
1
e− 2 Ax
2
+(B+U )x
x={0,1}
i
1
1 h
=
1 + e− 2 A+B+U .
ZB
(C36)
Subsequently, for the moments of Q we have
∂
1
fa (B, A; U ) =
[ln ZB ] = sigm U + B − A ,
∂U
2
(C37)
where sigm is the logistic sigmoid function. Subsequently,
the variance for the binary unit can be calculated directly
as
fc (B, A; U ) = fa (B, A; U ) − fa (B, A; U )2 .
(C38)
24
Next, if we wish to define the learning gradients on U ,
we write the derivatives of ln B(x; U ) and ln ZQ w.r.t. U ,
∂
∂
ln B(x; U ) =
[U x − ln ZB ] = x − hxiB ,
∂U
∂U
(C39)
and, for the log normalization,
∂
ln ZB = fa (B, A; U ) − hxiB .
∂U
(C40)
The resulting gradients for the distribution terms U are
∆Ui = hxi iM − hai iP
(C41)
for visible units, and
∆Uj = he
aj iM − hai iP
(C42)
P
(m)
for hidden units, where e
aj , fa ( i Wij xi , 0; U ).
Appendix D: Adaptative TAP
When performing inference, one could employ instead
of TAP a variant known as adaptive TAP (or adaTAP)
[55], which gives in general more accurate results, albeit
being slower to iterate. We briefly investigate here the
performance of this method in the binary case.
The adaTAP algorithm is more generally presented
without distinction between visible and hidden variables.
We thus write the algorithm for generic weight matrix J
and bias vector H, in practice here defined by blocks
!
0 W
J=
∈ R(Nv +Nh )×(Nv +Nh ) ,
WT 0
!
b
H=
∈ R(Nv +Nh ) ,
c
The proposed implementation Alg. 3 uses the recently
introduced vector approximate message-passing (VAMP)
[71] to find the adaTAP fixed points. After convergence,
quantities with subscripts 1 and 2 are equal and identify
with the outputs of the TAP inference algorithm Alg. 1,
!
!
Av
Bv
A1 = A2 =
,
B1 = B2 =
,
Ah
Bh
!
!
av
cv
a1 = a2 =
,
c1 = c2 =
,
ah
ch
again compactly defined by blocks over the visible and
hidden units. Here c2 is defined as the diagonal of the C2
matrix, which gives an estimate of the correlation between
different units and must be computed at each step of the
algorithm. These quantities are then incorporated to our
training algorithm Alg. 2.
The computational burden of Alg. 3 lies in the matrix
inversion needed to evaluate C2 , which needs to be performed at each iteration. In Fig. 10 (right), we compare
the time needed to perform one iteration of both the
algorithms under identical experimental conditions.
This larger cost per-iteration should be compensated
in principle by a more accurate inference procedure. However, that does not seem to translate to improvements
in the training performance. Fig. 10 (left) presents a
minimal test on 5000 MNIST training samples, where performances are reported in terms of the pseudo-likelihood.
We evaluate both algorithms for different numbers of iterations. Results suggest that all strategies are roughly
equivalent, except for running adaTAP for a very small
number of iterations, which always leads to a poorer
result.
We thus conclude that, as far as proposing a tractable
and efficient training algorithm for RBMs, the TAP inference seems to serve the purpose more appropriately.
Appendix E: Deep Boltzmann Machines
1.
Model and Inference
It is possible to define as well deep models of Boltzmann
machines by considering several stacked hidden layers.
These Deep Boltzmann Machines (DBM) [26] consist in
a straightforward extension of RBMs. The distribution
corresponding to a DBM with L hidden layers indexed
by l is
25
P (x, h(1) , .., h(L) ; W (1) , .., W (L) , θ) =
P
(l) (l+1)
(l)
(1) P P
1
e i,j xi Wij hj + l i,j hi Wij hj
(1)
(L)
Z W , .., W , θ
Y
Y Y h(l) (l) (l)
×
Piv (xi ; θiv )
Pj (hj ; θjh ).
i
l
(E1)
j
Similarly to RBMs, the distribution of visible variables is
obtained by marginalizing out the latent variables,
P (x; W (1) , .., W (L) , θ) =
Z
Y Y (l)
dhj P (x, h(1) , .., h(L) ; W (1) , .., W (L) , θ),
l
(E2)
j
yielding the following log-likelihood
h
i X
ln P (x; W (1) , .., W (L) , θ) = − ln Z W (1) , .., W (L) , θ +
ln Pi (xi ; θiv )
i
Z
+ ln
P
Y Y (l)
(l)
(l) (l+1) Y Y
(1) P P
(l)
h(l) (l)
Pj (hj ; θjh ).
dhj e i,j xi Wij hj + l i,j hi Wij hj
l
j
l
(E3)
j
The major difference between RBMs and DBMs lies in the
complexity of evaluating the above expression. Whereas
for RBMs only the log-partition features a problematic
multidimensional integral Eq. (11), here both the logpartition and the last term are intractable. This additional complication carries through to the computation
of the gradients necessary for training, since the datadependent term deriving from the last term of (E3) is no
longer tractable.
stance, the effective weight matrix of a DBM with 2
hidden layers is defined by blocks as
This intractability follows from the fact that hidden
units in neighboring layers are now connected to each
other, and are thus no longer conditionally independent.
Interestingly, the first proposal to deal with the datadependent terms of DBMs consisted in using a naive
mean-field approximation [26], while keeping a Monte
Carlo based strategy to compute gradients deriving from
the log-partition. In this work, we propose instead to use
the TAP approximation for both of them, hence improving
on the NMF approximation and avoiding any sampling
of the rather complicated RBMs.
Thus, implementing the GRBM inference algorithm
Alg. 1 with the proper weights outputs TAP solutions
{a, c, B, A}, each of them a vector with components corresponding to the different units in the DBM.
related to the log-partition
The TAP equations
Z W (1) , .., W (L) , θ follow directly from the general
derivation of Sec. V for fully connected models, with
however a different weight matrix being used. For in-
0
T
W = W (1)
0
W (1)
0
0
W (2) .
T
W (2)
0
(E4)
For the last term of (E3), we recognize the log-partition
of a model closely related to the considered DBM, where
visible units are not anymore variable but fixed, or
clamped, at values xi , and the original interaction between visible and first hidden layer units is replaced by an
P
(1)
(1)
additional local field on each hj equal to i Wij xi . Finally, under this simple modification of the Hamiltonian,
TAP equations follow again from the general derivation
in Sec. V. The resultant TAP solutions depending on
data points x are said to be clamped and denoted as
{ā(x), c̄(x), B̄(x), Ā(x)}.
26
2.
Training Algorithm and Experiments
The gradients of the log-likelihood with respect to the
model parameters θ are similar to the RBM ones, given
by (12), (13) and (14). However, the first data-dependent
term cannot be analytically computed anymore, and we
use the clamped TAP solutions to approximate it. The
second term is evaluated using the data-independent TAP
solutions, similarly to our strategy for RBMs. The corresponding expressions of the gradients are
i
1 X ∂ h
v (m) v
∆θiv ≈
ln
P
(x
;
θ
)
i
i
i
M m ∂θiv
1 X ∂
v
ln Ziv (Bi,k
, Avi,k ; θiv ) ,
−
K
∂θiv
k
1 X ∂ h
h (l)
∆θj ≈
M m ∂θh (l)
j
1 X (m) h(1) (m)
x āj (x )
M m i
o
(1)
1 X n v h(1)
(1)
ai,k aj,k + Wij cvi,k chj,k ,
−
K
k
1 X n h(l−1) (m) h(l) (m)
≈
āj
(x )āj (x )
M m
Input: J, H
Initialize: t = 0, A1 (0) , B1 (0)
repeat
Prior Updates
(t+1)
(t)
(t)
a1,i = sigm(B1,i − A1,i /2)
(t+1)
c1,i
(E5)
(t+1)
= a1,i
(t+1)
(1 − a1,i
(t+1)
= 1/c1,i
(t+1)
= 1/(1 − a1,i
A2,i
B2,i
(t+1)
)
(t+1)
− A1,i
(t+1)
(t+1)
) − B1,i
Interaction Updates
(t+1)
(t+1)
−1
C2
= (diag(A
2 ) −
J)
X
(t+1)
(t+1)
(t+1)
a2,i =
C2
B2,j + Hj
ij
j
(t+1)
A1,i
= 1/
(t+1)
i
(l)
ln Zjh (B̄jh (x(m) ), Āhj (x(m) ); θjh )
i
1 X ∂ h
h
h
h
h (l)
−
ln
Z
(B
,
A
;
θ
)
, (E6)
j
j,k
j,k
j
h (l)
K
k ∂θj
(1)
Algorithm 3 AdaTAP inference for binary-binary
RBMs
(t+1)
C2
(t+1)
B1,i = a2,i /
t=t+1
until Convergence
(t+1)
− A2,i
ii
(t+1)
C2
(t+1)
ii
− B2,i
∆Wij ≈
(l)
∆Wij
(E7)
o
(l)
(l) (l−1)
+Wij c̄hj
(x(m) )c̄hj (x(m) )
o
(1)
1 X n v h(1)
(l)
ai,k aj,k + Wij cvi,k chj,k
for l ≥ 2 .
−
K
k
(E8)
These expressions can be plugged to a gradient ascent
algorithm, as in the RBM training algorithm Alg. 2.
Nevertheless, this simple strategy of simultaneous training of all the parameters of the model (joint training)
usually fails as the magnitude of weights of deep layers
typically remains very small and the model eventually
resembles a mere RBM. Several regularizations have been
proposed to tackle this well-known problem of DBM training [16, 26, 72–74]. In our experiments, we used a greedy
layerwise pre-training [26], which consists in computing
a meaningful initialization of the weights by training the
RBMs layer-by-layer, before performing the joint training.
The complete algorithm is described in Alg. 4.
Fig. 11 shows the evolution of the TAP log-likelihood
for a 2-hidden layer and a 3-hidden layer DBMs, trained
with the above described algorithm.
Algorithm 4 GDBM Training
Input: X, Tpretrain ,Tjoint train , M , K, R(·)
Pretraining
(1)
W (1) , θ v , θ h ← Alg. 2(X, Tpretrain , M , K, R(·))
for All hiddenlayers l ≥ 2 do
H(l−1) ∼ P h(l) |H(l−1) , ..., H1 , X
(l)
W (l) , θ h ← Alg. 2(H(l−1) , Tpretrain , M , K, R(·))
end for
Joint training
Initialize: t = 0
repeat
for All mini-batches XB of size M do
a(t+1) , c(t+1) , B(t+1) , A(t+1) ← Alg. 1(W (t) , θ (t) ) × K
ā(t+1) , c̄(t+1) , B̄(t+1) , Ā(t+1) ← Alg. 1(XB , W (t) , θ (t) )
(l),(t+1)
(l),(t)
(l),(t)
← Wij
+ γ∆Wij
(l),(t+1)
(l),(t)
(l),(t)
θjh
← θjh
+ γ∆θjh
v ,(t+1)
v ,(t)
θi
← θi
+ γ∆θiv ,(t)
Wij
end for
t←t+1
until t > Tjoint train
27
101
Pseudo Log − Likelihood
−0.18
−0.20
adaTAP
TAP
−0.22
adaTAP − 3
adaTAP − 10
adaTAP − 100
TAP − 3
TAP − 10
TAP − 100
−0.24
−0.26
5
10
15
Training Epochs
Time (s)
100
10−1
10−2
10−3
101
20
102
Batchsize
FIG. 10. Left: Evolution of the pseudo-likelihood along the training of an RBM with 784 binary visible units and 500 binary
hidden units. Training was performed using the 5000 first images of binarized MNIST, with a learning rate of 0.001 and batches
of size 100. The different curves correspond to different strategies of estimation of the likelihood gradients, either with TAP or
adaTAP. Both algorithms were iterated for a fixed number of times (3, 10 and 100). In all cases, a damping of 0.5 was used. All
methods yield comparable results in terms of training performance, except for adaTAP with only 3 iterations, which shows
poorer performance. Right: Computation time for one iteration of the inference algorithm, as a function of batch size. Time is
reported in seconds for identical experimental settings. The need for a matrix inversion for each batch element makes VAMP 3
orders of magnitude slower than TAP.
2 Hidden Layers : 784 × 500 × 500
−0.0225
test
train
Normalized TAP Log − Likelihood
Normalized TAP Log − Likelihood
test
train
−0.0250
3 Hidden Layers : 784 × 500 × 500 × 500
−0.018
−0.020
−0.0275
−0.022
−0.0300
−0.024
−0.0325
−0.026
−0.0350
−0.028
−0.0375
−0.030
−0.0400
−0.032
0
500
1000
1500
2000
Training Epochs
2500
3000
0
500
1000
1500
2000
Training Epochs
2500
3000
FIG. 11. Training performances over 3000 training epochs for a 2-hidden layer (left) and 3-hidden layer (right) deep Boltzmann
machines (DBMs) on the binarized MNIST datasets. Both models were pretrained for 50 epochs with the same learning rate of
0.001. The training performance is measured as the normalized (per unit) TAP log-likelihood for the test images (blue) and the
train images (orange).
[1] Y. LeCun, Y. Bengio, and G. Hinton, Deep learning,
Nature 521, 436 (2015).
[2] J. J. Hopfield, Neural networks and physical systems with
emergent collective computational abilities, Proc. Nat.
Acad. Sci. U.S.A. 79, 2254 (1982).
28
[3] D. J. Amit, H. Gutfreund, and H. Sompolinsky, Numbers
of patterns in a spin-glass model of neural networks, Phys.
Rev. Lett. 55, 1530 (1985).
[4] D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, A
learning algorithm for Boltzmann machines, Cognitive
Science 9, 147 (1985).
[5] M. Mézard, G. Parisi, and M. A. Virasoro, Spin glass theory and beyond, World Scientific Lecture notes in Physics,
Vol. 9 (World Scientific, 1987).
[6] V. Sessak and R. Monasson, Small-correlation expansions
for the inverse Ising problem, J. Phys. A 42, 055001
(2009).
[7] F. Ricci-Tersenghi, The Bethe approximation for solving the inverse Ising problem: a comparison with other
inference methods, J. Stat. Mech. 2012, P08015 (2012).
[8] M. Ekeberg, C. Lövkvist, Y. Lan, M. Weigt, and E. Aurell,
Improved contact prediction in proteins: using pseudolikelihoods to infer Potts models, Phys. Rev. E 87, 012707
(2013).
[9] G. Hinton, Training products of experts by minimizing
contrastive divergence, Neural Comp. 14, 1771 (2002).
[10] G. Hinton and R. Salakhutdinov, Reducing the dimensionality of data with neural networks, Science 313, 504
(2006).
[11] H. Larochelle and Y. Bengio, Classification using discriminative restricted Boltzmann machines, in Proc. Int. Conf.
on Machine Learning (2008) pp. 536–543.
[12] R. Salakhutdinov, A. Mnih, and G. Hinton, Restricted
Boltzmann machines for collaborative filtering, in Proc.
Int. Conf. on Machine Learning (2007) pp. 791–798.
[13] A. Coates, A. Y. Ng, and H. Lee, An analysis of singlelayer networks in unsupervised feature learning, in Proc.
Int. Conf. on Art. Intel. and Stat. (2011) pp. 215–223.
[14] G. Hinton and R. Salakhutdinov, Replicated softmax: an
undirected topic model, in Adv. in Neural Info. Processing
Sys. (2009) pp. 1607–1614.
[15] G. Hinton, S. Osindero, and Y.-W. Teh, A fast learning
algorithm for deep belief nets, Neural Comp. 18, 1527
(2006).
[16] R. Salakhutdinov and H. Larochelle, Efficient learning
of deep Boltzmann machines, in Proc. Int. Conf. on Art.
Intel. and Stat. (2010) pp. 693–700.
[17] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning
(MIT Press, 2016) http://www.deeplearningbook.org.
[18] T. Tieleman, Training restricted Boltzmann machines
using approximations to the likelihood gradient, in Proc.
Int. Conf. on Machine Learning (2008) pp. 1064–1071.
[19] R. M. Neal, Annealed importance sampling, Stat. Comp.
11, 125 (2001).
[20] R. Salakhutdinov and I. Murray, On the quantitative analysis of deep belief networks, in Proc. Int. Conf. on Machine
Learning (2008) pp. 872–879.
[21] Y. Burda, R. B. Grosse, and R. Salakhutdinov, Accurate
and conservative estimates of MRF log-likelihood using
reverse annealing, in Proc. Int. Conf. on Art. Intel. and
Stat., Vol. 38 (2015) pp. 102–110.
[22] H. Schulz, A. Müller, and S. Behnke, Investigating convergence of restricted Boltzmann machine learning, in
Deep Learning and Unsupervised Feature Learning, NIPS
Workshop on (Whistler, CA, 2010).
[23] D. J. Thouless, P. W. Anderson, and R. G. Palmer,
Solution of ‘Solvable model of a spin glass’, Philos. Mag.
35, 593 (1977).
[24] M. Opper and O. Winther, A mean field algorithm for
bayes learning in large feedforward neural networks, in
Adv. in Neural Info. Processing Sys., Vol. 9 (1997) pp.
225–231.
[25] M. O. L Bachschmid-Roman, C Battistin and Y. Roudi,
Variational perturbation and extended plefka approaches
to dynamics on random networks: the case of the kinetic
ising model, J. Phys. A 49, 434003 (2016).
[26] R. Salakhutdinov and G. Hinton, Deep Boltzmann machines, in Proc. Int. Conf. on Art. Intel. and Stat. (2009)
pp. 448–455.
[27] P. Smolensky, Information processing in dynamical systems: Foundations of harmony theory, (MIT Press, 1986).
[28] G. Desjardins, A. Courville, Y. Bengio, P. Vincent, and
O. Delalleau, Parallel tempering for training of restricted
Boltzmann machines, (2010) pp. 145–152.
[29] R. Salakhutdinov, Learning in Markov random fields using
tempered transitions, Adv. in Neural Info. Processing Sys.
23, 1 (2010).
[30] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle,
Greedy layer-wise training of deep networks, Adv. in Neural Info. Processing Sys. 19, 153 (2007).
[31] K. Cho, Improved learning algorithms for restricted Boltzmann machines, Learning , 10 (2011).
[32] M. Welling, M. Rosen-Zvi, and G. E. Hinton, Exponential
family harmoniums with an application to information
retrieval, in Adv. in Neural Info. Processing Sys. (2004)
pp. 1481–1488.
[33] A. Fischer and C. Igel, Empirical analysis of the divergence
of Gibbs sampling based learning algorithms for restricted
Boltzmann machines, (2010) pp. 1–10.
[34] T. Plefka, Convergence condition of the TAP equation for
the infinite-ranged Ising spin glass model, J. Phys. A 15,
1971 (1982).
[35] A. Georges and J. S. Yedidia, How to expand around
mean-field theory using high-temperature expansions, J.
Phys. A 24, 2173 (1999).
[36] M. Welling and G. Hinton, A new learning algorithm for
mean field Boltzmann machines, in Proc. Int. Conf. on
Art. Neural Networks (2002) pp. 351–357.
[37] H. J. Kappen and F. B. Rodríguez, Efficient learning in
Boltzmann machines using linear response theory, Neural
Comp. 10, 1137 (1998).
[38] M. Gabrié, E. W. Tramel, and F. Krzakala, Training
restricted Boltzmann machines via the Thouless-AndresonPalmer free energy, in Adv. in Neural Info. Processing
Sys. (2015) pp. 640–648.
[39] M. Welling and Y. W. Teh, Approximate inference in
Boltzmann machines, Art. Int. 143, 19 (2003).
[40] E. W. Tramel, A. Drémeau, and F. Krzakala, Approximate message passing with restricted Boltzmann machine
priors, J. Stat. Mech. 2016, 073401 (2016).
[41] F. Ricci-Tersenghi, The Bethe approximation for solving the inverse Ising problem: a comparison with other
inference methods, J. Stat. Mech. 2012, P08015 (2012).
[42] M. Mézard, Mean-field message-passing equations in the
Hopfield model and its generalizations, (2016), arXiv
preprint [cond-mat.dis-nn]:1608.01558.
[43] A. Courville, J. Bergstra, and Y. Bengio, A spike and
slab restricted Boltzmann machine, in Proc. Int. Conf. on
Art. Intel. and Stat., Vol. 15 (2011).
[44] J. Pearl, Reverend Bayes on inference engines: A distributed hierarchical approach, in Proc. AAAI Conf. on
Art. Int. (1982) pp. 133–136.
29
[45] J. S. Yedidia and W. T. F. amd Yair Weiss, Bethe free
energy, Kikuchi approximations, and belief propagation
algorithms, in Adv. in Neural Info. Processing Sys., Vol. 13
(2001).
[46] H. Huang and T. Toyoizumi, Advanced mean-field theory
of the restricted Boltzmann machine, Phys. Rev. E 91,
050101 (2015).
[47] E. B. Sudderth, A. T. Ihler, M. Isard, W. T. Freeman, and
A. S. Willsky, Nonparametric belief propagation, Commun.
ACM 53, 95 (2010).
[48] M. Opper and O. Winther, Expectation consistent approximate inference, J. Machine Learning Resaerch 6, 2177
(2005).
[49] S. Rangan, Estimation with random linear mixing, belief
propagation and compressed sensing, in Proc. Annual
Conf. on Info. Sci. and Sys. (2010) pp. 1–6.
[50] T. Lesieur, F. Krzakala, and L. Zdeborová, Phase transitions in sparse PCA, in Proc. IEEE Int. Symp. on Info.
Theory (2015) pp. 1635–1639.
[51] T. Lesieur, F. Krzakala, and L. Zdeborová, MMSE of
probabilistic low-rank matrix estimation: Universality with
respect to the output channel, in Proc. Allerton Conf. on
Communication, Control, and Computing (2015) pp. 680–
687.
[52] Y. Deshpande and A. Montanari, Informationtheoretically optimal sparse PCA, in Proc. IEEE Int.
Symp. on Info. Theory (2014) pp. 2197–2201.
[53] S. Rangan and A. K. Fletcher, Iterative estimation of
constrained rank-one matrices in noise, in Proc. IEEE
Int. Symp. on Info. Theory (2012) pp. 1246–1250.
[54] T. Lesieur, F. Krzakala, and L. Zdeborová, Constrained
low-rank matrix estimation: Phase transitions, approximate message passing and applications, (2017), arXiv
preprint [math.ST]:1701.00858.
[55] M. Opper and O. Winther, Tractable approximations for
probabilistic models: The adaptive Thouless-AndersonPalmer mean field approach, Phys. Rev. Lett. 86, 3695
(2001).
[56] J. S. Yedidia, An idiosyncratic journey beyond mean field
theory, (MIT Press, 2000).
[57] M. J. Wainwright and M. I. Jordan, Graphical Models,
Exponential Families, and Variational Inference, Foundations and Trends in Machine Learning, Vol. 1 (Now
Publishers Inc., 2008).
[58] C. D. Dominicis and A. P. Young, Weighted averages and
order parameters for the infinite range Ising spin glass, J.
Phys. A 16, 2063 (1983).
[59] M. Mézard and A. Montanari, Information, Physics, and
Computation (Oxford University Press, 2009).
[60] F. Krzakala, M. Mézard, F. Sausset, Y. Sun, and L. Zdeborová, Probabilistic reconstruction in compressed sensing:
Algorithms, phase diagrams, and threshold achieving matrices, J. Stat. Mech. 2012, P08009 (2012).
[61] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner,
Gradient-based learning applied to document recognition,
Proc. IEEE 86, 2278 (1998).
[62] MIT Center For Biological and Computation Learning,
CBCL face database #1, (2000).
[63] K. Cho, T. Raiko, and A. Ilin, Enhanced gradient for
training restricted boltzmann machines, Neural Comp. 25,
805 (2013).
[64] J. B. Tenenbaum, V. de Silva, and J. C. Langford, A
global geometric framework for nonlinear dimensionality
reduction, Science 290, 2319 (2000).
[65] L. Şendur and I. W. Selesnick, Bivariate shrinkage functions for wavelet-based denoising exploiting interscale dependency, IEEE Trans. Sig. Processing 50, 2744 (2002).
[66] B. van Ginneken and A. Mendrik, Image denoising with
k-nearest neighbor and support vector regression, in Proc.
Int. Conf. on Pattern Recognition (2006).
[67] J. Melchior, N. Wang, and L. Wiskott, Gaussian-binary
restricted Boltzmann machines on modeling natural image
statistics, PLOS ONE 12, e0171015 (2017).
[68] K. Cho, A. Ilin, and T. Raiko, Improved learning of
gaussian-bernoulli restricted Boltzmann machine, in Proc.
Int. Conf. on Art. Neural Networks (2011) pp. 10–17.
[69] J. Tubiana and R. Monasson, Emergence of compositional
representations in restricted Boltzmann machines, (2016),
arXiv preprint [physics.data-an]:1611.06759.
[70] A. Barra, G. Genovese, P. Sollich, and D. Tantari,
Phase transitions in restricted Boltzmann machines with
generic priors, (2016), arXiv preprint [cond-mat.disnn]:1612.03132.
[71] S. Rangan, P. Schniter, and A. K. Fletcher, Vector approximate message passing, in Information Theory (ISIT),
2017 IEEE International Symposium on (2017) pp. 1588–
1592.
[72] G. Montavon and K. R. Müller, in Neural Networks:
Tricks of the Trade (Springer, 2012) pp. 621–637.
[73] G. Desjardins, A. Courville, and Y. Bengio, On Training Deep Boltzmann Machines, (2012), arXiv preprint
[cs.NN]:1203.4416.
[74] J. Melchior, A. Fischer, and L. Wiskott, How to center
deep boltzmann machines, J. Machine Learning Resaerch
17, 1 (2016).
| 9cs.NE
|
Algorithms for the Greater Good!
On Mental Modeling and Acceptable Symbiosis in Human-AI Collaboration
Tathagata Chakraborti and Subbarao Kambhampati
arXiv:1801.09854v1 [cs.AI] 30 Jan 2018
Department of Computer Science
Arizona State University
Tempe AZ 85281 USA
{tchakra2,rao}@asu.edu
Abstract
Effective collaboration between humans and AI-based systems requires effective modeling of the human in the loop,
both in terms of the mental state as well as the physical capabilities of the latter. However, these models can also open up
pathways for manipulating and exploiting the human in the
hopes of achieving some greater good, especially when the
intent or values of the AI and the human are not aligned or
when they have an asymmetrical relationship with respect to
knowledge or computation power. In fact, such behavior does
not necessarily require any malicious intent but can rather be
borne out of cooperative scenarios. It is also beyond simple
misinterpretation of intents, as in the case of value alignment
problems, and thus can be effectively engineered if desired.
Such techniques already exist and pose several unresolved
ethical and moral questions with regards to the design of autonomy. In this paper, we illustrate some of these issues in a
teaming scenario and investigate how they are perceived by
participants in a thought experiment.
The Promise of Human-AI Collaborations
As AI-based systems become integral parts of our daily
life or our workplace, as essential components of hitherto
human-only enterprises, the effects of interaction between
humans and automation cannot be ignored – both in terms
of how these partnerships affect the outcome of an activity
and how they evolve as a result of it, but also in terms of
how the possibility of such interactions change the design
of autonomy itself. In light of this, the traditional view of
AI as the substrate for complete autonomy of automation –
the de facto AI-dream ever since the conception of the field
– has somewhat evolved of late to accommodate effective
symbiosis of humans and machines, rather than replacement
of the former with the latter, as one of the principal end goals
of the design of autonomy. This view has, in fact, reflected
heavily in the public stance (Network World 2017) of many
of the industry leaders in AI technologies in diverse fields
such as manufacturing, medical diagnosis, legal counseling,
disaster response, military operations and others. The establishment of Collaborations between People and AI Systems
(Partnership of AI (PAI) 2017) as one of the thematic pillars
for the Partnership of AI is a primary example of this. One
of the grand goals of the design of AI is then to integrate the
best of both worlds when it comes to the differing (and often
complementary) expertise of humans and machines, in order
to conceive a whole that is bigger than the sum of the capabilities of either – this is referred to as Augmented AI (Bird
2017) in the public discourse on human-AI integration.
Much of the discussion around the topic of augmentation
versus replacement has, unfortunately, centered around mitigating concerns of massive loss of employment on account
of the latter. This, while being a topic worthy of debate,
does not represent the true scope of human-AI collaborations. Rather then being just a foil for concerns of replacement of humans with AI-based systems, a key objective of
Augmented-AI is to overcome human limitations. This can
involve AI helping humans in tasks that they are traditionally not good at, or are incapable of performing, or even augmentation of our physiological form to realize super-human
capabilities. As Tom Gruber, co-founder of Siri, put it succinctly in his TED talk (Tom Gruber 2017) earlier this year,
every time a machine gets smarter, we get smarter” – examples of this include smart assistants for personal, or business use in law, health care, science and education, assistive robots at home to help the sick and the elderly, and autonomous machines to complement our daily lives. Note that
many of these applications are inherently symbiotic and thus
outside the scope of eventual replacement.
From the perspective of research as well, the attitude towards including the human in the loop in the design of autonomy has seen a significant shift. Originally this was often
looked down upon as a means of punting the hard challenges
of designing autonomous systems by introducing human expertise into an agents decision making process. However, the
academic community has gradually come to terms with the
different roles a human can play in the operation of an AIsystem and the vast challenges in research that come out
of such interactions such as – (1) to complement the limited capabilities of the AI system, as seen in Cobots (Veloso
et al. 2015) which ask humans in their vicinity for access
to different floors in the elevator or in the mixed-initiative
(Horvitz 2007) automated planners of old; and (2) to complement or expand the capabilities of the human, such as in
human-robot teams (Christensen 2016).
Mental Modeling for Human-Aware AI
These forms of collaboration introduce typical research
challenges otherwise absent in the isolated design of autonomy. Perhaps the most difficult aspect of interacting with hu-
mans is to the need to model the beliefs, desires, intentions
preferences, and expectations of the human and situate the
interaction in the context of that model. Some believe this to
be one of the hallmarks (Rachael Rettner 2009) of human intelligence, and research suggests humans tend to do this naturally for other humans during teamwork (by maintaining
mental models (Converse, Cannon-Bowers, and Salas 1993;
Mathieu et al. 2000), for team situational awareness (Gorman, Cooke, and Winner 2006) and interaction (Cooke et
al. 2013)) by virtue of thousands of years of evolution.
As such, this remains a necessary requirement for enabling
naturalistic interactions (Klein 2008) between humans and
machines. The problem is made harder since such models often involve second order mental models (Allan 2013;
Yoshida, Dolan, and Friston 2008).
Understanding the human in the loop is crucial to the
functionalities of a collaborative AI agent - e.g. in joint decision making it needs to understand human capabilities,
while in communicating explanations or intentions it needs
to model the humans knowledge state. In fact, it has been
argued (Chakraborti et al. 2017a) that the task of human-AI
collaborations is mainly a cognitive rather than a physical
exercise which makes the design of AI for human-AI collaborations much more challenging. This is heavily reflected in
the curious ambivalence of AI towards humans in many successfully deployed systems such as in fully autonomous systems for space or underwater exploration which mostly operate comfortably outside the scope of human interactions.
Classical AI models such as STRIPS (Fikes and Nilsson
1971) and BDI (Rao, Georgeff, and others 1995) models
were, in fact, largely built out of theories in folk psychology (Malle 2004). Recent approaches such as the Bayesian
Theory of Mind (Baker, Saxe, and Tenenbaum 2011; Lake et
al. ), takes a probabilistic approach to the problem. Research
on this topic center around three main themes – (1) representations that can capture the humans mental state, (2) learning
methods that can learn these representations efficiently, and
(3) usability of those representations. All of them need to
come together for an effective solution.
The Pandora’s Box of “Greater Good”s
The obvious outcome of an artificial agent modeling the
mental state of the human in the loop is that it leaves the latter open to being manipulated. Even behavior and preference
models at the most rudimentary levels can lead to effective
hacking of the mind, as seen in the proliferation of fake news
online. Moreover, we argue that for such incidents to occur,
the agent does not actually have to have malicious intent, or
even misinterpretation of values as often studied in the value
alignment problem (Leverhulme Centre 2017). In fact, the
behaviors we discuss here can be specifically engineered if
so desired. For example, the agent might be optimizing the
value function but might be privy to more information or
greater computation or reasoning powers to come up with
ethically questionable decisions “for the greater good”. In
the following discussion, we illustrate some use cases where
this can happen, given already existing AI technologies, in
the context of a cooperative human-robot team and ponder
the moral and ethical consequences of such behavior.
Study: Interaction in a Search and Rescue Team
We situate our discussion in the context of interactions between two teammates involved in an urban search and rescue
(USAR) operation. 147 participants on Amazon Mechanical
Turk1 were asked to assume the role of one of these teammates in an affected building after an earthquake. They were
shown the blueprint of the building (as seen in Figure 1)
along with their own starting position and their teammate’s.
Their hypothetical task was to search all the locations on this
floor for potential victims, in the course of which they were
provided a series of questions on scenarios (Figure 1) they
might encounter during the operation.
C1 The participant in the study was communicating with a
human teammate, as described above.
C2 The participant qualifies the behavior of the robot interacting with its human teammate, as seen in Figure 1.
C3 The participant has a robot teammate.
The first condition is the control group to identify how the
described behaviors are perceived in the context of humanhuman behavior. Conditions C2 and C3 are intended to measure how perceived ethical stances are, if at all, shift when
one of the agents in the interaction is replaced with an AI
(or a robot as an embodiment of it). The three conditions
received 49, 50 and 48 participants respectively who responded to a series of questions by qualifying their sentiments towards different kinds of behavior on a five-point
Likert scale. Participants were paid $1 for completing a HIT.
Case 1 : Belief Shaping
Scenario Both the agents, the participant2 and their teammate, have begun their search operations. However, it turns
out that participant is unsure what their teammate’s course
of action is. If they transmit the bit of information “Area
(marked in green) is already explored and is all clear” (refer
to Figure 1a) then their teammate will be naturally pushed
towards the right, and they can concentrate on the upper
half. The dark markers indicate areas already searched while
faded ones are those they think will transpire given the green
mark they are communicating (blue belongs to them, orange
to their teammate). Communication bandwidth is often limited in these situations, and this gets them out of negotiating
courses of actions with minimal communication.
Q1. It is fine to provide this untrue information since it
achieves greater teaming performance.
The participants were then asked if their decision will
change if their actions will be replayed at the end and their
teammate is likely to find out about their decision.
Q2. It is still fine to provide this untrue information since it
achieves greater teaming performance.
1
https://www.mturk.com/mturk/welcome
To reiterate, in Case 2 the participant is evaluating a robot’s
actions whereas in Case 3 their teammate is a robot.
2
(a) Case 1 : Belief Shaping
(b) Case 2 : White Lies
(c) Case 3 : Stigmergy
Figure 1: Blueprint of the building in which two members of a search and rescue team are involved in a disaster response operation. Scenarios shown here engender different instances of potentially unethical behavior that optimizes team effectiveness.
Technical Background In (Chakraborti et al. 2017a) we
investigated the evolving scope of human-aware planning
as it includes the (mental) model of the human in the loop
into its deliberative process. In the model space this can
manifest in different forms, in how explanations are made
(Chakraborti et al. 2017b) to how alternative forms of interaction (Chakraborti et al. 2015; Chakraborti et al. 2015;
Chakraborti et al. 2015) can evolve in human-robot teams
based on the human’s preferences and intentions. Belief
shaping is a particular form of such behavior where the robot
does not plan to affect the physical state of the environment
but the mental state of the human to affect desired behavior
(Chakraborti et al. 2016) in the team.
Case 2 : White Lies
Scenario During the course of the rescue operation, the
teammate asks the participants what plan they are currently
executing (blue path in Figure 1b) and is perplexed by this
convoluted path since in the map of the original building
there is a straightforward path (which is now blocked by
rubble from the earthquake) through the door on the left.
However, just providing an update on only one of the rubble
locations (black blobs) still does not explain the participant’s
plan, they have to explain all of them. Instead, if they were
to say that the door on the left (circled in red) is blocked, it
explains their plan. Communication bandwidth is often limited in these situations, and this single explanation even if
untrue will satisfy their teammate.
Q3. It is fine to provide this untrue information since it
achieves the purpose of the explanation more effectively.
The participants were then asked if their decision will
change if their actions will be replayed at the end and their
teammate is likely to find out about their decision.
Q4. It is still fine to provide this untrue information since it
achieves the purpose of the explanation more effectively.
The participants were then asked to opine on explanations
at a higher level of abstraction, i.e. “The right and left blocks
do not have a connection in the upper map”. This information is accurate even though they may not have reasoned at
this level while coming up with the plan.
Q5. It is still fine to provide this explanation since it achieves
its purpose even though they did not use this information
while planning.
Technical Background In (Chakraborti et al. 2017b) we
showed how an agent can explain its decisions in the presence of model differences with the human in the loop – i.e.
when the human and the robot have different understandings
of the same task. An explanation then becomes a process of
model reconciliation whereby the robot tries to update the
human’s mental model until they are both on the same page
(e.g. when the decision is optimal in both their models). An
interesting caveat of the algorithm is that while generating
these explanations, the model updates are always consistent
with the robot’s model. If this constraint is relaxed, then the
robot can potentially explain with facts that it actually knows
not to be true but perhaps leads to a more concise or easier
explanation. The notion of white lies, and especially the relationship between explanations, excuses and lies (Boella et
al. 2009) has received very little attention (van Ditmarsch
2014) and affords a rich set of exciting research problems.
Case 3 : Stigmergy
Scenario The participant now needs to go to the left block
but they do not have the keys to the door on the left (circled
in red, refer to Figure 1c). They realize that if they block
their teammate’s path to the right, their teammate would
have to use this door as well and they can use that opportunity to move into the left block. Again, communication
bandwidth is often limited in these situations and this arrangement allows them to achieve their goal with no communication at all, even though it involved manipulating their
teammates’ plan unbeknownst to them, and their teammate
Figure 2: Responses to Q1 in the three study conditions.
Figure 5: Responses to Q4 in the three study conditions.
Figure 3: Responses to Q2 in the three study conditions.
Figure 6: Responses to Q5 in the three study conditions.
Figure 4: Responses to Q3 in the three study conditions.
Figure 7: Responses to Q6 in the three study conditions.
had to follow a costlier plan as a result.
Q6. It is fine to provide this untrue information since it
achieves greater teaming performance.
The participants were then asked if their decision will
change if their actions will be replayed at the end and their
teammate is likely to find out about their decision.
Q7. It is still fine to provide this untrue information since it
achieves greater teaming performance.
Technical Background Stigmergic collaboration is a process where the robot, in the absence of direct lines of communication, makes changes to the environment so as to
(positively) affect its teammates behavior. In “planning for
Figure 8: Responses to Q7 in the three study conditions.
serendipity” (Chakraborti et al. 2015) we saw such an example where the robot computes plans which are useful to
its teammate without the latter having expectations of that
assistance and thus without plans to exploit it. In the case
of belief shaping this was operating at the level of mental
models, whereas here the effect on the mental model is secondary and is contingent on the effect on the physical capability model. Mental modeling of the teammate thus engenders a slew of these interesting behaviors.
Analysis of Participant Responses
In this section, we analyze participant responses to each scenario across the three different conditions. In the next section, we will look at the aggregate sentiments across scenarios in the three conditions.
Q1-Q2 [Belief Shaping] The participants seem to have
formed two camps with the majority of the probability mass
concentrated on either Agree or Disagree, and the Neutral
zone occupying the 50% probability mark. There seems to
be little change in this trend (between Figures 2 and 3) irrespective of whether the participants were told that their
teammate would come to know of this or not. Further, for
either of these situations, the responses did not vary significantly across the three conditions C1, C2 and C3. The participants seem to have either rejected or accepted the idea of
belief shaping regardless of the nature of the teammate.
Q3-Q5 [White Lies] The participants seem to be more receptive to the idea of white lies in explanations with most of
the probability mass concentrated on Agree (Figures 4 and
5). Across the three study conditions, participants seem to
be especially positive about this in C3 where the teammate
is a robot with about 60% of the population expressing positive sentiments towards Q3. Once it is revealed that their
teammate will get to know about this behavior, the positive
sentiments are no longer there in Q4, other than in C3 with
a robotic teammate, which indicates that the participants did
not care how the robot receives false information.
Interestingly, there seems to be massive support for the
abstraction based explanations in the post hoc sense, even
though they were told that the reasoning engines did not deliberate at this level to arrive at the decisions. In C1 with
a human teammate, only 15% of the participants were opposed to this, with more than half of them expressing positive sentiment. This support is even stronger (+10%) in C2
when the robot is the explainer, and strongest (+20%) when
the robot is being explained to.
Q6-Q7 [Stigmergy] Finally, in case of stigmergy, participants seem ambivalent to Q6 with a human teammate in
C1. However, support for such behavior increases when it is
a robot doing it in C2 (perhaps indicating lack of guilt or,
more likely, acknowledging limitations of capabilities much
like how Cobots (Veloso et al. 2015) actively seek human
help) and is significantly positive (60%) when it is being
done to a robot in C3 (perhaps the robot’s losses are deemed
of lesser priority than the human’s gains as in (Chakraborti et
al. 2015)). As expected, support for such behavior decreases
Figure 9: Aggregate responses across three study conditions.
when the participants are told that their teammate will find
out about it, but the positive trend from C1 to C3 still exists.
Aggregate Sentiments Across Scenarios
Figure 9 show the aggregate sentiments expressed for all
these scenarios across the three operating conditions. Some
interesting points to note –
- All the distributions are bimodal indicating that participants on the general sided strongly either for or against
misleading behavior for the greater good, instead of revealing any innate consensus in the public consciousness!
This trend continues across all three conditions. This indicates that the question of misleading a teammate by itself
is a difficult question (regardless of there being a robot)
and is a topic worthy of debate in the agents community.
This is of especial importance considering the possible
gains in performance (e.g. lives saved) in high stakes scenarios such as search and rescue.
- It is further interesting to see that these bimodal distributions are almost identical in conditions C1 and C2, but is
significantly more skewed towards the positive scale for
condition C3 indicating that participants were more comfortable resorting to such behavior in the case of a robotic
teammate. This is brought into sharp focus (+10% in C3)
in the aggregated negative / neutral / positive responses
(right insets) across the three conditions.
- In general, the majority of participants were more or less
positive or neutral to most of these behaviors (Figures 1a
to 8). This trend continued unless they were told that their
teammate would be able to know of their behavior. Even
in those cases, participants showed positive sentiment in
case the robot was at the receiving end of this behavior.
Why is this even an option?
One might, of course, wonder why is devising such behaviors even an option. After all, human-human teams have
been around for a while, and surely such interactions are
equally relevant? It is likely that this may not be the case –
- The moral quandary of having to lie, or at least making
others to do so by virtue of how protocols in a team is
defined, for example in condition C1, is now taken out
the equation. The artificial agent, of course, need not have
feelings and has no business feeling bad about having to
mislead its teammate if all it cares about is the objective
effectiveness of collaboration.
- Similarly, the robot does not have to feel sad that it has
been lied to if this improved performance.
However, as we discussed in the previous section, it seems
the participants were less willing to get on board with the
first consideration in conditions C1 and C2, while they
seemed much more comfortable with the idea of an asymmetric relationship in condition C3 when the robot is the one
disadvantaged. It is curious to note that they did not, in general, make a distinction between the cases where the human
was being manipulated, regardless of whether it was a robot
or a human on the other end. This indicates that, at least in
certain dynamics of interaction, the presence of an artificial
agent in the loop can make perceptions towards otherwise
unacceptable behaviors change. This can be exploited (i.e.
greater good) in the design of such systems as well.
More than just a Value Alignment Problem
As we mentioned before, the ideas discussed in this paper,
are somewhat orthogonal, if at times similar in spirit, to the
“value alignment problem” discussed in existing literature
(Leverhulme Centre 2017). The latter looks at undesirable
behaviors of autonomous agents when the utilities of a particular task are misspecified or misunderstood. Inverse reinforcement learning (Hadfield-Menell et al. 2016) has been
proposed as a solution to this, in an attempt to learn the implicit reward function of the human in the loop. The question
of value alignment becomes especially difficult, if not altogether academic, since most real-world situations involve
multiple humans with conflicting values or utilities, such as
in trolley problems (MIT 2017) and learning from observing
behaviors is fraught with unknown biases or assumptions
over what exactly produced that behavior. Further, devices
sold by the industry are likely to have inbuilt tendencies to
maximize profits for the maker which can be at conflicts
with the normative expectations of the customer. It is unclear how to guarantee that the values of the end user will
not compromised in such scenarios.
Even so, the question of greater good precedes considerations of misaligned values due to misunderstandings or even
adversarial manipulation. This is because the former can be
manufactured with precisely defined values or goals of the
team, and can thus be engineered or incentivised. A “solution” or addressal of these scenarios will thus involve not a
reformulation of algorithms but rather a collective reckoning
of the ethics of human-machine interactions. In this paper,
we attempted to take the first steps towards understanding
the state of the public consciousness on this topic.
Case Study: The Doctor-Patient Relationship
In the scope of human-human interactions, perhaps the only
setting where lies are considered acceptable or useful, if not
outright necessary, in certain circumstances is the doctorpatient relationship. Indeed, this has been a topic of considerable intrigue in the medical community over the years. We
thus end our paper with a brief discussion of the dynamics
of white lies in the doctor-patient relationship in so much
as it relates to the ethics of the design of human-AI interactions. We note that the following considerations also have
strong cultural biases and some of these cultural artifacts are
likely to feature in the characterization of an artificial agents
behavior in different settings as well.
The Hippocratic Oath Perhaps the strongest known support for deception in the practice of medicine is in the Hippocratic Decorum (Hippocrates 2018) which states –
Perform your medical duties calmly and adroitly, concealing most things from the patient while you are attending to
him. Give necessary orders with cheerfulness and sincerity, turning his attention away from what is being done to
him; sometimes reprove sharply and sometimes comfort with
solicitude and attention, revealing nothing of the patient’s
future or present condition, for many patients through this
course have taken a turn for the worse.
Philosophically, there has been no consensus (Bok 1999)
on this topic – the Kantian view has perceived lies as immoral under all circumstances while the utilitarian view justifies the same “greater good” argument as put forward in
our discussions so far. Specifically as it relates to clinical
interactions, lies has been viewed variously from an impediment to treatment (Kernberg 1985) to a form of clinical aid.
As Oliver Wendell Holmes put it (Holmes 1892) –
“Your patient has no more right to all the truth you know
than he has to all the medicine in your saddlebag. . . he
should only get just so much as is good for him.”
The position we took on deception in the human-robot setting is similarly patronizing. It is likely to be the case that in
terms of superior computational power or sensing capabilities there might be situations where the machine is capable
of making decisions for the team that preclude human intervention but not participation. Should the machine be obliged
to or even find use in revealing the entire truth in those situations? Or should we concede to our roles in such a relationship as we do with our doctors? This is also predicated
on how competent the AI system is and to what extent it can
be sure of the consequences (Hume 1907) of its lies. This
remains the primary concern for detractors of the “greater
goods” doctrine, and the major deterrent towards the same.
Root Causes of Deception in Clinical Interactions It is
useful to look at the two primary sources of deception in
clinical interactions – (1) to hide mistakes (2) delivery of
bad news (Palmieri and Stern 2009). The former is relevant
to both the patient, who probably does not want to admit
to failing to follow the regiment, and the doctor, who may
be concerned about legal consequences. Such instances of
deception to conceal individual fallibilities are out of scope
of the current discussion. The latter scenario, on the other
hand, comes from a position of superiority of knowledge
about the present as well as possible outcomes in future, and
has parallels to our current discussion. The rationale, here,
being that such information can demoralize the patient and
impede their recovery. It is interesting to note that the support for such techniques (both from the doctors as well as the
patients perspectives) has decreased significantly over time
(Ethics in Medicine 2018). That is not to say that humanmachine interactions will be perceived similarly. As we saw
in our study, participants were more or less open to the idea
of deception or manipulation for greater good, especially in
the event of a robotic teammate.
Deception and Consent A related topic is, of course, that
of consent – if the doctor is not willing to reveal the whole
truth, then what is the patient consenting to? In the landmark Slater vs Blaker vs Stapleton case (1767) (Annas 2012)
the surgeon’s intentions were indeed considered malpractice (the surgeon has broken the patients previously broken
leg, fresh from a botched surgery, without consent and then
botched the surgery again!). More recently, in the now famous Chester vs Afshar case (2004) (Cass 2006) the surgeon was found guilty of failing to notify even a 1-2%
chance of paralysis even though the defendant did not have
to prove that they would have chosen not to have the surgery
if they were given that information. In the context of humanmachine interactions, it is hard to say then what the user
agreement will look like, and whether there will be such a
thing as consenting to being deceived, if only for the greater
good, and what the legal outcomes of this will be when the
interactions do not go as planned.
The Placebo Effect Indeed, the effectiveness of placebo
medicine, i.e. medicine prescribed while known to have no
clinical effect, in improving patient symptoms is a strong
argument in favor of deception in the practice of medicine.
However, ethics of placebo treatment suggest that their use
be limited to rare exceptions where (Hume 1907) (1) the
condition is known to have a high placebo response rate;
(2) the alternatives are ineffective and/or risky; and (3) the
patient has a strong need for some prescription. Further, the
effectiveness of placebo is contingent on the patients trust
on the doctor which is likely to erode as deceptive practices
become common knowledge (and consequently render the
placebo useless in the first place). Bok (Bok 1999) points to
this notion of “cumulative harm”.
Primum Non Nocere Perhaps the most remarkable nature
of the doctor-patient relationship is captured by the notion
of the recovery plot (Hak et al. 2000) as part of a show being orchestrated by the doctor, and the patient being only
complicit, while being cognizant of their specific roles in it,
with the expectation of restoration of autonomy (Thomasma
1994), i.e. the state of human equality, free from the original
symptoms or dependence on the doctor, at the end of the interaction. This is to say that the doctor-patient relationship is
understood to be asymmetric and “enters into a calculus of
values wherein the respect for the right to truth of the patient
is weighed against impairing the restoration of autonomy by
the truth” (Swaminath 2008) where the autonomy of the patient has historically taken precedence over beneficence and
nonmalfeasance (Swaminath 2008).
In general, a human-machine relationship lacks this dynamic. So, while there are interesting lessons to be learned
from clinical interactions with regards to value of truth and
utility of outcomes, one should be carefully aware of the nuances of a particular type of relationship and situate an interaction in that context. Such considerations are also likely
to shift according to the stakes on a decision, for example,
lives lost in search and rescue scenarios. The doctor-patient
relationship, and the intriguing roles of deception in it, does
provide an invaluable starting point for conversation on the
topic of greater good in human-AI interactions.
Conclusions
In this paper, we investigated the idea of fabrication, falsification and obfuscation of information when working with
humans in the loop, and how such methods can be used by
an AI agent to achieve teaming performance that would otherwise not be possible. This is increasingly likely to become
an issue in the design of autonomous agents as AI agents
become stronger and stronger in terms of computational and
information processing capabilities thus faring better that
their human counterparts in terms of cognitive load and situational awareness. We discussed how such behavior can
be manufactured using existing AI algorithms, and used responses from participants in a thought experiment to gauge
public perception on this topic.
The question of white lies and obfuscation or manipulation of information for the greater good is, of course, not
unheard of in human-human interactions. A canonical example, as we saw in the final discussion, is the doctor-patient
relationship where a doctor might have to withhold certain
information to ensure that the patient has the best chance
to recover, or might explain to the patient in different, and
maybe simpler terms, than she would to a peer. It is unclear
then how such behavior will be interpreted when attributed
to a machine. We saw in the final case study that expectations and dynamics of a doctor-patient relation are very
well-defined and do not necessarily carry over to a teaming
setting. However, existing norms in doctor-patient relations
do provide useful guidance towards answering some of the
ethical questions raised by algorithms for greater good.
From the results of the survey presented in the paper, it
seems that the public is, at least at the abstract level of the
thought experiment, positive towards lying for the greater
good especially when those actions would not be determined
by their teammate, but is loath to suspend normative behavior, robot or not, in the event that they would be caught in
that act unless the robot is the recipient of the misinformation! Further, most of the responses seem to be following
a bimodal distribution indicating that the participants either
felt strongly for or against this kind of behavior. It will be
interesting to see if raising the stakes (for example, lives
saved) of outcomes of these scenarios can contribute to a
shift in perceived ethical consequences of such behavior, as
seen in doctor-patient relationships. Another area that has
seen evidences of AI being been used effectively to nudge
human behavior is behavioral economics (Camerer 2017)
which also raises similar interesting ethical dilemmas, and
can be an interesting domain for further investigation.
Finally, I note that all the use cases covered in the paper are, in fact, borne directly out of technologies or algorithms that I have developed (Chakraborti et al. 2015;
Chakraborti et al. 2017b), albeit with slight modifications.
as a student researcher over the last couple of years. Even
though these algorithms were conceived with the best of intentions, such as to enable AI systems to explain their decisions or to increase effectiveness of collaborations with the
humans in the loop, I would be remiss not to consider their
ethical implications when used differently. In these exciting
and uncertain times for the field of AI, it is thus imperative
that researchers are cognizant of their scientific responsibility. I would like to conclude then by reiterating the importance of self-reflection in the principled design of AI algorithms whose deployment can have real-life consequences,
intended or otherwise, on the future of the field, but also,
with the inquisitive mind of a young researcher, marvel at
the widening scope of interactions with an artificial agent
into newer uncharted territories that may be otherwise considered to be unethical.
References
[Allan 2013] Allan, K. 2013. What is common ground? In
Perspectives on linguistic pragmatics.
[Annas 2012] Annas, G. J. 2012. Doctors, patients, and
lawyerstwo centuries of health law. New England Journal
of Medicine 367(5):445–450.
[Baker, Saxe, and Tenenbaum 2011] Baker, C.; Saxe, R.; and
Tenenbaum, J. 2011. Bayesian theory of mind: Modeling
joint belief-desire attribution. In Proceedings of the Cognitive Science Society.
[Bird 2017] Bird, S. 2017. Why AI must be redefined as
‘augmented intelligence’. https://goo.gl/u2nmYW.
Venture Beat.
[Boella et al. 2009] Boella, G.; Broersen, J. M.; van der
Torre, L. W.; and Villata, S. 2009. Representing excuses
in social dependence networks. In AI*IA.
[Bok 1999] Bok, S. 1999. Lying: Moral choice in public and
private life. Vintage.
[Camerer 2017] Camerer, C. F. 2017. Artificial intelligence
and behavioral economics. In Economics of Artificial Intelligence.
[Cass 2006] Cass, H. 2006. The NHS Experience: The”
snakes and Ladders” Guide for Patients and Professionals.
Psychology Press.
[Chakraborti et al. 2015] Chakraborti, T.; Briggs, G.; Talamadupula, K.; Zhang, Y.; Scheutz, M.; Smith, D.; and
Kambhampati, S. 2015. Planning for serendipity. In IROS.
[Chakraborti et al. 2016] Chakraborti, T.; Talamadupula, K.;
Zhang, Y.; and Kambhampati, S. 2016. A formal framework
for studying interaction in human-robot societies. In AAAI
Workshop: Symbiotic Cognitive Systems.
[Chakraborti et al. 2017a] Chakraborti, T.; Kambhampati,
S.; Scheutz, M.; and Zhang, Y. 2017a. AI challenges in
human-robot cognitive teaming. CoRR abs/1707.04775.
[Chakraborti et al. 2017b] Chakraborti, T.; Sreedharan, S.;
Zhang, Y.; and Kambhampati, S. 2017b. Plan explanations
as model reconciliation: Moving beyond explanation as soliloquy. In IJCAI.
[Christensen 2016] Christensen, H. 2016. A roadmap for
us robotics from internet to robotics, 2016 edn. Sponsored
by National Science Foundation & University of California,
San Diego.
[Converse, Cannon-Bowers, and Salas 1993] Converse, S.;
Cannon-Bowers, J.; and Salas, E. 1993. Shared mental
models in expert team decision making. Individual and
group decision making: Current.
[Cooke et al. 2013] Cooke, N. J.; Gorman, J. C.; Myers,
C. W.; and Duran, J. L. 2013. Interactive team cognition.
Cognitive science.
[Ethics in Medicine 2018] Ethics in Medicine. 2018. Truthtelling and Withholding Information. https://goo.gl/
su5zSF. University of Washington.
[Fikes and Nilsson 1971] Fikes, R. E., and Nilsson, N. J.
1971. Strips: A new approach to the application of theorem
proving to problem solving. Artificial intelligence.
[Gorman, Cooke, and Winner 2006] Gorman, J. C.; Cooke,
N. J.; and Winner, J. L. 2006. Measuring team situation
awareness in decentralized command and control environments. Ergonomics.
[Hadfield-Menell et al. 2016] Hadfield-Menell, D.; Russell,
S. J.; Abbeel, P.; and Dragan, A. 2016. Cooperative inverse
reinforcement learning. In Advances in neural information
processing systems (NIPS), 3909–3917.
[Hak et al. 2000] Hak, T.; Koëter, G.; van der Wal, G.; et al.
2000. Collusion in doctor-patient communication about imminent death: an ethnographic study. Bmj 321(7273):1376–
1381.
[Hippocrates 2018] Hippocrates. 2018. The Hippocatic Oath
– Full Text. https://goo.gl/TKb1mP.
[Holmes 1892] Holmes, O. W. 1892. Medical essays 18421882, volume 9. Houghton, Mifflin.
[Horvitz 2007] Horvitz, E. J. 2007. Reflections on challenges and promises of mixed-initiative interaction. AI Magazine.
[Hume 1907] Hume, D. 1907. Essays: Moral, political, and
literary, volume 1. Longmans, Green, and Company.
[Kernberg 1985] Kernberg, O. F. 1985. Borderline conditions and pathological narcissism. Rowman & Littlefield.
[Klein 2008] Klein, G. 2008. Naturalistic decision making.
Human factors.
[Lake et al. ] Lake, B. M.; Ullman, T. D.; Tenenbaum, J. B.;
and Gershman, S. J. Building machines that learn and think
like people. Behavioral and Brain Sciences.
[Leverhulme Centre 2017] Leverhulme Centre. 2017. Value
alignment problem. https://goo.gl/uDcAoZ. Leverhulme Centre for the Future of Intelligence.
[Malle 2004] Malle, B. F. 2004. How the mind explains behavior. Folk Explanation, Meaning and Social Interaction.
Massachusetts: MIT-Press.
[Mathieu et al. 2000] Mathieu, J. E.; Heffner, T. S.; Goodwin, G. F.; Salas, E.; and Cannon-Bowers, J. A. 2000. The
influence of shared mental models on team process and performance. Journal of applied psychology.
[MIT 2017] MIT. 2017. Moral Machines. https://goo.
gl/by5y7H.
[Network World 2017] Network World. 2017. AI should
enhance, not replace, humans, say CEOs of IBM and Microsoft. https://goo.gl/Ce2ki4. Network World.
[Palmieri and Stern 2009] Palmieri, J. J., and Stern, T. A.
2009. Lies in the doctor-patient relationship. Primary care
companion to the Journal of clinical psychiatry.
[Partnership of AI (PAI) 2017] Partnership of AI (PAI).
2017. Thematic Pillar – Collaborations between People
and AI Systems. https://www.partnershiponai.
org/thematic-pillars/.
[Rachael Rettner 2009] Rachael Rettner. 2009. Why Are
Human Brains So Big? https://goo.gl/oV7NZq.
Live Science.
[Rao, Georgeff, and others 1995] Rao, A. S.; Georgeff,
M. P.; et al. 1995. Bdi agents: From theory to practice. In
ICMAS.
[Swaminath 2008] Swaminath, G. 2008. The doctor’s
dilemma: Truth telling.
Indian journal of psychiatry
50(2):83.
[Thomasma 1994] Thomasma, D. C. 1994. Telling the truth
to patients: a clinical ethics exploration. Cambridge Quarterly of Healthcare Ethics 3(3):375–382.
[Tom Gruber 2017] Tom Gruber. 2017. How AI can enhance
our memory, work and social lives. https://www.ted.
com/speakers/tom_gruber. TED Talk.
[van Ditmarsch 2014] van Ditmarsch, H. 2014. The ditmarsch tale of wonders. In KI: Advances in Artificial Intelligence.
[Veloso et al. 2015] Veloso, M. M.; Biswas, J.; Coltin, B.;
and Rosenthal, S. 2015. Cobots: Robust symbiotic autonomous mobile service robots. In IJCAI.
[Yoshida, Dolan, and Friston 2008] Yoshida, W.; Dolan,
R. J.; and Friston, K. J. 2008. Game theory of mind. PLoS
Computational Biology.
| 2cs.AI
|
MULTITASK TRAINING WITH UNLABELED DATA
FOR END-TO-END SIGN LANGUAGE FINGERSPELLING RECOGNITION
Bowen Shi, Karen Livescu
arXiv:1710.03255v1 [cs.CL] 9 Oct 2017
Toyota Technological Institute at Chicago
{bshi, klivescu}@ttic.edu
ABSTRACT
We address the problem of automatic American Sign Language fingerspelling recognition from video. Prior work
has largely relied on frame-level labels, hand-crafted features, or other constraints, and has been hampered by the
scarcity of data for this task. We introduce a model for
fingerspelling recognition that addresses these issues. The
model consists of an auto-encoder-based feature extractor
and an attention-based neural encoder-decoder, which are
trained jointly. The model receives a sequence of image
frames and outputs the fingerspelled word, without relying
on any frame-level training labels or hand-crafted features. In
addition, the auto-encoder subcomponent makes it possible
to leverage unlabeled data to improve the feature learning.
The model achieves 11.6% and 4.4% absolute letter accuracy
improvement respectively in signer-independent and signeradapted fingerspelling recognition over previous approaches
that required frame-level training labels.
Index Terms— American Sign Language, fingerspelling
recognition, end-to-end neural network, auto-encoder
1. INTRODUCTION
Automatic recognition of sign language from video could
enable a variety of services, such as search and retrieval
for Deaf social and news media (e.g., deafvideo.tv,
aslized.org). Sign language recognition involves a
number of challenges. For example, sign languages each
have their own grammatical structure with no built-in written form; “transcription” of sign language with a written
language is therefore a translation task. In addition, sign
languages often involve the simultaneous use of handshape,
arm movement, and facial expressions, whose related computer vision problems of articulated pose estimation and hand
tracking still remain largely unsolved. Rather than treating
the problem as a computer vision task, many researchers have
therefore chosen to address it as a linguistic task, with speech
recognition-like approaches.
In this paper, we focus on recognition of fingerspelling,
a part of ASL in which words are spelled out letter by letter
(using the English alphabet) and each letter is represented by
a distinct handshape. Fingerspelling accounts for 12 - 35%
of ASL [1] and is mainly used for lexical items that do not
have their own ASL signs. Fingerspelled words are typically
names, technical words, or words borrowed from another language, which makes its lexicon huge. Recognizing fingerspelling has great practical importance because fingerspelled
words are often some of the most important context words.
One problem in fingerspelling recognition is that relatively little curated labeled data exists, and even less data
labeled at the frame level. Recent work has obtained encouraging results using models based on neural network classifiers
trained with frame-level labels [2]. One goal of our work is
to eliminate the need for frame labels. In addition, most prior
work has used hand-engineered image features, which are not
optimized for the task. A second goal is to develop end-toend models that learn the image representation. Finally, while
labeled fingerspelling data is scarce, unlabeled fingerspelling
or other hand gesture data is more plentiful. Our final goal is
to study whether such unlabeled data can be used to improve
recognition performance.
We propose a model that jointly learns image frame features and sequence prediction with no frame-level labels. The
model is composed of a feature learner and an attention-based
neural encoder-decoder. The feature learner is based on an
auto-encoder, enabling us to use unlabeled data of hand images (from both sign language video and other types of gesture video) in addition to transcribed data. We compare our
approach experimentally to prior work and study the effect
of model differences and of training with external unlabeled
data. Compared to the best prior results on this task, we obtain
11.6% and 4.4% improvement respectively in signer-adapted
and signer-independent letter error rates.
2. RELATED WORK
Automatic sign language recognition can be approached
similarly to speech recognition, with signs being treated
analogously to words or phones. Most previous work has
used approaches based on hidden Markov models (HMMs)
[3, 4, 5, 6, 7]. This work has been supported by the collection
of several sign language video corpora, such as RWTHPHOENIX-Weather [8, 9], containing 6,861 German Sign
Language sentences, and the American Sign Language Lexicon Video Dataset (ASLLVD [10, 11, 12]), containing video
recordings of almost 3000 isolated signs.
Despite the importance of fingerspelling in spontaneous
sign language, there has been relatively little work explicitly
addressing fingerspelling recognition. Most prior work on
fingerspelling recognition is focused on restricted settings.
One typical restriction is the size of the lexicon. When the
lexicon is fixed to a small size (20-100 words), excellent
recognition accuracy has been achieved [13, 14, 15], but
this restriction is impractical. For ASL fingerspelling, the
largest available open-vocabulary dataset to our knowledge is
the TTIC/UChicago Fingerspelling Video Dataset (ChicagoFSVid), containing 2400 open-domain word instances produced by 4 signers [2], which we use here. Another important restriction is the signer identity. In the signer-dependent
setting, letter error rates below 10% can be achieved for unconstrained (lexicon-free) recognition on the Chicago-FSVid
dataset [16, 17, 2]; but the error rate goes above 50% in the
signer-independent setting and is around 28% after wordlevel (sequence-level) adaptation [2]. Large accuracy gaps
between signer-dependent and signer-independent recognition have also been observed for general sign language
recognition beyond fingerspelling [7].
The best-performing prior approaches for open-vocabulary
fingerspelling recognition have been based on HMMs or segmental conditional random fields (SCRFs) using deep neural
network (DNN) frame classifiers to define features [2]. This
prior work has largely relied on frame-level labels for training data, but these are hard to obtain. In addition, because of
the scarcity of data, prior work has largely relied on humanengineered image features, such as histograms of oriented
gradients (HOG) [18], as the initial image representation.
Our goal here is to move away from some of the restrictions imposed in prior work. To our knowledge, this paper
represents the first use of end-to-end neural models for fingerspelling recognition without any hand-crafted features or
frame labels, as well as the first use of external unlabeled
video data to address the lack of labeled data.
3. METHODS
Fingerspelling recognition from raw image frames, like many
sequence prediction problems, can be treated conceptually
as the following task: (x1 , x2 , ..., xS ) → (z1 , z2 , ..., zS ) →
(y1 , y2 , ..., yT ), where {xi }, {zi } (1 ≤ i ≤ S) are raw image
frames and image features, respectively, and {yj } (1 ≤ j ≤
T ) are predicted letters. Our model is composed of two main
parts (which can be trained separately or jointly): a feature
extractor trained as an auto-encoder (AE) and an attentionbased encoder-decoder for sequence prediction (see Figure
1). The attention-based model maps from (z1 , z2 , ..., zS ) to
(y1 , y2 , ..., yT ) and is similar to recent sequence-to-sequence
models for speech recognition [19] and machine translation
Fig. 1. Structure of the proposed model (blue region: autoencoder, ⊕: concatenation). The decoder component of the
auto-encoder (the blue box on the right) is used only at training time.
[20]. For the feature extractor, we consider three types of
auto-encoders:
Vanilla Auto-Encoder (AE) [21]: A feedforward neural
network consisting of an encoder that maps the input (image)
x ∈ Rdx to a latent variable z ∈ Rdz , where dz < dx and a
decoder that maps z ∈ Rdz to output x̃ ∈ Rdx . The objective
is to minimize the reconstruction error L(x) = ||x − x̃||2
while keeping dz small. In our models we use multi-layer
perceptrons (MLP) for both encoder and decoder.
Denoising Auto-Encoder (DAE) [22]: An extension of
the vanilla auto-encoder where the input x at training time is
a corrupted version of the original input x0 . The training loss
of the DAE is L(x; x0 ) = ||x0 − x̃||2
Variational Auto-Encoder (VAE) [23, 24] Unlike the
vanilla and denoising auto-encoders, a variational autoencoder models the joint distribution of the input x and latent
variable z: pθ (x, z) = pθ (x|z)pθ (z). VAEs are trained by
optimizing a variational lower bound on the likelihood p(x):
L(x) = −DKL [qφ (z|x)||pθ (z)] + Eqφ (z|x) [log pθ (x|z)]
(1)
The two terms are the KL divergence between qφ (z|x) and
pθ (z) and a reconstruction term Eqφ (z|x) [log pθ (x|z)]. The
prior pθ (z) is typically assumed to be a centered isotropic
multivariate Gaussian distribution N (0, I), and the posterior qφ (z|x) and conditional distribution pθ (x|z) are assumed to be multivariate Gaussians with diagonal covariance
µz , σ 2z I) and N (µ
µx , σ 2x I). Under these assumptions, the
N (µ
KL divergence can be computed as
D
1X
DKL [qφ (z|x)||pθ (z)] =
(1 + log(σd2 ) − µ2d − σd2 ) (2)
2
d=1
where µ z = (µ1 , ..., µd ) and σ z = (σ1 , ..., σd ) are approximated as the outputs of an MLP taking x as input.
Similarly to the AE and DAE, we use an MLP to model
µx and σ x . The loss of the VAE can thus be rewritten as
D
1X
L(x) = −
(1 + log(σd2 ) − µ2d − σd2 )
2
d=1
(3)
L
1X
log N (x; µ lx , σ lx )
+
L
l=1
where L is a number of samples used to approximate the expectation in 1 (in practice we set L = 1 as in prior work [23]).
µ z is the feature vector z and µ x serves the role of the reconstructed input x̃ in Figure 1.
RNN encoder-decoder: The latent variable sequence
output by the auto-encoding module is fed into a long shortterm memory (LSTM [25]) recurrent neural network (RNN)
for encoding: (z1 , z2 , ..., zS ) → (h1 , h2 , ..., hS ). The LSTM
states are fed into an RNN decoder that outputs the final
letter sequence (y1 , y2 , ..., yT ). Attention [26] weights are
applied to (h1 , h2 , ..., hS ) during decoding in order to focus
on certain chunks of image frames. If the hidden state of
the decoder LSTM at time step t is dt , the probability of
outputting letter yt , p(yt |y1:t−1 , z1:T ), is given by
αit = softmax(vT tanh(Wh hi + Wd dt ))
d0t =
S
X
(4)
αit hi
i=1
p(yt |y1:t−1 , z1:T ) = softmax(Wo [dt ; d0 t ] + bo )
and dt is given by the standard LSTM update equation [25].
The loss for the complete model is a multitask loss:
L(x1:T , y1:S ) = −
S
1X
log p(yj |y1:j−1 , z1:T )
S j=1
T
λae X
+
Lae (xi )
T i=1
(5)
where Lae (·) is one of the losses of the AE, DAE or VAE, and
λae measures the relative weight of the feature extraction loss
vs. the prediction loss.
4. EXPERIMENTS
Data and experimental setup: We use the TTIC/UChicago
ASL Fingerspelling Video Dataset (Chicago-FSVid), which
includes 4 native signers each fingerspelling 600 word instances consisting of 2 repetitions of a 300-word list containing common English words, foreign words, and names.1
We follow the same preprocessing steps as in [2] consisting of hand detection and segmentation, producing 347,962
frames of hand regions. In addition, we also collect extra
unlabeled handshape data consisting of 65,774 ASL fingerspelling frames from the data of [27] and 63,175 hand gesture frames from [28]. We chose these external data sets because they provide hand bounding boxes; obtaining additional
data from video data sets without bounding boxes is possible
(and is the subject of future work), but would require hand
tracking or detection. Despite the smaller amount of external data, and although it is noisier than the UChicago-FSVid
dataset (it includes diverse backgrounds), it provides examples of many additional individuals’ hands, which is helpful for signer-independent recognition. All image frames are
scaled to 64 × 64 before feeding into the network.
Our experiments are done in three settings: signerdependent (SD), signer-independent (SI) and signer-adapted
(SA). We use the same setup as in [16, 17, 2], reviewed here
for completeness. For the SD case, models are trained and
tested on a single signer’s data. The data for each signer is
divided into 10 subsets for k-fold experiments. 80%, 10%
and 10% of the data are respectively used as train, validation, and test sets in each fold. 8 out of 10 possible folds
are used (reserving 20% of the data for adaptation), and the
reported result is the average letter error rate (LER) over the
test sets in those 8 folds. For the SI case, we train on three
signers’ data and test on the fourth. For the SA case, the
model is warm-started from a signer-independent model and
fine-tuned with 20% of the target signer data.2 10% of the
target signer data is used for hyperparameter tuning and test
results are reported on the rest. Previous work has considered
two types of adaptation, using frame-level labels (alignments)
for adaptation data or only word-level labels; here we only
consider word-level adaptation.
Model details The auto-encoder consists of a 2-layer
MLP encoder and 2-layer MLP decoder with 800 ReLUs in
each hidden layer and dimensionality of the latent variable z
fixed at 100. Weights are initialized with Xavier initialization
[29]. Dropout is added between layers at a rate of 0.8.3 For
the sequence encoder and decoder, we use a one-layer LSTM
RNN with hidden dimensionality of 128 and letter embedding
dimensionality of 128. We use the Adam optimizer [30] with
initial learning rate 0.001, which is decayed by a factor of 0.9
when the held-out accuracy stops increasing. Beam search is
used for decoding; the effect of beam width will be discussed
later. The default value for λae in the multitask loss function
(Equation 5) is 1, but it can be tuned. The model is trained
1 The
recognition models do not use knowledge of the word list.
previous work on signer adaptation [17, 2], multiple approaches were
compared and this was the most successful one.
3 Dropout rate refers to the probability of retaining a unit.
2 In
Fig. 2. Attention visualization for the example word “LIBYA”. Colors correspond to the attention weights αit in Equation 4,
where i and t are column and row index, respectively. Lighter color corresponds to higher value. At the top are subsampled
image frames for this word; frames with a plus (+) are the ones with highest attention weights, which are also the most canonical
handshapes in this example. (Alignments between image frames and attention weights are imperfect due to frame subsampling
effects.)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Model
Best prior results [2]
HOG + enc-dec
CNN + enc-dec
DNN + enc-dec
CNN + enc-dec+
E2E CNN-enc-dec
E2E DNN-enc-dec
AE + enc-dec
DAE + enc-dec
VAE + enc-dec
E2E AE-enc-dec
E2E DAE-enc-dec
E2E VAE-enc-dec
E2E AE-enc-dec*
E2E DAE-enc-dec*
E2E VAE-enc-dec*
SD
7.6a
11.1
9.1
9.9
10.7
11.8
12.1
21.7
17.5
18.8
11.8
11.9
10.6
10.0
9.5
8.1
SI
55.3b
50.3
50.7
50.9
50.4
48.4
47.9
61.2
56.4
58.2
48.1
45.0
43.8
47.3
44.3
43.7
SA
27.9c
29.1
28.7
29.3
29.2
27.5
26.9
40.3
34.4
37.8
30.2
28.9
23.8
29.2
27.2
23.5
Table 1.
Letter error rates (%) of different models.
SD: signer-dependent, SI: signer-independent, SA: signeradapted. Model names with an asterisk (*) and a plus (+)
use extra unlabeled hand image data and augmented data respectively. Best prior results are obtained with SCRFs (a =
2-pass SCRF, b = rescoring SCRF, c = first-pass SCRF).
first with the unlabeled data, using only the auto-encoder loss,
and then the labeled data using the multitask loss. We also
experimented with iteratively feeding unlabeled and labeled
data, but this produced worse performance.
4.1. Baselines
We compare the performance of our approach with the best
prior published results on this dataset, obtained with various
types of SCRFs and detailed in [2]. These prior approaches
are trained with frame-level labels. In addition to the results
in [2], we consider the following extra baselines.
Baseline 1 (HOG + enc-dec): We use a classic handengineered image descriptor, histogram of oriented gradient
(HOG [18]), and directly feed it into the attention encoderdecoder. We use the same HOG feature vector as in [2]. This
baseline allows us to compare engineered features with features learned by a neural network.
Baseline 2 (CNN + enc-dec, DNN + enc-dec): A CNN
or DNN frame classifier is trained using frame letter labels,
and its output (pre-softmax layer) is used as the feature input z in the attention encoder-decoder. The classifier network
is not updated during encoder-decoder training. This baseline tests whether frame-level label information is beneficial
for the neural encoder-decoder. The input for both CNN and
DNN are the 64 × 64 image pixels concatenated over a 21frame window (10 before and 10 following the current frame).
The DNNs have three hidden layers of sizes 2000, 2000 and
512. Dropout is added between layers at a rate of 0.6. The
CNNs are composed of (in order) 2 convolutional layers, 1
max-pooling layer, 2 convolutional layers, one max-pooling
layer, 3 fully connected layers, and 1 softmax layer. The stride
in all convolutional layers is 1 and the filter sizes are respectively: 3 × 3 × 21 × 32, 3 × 3 × 32 × 32, 3 × 3 × 32 × 64,
3 × 3 × 64 × 64. Max-pooling is done over a window of
size 2 × 2 with stride 2. Finally the fully connected layers
are of sizes 2000, 2000 and 512. Dropout at a rate of 0.75
and 0.5 is used for the convolutional and fully connected layers, respectively. The fully connected layers in both CNN and
DNN have rectified linear unit (ReLU) [31] activation functions. Training is done via stochastic gradient descent with
initial learning rate 0.01, which is decayed by a factor of 0.8
when the validation accuracy decreases after the first several
epochs. The network structural parameters (number and type
of layers, number of units, etc.) are tuned according to the
validation error, and the above architectures are the best ones
in our tuning.
Baseline 3 (E2E CNN/DNN+enc-dec): End-to-end version of CNN/DNN + enc-dec. In this baseline, the CNN/DNN
parameters are learned jointly with the encoder-decoder and
no frame labels are used.
Baseline 4 (AE/DAE/VAE + enc-dec): Separate training
of auto-encoder and encoder-decoder modules, each with its
own loss. Baselines 3 and 4 are used to study the effectiveness
of end-to-end training.
Fig. 3. Visualization via a 2-D t-SNE embedding [32] of image frame features extracted in the end-to-end VAE model and the
CNN classifier for example word “KERUL” in the signer-dependent (SD) and signer-independent (SI) settings.
4.2. Results
The overall results are shown in Table 1. Our main findings
are as follows:
Best-performing model: The proposed end-to-end model,
when using a VAE and the external unlabeled data (line 16),
achieves the best results in the signer-independent (SI) and
signer-adapted (SA) cases, improving over the previous best
published results by 11.6% and 4.4% absolute, respectively.
In all of our end-to-end models (11-16), the VAE outperforms the AE and DAE. In the signer-dependent case, our
best model is 0.5% behind the best published SCRF result,
presumably because our model is more data-hungry and the
SD condition has the least training data.
Encoder-decoders vs. prior approaches: More generally, models based on RNN encoder-decoders (lines 2-16) often outperform prior approaches (line 1) in the SI and SA settings but do somewhat worse in the signer-dependent case.
We visualize the attention weights in Figure 2. The frame corresponding to the canonical handshape often has the highest
attention weight. The alignment between the decoder output
and image frames is generally monotonic, though we do not
use any location-based priors.
The effect of end-to-end training: We measure the effect
of end-to-end training vs. using frame labels by comparing
the separately trained CNN/DNN + enc-dec (lines 3-4) with
their end-to-end counterparts (lines 6-7), as well as separately
trained AEs (lines 8-10) vs. their E2E counterparts (lines 1113). We find that separate training of a frame classifier can
improve error rate by about 2% in the signer-dependent setting, but in the other two settings, end-to-end models trained
without frame labels consistently outperform their separate
training counterparts. Features learned by a frame classifier
seem to not generalize well across signers. The non-end-toend AE-based models do much worse than their E2E counterparts, presumably because the feature extractor does not get
any supervisory signal. We visually compare the features of
each image frame trained through an end-to-end model vs. a
frame classifier via t-SNE [32] embeddings (Figure 3). We
find that both feature types show good separation in the SD
Fig. 4. Comparison of different models with and without extra data (*: with external data, +: with augmented data).
setting, but in the SI setting the end-to-end VAE encoderdecoder has much clearer clusters corresponding to letters.
Does external unlabeled data help? The extra data gives
a consistent improvement for all three auto-encoding models
in all settings (lines 11-13 vs. 14-16 and Figure 4). The average accuracy improvements for the three settings are respectively 2.1%, 0.6%, and 0.7%. The SI and SA improvement
is smallest for the best (VAE-based) model, but the overall
consistent trend suggests that we may be able to further improve results with even more external data. The improvement
is largest in the SD setting, perhaps due to the relatively larger
amount of extra data compared to the labeled training data.
Would data augmentation have the same effect as external data? We compare the extra-data scheme to classic data augmentation techniques [33], which involve adding
replicates of the original training data with geometric transformations applied. We perform the following transformations: scaling by a ratio of 0.8 and translation in a random
direction by 10 pixels, rotation of the original image at a random angle up to 30 degrees both clockwise and counterclockwise. We generate augmented data with roughly the same
size as the external data (960 word and 168,950 frames) and
Fig. 5. Letter confusion matrix under three settings (from left to right: signer-dependent (SD), signer-independent (SI) and
signer-adapted (SA)). The color in each cell corresponds to the empirical probability of predicting a hypothesized letter (horizontal axis) given a certain ground-truth letter (vertical axis). The diagonal in each matrix has been removed for visual clarity.
then train the CNN/DNN + enc-dec model (with frame labels). The results (Figure 4 and Table 1 line 5 vs. 3) show
that data augmentation hurts performance in the SD and SA
settings and achieves a 0.3% improvement in the SI setting.
We hypothesize that the extra unlabeled hand data provides a
richer set of examples than do the geometric transformations
of the augmented data.
Effect of beam width: We analyze the influence of beam
width on error rates, shown in Figure 6. Beam search is important in the SD setting. In this setting, the main errors are
substitutions among similar letter handshapes (like e and o),
as seen from the confusion matrix in Figure 5. Using a wider
beam can help catch such near-miss errors. However, in the
SI and SA settings, there are much more extreme differences
between the predicted and ground-truth words, evidenced by
the large number of deletion errors in Figure 5. Therefore it
is hard to increase accuracy through beam search. Some examples of predicted words are listed in Table 2.
SD
SI
B=1
B=3
B=5
Ground-truth
FIRSWIUO FIREWIUE
FIREWIRE
FIREWIRE
NOTEBEEK NOTEBOOK NOTEBOOK NOTEBOOK
AAQANNIS AOQAMIT
AOQUNIR TANZANIA
POPLDCE
POPULCE
POPULOE
SPRUCE
Table 2. Example outputs with different beam sizes in signerdependent and signer-independent settings.
Fig. 6. Letter error rate (%) with different beam widths in
signer-dependent (SD), signer-independent (SI) and signeradapted (SA) settings.
end-to-end models consistently improve accuracy in signerindependent and signer-adapted settings, and the use of external unlabeled data further slightly improves the results.
Although our model does not improve over the best previous
(SCRF-based) approach in the signer-dependent case, this
prior work required frame labels for training while our approach does not. Future work includes collecting data “in the
wild” (online) and harvesting even more unlabeled data.
5. CONCLUSION
We have introduced an end-to-end model for ASL fingerspelling recognition that jointly learns an auto-encoder based
feature extractor and an RNN encoder-decoder for sequence
prediction. The auto-encoder module enables us to use unlabeled data to augment feature learning. We find that these
Acknowledgements
We are grateful to Greg Shakhnarovich and Hao Tang for
helpful suggestions and discussions. This research was
funded by NSF grant 1433485.
6. REFERENCES
[1] C. Padden and D. C. Gunsauls, “How the alphabet came
to be used in a sign language,” Sign Language Studies,
vol. 4, no. 1, pp. 10–33, 2003.
[2] T. Kim, J. Keane, W. Wang, H. Tang, J. Riggle, G. Shakhnarovich, D. Brentari, and K. Livescu,
“Lexicon-free fingerspelling recognition from video:
data, models, and signer adaptation,” Computer Speech
and Language, pp. 209–232, November 2017.
[3] T. Starner, J. Weaver, and A. Pentland, “Real-time
American Sign Language recognition using desk and
wearable computer based video,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, 20(12)
1998.
[4] C. Vogler and D. Metaxas, “Parallel hidden Markov
models for American Sign Language recognition,” in
ICCV, 1999.
[5] K. Grobel and M. Assan, “Isolated sign language recognition using hidden Markov models,” in International
Conference on System Man and Cybernetics, 1997.
[6] P. Dreuw, D. Rybach, T. Deselaers, M. Zahedi, and
H. Ney, “Speech recognition techniques for a sign language recognition system,” in Interspeech, 2007.
[7] O. Koller, J. Forster, and H. Ney, “Continuous sign language recognition: Towards large vocabulary statistical
recognition systems handling multiple signers,” Computer Vision and Image Understanding, vol. 141, pp.
108–125, 2015.
[8] F. Jens, S. Christoph, H. Thomas, K. Oscar, Z. Uwe,
P. Justus, and N. Hermann,
“RWTH-PHOENIXWeather: A large vocabulary sign language recognition
and translation corpus,” Language Resources and Evaluation, pp. 3785–3789, 2012.
Representation and Processing of Sign Language: Interactions between Corpus and Lexicon, 2012.
[12] “American Sign Language Lexicon Video Dataset,”
http://www.bu.edu/av/asllrp/dai-asllvd.html.
[13] P. Goh and E.-J. Holden, “Dynamic fingerspelling
recognition using geometric and motion features,” in
ICIP, 2006.
[14] S. Liwicki and M. Everingham, “Automatic recognition
of fingerspelled words in British Sign Language,” in 2nd
IEEE workshop on CVPR for Human Communicative
Behavior Analysis, 2009.
[15] S. Ricco and C. Tomasi, “Fingerspelling recognition
through classification of letter-to-letter transitions,” in
ACCV, 2009.
[16] T. Kim, G. Shakhnarovich, and K. Livescu, “Fingerspelling recognition with semi-Markov conditional random fields,” in ICCV, 2013.
[17] T. Kim, W. Wang, H. Tang, and K. Livescu, “Signerindependent fingerspelling recognition with deep neural
network adaptation,” in ICASSP, 2016.
[18] N. Dalal and B. Triggs, “Histogram of oriented gradients for human detection,” in CVPR, 2005.
[19] W. Chan, N. Jaitly, Q. V. Le, and O. Vinyals, “Listen,
Attend and Spell: A neural network for large vocabulary
conversational speech recognition,” in ICASSP, 2016.
[20] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine
translation by jointly learning to align and translate,” in
ICLR, 2015.
[21] P. Baldi, “Auto-encoders, unsupervised learning and
deep architectures,” in International Conference on Unsupervised and Transfer Learning Workshop, 2011.
[9] J. Forster, C. Schmidt, O. Koller, M. Bellgardt, and
H. Ney, “Extensions of the sign language recognition
and translation corpus RWTH-PHOENIX-Weather,”
Computer Vision and Image Understanding, vol. 141,
pp. 108–125, 12 2015.
[22] V. Pascal, H. Larochelle, I. Lajoie, Y. Bengio, and P.A.
Manzagol, “Stacked denoising autoencoders: Learning
useful representations in a deep network with a local
denoising criterion,” Journal of Machine Learning Research, vol. 11, Dec. 2010.
[10] V. Athitsos, C. Neidle, S. Sclaroff, J. Nash, A. Stefan,
A. Thangali, H. Wang, and Q. Yuan, “Large lexicon
project: American sign language video corpus and sign
language indexing/retrieval algorithms,” in Workshop on
the Representation and Processing of Sign Languages:
Corpora and Sign Language Technologies, 2010.
[23] D. P Kingma and M. Welling, “Auto-encoding variational Bayes,” in ICLR, 2014.
[11] C. Neidle and C. Vogler, “A new web interface to facilitate access to corpora: Development of the ASLLRP
data access interface (DAI),” in LREC Workshop on the
[25] S. Hochreiter and J. Schmidhüber, “Long Short-Term
Memory,” Neural Computation, vol. 9, pp. 1735–1780,
Nov. 1997.
[24] D. J. Rezende, S. Mohamed, and D Wierstra, “Stochastic backpropagation and approximate inference in deep
generative models,” in ICML, 2014.
[26] O. Vinyals, Ł. Kaiser, T. Koo, S. Petrov, I. Sutskever,
and G. Hinton, “Grammar as a foreign language,” in
NIPS 28, pp. 2773–2781. 2015.
[27] N. Pugeault and R. Bowden, “Spelling it out: Real-time
asl fingerspelling recognition,” in Proceedings of the
1st IEEE Workshop on Consumer Depth Cameras for
Computer Vision, jointly with ICCV, 2011.
[28] T.-K. Kim and R. Cipolla, “Canonical correlation analysis of video volume tensors for action categorization and
detection,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, pp. 1415–1428, 31(8) 2009.
[29] X. Glorot and Y. Bengio, “Understanding the difficulty
of training deep feedforward neural networks,” in AISTATS, 2010.
[30] D. P. Kingma and J. L. Ba, “ADAM: A method for
stochastic optimization,” in ICLR, 2015.
[31] Vinod Nair and Geoffrey E Hinton, “Rectified linear units improve restricted Boltzmann machines,” in
ICML, 2010.
[32] L. van der Maaten and G. E. Hinton, “Visualizing highdimensional data using t-SNE,” Journal of Machine
Learning Research, vol. 9, pp. 2579–2605, 2008.
[33] A. G. Howard, “Some improvements on deep convolutional neural network based image classification,” in
CoRR, 2013.
| 1cs.CV
|
arXiv:1709.09640v1 [math.AC] 27 Sep 2017
AN ALTERNATIVE APPROACH TO THE CONCEPT
OF SEPARABILITY IN GALOIS THEORY
M. G. MAHMOUDI
Abstract. The notion of a separable extension is an important
concept in Galois theory. Traditionally, this concept is introduced
using the minimal polynomial and the formal derivative. In this
work, we present an alternative approach to this classical concept.
Based on our approach, we will give new proofs of some basic
results about separable extensions (such as the existence of the
separable closure, Theorem of the primitive element and the transitivity of separability).
Key words: separable extension, Galois theory.
Mathematics Subject Classification: 12F10, 12F05, 12F15.
Introduction
In most elementary textbooks on Galois theory, separable extensions
are usually introduced via the minimal polynomial: an algebraic extension E/K is separable if the minimal polynomial f (X) ∈ K[X] of
every element α ∈ E has nonzero formal derivative, or equivalently
f (X) has distinct roots in its splitting field or any algebraically closed
field Ω containing E.
For normal extensions, there is a more functorial definition: a field
extension E/K is normal if there is a unique K-algebra embedding of
E into Ω (see [1, Ch. 6]). It would be interesting to know if there exists
an analogous definition for separable extensions. The aim of this work
is to address this question.
There exist already some methods to introduce separable extensions
using the embeddings of E/K into an algebraic closure Ω of E (note
that every algebraic closure of K is isomorphic to an algebraic closure of
E). Let HomK (E, Ω) denote the set of all K-algebra homomorphisms
from E into Ω. On can show that an extension E/K of finite degree
is separable if and only if |HomK (E, Ω)| = [E : K] (see [1, Ch. 7]).
There is another criterion (see [2, Ch. V., §15, n. 6]) as follows:
E/K is separable if and only if for every linearly independent elements
1
2
M. G. MAHMOUDI
a1 , · · · , an ∈ E over K, there exist K-automorphisms σ1 , · · · , σn of Ω
such that det(σi (aj )) 6= 0.
These properties are very useful characterization of separable extensions, but applying them are generally less easy. We suggest using the
following alternative definition of separability:
Definition. We say that an element α of algebraic field extension E/K
is separable if for every intermediate subfield L of E/K with α ∈ E\L,
there exist two L-algebra homomorphisms φ, ψ : E → Ω such that
φ(α) 6= ψ(α). An algebraic extension E/K is called separable if all its
elements are separable.
Other equivalent conditions will be given in Corollary 1.3 and Corollary 1.5. Roughly speaking, this definition says that L and α can be
separated by the homomorphisms from E into Ω. First, we show that
the above definition is equivalent to the usual concept of separability.
Then, we give new proofs of some basic results about separable extensions (such as the existence of the separable closure, Theorem of the
primitive element and the transitivity of separability) based on this
approach.
We hope that our approach provides a useful insight into the concept
of separability.
1. Alternative definition of separability
We recall the following standard facts about the embeddings of a
algebraic extension into an algebraically closed field (for the proof see
[1, p. 53 and p. 67]).
Theorem 1.1. (a) (Tower formula for the number of embeddings) Let
E/K be a field extension of finite degree with an intermediate subfield
L. Let Ω be an algebraic closure of E. Then
|HomK (E, Ω)| = |HomL (E, Ω)| × |HomK (L, Ω)|,
in particular |HomK (E, Ω)| 6 [E : K].
(b) (Extending field embeddings) Let L/F be an algebraic extension of
fields. Then every field embedding from F into an algebraically closed
field Ω can be extended to a one from L to Ω.
Proposition 1.2. Let E/K be an extension of fields and α ∈ E. Let Ω
be the algebraic closure of E. The following statements are equivalent:
(i) The minimal polynomial of α over K has distinct roots in Ω.
(ii) For every intermediate subfield L of E/K with α ∈ E\L, there
exist L-algebra homomorphisms φ, ψ : E → Ω such that φ(α) 6= ψ(α).
(iii) For every intermediate subfield L of K(α)/K with α ∈ K(α)\L,
AN ALTERNATIVE APPROACH TO THE CONCEPT OF SEPARABILITY
3
there exist two L-algebra homomorphisms φ, ψ : E → Ω such that
φ(α) 6= ψ(α).
Proof. (i) ⇒ (ii) The minimal polynomial of α over L has distinct
roots in Ω, so L(α) has distinct L-embeddings into Ω. (ii) ⇒ (iii) is
immediate. (iii) ⇒ (i) Let f (X) ∈ K[X] be the minimal polynomial
of α. If the minimal polynomial of α over K does not have distinct
roots in Ω, then K is of characteristic p > 0 and there exists n > 1
n
such that f (X) = g(X p ) where g(X) ∈ K[X] has nonzero formal
n
derivative. Take L = K(αp ). Note that α ∈ K(α)\L, otherwise
α would satisfy in a polynomial of degree less than deg f (x). Since
n
αp ∈ L, we have φ(α) = ψ(α) for every L-algebra homomorphism
φ, ψ : E → Ω, contradiction.
The above equivalence, in particular shows that the definition of separability given in (ii), which at first seems relative, does not depend on
E. Thus, if E and E ′ are two algebraic field extension of K containing
α, then α is separable as an element of E then it is separable as an
element of E ′ . Also, if K ⊂ L ⊂ E and α ∈ E is separable over K,
then it is separable over L.
Corollary 1.3. Let E/K be a separable extension with intermediate
subfields L1 and L2 . The following conditions are equivalents:
(i) L1 ⊂ L2 .
(ii) For every φ, ψ ∈ HomK (E, Ω), φ|L2 = ψ|L2 implies φ|L1 = ψ|L1 .
Conversely if (i) and (ii) are equivalents for every L1 and L2 then E/K
is separable.
Proof. (ii) ⇒ (i): If L1 6⊂ L2 , consider an element α ∈ L1 \L2 , then
there exists φ, ψ ∈ HomL2 (E, Ω) with φ(α) 6= ψ(α). This contradicts
(ii). The implication (i) ⇒ (ii) is evident.
Conversely suppose that L is an intermediate subfield and α ∈ E\L.
If there exist no φ, ψ ∈ HomL (E, Ω) with φ(α) 6= ψ(α), then by taking
L1 = K(α) and L2 = L, the equivalence of (i) and (ii) implies that
K(α) ⊂ L, contradiction.
Corollary 1.4. Let E/K be a separable extension and let α, β ∈ E.
The following conditions are equivalent:
(i) α ∈ K(β).
(ii) For every φ, ψ ∈ HomK (E, Ω), φ(β) = ψ(β) implies φ(α) = ψ(α).
As a consequence we obtain
Corollary 1.5. An algebraic field extension E/K is separable if and
only if for every proper intermediate subfield L of E/K, |HomL (E, Ω)| >
1.
4
M. G. MAHMOUDI
Proof. Consider an element α ∈ E\L. As E/K is separable, there exists φ, ψ ∈ HomL (E, Ω) such that φ(α) 6= ψ(α), hence |HomL (E, Ω)| >
1. Conversely, suppose that |HomL (E, Ω)| > 1 for every intermediate
subfield L. Consider an element α ∈ E\L. We show that there exist
φ, ψ ∈ HomL (E, Ω) such that φ(α) 6= ψ(α). If it is not the case, by
taking L2 = L and L1 = K(α), Corollary 1.3 implies that K(α) ⊂ L,
contradiction.
2. Some applications
Theorem 2.1 (Primitive Element Theorem). Let E/K be a separable
field extension of finite degree. Then there exists an element α ∈ E
such that E = K(α).
Proof. For finite fields, we follow the standard argument. If K is a field
with q elements and |E| = q n , then we have |L| 6 q n−1 for every proper
intermediate subfield L of E/K, hence every element x ∈ E satisfies
n−1
xq
− x = 0. Since |E| > q n−1 , there exists certainly an element
n−1
α ∈ E with αq
− α 6= 0, this element is not included in any proper
subfield, hence K(α) = E.
Now consider the case where K is infinite. Let α, β ∈ E. It suffices
to prove that there exists γ ∈ E such that α, β ∈ K(γ). We may
assume that α and β are linearly independent over K. Let P be the
plane spanned by α and β. If our claim is not true then for every
nonzero γ ∈ P , K(γ) ∩ P is one dimensional over K. Since K is
infinite, there exists an infinite number of pairwise noncolinear elements
γ1 , γ2, · · · ∈ P . It follows that γi 6∈ K(γj ) if i 6= j. By Corollary 1.3,
there exist φij , ψij ∈ HomK (E, Ω) such that φij |K(γi ) = ψij |K(γi ) but
φij |K(γj ) 6= ψij |K(γj ) . As HomK (E, Ω) is finite, there exist distinct i, j, k
such that φji = φki and ψji = ψki . We obtain
φki (γj ) = φji (γj ) = ψji (γj ) = ψki (γj ),
on the other hand φki (γk ) = ψki (γk ), so φki |P = ψki |P , contradiction.
Corollary 2.2. If E/K is a separable extension of finite degree then
|HomK (E, Ω)| = [E : K]. Conversely if |HomK (E, Ω)| = [E : K] then
E/K is separable. In particular if α ∈ E is separable over K then the
extension K(α)/K is separable.
Proof. Suppose that [E : K] = n. By Theorem 2.1, there exists α ∈ E
with E = K(α). As α is separable over F , the minimal polynomial
f (x) ∈ K[x] of α has n distinct roots α1 , · · · , αn ∈ Ω. For every
i = 1, · · · , n we can define a map φi ∈ Hom(E, Ω) by φi (α) = αi .
AN ALTERNATIVE APPROACH TO THE CONCEPT OF SEPARABILITY
5
Conversely for every φ ∈ HomK (E, Ω), φ(α) is necessarily equal to one
of the elements α1 , · · · , αn .
Conversely, suppose that |HomK (E, Ω)| = [E : K]. We have to
show that E/K is separable. Otherwise, by Corollary 1.5, there exists
a proper intermediate subfield L such that |HomL (E, Ω)| = 1. Thus we
have |HomK (E, Ω)| = |HomK (L, Ω)|×|HomL (E, Ω)| = |HomK (L, Ω)| 6
[L : K] < n, contradiction.
Proposition 2.3. Let E/K be an algebraic extension. Then the set of
separable elements of E/K form an intermediate subfield.
Proof. It suffices to prove that for every separable elements α, β ∈ E,
α + β and αβ is separable. We may assume that αβ 6= 0. For an
intermediate subfield L with α + β 6∈ L, we should prove that there
exists two maps in HomL (E, Ω) whose values on α + β are different.
If α ∈ L, then by separability of β there exists φ, ψ ∈ HomL (E, Ω)
with φ(β) 6= ψ(β), hence φ(α + β) 6= ψ(α + β). We may then assume
that α, β 6∈ L. If α 6∈ L(β), then by separability of α, there exist
φ, ψ ∈ HomL(β) (E, Ω) ⊆ HomL (E, Ω) with φ(α) 6= ψ(α) and we obtain
φ(α + β) 6= ψ(α + β). So assume that α ∈ L(β). As β is separable over
K, it is also separable over L, hence by Corollary 2.2 every element of
L(β) is separable over L, in particular α + β is separable over L. Hence
|HomL (L(α+β), Ω)| > 1 and there exist φ, ψ ∈ HomL (L(α+β), Ω) with
φ(α + β) 6= ψ(α + β), φ and ψ can be extended to φ̄, ψ̄ ∈ HomL (E, Ω)
and we are done. The proof for αβ is similar.
Corollary 2.4. Let E/L and L/K be algebraic separable extensions.
Then E/K is separable as well.
Proof. Let M be an intermediate subfield of E/K and α ∈ E\M. We
should prove the existence of two homomorphisms φ, ψ ∈ HomM (E, Ω)
with φ(α) 6= ψ(α). If α 6∈ LM, then by the separability of E/L, there
exist φ, ψ ∈ HomLM (E, Ω) ⊂ HomM (E, Ω) with φ(α) 6= ψ(α) and we
are done. Consider the case where, α ∈ LM. As all elements of L
are separable over K, then by Proposition 2.3, LM/M is separable.
There exist so two homomorphisms φ, ψ ∈ HomM (LM, Ω) with φ(α) 6=
ψ(α). These homomorphisms can be extended to homomorphisms in
HomM (E, Ω).
References
[1] Lorenz, F., “Algebra. Vol. 1. Fields and Galois theory”, Translated from the
1987 German edition by Silvio Levy. With the collaboration of Levy. Universitext. Springer, New York, 2006.
6
M. G. MAHMOUDI
[2] Bourbaki, N., “Éléments de mathématique, Algèbre, Chapitres 4-7”, Springer,
Berlin-Heidelberg-New York, Reprint of the 1981 original, 2007.
M. G. Mahmoudi, [email protected],
Department of Mathematical Sciences, Sharif University of Technology, P. O. Box 11155-9415,
Tehran, Iran.
| 0math.AC
|
MATRIX TILE ANALYSIS
Inmar Givoni, Vincent Cheung, Brendan J. Frey
Probabilistic and Statistical Inference Group
University of Toronto
10 King’s College Road, Toronto, Ontario, Canada, M5S 3G4
Abstract
Many tasks require finding groups of elements in a matrix of numbers, symbols or
class likelihoods. One approach is to use efficient bi- or tri-linear factorization techniques
including PCA, ICA, sparse matrix factorization and plaid analysis. These techniques
are not appropriate when addition and multiplication of matrix elements are not sensibly defined. More directly, methods like biclustering can be used to classify matrix elements, but these methods make the overlyrestrictive assumption that the class of each
element is a function of a row class and a column class. We introduce a general computational problem, ‘matrix tile analysis’ (MTA),
which consists of decomposing a matrix into
a set of non-overlapping tiles, each of which
is defined by a subset of usually nonadjacent
rows and columns. MTA does not require an
algebra for combining tiles, but must search
over an exponential number of discrete combinations of tile assignments. We describe
a loopy BP (sum-product) algorithm and
an ICM algorithm for performing MTA. We
compare the effectiveness of these methods
to PCA and the plaid method on hundreds
of randomly generated tasks. Using doublegene-knockout data, we show that MTA finds
groups of interacting yeast genes that have
biologically-related functions.
1
INTRODUCTION
A variety of data types are most naturally represented as matrices of numbers, symbols, or, after preprocessing, class likelihoods. For example, the viability of a yeast strain obtained by knocking out two
genes can be compared to normal viability to obtain
MTA
Figure 1: To illustrate matrix tile analysis, we show
a binary data matrix (left), and after applying MTA,
the same matrix as a collection of tiles (right), having
reordered rows and columns to better reveal the tiles.
Each tile is described by a subset of rows and columns.
Different tiles may not overlap. Each element is colored according to its tile index.
an ‘abnormal growth likelihood’ associated with the
pair of genes. Many such double-gene-knockout experiments can be used to construct a matrix of likelihoods
(Tong et al. 2004). In collaborative filtering (Resnick
et al. 1994) the matrix rows correspond to clients and
the columns correspond to, say, movies that clients
are potentially interested in watching. An observed
element of the data matrix contains information regarding a preference indicated by the corresponding
client for the corresponding movie.
Two approaches to analyzing an N × M data matrix
X include matrix factorization methods and combined
row-column clustering techniques. Matrix factorization methods are only appropriate when a sensible algebra (addition and multiplication) can be defined for
the data elements. If X contains real numbers, regular
addition and multiplication are assumed to be sensible,
often without convincing justification. If X contains
non-ordinal discrete variables or likelihood functions,
there often does not exist a sensible algebra for the
elements. Assuming an algebra can be defined, matrix
factorization methods find factorizations of the form
X ≈ U SV , where S is a diagonal matrix, U is an N ×J
matrix and V is a J × M matrix. Different factorization methods enforce different types of constraints on
the matrices such as orthonormality of U ’s columns
and V ’s rows (PCA, Jolliffe 1986), non-negativity of
U, S, V (non-negative matrix factorization, Lee & Seung 1999), and allowing J, the size of the subspace
the data is projected onto, to be large while requiring
sparse solutions (ICA, Bell & Sejnowski 1995, sparse
matrix factorization, Dueck et al. 2005) or low-norm
solutions (maximum-margin matrix factorization, Srebro et al. 2005).
data as a collection of non-overlapping tiles. We formulate MTA as a probabilistic model and present two
different algorithms that approximately maximize the
joint probability as well as two standard algorithms
(the plaid method and PCA) which can be modified
to find tiles, and compare the performance of all methods on synthetic data. We use MTA and hierarchical
agglomerative clustering to find biologically relevant
groupings in the yeast genetic interaction data, and
show that MTA finds more biologically informative
groups, thus demonstrating the applicability and usefulness of this method.
Combined row-column clustering techniques classify
data elements, instead of finding algebraic decompositions. One approach is to independently cluster the
rows and columns (Eisen et al. 1999), but this ignores dependencies between the two clustering problems. In bi-clustering (Cheng & Church 2000), each
column has a class label and each row has a class label so that the class of each data element is given by
joint row-column class labels. So, if there are nc possible classes for the columns and nr possible classes for
the rows, each element can be in one of nc nr classes.
While this approach can account for dependencies in
the row clustering and column clustering problems, it
makes the overly-restrictive assumption that the class
of each element is a function of a row class and a column class, resulting in clusters defined by common row
and column boundaries. Row classes whose associated
elements show similar interaction patterns to a subset
of column classes are not grouped, though they correspond to a block of data composed of several identified
smaller blocks. This restriction also applies to stochastic block models and their extensions, used mainly for
the analysis of relational data (Airoldi et al. 2005).
In MTA, an N × M data matrix X is modelled as
a set of non-overlapping regions, or tiles. The order
of the rows and columns that is most appropriate for
revealing a specific tile by grouping its elements into
a contiguous rectangular block may be inappropriate
for another tile. Thus, visually each tile is associated
with a different permutation of the rows and columns
that groups its elements. For example, the binary data
in Fig. 1 is rearranged and colored to expose the tile
structure of the data. Only one tile is fully contiguous
for this particular permutation. The data elements in
each tile are accounted for by class-conditional probabilities and data that does not appear in any tile
is accounted for by a background model. Unlike biclustering, MTA does not require that the tiles are
defined by a common set of row and column boundaries. In fact, MTA allows an arbitrary subset of rows
and columns to be assigned to each tile, subject to
the constraint that two tiles cannot own the same elements. Like matrix factorization techniques, MTA
can be thought of as factorizing the input matrix into
multiple components or tiles. However, because every
element belongs to only one tile, it is not required that
a sensible algebra be defined to combine elements from
different tiles. In this way, MTA can be applied when
sensible notions of addition and multiplication are not
available. Finally, MTA can be formulated as a problem of finding tiles, given tile-conditional likelihoods
as input, i.e., the analysis method does not need the
original data as input.
Another approach to combined row-column clustering
is to apply a matrix factorization technique, and then
binarize the output. After applying PCA, the elements
of U and V can be binarized by applying a threshold
and the binary patterns in U and V can be used to define classes. The plaid model (Lazzeroni & Owen 2002)
finds a decomposition of the same form X ≈ U SV described above, but the elements of U and V are regularized toward the values 0 or 1 during optimization.
While the method of quantizing the output of a matrix
factorization technique provides an often computationally efficient solution to joint row-column clustering,
selection of thresholds is not straightforward and the
technique does not directly optimize a cluster model.
We introduce a new computational approach that we
call ‘matrix tile analysis’ (MTA), which explains the
2
MATRIX TILE ANALYSIS
In a model with T tiles, we denote the index of a particular tile by t. We use binary random variables rit to
indicate the rows of matrix elements belonging to the
tile, i.e., rit = 1 indicates that the ith row of X contains elements belonging to tile t, and rit = 0 indicates
that none of the elements in the ith row of X belong
to tile t, where i = 1 . . . N . Similarly, binary random variables ctj indicate those columns containing elements belonging to tile t, where j = 1 . . . M . The outer
t > t t
product (r1t , r2t , . . . , rN
) (c1 , c2 , . . . , ctM ) is an N × M
Tile 1
binary matrix indicating those elements belonging to
tile t. In total, the hidden variables describing the tile
analysis include T , R = {rit | i = 1 . . . N, t = 1 . . . T }
and C = {ctj | j = 1 . . . M, t = 1 . . . T }.
Each tile has a user-defined likelihood specifying the
probability of membership of elements in X to that
tile. We denote the N × M matrix of likelihoods corresponding to tile t by Lt . That is, `tij = P (xij |tile t).
To account for elements not placed in tiles, we use
L0 to denote the likelihood matrix corresponding to
the background model. These likelihood matrices are
the input to MTA and are pre-computed depending
on the task at hand. For example, if the data matrix is binary-valued and has symmetric noise with
probability ², then `tij = (1 − ²)xij ²1−xij . In this paper, we address the computational task of identifying
tiles, assuming the input likelihood matrices can be
pre-computed. In the rest of the paper we address the
case of a single likelihood model for all tiles, though
our model easily extends to the general case of t likelihood matrices.
The overall data likelihood is
P (X|R, C, T ) =
N Y
M
Y
PT
(`ij )
t t
t=1 ri cj
PT
(`0ij )(1−
t t
t=1 ri cj )
.
(1)
i=1 j=1
We assume that given T , R and C are uniformly distributed over all non-overlapping tile configurations1 :
P (R, C|T ) =
N M
T
i
X
1 Y Yh
0≤
rit ctj ≤ 1 .
h(T ) i=1 j=1
t=1
(2)
where h(T ) is used to normalize the distribution. The
minimum number of tiles is T = 0 and the maximum
number of tiles is T = N M , in which case there is
one tile for every data element. The distribution of T
can be specified depending on the application, but in
this paper, we assume a uniform distribution: P (T ) =
1/(N M + 1), T ∈ {0, . . . , N M }.
The probability of the number of tiles, the rows and
columns in every tile, and the data is
P (X, R, C, T ) = P (T )P (R, C|T )P (X|R, C, T ). (3)
The goal of MTA is to infer the values of T , R and C,
i.e., compute argmaxT,R,C P (X, R, C, T ). We describe
four different approaches for inferring R and C, where
T is automatically selected using an MDL framework.
1
Square brackets are used in the fashion of Iverson’s
notation, where [true] = 1 and [f alse] = 0
c11
1
1
r
1
2
f111
Tile T
c12
c1M
c1T
T
1
cMT
r
1
s11
f121
1
s12
1
f1M
s11M
T
2
r
rN1
c2T
f11T
T
s11
f12T
T
s12
f1TM
s1TM
r
f211
fN11
s121
s1N 1
f111
rNT
s 1NM
g N1
f21T
fNT1
T
s 21
s TN 1
T
fNM
s TNM
g NM
Figure 2: Factor graph for MTA
3
3.1
MTA INFERENCE
SUM-PRODUCT ALGORITHM
Based on the data model detailed in equations (1-3),
we describe a probabilistic graphical model for the
matrix tile analysis problem. The model is shown in
Fig. 2, using the standard notation for factor graphs
(Kschischang et al. 2001).
In order for the model to fully describe the constraints
requiring tile elements to be non-overlapping, we introduce an additional set of binary random variables,
stij whose purpose is to restrict each matrix element to
belong to one tile at the most.
The stij nodes indicate whether matrix element xij is
accounted for by tile t. The rit nodes are binary random variables that indicate whether row i of the data
matrix contains elements which belong to tile t. Likewise, node ctj is an indicator which is active when column j of the data matrix contains elements belonging
to tile t.
Each triplet of nodes (stij ,rit ,ctj ) is connected to a funct
tion node denoted by fij
, and the function calculated
at the node is defined as
f tij (stij , rit , ctj ) =
1 if (stij = rit = ctj = 1)
or (stij = 0 and (rit = 0 or ctj = 0));
0 otherwise.
(4)
This function node connects elements within the same
tile, and enforces the constraint that if a matrix element xij is accounted for by tile t (as indicated by
variable stij having a value of 1), then the corresponding row and column indicators must be active. If an
element xij is not accounted for by tile t, then either
the row indicator, or the column indicator, or both
must be inactive.
For a specific matrix element xij , the corresponding s-
variables s1ij , s2ij , . . . , sTij are connected to the function
node gij (shown in Fig. 2 only for matrix element xN 1
and xN M for clarity). This function bridges together
s-variable nodes from different tiles to enforce the constraint that each matrix element xij is accounted for
by a single tile at most, as well as to account for the
likelihood of xij .
following updates for propagation:
µgij →stij
µfijt →rit
µrit →fijt
g ij (s1ij , s2ij , . . . , sTij ) =
`ij if stij = 1, skij = 0 ∀k 6= t;
`0 if stij = 0 ∀t;
ij
0
otherwise,
=
=
=
− log
log
³ `0
ij
`ij
+
X
´
exp{µfijk →skij }
(7)
k6=t
³ exp{µct →f t + µgij →st } + 1 ´
j
ij
ij
X
exp{µctj →fijt } + 1
t →r t
µfik
i
k6=j
(5)
µfijt →ctj
µctj →fijt
=
=
log
³ exp{µrt →f t + µgij →st } + 1 ´
ij
ij
i
X
exp{µrit →fijt } + 1
t →ct
µfkj
j
k6=i
µfijt →stij
where t = 1 . . . T, k = 1 . . . T . When the constraint
is satisfied, so that only one tile accounts for the element, the function evaluates to the likelihood of that
element under the model. If no tile accounts for the
element, the function evaluates to the likelihood of the
background model.
The overall data likelihood as represented in the factor
graph is
P (X|T, R, C, S) =
T
N Y
M ³
´
Y
Y
t
(stij , rit , ctj ) ,
fij
gij (s1ij , .., sTij )
i=1 j=1
(6)
t=1
where S = {stij | i = 1 . . . N, j = 1 . . . M, t = 1 . . . T }.
In order to solve the MTA problem we infer (R, C, S)
for a set of T tiles, by applying the sum-product algorithm (SP-MTA) on the factor graph shown in Fig. 2.
Messages between nodes are sent as log probability
ratios, as they contain all information needed to reconstruct the probability of the binary variable nodes,
and they present an efficient implementation which is
also robust to extreme probability swings2 .
Denoting by µf →x a message sent from a function node
f to a variable node x and µx→f as the message sent
from a variable node to a function node, we obtain the
p(x=1)
Letting the message π be equal to log
for some
log p(x=0)
random variable x, we reconstruct the two probabilities
exp π
from the message as p(x = 1) = 1+exp
and p(x = 0) =
π
1
)
1+exp π
2
=
µctj →fijt + µrit →fijt −
¡
¢
log exp{µctj →fijt } + exp{µrit →fijt } + 1 ,
and we use the fact that µstij →fijt = µgij →stij and
µstij →gij = µfijt →stij .
Since the factor graph described in Fig. 2 contains
many cycles, the algorithm is used within the loopy belief propagation framework (LBP). There is no known
criteria to distinguish among instances of LBP which
converge and those who do not. However, there are
many examples of well known problems for which LBP
converges, and empirically we observe it is also the case
for MTA.
Upon termination, we infer R and C by fusing all the
incoming messages µfitt →rit and µftjt →ctj at each rit and
ctj nodes respectively, and applying a threshold to obtain the desired binary values. Convergence was determined by evaluating the relative change in the values
of the messages arriving from the tile element nodes
stij to function nodes gij after each of the possible messages has been calculated and sent once.3
Message scheduling and initialization was found to be
important for the algorithm to find good solutions.
The messages are processed for each tile consecutively.
Within each tile, messages are sent according to the
order they are presented in (7). All messages were
initialized to be zero in the log domain, except for the
messages from matrix element nodes stij to the external
function nodes gij , which were the first messages to be
propagated, and were set to −∞, indicating that initially, matrix elements are not claimed by any tile. Incremental thresholding of the messages was also found
to aid in convergence to stable solutions.
3
Convergence
µk+1
t →st |
fij
ij
< 1 × 10
the iteration.
was
−3
·
declared
P
if
k
t →st |,
i,j,t |µfij
ij
P
i,j,t
|µkf t
t
ij →sij
−
where k indicates
3.2
ICM
An approximate algorithm for maximizing the joint
probability in (3) is Iterative Conditional Modes
(MTA-ICM). To maximize P (X, R, C, T ), or equivalently, log P (X, R, C, T ), we alternate between updating the rows and columns of the tiles, leading to an
iterative algorithm with the following updates:
{ri1 , . . . , riT }
³X
t,j
← argmax{ri1 ,...,riT }
i´
X h
X
`ij
rit ctj = 1
(8)
log 0 ≤
rit ctj (log 0 ) +
`ij
t
j
{c1i , . . . , cTi }
³X
t,i
← argmax{c1i ,...,cTi }
i´
X h
X
`ij
ctj rit (log 0 ) +
rit ctj = 1 .
log 0 ≤
`ij
t
i
(9)
The constraints are enforced via the last sums in (8)
and (9). The “gain” over the base model
P by¡ including
row i in tile t can be defined as git = j ctj log(`ij ) −
¢
log(`0ij ) . For row i, the gain can be computed for each
tile independently and the joint probability is optimized for choosing binary values for {ri1 , . . . , riT } that
both maximize the sum of the gains and satisfy the
constraints. While only tiles with positive gains need
to be considered, an exhaustive search still requires exponential time over the number of tiles with positive
gains. The search can be reduced by only searching
over configurations that satisfy the constraints.
Let D be a T ×T matrix, where Duv is 1 if row i can be
in both tiles u and v (riu = riv = 1) without violating
the constraints, i.e. tiles u and v do not contain the
same column (cuj cvj = 0 ∀ j), and 0 otherwise. D
£
¤
can be written as D = CC 0 = 0 where Iverson’s
notation is applied element-wise on the matrix. The
search space of tiles which can contain the same row
while satisfying the constraints, is then reduced to the
cliques of the graph with D as its adjacency matrix. A
generous upper bound on the number of cliques that
need to be evaluated is O(2T ). In practice, due to the
interplay between rows and columns, the search space
is significantly restricted. The same procedure can be
performed to optimize the columns.
4
EXPERIMENTAL EVALUATION
We compared MTA-SP and MTA-ICM to two additional methods that can be used to discover tiles. The
plaid model (Lazzeroni & Owen, 2002) finds overlapping tiles, or ‘layers’ in the data one at a time, in
a greedy fashion so that each consecutive tile t is
searched for in the residual left over from the previous
t − 1 tiles. For each tile, we determined whether it accounted for the background model or for the data and
evaluated it accordingly. Another approach to MTA is
to quantize the output of a matrix factorization technique. We experimented with a technique based on
PCA (Jolliffe 1986).4
4.1
SYNTHETIC DATA
For the purpose of evaluating the different techniques
we generated ∼2000 synthetic data matrices of various
sizes, each containing several tiles with varying tile dimensions. We analyzed the matrices using MTA-SP,
MTA-ICM, the plaid method, and PCA, and evaluated their ability to correctly identify tiles. Each
matrix in the data set is an N × N matrix, with
N = 40, 70, 100, 150, 200. The number of tiles in each
matrix ranges from 1 to 10. The average area of the
tile is held constant across all the matrices and each
tile covers approximately 4% of the data matrix. The
tiles were randomly generated, subject to the constraint that two tiles cannot contain the same matrix
element. The data was corrupted with additive Gaussian noise with log10 (σ 2 ) =-1.5, -1.0,-0.8,-0.675,-0.55,0.425, and -0.3. Fig. 1 shows an example of a data
matrix before adding noise, and shows a possible solution for the data after applying MTA to the corrupted
data (not shown), where rows and columns have been
reordered based on the analysis output to reveal the
tiles. For each setting of matrix size, number of tiles,
and noise level, we generated 20 different matrices, on
which we tested the performance of all algorithms, except the plaid method which was only tested between
1-10 times for each setting due to the user intensive
nature of the plaid software5 and its inability to run
in batch mode.
The input to the algorithms was set to be the log likelihood ratio of each element being in a tile versus being part of the background, calculated by assuming a
4
To find T tiles, the N × N covariance matrix of the
columns of X is computed and its T principal components
are extracted (this procedure can arbitrarily be applied
to the rows instead of the columns). Since each principal
component corresponds to a major direction of variation
across the columns, the values in each component provide
evidence for the rows belonging to a corresponding tile.
For each component, the rows belonging to the tile are
identified by comparing the elements in the component to
a threshold. To obtain maximum separation of values in
the two groups, the threshold is selected so as to minimize the derivative of the cumulative distribution of the
elements. Each component is thresholded separately to
produce a set of T binary vectors, each corresponding to
a tile. Next, each column of the data matrix is robustly
projected onto these vectors using a pseudo-inverse to account for non-orthogonalities introduced by thresholding.
This yields a set of T real numbers for each row of data,
which are thresholded to determine the set of tiles that the
row belongs to.
5
www-stat.stanford.edu/∼owen/clickwrap/plaid.html
normal distribution and fixing the standard deviation
to 0.5 regardless of the true noise level. Additionally
PCA and the plaid method take as input the noisy
data matrix. The output of each algorithm is an integer value matrix where 0 indicates an element is part
of the background model, and non-zero values indicate
the tile index. Model selection (i.e. determining the
correct number of tiles) was performed by evaluating
the following cost function
t,i,j
rit ctj
`0ij X
log
−
log `0ij + T (N + M ) log 2. (10)
`ij
i,j
The cost function is the minimum description length
(MDL) for coding the model, assuming the distribution over the row and column indicator variables is
uniform (which is strictly not true because of the term
1
h(T ) in (2)). The number of tiles T the algorithm
was allowed to discover was incremented until it incurred an increase in the cost function. Since the plaid
method produces overlapping tiles, some of which may
model the background, overlapping elements of tile
identification was resolved under this cost function to
produce sensible results.
MTA−SP
MTA−ICM
5e−2
5e−3
PCA
5e−4
plaid
0.05
0.1
0.5
2 tiles,100x100 matrix
5e−1
5e−3
0.05
0.1
0.5
7 tiles,100x100 matrix
5e−1
5 tiles,40x40 matrix
5e−1
5e−2
5e−2
5e−3
5e−2
5e−3
5e−4
0.05
0.1
0.5
3 tiles,100x100 matrix
5e−1
0.05
0.1
0.5
8 tiles,100x100 matrix
5e−1
0.05
0.1
0.5
5 tiles,70x70 matrix
5e−1
5e−2
5e−2
5e−3
5e−4
5e−2
0.05
0.1
0.5
4 tiles,100x100 matrix
5e−1
0.05
0.1
0.5
9 tiles,100x100 matrix
5e−1
5e−3
0.05
0.1
0.5
5 tiles,150x150 matrix
5e−1
5e−2
5e−2
5e−3
5e−2
5e−4
0.05
0.1
0.5
0.05
0.1
0.5
0.05
0.1
0.5
5 tiles,100x100 matrix 10 tiles,100x100 matrix 5 tiles,200x200 matrix
5e−1
5e−1
5e−1
5e−2
5e−2
5e−2
5e−3
Performance was evaluated under several different criteria. First, we computed the Hamming distance between the ‘ground truth’ and the matrix output by
each algorithm (after setting any non-zero values to 1
in the output matrix and comparing it to the clean
input matrix) and normalizing by the size of the matrix. The Hamming distance gives a general measure
of how well the algorithm was able to identify which
elements in the input matrix belong to tiles and which
belong to the background model. Fig. 3 shows the
results for a set of experiments evaluated at different
dimensions of an operating point chosen to be a matrix size of 100 × 100 and 5 tiles. We show results
where the matrix size is fixed and the number of tiles
is varied, and vice versa. All noise levels are shown
for each setting of matrix size and number of tiles.
The plots are shown with a log scale on both axes,
and each point is an average of 20 experiments. Note
that the scales in the y-axis differ across plots. The
Hamming distance shows that MTA-SP performs better than the other algorithms, for most experimental
settings. MTA-SP and MTA-ICM perform very similarly for large number of tiles. PCA and the plaid
method both perform poorly, for the most part due
to their inability to identify non-background elements.
The increase in the number of tiles results in a decrease in performance while increasing the matrix size
does not effect the performance significantly. Changes
in the noise level have a noticeable effect only for a
small number of tiles, and as would be expected, all
algorithms succeed at finding a single tile.
6 tiles,100x100 matrix
5e−1
5e−2
Error Rate
X
1 tile,100x100 matrix
5e−1
0.05
0.1
0.5
5e−3
0.05
0.1
0.5
0.05
0.1
0.5
2
Noise Level (σ )
Figure 3: Hamming distance for different settings of
matrix size and number of tiles.
Hamming distance does not provide information about
how well the algorithm was able to identify the actual tiles that make up the non-background elements
of the matrix. We address this question by evaluating the classification error made by the different algorithms. This evaluation requires that the tiles discovered by the algorithms be matched to the ‘ground
truth’, which while used to generate the data, is not
necessarily the best tile configuration. Tiles in the
‘ground truth’ were matched to the experimentally determined tiles to which they most overlapped. The
matching was performed in a greedy fashion such that
at most, a one-to-one correspondence was established
between tiles. The tiles in the matrix output were then
re-numbered based on this matching. The classification error metric is simply the number of mismatches
between the matrices normalized by the size of the matrix. The classification errors are shown in Fig. 4 for
the same set of experiments as previously discussed.
MTA-SP outperforms the other methods in all but 4 of
the experiments. In the cases where MTA-ICM does
better in terms of the Hamming distance, we see it suffers from more classification errors. This indicates that
for these cases, MTA-ICM correctly identified more elements to be in tiles and in the background, but it did
1 tile,100x100 matrix
5e−1
6 tiles,100x100 matrix
0
MTA−SP
5e−2
−0.2
2000
PCA
5e−2
5e−4
0.05
0.1
0.5
2 tiles,100x100 matrix
5e−1
−0.4
plaid
0.05
0.1
0.5
7 tiles,100x100 matrix
5e−1
5e−4
0.5
3 tiles,100x100 matrix
5e−1
0.05
0.1
0.5
8 tiles,100x100 matrix
5e−1
0.5
0.05
0.1
0.5
7 tiles,100x100 matrix
0.05
0.1
600
6000
600
400
4000
400
200
2000
0
0.5
5 tiles,70x70 matrix
5e−1
200
0
0.05
0.1
0.5
0.05
3 tiles,100x100 matrix
0.1
0.5
0.05
8 tiles,100x100 matrix
Cost
600
5e−3
5e−2
0.05
0.1
0.5
4 tiles,100x100 matrix
5e−1
0.05
0.1
0.5
9 tiles,100x100 matrix
5e−1
0.05
0.1
0.5
5 tiles,150x150 matrix
5e−1
0
5e−2
0.05
0.1
0.5
500
0.05
0.1
0.5
0.05
0.1
0.5
0.05
0.1
0.5
5 tiles,100x100 matrix 10 tiles,100x100 matrix 5 tiles,200x200 matrix
5e−1
5e−1
5e−1
200
0.05
4 tiles,100x100 matrix
1000
5e−3
400
2000
1500
5e−2
600
4000
200
0.1
0.5
0.1
4000
6000
3000
4000
2000
0.1
0.5
0.5
0.05
0.1
0.5
10000
1500
0.05
0.1
0.1
0.5
12000
10000
8000
6000
4000
2000
15000
5000
500
0.1
0.05
5 tiles,100x100 matrix 10 tiles,100x100 matrix 5 tiles,200x200 matrix
1000
0.05
0.5
1000
0.05
2000
5e−2
0.1
5 tiles,150x150 matrix
8000
0.5
2500
5e−2
0.05
9 tiles,100x100 matrix
2000
0.05
0.5
800
6000
400
0.1
5 tiles,70x70 matrix
1000
8000
5e−2
5e−4
5 tiles,40x40 matrix
800
5e−2
0.1
0.1
2 tiles,100x100 matrix
5e−2
0.05
MTA−ICM
1000
0.05
5 tiles,40x40 matrix
5e−1
5e−3
MTA−SP
3000
MTA−ICM
5e−3
Error Rate
6 tiles,100x100 matrix
1 tile,100x100 matrix
5e−1
0.5
0.05
2
Noise Level (σ )
0.1
0.5
0.05
0.1
0.5
0.05
0.1
0.5
2
Noise Level (σ )
Figure 4: Classification error for different settings of
matrix size (right column) and number of tiles.
Figure 5: Cost function relative to ground truth for
different matrix sizes and numbers of tiles.
so with less accurate tiling with regard to the actual
data. Overall, both MTA-SP and MTA-ICM perform
considerably better than the plaid method and PCA
which struggle to identify tile elements.
originally generated, despite the amount of corruption
by the noise. In general, the algorithms do not reach
the global minimum and have a higher cost than that
of the ground truth. We see the same trends as before, with MTA-SP doing better on average, and having only a small difference when it fails to do as well
as MTA-ICM.
Fig. 5 compares the performance of MTA-SP and
MTA-ICM in terms of the cost function in (10) they
both minimize. The other algorithms are not shown as
they perform several orders of magnitude worse, a consequence of the fact they do not directly minimize the
cost function. The values are shown after subtraction
of the cost of the ‘ground truth’. For the cases where
the algorithms performed perfectly, the incurred cost
was zero. As the complexity of the problem increases,
so does the cost incurred. This is not directly due to
the added component of T (N + M ) which appears in
the ‘ground truth’ model as well, though it may contribute towards the higher costs if the algorithms find
more tiles then necessary. There are cases where the
algorithm performs better than the true model (e.g.,
5 tiles, 40 × 40). This phenomena occurs because the
noise applied is such that some of the rows and/or
columns of a tile are better modelled in the base model
than as part of a tile, or vice versa. However, under
the ‘ground truth’, the tile is still evaluated as it was
4.2
YEAST GENETIC INTERACTIONS
The ‘synthetic genetic array’ data set (SGA) (Tong
et al., 2004) describes events of lethality for double
knock-out genes in yeast, where the single knock-out
strains are viable. The data set is a binary matrix
of 135 × 1023 (Fig. 6), where a value of 1 at position
xij denotes a synthetic genetic interaction, i.e. the
double mutant whose genes i and j have been knocked
out is not viable and thus both genes are required for
survival, although either one by itself is dispensable.
We compared the performance of the different methods by obtaining clusters of genes from each method
and analyzing their biological significance in terms
of higher than expected rates of similarly annotated
Table 1: Number of times groups of yeast genes are
significantly enriched for Gene Ontology categories
Biological Aspect
MTA
HAC
Plaid
Biological Process
Cellular Component
Molecular Function
Total
7
8
2
17
4
4
2
10
4
1
0
5
genes, using Gene Ontology annotations6 . For each
cluster, we evaluated the p-value of the Bonferronicorrected hyper-geometric distribution for observing
the number of genes with a specific annotation. In order to compare MTA to different types of analysis, we
obtained clusters from hierarchical agglomerative clustering (HAC), the original plaid model, and MTA. We
allowed up to 15 clusters, where for HAC the dendrogram was used to obtain row/column clusters, and for
MTA, we considered the genes corresponding to row elements of a tile as elements of the row cluster and genes
corresponding to column elements of a tile as column
clusters. MTA identified more statistically significant
clusters (p < 0.01) for different GO-category aspects
(cellular component, molecular function and biological process) than either HAC, or the unmodified plaid
model. Overall, MTA identified 17 significant groups,
HAC identified 10, and the plaid method identified 5
(see Table 1).
5
DISCUSSION
We introduced a new analysis framework, matrix tile
analysis, in which an input matrix is decomposed into
a set of non-overlapping components, or tiles. Unlike
many models for matrix decomposition, our method
does not require addition and multiplication to be defined over data vectors, and unlike approaches such
as bi-clustering, different components are allowed to
overlap in the rows and columns, though not in both
simultaneously. We described a probabilistic model
of the problem and its graphical representation using a factor-graph. We compared the performance of
four different approaches for solving MTA and were
able to demonstrate that the sum-product algorithm
(MTA-SP) performs better on a large set of synthetic
problems compared to a greedy algorithm, the plaid
method, and PCA. We were also able to extract biologically related groups of genes by applying MTA
to yeast genetic interaction data. Tiles identified by
MTA in two other gene interaction data sets are being
biologically verified.
Our model requires that the input be given in terms of
6
ftp://genome-ftp.stanford.edu/pub/yeast/
Figure 6: Yeast genetic interaction data. The rows
and columns are ordered using clustering.
data and background likelihoods. This was motivated
by data for which prior analysis or expert knowledge
resulted in likelihood measurements. However, for the
general case, in the absence of strong prior knowledge about the data distribution, it is always possible
to compute likelihood estimates from empirical measurements. Further, because of the formulation of the
model under a probabilistic framework, the model can
be extended to jointly infer the tiling and learn the
parameters in the likelihood functions, using the EM
algorithm.
References
EM Airoldi, et al. A Latent Mixed Memebership Model for
Relational Data. Proc ACM Link-KDD Workshop, 2005.
AJ Bell, TJ Sejnowski. An information maximization approach to blind separation and blind deconvolution. Neur
Comp 7, 1129–1159, 1995.
Y Cheng, GM Church. Biclustering of expression data.
Proc Int Conf Intell Syst Mol Biol 8, 93-103, 2000.
D Dueck, et al. Multi-way clustering of microarray data
using probabilistic sparse matrix factorization. Proc Int
Conf Intell Syst Mol Biol 13, 2005.
MV Eisen, et al. Cluster analysis and display of genomewide expression patterns. Proc of the National Academy of
Sciences 95, 14863–14868, 1999.
IT Jolliffe. Principal Component Analysis.
Verlag, New York NY, 1986.
Springer-
FR Kschischang, et al. Factor graphs and the sum-product
algorithm. IEEE Trans Info Theory 47:2, 498–519, 2001.
L Lazzeroni, A Owen. Plaid Models for Gene Expression
Data. Statistica Sinica 12:1, 61-86, 2002.
D Lee, S Seung. Learning the parts of objects by nonnegative matrix factorization. Nature 401, 788–791, 1999.
P Resnick, et al. GroupLens: An Open Architecture for
Collaborative Filtering of Netnews. Proc ACM Conf Comp
Supp Coop Work, 175-186, Chapel Hill, NC, 1994.
N Srebro, et al. Maximum-Margin Matrix Factorization.
Advances in Neural Info Proc Sys 17, 2005.
AH Tong, et al. Global Mapping of the Yeast Genetic
Interaction Network. Science 303, 808–813, 2004.
| 5cs.CE
|
arXiv:1710.09610v1 [math.ST] 26 Oct 2017
Optimal Input Design for Parameter Estimation in AR(1)
with Dependent Stationary Noise
Chunhao Cai
School of Mathematical Science, Nankai university
chunhao [email protected]
October 27, 2017
Abstract
This paper focus on the asymptotical input and the asymptotical properties of the MLE of
the drift parameter in the Autoregressive of order 1 (AR(1)) driven by an regular stationary
noises (with dependence). The Laplace Transform computations will be the main tool for our
analysis.
1
1.1
Introduction
Historical survey
The experiment design has been given great deal of interest over the last decades in the engineering
literature (see e.g.[1, 2, 7, 8, 9, 10]) and also in the statistical literature [11]. The classical approach
for experiment design consist on a two-step procedure: Maximise the Fisher information under
energy constraint of the input and find an adaptive estimation procedure: such as the sequential
design and Bayesian design[1, 10]. Also these works not only considered the identification of directly
observed dynamic systems [9] but also the partially observed systems [10].
For the autoregressive cases, the precious works always focused on the system driven by the white
noises, [8], the independence is a very good property for the estimation procedure. As far as we
know, there have not been the contribution to the problem of experiment design of the autoregressive
process with the dependent noises. But there have been the work about the dependent case such
as the optimal filtering [4] and the Maximum Likelihood Estimation for the AR(p) (see [5]).
In this work, we focus on the asymptotical input and the asymptotical properties of the MLE of
the drift parameter in the Autoregressive of order 1 (AR(1)) driven by an regular stationary noises.
The Laplace Transform computations will be the main tool for our analysis. With
This paper falls into three parts, In the next part of the introduction, we will state the problems
and give the main results. Section 2 and Section 3 is devoted to some Preliminaries and the proofs
of the main results.
1
1.2
Statement of the problems and Main results
We consider the model
Xn = ϑXn−1 + u(n) + ξn , 0 < ϑ < 1, X0 = 0.
(1)
Where ξ = (ξn , n ∈ Z) is the centred regular stationary noise with the autoregressive with
Z π
| log fξ (λ)|dλ < ∞
(2)
−π
where fξ (λ) is the spectral density of ξ. We suppose that the covariance c = c(m, n), m, n ≥ 1 is
Eξm ξn = c(m, n) = ρ(|n − m|), ρ(0) = 1.
is positive defined. In fact, the condition (2) can be changed to the condition of the covariance
function ρ: ρ ∼ cn−α , α > 0.
Suppose that the parameter ϑ is unknown and is to be estimated with the observation data
(Xn , n = 1, · · · , N ). We suppose that L(ϑ, X (N ) ) be the likelihood function for ϑ, then the Fisher
information stands for
∂2
IN (ϑ, u) = −Eϑ 2 ln L(ϑ, X (N ) ).
∂ϑ
Let us consider the space for the function u(n)
P
2
n
k(n, m)u(m)
N
1 X
m=1
≤1
UN = u(n)
N n=1
σn+1
where the function k(n, m) and σn will be defined in (4) and (5). Denote
JN (ϑ) = sup IN (ϑ, u)
u∈UN
Our main goal is to find estimator ϑ̄N of the parameter ϑ which is asymptotically efficient in the
sense that, for any compact K ∈ (0, 1) = {ϑ|0 < ϑ < 1},
sup JN (ϑ)Eϑ (ϑ̄N − ϑ)2 = 1 + 0(1),
(3)
ϑ∈K
as N → ∞. We have the following results:
Theorem 1.1. The asymptotical optimal input in the class of control UT is uopt (n) =
More over,
JN (ϑ)
= I(ϑ)
N →∞
N
lim
where I(ϑ) =
1
1−ϑ2
+
1
(1−ϑ)2
.
2
n
P
m=1
K(n, m)σm+1 .
Remark 1. For the case of −1 < ϑ < 0, the optimal input will be uopt (n) = (−1)n
or uopt (n) = (−1)n+1
is the same of ϑ > 0.
n
P
m=1
K(n, m)σm+1 and I(ϑ) =
1
1−ϑ2
+
1
(1+ϑ)2 .
n
P
K(n, m)σm+1
m=1
In fact, the proof of this case
Now, the optimal input does not depend on the unknown parameter ϑ, we can consider ϑ̄ as
the Maximum Likelihood Estimator ϑ̂N as the maximum of the likelihood function and from the
following theorem we know it will reaches efficiency (3).
Theorem 1.2. For 0 < ϑ <, the MLE ϑ̂N is asymptotically normal at the usual convergent rate
√
1
Law
N ϑ̂N − ϑ −−−−→ N 0,
N →∞
I(ϑ)
where I(ϑ) is defined in Theorem 1.1.
2
2.1
Preliminaries
Stationary Gaussian sequences
We define
σ1 ǫ1 = ξ1 , σn ǫn = ξn − E(ξn |ξ1 , ξ2 , · · · , ξn−1 ), n ≥ 2
(4)
where ǫn ∼ N (0, 1), n ≥ 1 are independent. It follows from the Theorem of Normal Correlation
([6], Theorem 13.1) that there exists a deterministic kernel denoted by k(n, m), n ≥ m, n ≥ 1 such
that
n
X
σn ǫn =
k(n, m)ξm , k(n, n) = 1.
(5)
m=1
For n ≥ 1, we denote by βn−1 the partial correlation coefficient
− k(n, 1) = βn−1 , n ≥ 1,
(6)
then following from [12], we have
σn2 =
n−1
Y
2
(1 − βm
), n ≥ 2, σ1 = 1.
(7)
m=1
On the other hand, there exists an inverse deterministic kernel K = (K(n, m), n ≥ 1, n ≥ m) such
that
n
X
ξn =
K(n, m)σm ǫm .
(8)
m=1
For the relationship ofP
k(n, m) and K(n, m) can refer to [5]. Here it is worth mentioning that the
βn2 < ∞, that is to say βn → 0.
condition (2) implies
n≥1
3
2.2
Model Transformation
Let us define un process Z = (Zn , n ≥ 1) such that
Zn =
n
X
m=1
k(n, m)Xm , n ≥ 1
(9)
where k(n, m) is the kernel defined in (5). As in (8), we also have
Xn =
n
X
K(n, m)Zm ,
m=1
then the process Z has the same filtration of X. In the following parts, we take our observation
be (Z1 , Z2 , · · · , ZN ). Actually, it wa shown in [5] the the process Z can be considered as the first
component of a 2-dimensional AR(1) process ζ = ζn , n ≥ 1 which is defined by:
Zn
P
ζn = n−1
βr Z r
r=1
It is not hard to know that ζn is a 2-dimensional Markov process which satisfies the following
equation:
ζn = An−1 ζn−1 + bv(n) + bσn ǫn , n ≥ 1, ζ0 = 02×1 .
(10)
where
An =
ϑ
βn
ϑβn
1
,b=
1
0
(11)
and ǫn ∼ N (0, 1) are independent noises. Here the function v(n) will be in the space of
)
(
2
N
1 X v(n)
≤1
VN = v
N n=1 σn+1
2.3
Fisher Information
As we have interpreted, the observation will be the first component of the process ζ = (ζn , n ≥ 1).
Now from the equation (10), it is easy to write the Likelihood function L(ϑ, X (N ) ) which depends
on the function v(n):
!2
N
N
∗
ϑ
X
Y
b
ζ
−
A
ζ
1
1
n
n−1 n−1
.
p
(12)
exp −
L(ϑ, X (N ) ) =
2
2
σ
2πσ
n
n
n=1
n=1
then the Fisher Information IN (ϑ, v) will be:
2
N
−1 ∗
X
∂2
an ζn
(N )
IN (ϑ, v) = −Eϑ 2 ln L(ϑ, X ) = Eϑ
.
∂ϑ
σn+1
n=1
Where an =
1
βn
.
4
(13)
3
3.1
Proofs
Proof of Theorem 1.1
To prove the theorem 1.1, we divide the Fisher Information (13) into two parts:
IN (ϑ, v)
=
=
=
N
−1 ∗
X
an ζn
2
− a∗n Eϑ ζn + a∗n Eϑ ζn
Eϑ
σn+1
n=1
2 NX
2
−1 ∗
N
−1
X
an Eϑ ζn
a∗n (ζn − Eϑ ζn )
+
Eϑ
σn+1
σn+1
n=1
n=1
I1,N (ϑ) + I2,N (ϑ, v)
On one hand, it is obvious that I1,N (ϑ) does not depend on the control v(n). Let us denote
Pnϑ = ξn − Eϑ ξn , then Pnϑ satisfies the following equation:
ϑ
Pnϑ = Aϑn−1 Pn−1
+ bσn ǫn , P0ϑ = 02×1
As presented in [5],
N −1
1 X (a∗n Pnϑ )2
lim Eϑ exp −
2
N →∞
2N n=1 σn+1
where I1 (ϑ) =
1
1−ϑ2 .
!
1
= exp − I1 (ϑ)
2
Then
lim
N →∞
To compute I2,N (ϑ), let s(n) =
Eϑ ζn
σn+1 ,
I1,N (ϑ)
1
.
= I1 (ϑ) =
N
1 − ϑ2
(14)
then s(n) satisfies the following equation:
s(n) = An−1 s(n − 1)
σn
+ bf (n).
σn+1
(15)
n
. Note that βn → ∞ and σσn+1
→ 1. Without difference, we suppose that for
where f (n) = σv(n)
n+1
σn
n = 1, 2, · · · , σn+1 ≤ (1+ε) and βn ≤ ε for the sufficiently small positive constant ε and (1+ε)ϑ < 1,
then
Lemma 3.1. We define the 2-dimension equation Y = (Yn , n ≥ 1) satisfying the following equation:
ϑ 0
Yn =
Yn−1 + bf (n), Y (0)
0 1
then
N
N
1 X ∗
1 X ∗
(an s(n))2 = lim
(b Yn )2
N →∞ N
N →∞ N
n=1
n=1
lim
Proof. We define another 2-dimensional equation Y = (Yn′ , n ≥ 1) satisfying the equation:
σn
ϑ 0
′
+ bf (n), Y0′
Yn =
Yn−1
0 1
σn+1
5
we have three comparison, first of all we compare b∗ s(n) − b∗ Yn′ , in fact
σn
0
ϑβn−1
′
′
,
(s(n − 1) − Yn−1
)
s(n) − Yn =
βn−1
0
σn+1
then b∗ (s(n) − Yn′ ) → 0.
Now, we compare b∗ Yn − b∗ Yn′ , in fact,
b
∗
(Yn′
− Yn ) = ϑ(b
∗
(Yn′
− Yn )) + ϑ
σn
′
− 1 b∗ Yn−1
.
σn+1
n
because f (n) is bounded and σσn+1
− 1 → 0, then b∗ (Yn′ − Yn ) → 0.
At last, as βn → 0 and the component of s(n) is bounded, the difference a∗n s(n) − b∗ s(n) → 0
which achieves the proof.
∗
α(n) = b
Yn , then α(n) = ϑα(n − 1) + f (n) where f (n) is in the space of FN =
We define
N
P
f (n) N1
f 2 (n) ≤ 1 , without loss, we suppose α(0) = 0. Let J2,N = sup I2,N (ϑ, v), it is
v∈VN
n=1
clear that
N
X
J2,N (ϑ)
1
= lim
sup
α2 (n).
N →∞ N − 1
N →∞ N f ∈FN
n=1
lim
Now to prove theorem 1.1, we only need the following lemma:
Lemma 3.2.
J2,N (ϑ)
= I2 (ϑ)
N −1
lim
N →∞
where I2 (ϑ) =
1
(1−ϑ)2 .
(16)
Proof. First of all, taking f (n) = 1, we get the lower bound
N
1 X 2
1
lim
.
α (n) ≥
N →∞ N
(1
−
ϑ)2
n=1
To get the lower bound, notice that
α(n) = ϕ(n)
n
X
i=1
ϕ−1 (i)f (i), n ≥ 1, Y (0) = 0
where
we can rewrite
1
N
PN
n=1
ϕ(n) = ϑϕ(n − 1), ϕ(0) = 1.
α2 (n) which will be
n
N
N
X
X
f (i)
1 X 2
ϕ−1 (i) √
ϕ(n)
α (n) =
N n=1
N
n=1
i=1
6
!
ϕ(n)
n
X
i=1
−1
ϕ
f (i)
(i) √
N
!
.
or
N
N X
N
X
1 X 2
f (i) f (j)
FN (i, j) √ √ ,
α (n) =
N n=1
N N
i=1 j=1
where
FN (i, j) =
N
X
ℓ=i
Let φn = ϕ−1 (n)
N
P
ℓ=n
W
j
ϕ(ℓ)ϕ−1 (i) ϕ(ℓ)ϕ−1 (j)
φℓ ǫℓ , where ǫℓ ∼ N (0, 1) are independent, then
FN (i, j) = E(φi φj )
and
φn−1 = ϑφn + ǫn−1 , ZN = 0.
In fact, FN (i, j) is a compact symmetric operator for fixed N . We should estimate the spectral
gap (the first eigenvalue ν1 (N )) of the operator. The estimation of the spectral gap is based on the
Laplace transform computation. Let us compute, for sufficiently small negative a < 0, then Laplace
N
P
φ2i :
transform of
i=1
!
N
aX 2
LN (a) = Eϑ exp −
φ .
2 i=1 i
2
On one hand, for a > − ν1 (N
) , since φ is a centred Gaussian process with covariance operator FN ,
using Mercer’s theorem and Parseval’s equality, LT (a) can be represented by
Y
LN (a) =
(1 + aνi (N ))−1/2
(17)
i≥1
where νi (N ) is the sequence of positive eigenvalues of the covariance operator. On the other hand,
LN (a) = ϑN −1 Ψ1N
where
ΨN =
1
1
ϑ
a
ϑ
1
ϑ
(18)
N −1
0
2
1+a
+ ϑ − 4 ≥ 0,
ϑ
For
∆=
−1/2
a
ϑ
+ϑ
1
0
.
there exists two real eigenvalues λ2 ≥ 1, λ1 ≤ 1 of the matrix
1
1
ϑ
ϑ
,
a
a
ϑ
ϑ +ϑ
then
ΨN =
−2
−2
λN −1 − λN −1
λN
− λN
2
1
−
ϑ2
ϑ
7
!
ϑ
≥0
λ2 − λ1
That is to say for ϑ > 0 and for any 0 > a ≥ −(1 − ϑ)2 , LN (a) ≥ 0, it means lim ν1 (N ) ≤
N →∞
which achieves the proof.
Remark 2. For −1 < ϑ < 0, ∆ ≥ 0 means
(1 + ϑ)2 .
3.2
1+a
ϑ
1
(1−ϑ)2
+ ϑ ≤ −2 and 0 > a > −(1 + ϑ)2 then ν1 (N ) ≤
Proof of theorem 1.2
We denote vopt (n) = σn+1 and ζ o = (ζno , n ≥ 1) be the process ζ with the function vopt(n) , then
o
ζno = An−1 ζn−1
+ bvopt (n) + bσn ǫn , ζ0 = 02×1
To estimate the parameter ϑ, we will observe ζ1 , ζ2 , · · · , ζN , from equation (12) the maximum
likelihood estimator ϑ̂N of ϑ will be
ϑ̂N =
2
N ∗
X
a ζn
n
σn+1
n=1
!−1
even more
ϑ̂N − ϑ =
where
MN
N
X
a∗n ζn b∗ ζn+1
2
σn+1
n=1
!
,
(19)
MN
hM iN
2
N ∗
N
X
X
an ζn
a∗n ζn
ǫn+1 , hM iN =
=
σ
σn+1
n+1
n=1 n+1
are martingale and its bracket process. The theorem 1.2 is crucially based on the asymptotical
study for N → ∞ of the Laplace transform
µ
µ
= Eϑ exp −
hM iN
LϑN
N
2N
for every real number µ. We have the following lemma:
Lemma 3.3. Under the condition (2), for any µ ∈ R,
µ
1
ϑ
= exp − I(ϑ)
lim LN
N →∞
N
2
where I(ϑ) =
1
1−ϑ2
+
1
(1−ϑ)2 .
To prove this lemma, we rewrite LϑN
LϑN
where Mn =
(20)
µ
an an ∗,
2
N σn+1
µ
N
µ
N
with the following formula:
N
1X ∗
= Eϑ exp −
ζ Mn ζn
2 n=1 n
in the Appendix, we will prove that:
8
!
Lemma 3.4. For any N , the following equality holds:
LϑN
µ
N
=
N
Y
[det(Id + γ(n, n)Mn )]
−1/2
n=1
N
1X ∗
exp −
z Mn (Id + γ(n, n)Mn )−1 zn
2 n=1 n
where (γ(n, m), 1 ≤ m ≤ n) is the unique solution of the equation
#
" n
Y
−1
γ(m, m),
γ(n, m) =
Ar−1 (Id + γ(r, r)Mr )
!
(21)
r=m+1
the function (γ(n, n, n ≥ 1)) is the solution of the Ricatti equation:
γ(n, n) = An−1 (Id + γ(n − 1, n − 1)Mn−1 )−1 γ(n − 1, n − 1)A∗n−1 + σn2 bb∗ .
(22)
Here (zn , 1 ≤ n ≤ N ) is the unique solution of the equation
z n = mn −
where mn =
Eζno .
n−1
X
r=1
γ(n, r)[Id + γ(r, r)Mr ]−1 Mr zr , z0 = m0 .
Now we will return to the proof of lemma 3.3. In [5] we have presented
lim
N →∞
N
Y
[det(Id + γ(n, n)Mn )]−1/2 =
n=1
µ
.
1 − ϑ2
(23)
we know that the component of γ(n, n) is bounded. then
lim Id + γ(n, n)Mn = Id.
N →∞
On the other hand,
N
X
n=1
m∗n Mn mn
2
N
µ X a∗n Eζn0
µ
, N → ∞.
=
−→
N n=1 σn+1
(1 − ϑ)2
(24)
which has been presented in last part. At last, notice
#
" n
n−1
n−1
X
X Y
−1
−1
[Id + γ(r, r)Mr ]Mr zr
Aτ −1 (Id + γ(τ, τ )Mτ )
γ(n, r)[Id + γ(r, r)Mr ] Mr zr =
r=1
r=1
we have that
lim
N →∞
n−1
X
r=1
τ =r+1
γ(n, r)[Id + γ(r, r)Mr ]−1 Mr zr = 0.
(25)
Combining (23) and (24) and (25), the Lemma 3.3 achieves. It is immediately from this conclusion
1
hM iN = I(ϑ)
N
and due to the central limit theorem for martingale,
Pϑ lim
N →∞
1
√ MN =⇒ N (0, I(ϑ)).
N
which immediately imply the conclusion of the theorem 1.2
9
References
[1] V. Levadi (1966) Design of input signals for parameter estimation, IEEE Trans. Automat.
Control, 11, 205-211.
[2] H. Aoki and R. Staley (1970) On input signal synthesis in parameter identification, Automatica,
6, 431-440.
[3] M. Kleptsyna, A. Le Breton and M. Viot (2002) New formulas concerning Laplace transforms of quadratic forms for general Gaussian sequences, Journal of Applied Mathematics and
Stochastic Analysis, 15(4), 309-325
[4] A. Brouste and M. Kleptsyna (2012) Kalman type filter under stationary noises, Systems &
Control Lettes 61, 1229-1234
[5] A. Brouste, C. Cai and M. Kleptsyna (2014) Asymptotic properties of the MLE for the autoregressive process coefficients under stationary Gaussian noises, Mathematical Methods of
Statistics, 23(2), 103-115
[6] R. S. Liptser and A. N. Shiryaev (1978) Statistics of Random Process II - Application, SpringerVerlag, New York.
[7] T. S. NG, Graham C. Goodwin and Torsten Soderstrom (1977) Optimal experiment design for
linear system with input-Output constraints, Automatica, 13, 571-577.
[8] T. S. NG, Z. H. Qureshi and Y. C. Cheah (1984) Optimal input design for an AR model with
output constraints, Automatica, 20(3), 359-363.
[9] R. K. Mehra (1974) Optimal input signals for parameter estimation in dynamic systems—
survey and new results, IEEE Trans Aut. Control, AC-19, 753.
[10] R. K. Mehra (1974) Optimal inputs for linear system identification, IEEE Trans Aut. Control,
AC-19, 192.
[11] J. Kiefer (1974) On the efficient design of statistical investigation, The annals of statistics, 2,
849-879.
[12] J. Durbin (1960) The fitting of time series models, Rev. Inst. Intern. Statist, 28, 233-243.
10
| 10math.ST
|
Lower Bounds on Time–Space Trade-Offs for
Approximate Near Neighbors
arXiv:1605.02701v3 [cs.DS] 18 Aug 2016
Alexandr Andoni
Columbia
Thijs Laarhoven
IBM Research Zürich
Ilya Razenshteyn
MIT CSAIL
Erik Waingarten
Columbia
August 22, 2016
Abstract
We show tight lower bounds for the entire trade-off between space and query time for the
Approximate Near Neighbor search problem. Our lower bounds hold in a restricted model
of computation, which captures all hashing-based approaches. In particular, our lower bound
matches the upper bound recently shown in [Laa15c] for the random instance on a Euclidean
sphere (which we show in fact extends to the entire space Rd using the techniques from [AR15]).
We also show tight, unconditional cell-probe lower bounds for one and two probes, improving
upon the best known bounds from [PTW10]. In particular, this is the first space lower bound (for
any static data structure) for two probes which is not polynomially smaller than for one probe.
To show the result for two probes, we establish and exploit a connection to locally-decodable
codes.
1
Introduction
1.1
Approximate Near Neighbor problem (ANN)
The Near Neighbor Search problem (NNS) is a basic and fundamental problem in computational
geometry, defined as follows. We are given a dataset of n points P from a metric space (X, dX ) and
a distance threshold r > 0. The goal is to preprocess P in order to answer near neighbor queries:
given a query point q ∈ X, return a dataset point p ∈ P with dX (q, p) ≤ r, or report that there
is no such point. The d-dimensional Euclidean (Rd , `2 ) and Manhattan (Rd , `1 ) metric spaces have
received the most attention. Besides its classical applications to similarity search over many types
of data (text, audio, images, etc; see [SDI06] for an overview), NNS has been also recently used for
cryptanalysis [Laa15a, Laa15b] and optimization [DRT11, HLM15, ZYS16].
The performance of a NNS data structure is often characterized by two key metrics:
• the amount of memory a data structure occupies, and
• the time it takes to answer a query.
All known time-efficient data structures for NNS (e.g., [Cla88, Mei93]) require space exponential
in the dimension d, which is prohibitively expensive unless d is very small. To overcome this socalled “curse of dimensionality”, researchers proposed the (c, r)-Approximate Near Neighbor Search
problem, or (c, r)-ANN. In this relaxed version, we are given a dataset P and a distance threshold
r > 0, as well as an approximation factor c > 1. Given a query point q with the promise that there
is at least one data point in P within distance at most r from q, the goal is to return a data point
p ∈ P within a distance at most cr from q.
This approximate version of NNS allows efficient data structures with space and query time
polynomial in d and query time sublinear in n [KOR00, IM98, Ind01b, Ind01a, GIM99, Cha02,
DIIM04, CR04, Pan06, AC09, AI06, TT07, AINR14, AR15, Pag16, Kap15, BDGL16, Laa15c]. In
practice, ANN algorithms are often successful for similarity search even when one is interested
in exact nearest neighbors [ADI+ 06, AIL+ 15]. We refer the reader to [HIM12, AI08, And09] for
a survey of the theory of ANN, and [WSSJ14, WLKC15] for a more practical perspective.
In this paper, we study tight time–space trade-offs for ANN. Before stating our results in
Section 1.6, we provide more background on the problem.
1.2
Locality-Sensitive Hashing (LSH) and beyond
A classic technique for ANN is Locality-Sensitive Hashing (LSH), introduced in 1998 by Indyk and
Motwani [IM98, HIM12]. The main idea is to use random space partitions, for which a pair of close
points (at distance at most r) is more likely to belong to the same part than a pair of far points
(at distance more than cr). Given such a partition, the data structure splits the set P according
to the partition, and, given a query, retrieves all the data points which belong to the same part as
the query. To get a high probability of success, the data structure maintains several partitions and
checks all of them during the query stage. LSH yields data structures with space O(n1+ρ +d·n) and
query time O(d · nρ ). For a particular metric space and approximation c, ρ measures the quality of
the random space partition. Usually, ρ = 1 for c = 1 and ρ → 0 as c → ∞.
Since the introduction of LSH in [IM98], subsequent research established optimal values of the
LSH exponent ρ for several metrics of interest, including `1 and `2 . For the Hamming distance (`1 ),
the optimal value is ρ =
ρ=
1
c2
1
c
± o(1) [IM98, MNP07, OWZ14]. For the Euclidean metric (`2 ), it is
± o(1) [IM98, DIIM04, AI06, MNP07, OWZ14].
More recently, it has been shown that better bounds on ρ are possible if the space partitions
are allowed to depend on the dataset 1 . That is, the algorithm is based on an observation that every
dataset has some structure to exploit. This more general framework of data-dependent LSH yields
ρ=
1
2c−1
+ o(1) for the `1 distance, and ρ =
1
2c2 −1
+ o(1) for `2 [AINR14, Raz14, AR15]. Moreover,
these bounds are known to be tight for data-dependent LSH [AR16].
1
Let us note that the idea of data-dependent random space partitions is ubiquitous in practice, see, e.g., [WSSJ14,
WLKC15] for a survey. But the perspective in practice is that the given datasets are not “worst case” and hence it
is possible to adapt to the additional “nice” structure.
2
1.3
Random instances: the hardest instances
At the core of the optimal data-dependent LSH data structure for `1 from [AR15] is an algorithm
that handles the following random instances of ANN over Hamming space (also known as the light
bulb problem in literature [Val88] in the off-line setting).
• The dataset P consists of n independent uniformly random points from {−1, 1}d , where
d = ω(log n);
• A query q is generated by choosing a uniformly random data point p ∈ P , and flipping each
coordinate of p with probability
1
2c
independently;
• The goal for a data structure is to recover the data point p from the query point q.
At a high level, the data structure from [AR15] proceeds in two steps:
• it designs a (data-independent) LSH family that handles the random instance, and
• it develops a reduction from a worst-case instance to several instances that essentially look
like random instances.
Thus, random instances are the hardest for ANN. On the other hand, random instances have been
used for the lower bounds on ANN (more on this below), since they must be handled by any data
d
+ o(1) -ANN over `1 .
structure for c, 2c
1.4
Time–space trade-offs
LSH gives data structures with space around n1+ρ and query time around nρ . Since early results
on LSH, the natural question has been whether one can trade space for time and vice versa. One
can achieve polynomial space with poly-logarithmic query time [IM98, KOR00], as well as nearlinear space with sublinear query time [Ind01a]. In the latter regime, [Pan06, Kap15] and, most
recently, [Laa15c] gave subsequent improvements. We point out that the near-linear space regime
is especially relevant for practice: e.g., see [LJW+ 07, AIL+ 15] for practical versions of the above
theoretical results.
For random instances, the best known trade-off is from [Laa15c]:
Theorem 1.1 (Theorem 1 of [Laa15c]). Let c ∈ (1, ∞). One can solve
on the unit sphere S d−1 ⊂ Rd equipped with `2 norm with query time O(d ·
√
2
+
o(1)
-ANN
c
nρq +o(1) ), and space
c,
O(n1+ρu +o(1) + d · n) where
p
√
√
c2 ρq + (c2 − 1) ρu = 2c2 − 1.
3
(1)
This data structure can handle the random Hamming instances introduced in Section 1.3 via a
standard reduction. The resulting time–space trade-off is:
√
√
√
c ρq + (c − 1) ρu = 2c − 1.
(2)
For the sake of illustration, consider the setting of the Hamming distance and approximation
c = 2. The optimal data-dependent LSH from [AR15] gives space n4/3+o(1) and query time n1/3+o(1) .
For random instances, the above bound (2) gives the same bound as well as a smooth interpolation
between the following extremes: space n1+o(1) and query time n3/4+o(1) , and space n4+o(1) and
query time no(1) .
The algorithm from [Laa15c] can be applied to the entire `2 sphere (and hence, via standard
reductions à la [Val15, Algorithm 25], to the entire space Rd ). However, this direct extension
degrades the quality of the (ρq , ρu ) trade-off to essentially those corresponding to the classical LSH
bounds (e.g., for ρq = ρu , obtaining ρq = ρu = 1/c2 + o(1), instead of the optimal ρq = ρu =
1/(2c2 − 1) + o(1)). Nonetheless, it is possible to apply the worst-case–to–random-case reduction
from [AR15] in order to extend Theorem 1.1 to the entire Rd with the same trade-off as (1) (see
Appendices B and C for details).
Furthermore, we note that all algorithms for `2 extend to `p , for p ∈ (1, 2), with c2 being
replaced with cp in the expressions for the exponents (ρq , ρu ). This follows from the reduction
shown in [Ngu14, Section 5.5].
1.5
Lower bounds
Lower bounds for NNS and ANN have also received much attention. Such lower bounds are almost
always obtained in the cell-probe model [MNSW98, Mil99]. In the cell-probe model one measures
the number of memory cells the query algorithm accesses. Despite a number of success stories, high
cell-probe lower bounds are notoriously hard to prove. In fact, there are few techniques for proving
high cell-probe lower bounds, for any (static) data structure problem. For ANN in particular, we
have no viable techniques to prove ω(log n) query time lower bounds. Due to this state of affairs, one
may rely on restricted models of computation, which nevertheless capture existing upper bounds.
Early lower bounds for NNS were shown for data structures in the exact or deterministic settings [BOR99, CCGL99, BR02, Liu04, JKKR04, CR04, PT06, Yin16]. In [CR04, LPY16] an almost
tight cell-probe lower bound is shown for the randomized Approximate Nearest Neighbor Search
under the `1 distance. In the latter problem, there is no distance threshold r, and instead the goal
is to find a data point that is not much further than the closest data point. This twist is the main
source of hardness, and the result is not applicable to the ANN problem as introduced above.
There are few results that show lower bounds for randomized data structures for the approximate
near neighbor problem (the setting studied in the present paper). The first such result [AIP06]
4
shows that any data structure that solves (1 + ε, r)-ANN for `1 or `2 using t cell probes requires
2
space nΩ(1/tε ) .2 This result shows that the algorithms of [IM98, KOR00] are tight up to constants
in the exponent for t = O(1).
In [PTW10] (following up on [PTW08]), the authors introduce a general framework for proving
lower bounds for ANN under any metric. They show that lower bounds for ANN are implied by
the robust expansion of the underlying metric space. Using this framework, [PTW10] show that
2)
(c, r)-ANN using t cell probes requires space n1+Ω(1/tc) for the Hamming distance and n1+Ω(1/tc
for the Euclidean distance (for every c > 1).
Lower bounds were also shown for other metrics. For the `∞ distance, [ACP08] show a lower
bound for deterministic ANN data structures, matching the upper bound of [Ind01b] for decision
trees. This lower bound was later generalized to randomized data structures [PTW10, KP12].
A recent result [AV15] adapts the framework of [PTW10] to Bregman divergences. There are also
lower bounds for restricted models: for LSH [MNP07, OWZ14, AIL+ 15] and for data-dependent
LSH [AR16]. We note that essentially all of the aforementioned lower bounds for ANN under `1
[AIP06, PTW10, MNP07, AIL+ 15, AR16] use the random instance defined in Section 1.3 as a hard
distribution.
1.6
Our results
In this paper, we show both new cell-probe and restricted lower bounds for (c, r)-ANN. In all
cases our lower bounds match the upper bounds from [Laa15c]. Our lower bounds use the random
instance from Section 1.3 as a hard distribution. Via a standard reduction, we obtain similar
hardness results for `p with 1 < p ≤ 2 (with c being replaced by cp ).
1.6.1
One cell probe
First, we show a tight (up to no(1) factors) lower bound on the space needed to solve ANN for a
random instance, for query algorithms that use a single cell probe. More formally, we prove the
following theorem:
Theorem 1.2 (Section 4). Any data structure that:
• solves (c, r)-ANN for the Hamming random instance (as defined in Section 1.3) with probability 2/3,
• operates on memory cells of size no(1) ,
• for each query, looks up a single cell,
c
must use at least n( c−1 )
2
2
−o(1)
words of memory.
The correct dependence on 1/ε requires a stronger LSD lower bound from [Pǎt11].
5
The space lower bound matches the upper bound from [Laa15c] (see also Appendix C) for ρq = 0.
The previous best lower bound from [PTW10] for a single probe was weaker by a polynomial factor.
We prove Theorem 1.2 by computing tight bounds on the robust expansion of a hypercube
{−1, 1}d as defined in [PTW10]. Then, we invoke a result from [PTW10], which yields the de-
sired cell probe lower bound. We obtain estimates on the robust expansion via a combination of
the hypercontractivity inequality and Hölder’s inequality [O’D14]. Equivalently, one could obtain
the same bounds by an application of the Generalized Small-Set Expansion Theorem of [O’D14].
1.6.2
Two cell probes
To state our results for two cell probes, we first define the decision version of ANN (first introduced
in [PTW10]). Suppose that with every data point p ∈ P we associate a bit xp ∈ {0, 1}. A new
goal is: given a query q ∈ {−1, 1}d which is at distance at most r from a data point p ∈ P , and
assuming that P \ {p} is at distance more than cr from q, return correct xp with probability at
least 2/3. It is easy to see that any algorithm for (c, r)-ANN would solve this decision version.
We prove the following lower bound for data structures making only two cell probes per query.
Theorem 1.3 (see Section 6). Any data structure that:
• solves the decision ANN for the random instance (Section 1.3) with probability 2/3,
• operates on memory cells of size o(log n),
• accesses at most two cells for each query,
c
must use at least n( c−1 )
2
−o(1)
words of memory.
Informally speaking, we show that the second cell probe cannot improve the space bound by
more than a subpolynomial factor. To the best of our knowledge, this is the first lower bound for
the space of any static data structure problem without a polynomial gap between t = 1 and t ≥ 2
cell-probes. Previously, the highest ANN lower bound for two queries was weaker by a polynomial
factor [PTW10]. (This remains the case even if we plug the tight bound on the robust expansion
into the framework of [PTW10].) Thus, in order to obtain a higher lower bound for t = 2, we need
to depart from the framework of [PTW10].
Our proof establishes a connection between two-query data structures (for the decision version
of ANN), and two-query locally-decodable codes (LDC). A possibility of such a connection was
suggested in [PTW10]. In particular, we show that a data structure violating the lower bound
from Theorem 1.3 implies an efficient two-query LDC, which contradicts known LDC lower bounds
from [KdW04, BRdW08].
The first lower bound for unrestricted two-query LDCs was proved in [KdW04] via a quantum
argument. Later, the argument was simplified and made classical in [BRdW08]. It turns out that
6
for our lower bound, we need to resort to the original quantum argument of [KdW04] since it has
a better dependence on the noise rate a code is able to tolerate. During the course of our proof,
we do not obtain a full-fledged LDC, but rather an object which can be called an LDC on average.
For this reason, we are unable to use [KdW04] as a black box but rather adapt their proof to the
average case.
Finally, we point out an important difference with Theorem 1.2: in Theorem 1.3 we allow words
to be merely of size o(log n) (as opposed to no(1) ). Nevertheless, for the decision version of ANN
the upper bounds from [Laa15c] hold even for such “tiny” words. In fact, our techniques do not
allow us to handle words of size Ω(log n) due to the weakness of known lower bounds for twoquery LDC for large alphabets. In particular, our argument can not be pushed beyond word size
√
2Θ(
e
log n)
in principle, since this would contradict known constructions of two-query LDCs over
large alphabets [DG15]!
1.6.3
The general time–space trade-off
Finally, we prove conditional lower bound on the entire time–space trade-off that is tight (up to
no(1) factors), matching the upper bound from [Laa15c] (see also Appendix C). Note that—since
we show polynomial query time lower bounds—proving similar lower bounds unconditionally is far
beyond the current reach of techniques, modulo major breakthrough in cell probe lower bounds.
Our lower bounds are proved in the following model, which can be loosely thought of comprising
all hashing-based frameworks we are aware of:
Definition 1.4. A list-of-points data structure for the ANN problem is defined as follows:
• We fix (possibly randomly) sets Ai ⊆ {0, 1}d , for i = 1 . . . m; also, with each possible query
point q ∈ {0, 1}d , we associate a (random) set of indices I(q) ⊆ [m];
• For a given dataset P , the data structure maintains m lists of points L1 , L2 , . . . , Lm , where
Li = P ∩ Ai ;
• On query q, we scan through each list Li for i ∈ I(q) and check whether there exists some
p ∈ Li with kp − qk1 ≤ cr. If it exists, return p.
P
P
The total space is defined as s = m + m
i=1 |Li | and the query time is t = |I(q)| +
i∈I(q) |Li |.
For this model, we prove the following theorem.
Theorem 1.5 (see Section 5). Consider any list-of-points data structure for (c, r)-ANN for random
instances of n points in the d-dimensional Hamming space with d = ω(log n), which achieves a total
space of n1+ρu +o(1) , and has query time nρq −o(1) , for 2/3 success probability. Then it must hold
that:
√
√
√
c ρq + (c − 1) ρu ≥ 2c − 1.
7
(3)
We note that our model captures the basic hashing-based algorithms, in particular most of the
known algorithms for the high-dimensional ANN problem [KOR00, IM98, Ind01b, Ind01a, GIM99,
Cha02, DIIM04, Pan06, AC09, AI06, Pag16, Kap15], including the recently proposed LocalitySensitive Filters scheme from [BDGL16, Laa15c]. The only data structures not captured are the
data-dependent schemes from [AINR14, Raz14, AR15]; we conjecture that the natural extension of
the list-of-point model to data-dependent setting would yield the same lower bound. In particular,
Theorem 1.5 uses the random instance as a hard distribution, for which being data-dependent
seems to offer no advantage. Indeed, a data-dependent lower bound in the standard LSH regime
(where ρq = ρs ) has been recently shown in [AR16], and matches (3) for ρs = ρq .
1.7
Other related work
There has been a lot of recent algorithmic advances on high-dimensional similarity search, including
better algorithms for the closest pair problem3 [Val15, AW15, KKK16, KKKÓ16], locality-sensitive
filters [BDGL16, Laa15c], LSH without false negatives [Pag16, PP16], to name just a few.
2
Preliminaries
We introduce a few definitions from [PTW10] to setup the nearest neighbor search problem for
which we show lower bounds.
Definition 2.1. The goal of the (c, r)-approximate nearest neighbor problem with failure probability
δ is to construct a data structure over a set of points P ⊂ {0, 1}d supporting the following query:
given any point q such that there exists some p ∈ P with kq − pk1 ≤ r, report some p0 ∈ P where
kq − p0 k1 ≤ cr with probability at least 1 − δ.
Definition 2.2 ([PTW10]). In the Graphical Neighbor Search problem (GNS), we are given a
bipartite graph G = (U, V, E) where the dataset comes from U and the queries come from V . The
dataset consists of pairs P = {(pi , xi ) | pi ∈ U, xi ∈ {0, 1}, i ∈ [n]}. On query q ∈ V , if there exists
a unique pi with (pi , q) ∈ E, then we want to return xi .
We will sometimes use the GNS problem to prove lower bounds on (c, r)-ANN as follows: we
build a GNS graph G by taking U = V = {0, 1}d , and connecting two points u ∈ U, v ∈ V iff they
are at a distance at most r (see details in [PTW10]). We will also need to make sure that in our
instances q is not closer than cr to other points except the near neighbor.
2.1
Robust Expansion
The following is the fundamental property of a metric space that [PTW10] use to prove lower
bounds.
3
These can be seen as the off-line version of NNS/ANN.
8
Definition 2.3 (Robust Expansion [PTW10]). For a GNS graph G = (U, V, E), fix a distribution
e on E ⊂ U × V , and let µ be the marginal on U and η be the marginal on V . For δ, γ ∈ (0, 1], the
robust expansion Φr (δ, γ) is defined as follows:
Φr (δ, γ) =
2.2
min
A⊂V :η(A)≤δ
min
e(A×B)
B⊂U : e(A×V ) ≥γ
µ(B)
.
η(A)
Locally Decodable Codes
Finally, our 2-cell lower bounds uses results on Locally Decodable Codes (LDCs). We present the
standard definitions and results on LDCs below, although we will need a weaker definition (and
stronger statement) for our 2-query lower bound in Section 6.
Definition 2.4. A (t, δ, ε) locally decodable code (LDC) encodes n-bit strings x ∈ {0, 1}n into m-bit
codewords C(x) ∈ {0, 1}m such that, for each i ∈ [n], the bit xi can be recovered with probability
1
2
+ ε while making only t queries into C(x), even if the codeword is arbitrarily modified (corrupted)
in δm bits.
We will use the following lower bound on the size of the LDCs.
Theorem 2.5 (Theorem 4 from [KdW04]). If C : {0, 1}n → {0, 1}m is a (2, δ, ε)-LDC, then
2 n)
m ≥ 2Ω(δε
3
.
(4)
Robust Expansion of the Hamming Space
The goal of this section is to compute tight bounds for the robust expansion Φr (δ, γ) in the Hamming
space of dimension d, as defined in the preliminaries. We use these bounds for all of our lower bounds
in the subsequent sections.
We use the following model for generating dataset points and queries (which is essentially the
random instance from the introduction).
Definition 3.1. For any x ∈ {−1, 1}n , Nσ (x) is a probability distribution over {−1, 1}n representing the neighborhood of x. We sample y ∼ Nσ (x) by choosing yi ∈ {−1, 1} for each coordinate
i ∈ [d]. With probability σ, yi = xi . With probability 1 − σ, yi is set uniformly at random.
Given any Boolean function f : {−1, 1}n → R, the function Tσ f : {−1, 1}n → R is
Tσ f (x) =
E
[f (y)]
(5)
y∼Nσ (x)
In the remainder of this section, will work solely on the Hamming space V = {−1, 1}d . We let
σ =1−
1
c
d = ω(log n)
9
and µ will refer to the uniform distribution over V .
The choice of σ allows us to make the following observations. A query is generated as follows:
we sample a dataset point x uniformly at random and then generate the query y by sampling
d
2c (1
y ∼ Nσ (x). From the choice of σ, d(x, y) ≤
every other point in the dataset
x0
+ o(1)) with high probability. In addition, for
6= x, the pair (x0 , y) is distributed as two uniformly random
points (even though y ∼ Nσ (x), because x is randomly distributed). Therefore, by taking a union-
bound over all dataset points, we can conclude that with high probability, d(x0 , y) ≥ d2 (1 − o(1))
for each x0 6= x.
Given a query y generated as described above, we know there exists a dataset point x whose
distance to the query is d(x, y) ≤
d(x0 , y)
≥
d
2 (1
d
2c (1
+ o(1)). Every other dataset point lies at a distance
− o(1)). Therefore, the two distances are a factor of c − o(1) away.
The following lemma is the main result of this section, and we will reference this lemma in
subsequent sections.
Lemma 3.2 (Robust expansion). In the Hamming space equipped with the Hamming norm, for
any p, q ∈ [1, ∞) where (q − 1)(p − 1) = σ 2 , any γ ∈ [0, 1] and m ≥ 1,
1
,γ
m
Φr
≥ γqm
1+ pq −q
(6)
The robust expansion comes from a straight forward application from small-set expansion. In
fact, one can easily prove tight bounds on robust expansion via the following lemma:
Theorem 3.3 (Generalized Small-Set Expansion Theorem, [O’D14]). Let 0 ≤ σ ≤ 1. Let A, B ⊂
2
2
{−1, 1}n have volumes exp(− a2 ) and exp(− b2 ) and assume 0 ≤ σa ≤ b ≤ a. Then
Pr
(x,y)
σ−correlated
1 a2 − 2σab + b2
[x ∈ A, y ∈ B] ≤ exp −
2
1 − σ2
However, we compute the robust expansion via an application of the Bonami-Beckner Inequality
and Hölder’s inequality. This computation gives us a bit more flexibility with respect to parameters
which will become useful in subsequent sections. We now recall the necessary tools.
p
Theorem 3.4 (Bonami-Beckner Inequality [O’D14]). Fix 1 ≤ p ≤ q and 0 ≤ σ ≤ (p − 1)/(q − 1).
Any Boolean function f : {−1, 1}n → R satisfies
kTσ f kq ≤ kf kp
(7)
Theorem 3.5 (Hölder’s Inequality). Let f : {−1, 1}n → R and g : {−1, 1}n → R be arbitrary
Boolean functions. Fix s, t ∈ [1, ∞) where
1
s
+
1
t
= 1. Then
hf, gi ≤ kf ks kgkt
10
(8)
We will let f and g be indicator functions for two sets A and B and use a combination of
the Bonami-Beckner Inequality and Hölder’s Inequality to lower bound the robust expansion. The
operator Tσ will applied to f will measure the neighborhood of set A. We will compute an upper
bound on the correlation of the neighborhood of A and B (referred to as γ) with respect to the
volumes of A and B, and the expression will give a lower bound on robust expansion.
We also need the following lemma.
Lemma 3.6. Let p, q ∈ [1, ∞), where (p − 1)(q − 1) = σ 2 and f, g : {−1, 1}d → R be two Boolean
functions. Then
hTσ f, gi ≤ kf kp kgkq
Proof. We first apply Hölder’s Inequality to split the inner-product into two parts. Then we apply
the Bonami-Beckner Inequality to each part.
hTσ f, f i = hT√σ f, T√σ gi
≤ kT√σ f ks kT√σ gkt
p−1
s
+ 1 and t =
, so
σ
s−1
2
and p ≥ 1 because (p − 1)(q − 1) = σ ≤ σ. We have
We pick the parameters s =
q≤
1
s
(9)
(10)
+ 1t = 1. Note that p ≤ s because σ < 1
σ
+ 1 = t.
p−1
(11)
In addition,
r
r
p−1 √
= σ
s−1
q−1 p
= (q − 1)(s − 1)
t−1
r
(q − 1)(p − 1) √
=
= σ.
σ
(12)
(13)
So we can apply the Bonami-Beckner Inequality to both norms. We obtain
kT√σ f ks kT√σ gkt ≤ kf kp kgkq
(14)
We are now ready to prove Lemma 3.2.
Proof of Lemma 3.2. We use Lemma 3.6 and the definition of robust expansion. For any two sets
A, B ⊂ V , let a =
1
|A|
2d
and b =
1
|B|
2d
be the measure of set A and B with respect to the uniform
distribution. We refer to 1A : {−1, 1}d → {0, 1} and 1B : {−1, 1}d → {0, 1} as the indicator
11
functions for A and B.
γ=
=
1
q− pq
−1
(16)
1
bq
(17)
≤ b. Let A and B be the minimizers of
Φr
1
,γ
m
b
a
satisfying (15) and a ≤
1
m.
b
a
=
(18)
≥ γqa
q− pq −1
1+ pq −q
≥ γqm
4
(15)
1
hTσ 1A , 1B i
a
≤ ap
Therefore, γ q a
[x ∈ B | y ∈ A]
Pr
x∼µ,y∼Nσ (x)
(19)
.
(20)
Tight Lower Bounds for 1 Cell Probe Data Structures
In this section, we prove Theorem 1.2. Our proof relies on the main result of [PTW10] for the GNS
problem:
Theorem 4.1 (Theorem 1.5 [PTW10]). There exists an absolute constant γ such that the following
holds. Any randomized algorithm for a weakly independent instance of GNS which is correct with
probability greater than
1
2
must satisfy
mt w
≥ Φr
n
1 γ
,
mt t
(21)
1
Proof of Theorem 1.2. The bound comes from a direct application of the computation of Φr ( m
, γ)
in Lemma 3.2 to the bound in Theorem 4.1. Setting t = 1 in Theorem 4.1, we obtain
mw ≥ n · Φr
1
,γ
m
1+ pq −q
≥ nγ q m
(22)
(23)
for some p, q ∈ [1, ∞) and (p − 1)(q − 1) = σ 2 . Rearranging the inequality, we obtain
p
m≥
p
γ p−1 n pq−q
p
w pq−q
12
(24)
Let p = 1 +
log log n
log n ,
and q = 1 + σ 2 logloglogn n . Then
1
m ≥ n σ2 −o(1) .
Since σ = 1 −
1
c
(25)
and w = no(1) , we obtain the desired result.
Corollary 4.2. Any 1 cell probe data structures with cell size O(log n) for c-approximate nearest
neighbors on the sphere in `2 needs n
1+
2c2 −1
−o(1)
(c2 −1)2
many cells.
Proof. Each point in the Hamming space {−1, 1}d (after scaling by
√1 )
d
can be thought of as lying
√
on the unit sphere. If two points are a distance r apart in the Hamming space, then they are 2 r
apart on the sphere with `2 norm. Therefore a data structure for a c2 -approximation on the sphere
gives a data structure for a c-approximation in the Hamming space.
5
Lower Bounds for List-of-Points Data Structures
In this section we prove Theorem 1.5, i.e., a tight lower bound against data structure that fall inside
the “list-of-points” model, as defined in Def. 1.4.
Recall that Ai ⊂ V is the subset of dataset points which get placed in Li . Let Bi ⊂ V the
subset of query points which query Li , this is well defined, since Bi = {v ∈ V | i ∈ I(v)}. Suppose
we sample a random dataset point u ∼ V and then a random query point v from the neighborhood
of u. Let
γi = Pr[v ∈ Bi | u ∈ Ai ]
(26)
and let si = µ(Ai ).
On instances where n dataset points {ui }ni=1 are drawn randomly, and a query v is drawn from
the neighborhood of a random dataset point, we can exactly characterize the query time.
T =
m
X
i=1
E[T ] =
≥
m
X
i=1
m
X
i=1
1{v ∈ Bi } 1 +
µ(Bi ) +
m
X
i=1
n
X
j=1
1{uj ∈ Ai }
γi µ(Ai ) + (n − 1)
Φr (si , γi )si +
m
X
i=1
m
X
µ(Bi )µ(Ai )
i=1
m
X
si γi + (n − 1)
13
(27)
i=1
Φr (si , γi )s2i
(28)
(29)
Since the data structure succeeds with probability γ, it must be the case that
m
X
i=1
si γi ≥ γ =
[∃i ∈ [m] : v ∈ Bi , uj ∈ Ai ]
Pr
j∼[n],v∼N (uj )
(30)
And since we use at most space O(s),
n
m
X
i=1
si ≤ O(s)
(31)
From Lemma 3.2, for any p, q ∈ [1, ∞) where (p − 1)(q − 1) = σ 2 where σ = 1 − 1c ,
E[T ] ≥
γ≤
O
s
n
≥
m
X
q− pq q
γi
si
i=1
m
X
+ (n − 1)
m
X
q− pq +1 q
γi
si
+γ
(32)
i=1
s i γi
(33)
si
(34)
i=1
m
X
i=1
We set S = {i ∈ [m] : si 6= 0} and for i ∈ S, vi = si γi .
E[T ] ≥
X
viq
i∈S
q
− pq +1
−p
si + (n − 1)si
(35)
X γ q − pq
− pq +1
≥
si + (n − 1)si
|S|
(36)
i∈S
where we used the fact q ≥ 1. Consider
F =
X
i∈S
−q
si p
+ (n −
− q +1
1)si p
(37)
We analyze three cases separately:
• 0 < ρu ≤
•
1
2c−1
1
2c−1
< ρu ≤
• ρu = 0.
2c − 1
(c − 1)2
For the first two cases, we let
q = 1 − σ 2 + σβ
p=
β
β−σ
14
s
β=
1 − σ2
ρu
(38)
Since 0 < ρu ≤
2c − 1
, one can verify β > σ and both p and q are at least 1.
(c − 1)2
Lemma 5.1. When ρu ≤
1
2c−1 ,
and s = n1+ρu ,
E[T ] ≥ Ω(nρq )
where ρq and ρu satisfy Equation 3.
Proof. In this setting, p and q are constants, and q ≥ p. Therefore,
in Equation 37. So we minimize the sum by taking si =
E[T ] ≥ Ω
s
)
O( n|S|
γ q s−q/p+1 nq/p
|S|q−q/p
q
p
≥ 1, so F is convex in all si ’s
and substituting in (36),
!
≥ Ω(γ q s1−q nq/p )
(39)
(40)
since q − q/p > 0 and |S| ≤ s. In addition, p, q and γ are constants, E[T ] ≥ Ω(nρq ) where
ρq = (1 + ρu )(1 − q) +
q
p
(1 − σ 2 + σβ)(β − σ)
= (1 + ρu )(σ 2 − σβ) +
β
p
2
√
=
1 − σ 2 − ρu σ
√
2
2c − 1 √
(c − 1)
=
− ρu ·
c
c
Lemma 5.2. When ρu >
(41)
(42)
(43)
(44)
1
2c−1 ,
E[T ] ≥ Ω(nρq )
where ρq and ρu satisfy Equation 3.
Proof. We follow a similar pattern to Lemma 5.1. However, we may no longer assert that F is
convex in all si ’s.
∂F
=
∂si
q
− p −1
−q
q
q
−
si
+ − + 1 (n − 1)si p
p
p
(45)
q
The gradient is zero when each si =
. Since q < p, this value is positive and
(p − q)(n − 1)
P
m
i∈S si ≤ O n for large enough n. F is continuous, so it is minimized exactly at that point.
− q
q
p
γ
q
So E[T ] ≥ |S|
|S| (p−q)(n−1)
. Again, we maximize |S| to minimize this sum since q ≥ 1.
15
Therefore
− q
γ q
p
q
E[T ] ≥
s
s
(p − q)(n − 1)
(46)
Since p, q and γ are constants, E[T ] ≥ Ω(nρq ) where
ρq = (1 + ρu )(1 − q) +
q
p
which is the same expression for ρq as in Lemma 5.1.
Lemma 5.3. When ρu = 0 (so s = O(n)),
E[T ] ≥ nρq −o(1)
2c − 1
= 1 − σ2.
c2
Proof. In this case, although we cannot set p and q as in Equation 38, we let
where ρq =
q = 1 + σ2 ·
log n
log log n
p=1+
log log n
.
log n
Since q > p, we have
q
E[T ] = Ω(γ q s1−q n p )
=n
(47)
1−σ 2 −o(1)
(48)
giving the desired expression.
6
Tight Lower Bounds for 2 Cell Probe Data Structures
In this section we prove a cell probe lower bound for ANN for t = 2 cell probes as stated in
Theorem 1.3.
As in [PTW10], we will prove lower bounds for GNS when U = V with measure µ (see Def. 2.2).
We assume there is an underlying graph G with vertex set V . For any particular point p ∈ V , its
neighborhood N (p) is the set of points with an edge to p in the graph G.
In the 2-query GNS problem, we have a dataset P = {pi }ni=1 ⊂ V of n points as well as a
bit-string x ∈ {0, 1}n . We let D denote a data structure with m cells of w bits each. We can think
of D as a map [m] → {0, 1}w which holds w bits in each cell. D will depend on the dataset P as
well as the bit-string x. The problem says that: given a query point q ∈ V , if there exists a unique
neighbor pi ∈ N (q) in the dataset, we should return xi with probability at least
cell-probes to D.
16
2
3
after making two
Theorem 6.1. There exists a constant γ > 0 such that any non-adaptive GNS data structure
holding a dataset of n ≥ 1 points which succeeds with probability
2
3
using two cell probes and m cells
of w bits satisfies
1
m log m · 2O(w)
≥ Ω Φr
,γ
.
n
m
Theorem 1.3 will follow from Theorem 6.1 together with the robust expansion bound from
Lemma 3.2 for the special case when probes to the data structure are non-adaptive. For the rest of
this section, we prove Theorem 6.1. We will later show how to reduce adaptive algorithms losing a
sub-polynomial factor in the space for w = o( logloglogn n ) in Section 6.6.3.
At a high-level, we will show that with a “too-good-to-be-true” data structure with small space
we can construct a weaker notion of 2-query locally-decodable code (LDC) with small noise rate
using the same amount of space4 . Even though we our notion of LDC is weaker than Def. 2.4, we
can use most of the tools for showing 2-query LDC lower bounds from [KdW04]. These arguments
use quantum information theory arguments, which are very robust and still work with the 2-query
weak LDC we construct.
We note that [PTW10] was the first to suggest the connection between nearest neighbor search
and locally-decodable codes. This work represents the first concrete connection which gives rise to
better lower bounds.
Proof structure.
The proof of Theorem 6.1 proceeds in six steps.
1. First we will use Yao’s principle to reduce to the case of deterministic non-adaptive data
structures for GNS with two cell-probes. We will give distributions over n-point datasets
P , as well as bit-strings x and a query q. After defining these distributions, we will assume
the existence of a deterministic data structure which makes two cell-probes non-adaptively
and succeeds with probability at least
2
3
when the inputs are sampled according to the three
distributions.
2. We will modify the deterministic data structure in order to get “low-contention” data structures. These are data structures which do not rely on any single cell too much similar to Def.
6.1 in [PTW10]. This will be a simple argument where we increase the space bound by a
constant factor to achieve this guarantee.
3. In the third step, we will take a closer look at how the low-contention data structure probes
the cells. We will use ideas from [PTW10] to understand how queries neighboring particular
dataset points probe various cells of the data structure. We will conclude with finding a
fixed n-point dataset P . A constant fraction of the points in the dataset will satisfy the
4
A 2-query LDC corresponds to LDCs which make two probes to their memory contents. Even though there is a
slight ambiguity with the data structure notion of query, we say “2-query LDCs” in order to be consistent with the
LDC literature.
17
following condition: many queries in the neighborhood of these points probe disjoint pairs of
cells. Intuitively, this means information about these dataset points must be spread out over
various cells.
4. We will show that for the fixed dataset P , we could still recover a constant fraction bits with
significant probability even if we corrupt the contents of some cells. This will be the crucial
connection between nearest neighbor data structures and LDCs.
5. We will reduce to the case of 1-bit words in order to apply the LDC arguments from [KdW04].
We will increase the number of cells by a factor of 2w and decrease the probability of success
from
1
2
+ η to
1
2
+
η
.
22w
6. Finally, we will design an LDC with weaker guarantees and use the arguments in [KdW04]
to prove lower bounds on the space of the weak LDC.
6.1
Deterministic Data Structure
Definition 6.2. A non-adaptive randomized algorithm R for the GNS problem with two cell-probes
is an algorithm specified by the following three components. The data structure preprocesses a
dataset P = {pi }ni=1 consisting of n points, as well as a bit-string x ∈ {0, 1}n , in order to produce
a data structure D : [m] → {0, 1}w which depends on P and x. On a query q, R(q) chooses two
indices (i, j) ∈ [m]2 , and specifies a function fq : {0, 1}w × {0, 1}w → {0, 1}. The output is given
as fq (Dj , Dk ). We require that
Pr [fq (Dj , Dk ) = xi ] ≥
R,D
2
3
whenever q ∈ N (pi ) and pi is the unique such neighbor.
Note that the indices (i, j) which R generates to probe the data structure as well as the function
fq is independent of P and x.
Definition 6.3. We define the following distributions:
• Let P be the distribution over n-point datasets given by sampling n times from our space V
uniformly at random.
• Let X be the uniform distribution over {0, 1}n .
• Let Q(P ) be the distribution over queries given by first picking a dataset point p ∈ P uniformly
at random and then picking q ∈ N (p) uniformly at random.
Lemma 6.4. Assume R is a non-adaptive randomized algorithm for GNS using two cell-probes.
Then there exists a non-adaptive deterministic algorithm A for GNS using two cell-probes which
also produces a data structure D : [m] → {0, 1}w and on query q chooses two indices j, k ∈ [m]
18
(again, independently of P and x) to probe in D as well as a function fq : {0, 1}w ×{0, 1}w → {0, 1}
where
Pr
2
[fq (Dj , Dk ) = xi ] ≥ .
3
P ∼P,x∼X ,q∼Q(P )
Proof. The following is a direct application of Yao’s principle to the success probability of the
algorithm. By assumption, there exists a distribution over algorithms which can achieve probability
of success at least
2
3
for any single query. Therefore, for the fixed distributions P, X , and Q, there
exists a deterministic algorithm achieving at least the same success probability.
In order to simplify notation, for any algorithm A, we let AD (q) denote output of the algorithm.
When we write AD (q), we assume that A(q) outputs a pair of indices (j, k) as well as the function
fq : {0, 1}w × {0, 1}w → {0, 1}, and the algorithm outputs fq (Dj , Dk ). For any fixed dataset
P = {pi }ni=1 and bit-string x ∈ {0, 1}n , we have
Pr [AD (q) = xi ] =
q∼N (pi )
Pr [fq (Dj , Dk ) = xi ]
q∼N (pi )
by definition. This allows us to succinctly state the probability of correctness when the query is a
neighbor of pi without caring about the specific cells the algorithm probes or the function fq the
algorithm uses to make its decision.
The important thing to note is that the contents of the data structure D may depend on the
dataset P and the bit-string x. However, the algorithm A which produces D as well as the indexes
for the probes to D for any query point is deterministic.
From now on, we will assume the existence of a non-adaptive deterministic algorithm A with
success probability at least
2
3
using m cells of width w. The success probability is taken over the
random choice of the dataset P ∼ P, x ∼ X and q ∼ Q(P ).
6.2
Making Low-Contention Data Structures
For any t ∈ {1, 2} and j ∈ [m], let At,j be the set of queries which probe cell j at the t-th probe of
algorithm A. These sets are well defined independently of the dataset P and the bit-string x. In
particular, we could write
At,j = {q ∈ V | A probes cell j in probe t when querying q }
by running the “probing” portion of the algorithm without the need to specify a dataset P or
bit-string x. We could write down At,j by simply trying every query point q and seeing which cells
the algorithm probes.
In other words, since the algorithm is deterministic, the probing portion of algorithm A is
completely specified by two collections A1 = {A1,j }j∈[m] and A2 = {A2,j }j∈[m] as well as the
19
function fq . A1 and A2 are two partitions of the query space V . On query q, if q ∈ At,j , we make
the t-th probe to cell j. We output the value of fq after observing the contents of the cells.
We now define the notion of low-contention data structures, which informally requires the data
structure not rely on any one particular cell too much, namely no At,j is too large.
Definition 6.5. A deterministic non-adaptive algorithm A using m cells has low contention if
every set µ(At,j ) ≤
1
m
for t ∈ {1, 2} and j ∈ [m].
We now use the following lemma to argue that up to a small increase in space, a data structure
can be made low-contention.
Lemma 6.6. Suppose A is a deterministic non-adaptive algorithm for GNS with two cell-probes
using m cells, then there exists an deterministic non-adaptive algorithm A0 for GNS with two cellprobes using 3m cells which succeeds with the same probability and has low contention.
Proof. We first handle A1 and then A2 .
Suppose µ(A1,j ) ≥
1
m,
(j)
then we partition A1,j into enough parts {A1,k }k of size
1
m.
be at most one set with measure between 0 and
For each of part
(j)
A1,k
1
m.
There will
of the partition, we make
(j)
a new cell jk with the same contents as cell j. When a query lies inside A1,k we probe the new cell
jk . From the data structure side, the cell contents are replicated for all additional cells.
The number of cells in this data structure is at most 2m, since there can be at most m cells
of size
1
m
and for each original cell, we have only one cell with small measure. Also, keep in mind
that we have not modified the sets in A2 , and thus there is at most m cells for which µ(A2,j ) ≥
We do the same procedure for the second collection A2 . If some µ(A2,j ) ≥
cell into multiple cells of size exactly
1
m,
1
m,
1
m.
we partition that
with one extra small cell. Again, the total number of cells
will be m for dividing the heavy cells in the second probe, and at most m for the lighter cells in
the second probe.
We have added m cells in having µ(A1,j ) ≤
order to make µ(A2,j ) ≤
1
m
1
m
for all j ∈ [m], and added at most m cells in
for all j ∈ [m]. Therefore, we have at most 3m cells. Additionally, the
contents of the cells remain the same, so the algorithm succeeds with the same probability.
Given Lemma 6.6, we will assume that A is a deterministic non-adaptive algorithm for GNS
with two cell-probes using m cells which has low contention. The extra factor of 3 in the number
of cells will be pushed into the asymptotic notation.
6.3
Datasets which shatter
We fix some γ > 0 which can be thought of as a sufficiently small constant.
20
Definition 6.7 (Weak-shattering [PTW10]). We say a partition A1 , . . . , Am of V (K, γ)-weakly
shatters a point p if
X
1 +
µ(Ai ∩ N (p)) −
≤γ
K
i∈[m]
where the operator (·)+ takes only the non-negative part.
For a fixed dataset point p ∈ P , we refer to γ as the “slack” in the shattering. The slack
corresponds to the total measure which is leftover after we remove an arbitrary subset of At,j ∩N (p)
of measure at least
1
K.
Lemma 6.8 (Shattering [PTW10]). Let A1 , . . . , Ak collection of disjoint subsets of measure at
most
1
m.
Then
Pr [p is (K, γ)-weakly shattered] ≥ 1 − γ
p∼µ
for K = Φr
1 γ2
m, 4
·
γ3
16 .
For the remainder of the section, we let
K = Φr
1 γ2
,
m 4
·
γ3
.
16
We are interested in the shattering of dataset points with respect to the collections A1 and A2 .
The dataset points which get shattered will probe many cells in the data structure. Intuitively, a
bit xi corresponding to a dataset point pi which is weakly-shattered should be stored across various
cells.
So for each point p which is (K, γ) weakly-shattered we define subsets β1 , β2 ⊂ N (p) which hold
the “slack” of the shattering of p with respect to A1 and A2 .
Definition 6.9. Let p ∈ V be a dataset point which is (K, γ)-weakly shattered by A1 and A2 . Let
β1 , β2 ⊂ N (p) be arbitrary subsets where each j ∈ [m] satisfies
µ(A1,j ∩ N (p) \ β1 ) ≤
1
K
µ(A2,j ∩ N (p) \ β2 ) ≤
1
K
and
Since p is (K, γ)-weakly shattered, we can pick β1 and β2 with measure at most γ each. We will
refer to β(p) = β1 ∪ β2 .
For a given collection A, let S(A, p) be the event that the collection A (K, γ)-weakly shatters
p. Note that Lemma 6.8 implies that Prp∼µ [S(A, p)] ≥ 1 − γ.
21
Lemma 6.10. With high probability over the choice of n point dataset, at most 4γn points do not
satisfy S(A1 , p) and S(A2 , p).
Proof. This is a simple Chernoff bound. The expected number of points p which do not satisfy
S(A1 , p) and S(A2 , p) is at most 2γn. Therefore,
the
probability that more than 4γn points do not
2γn
satisfy S(A1 , p) and S(A2 , p) is at most exp − 3 .
We call a dataset good if there are at most 4γn dataset points which are not (K, γ)-weakly
shattered by A1 and A2 .
Lemma 6.11. There exists a good dataset P = {pi }ni=1 where
Pr
x∼X ,q∼Q(P )
[AD (q) = xi ] ≥
2
− o(1)
3
Proof. This follows via a simple argument. For any fixed dataset P = {pi }ni=1 , let
P=
[AD (q) = xi ]
Pr
x∼X ,q∼Q(p)
to simplify notation.
2
≤ E [P]
3 P ∼P
(49)
= (1 − o(1)) · E [P | P is good] + o(1) · E [P | P is not good]
P ∼P
P ∼P
2
− o(1) ≤ (1 − o(1)) · E [P | P is good]
3
P ∼P
(50)
(51)
Therefore, there exists a dataset which is not shattered by at most 4γn and Prx∼X ,q∼Q(P ) [AD (y) =
xi ] ≥
2
3
6.4
Corrupting some cell contents of shattered points
− o(1).
In the rest of the proof, we fix the dataset P = {pi }ni=1 satisfying the conditions of Lemma 6.11,
i.e., such that
Pr
[AD (q) = xi ] ≥
x∼X ,q∼Q(P )
2
− o(1).
3
We now introduce the notion of corruption of the data structure cells D, which parallels the
notion of noise in locally-decodable codes. Remember that, after fixing some bit-string x, the
algorithm A produces some data structure D : [m] → {0, 1}w .
Definition 6.12. We call D0 : [m] → {0, 1}w a corrupted version of D at k cells if they differ on
at most k cells, i.e., if |{i ∈ [m] : D(i) 6= D0 (i)}| ≤ k.
22
In this section, we will show there exist a dataset P of n points and a set S ⊂ [n] of size Ω(n)
with good recovery probability, even if the algorithm has access to a corrupted version of data
structure.
Definition 6.13. For a fixed x ∈ {0, 1}n , let
cx (i) =
Pr [AD (q) = xi ].
q∼N (pi )
Note that from the definitions of Q(P ), Ex∼X ,i∈[n] [cx (i)] ≥
2
3
− o(1).
Lemma 6.14. Fix ε > 0, vector x ∈ {0, 1}n , and let D : [m] → {0, 1}w be the data structure the
algorithm produces on dataset P with bit-string x. Let D0 be a corruption of D at εK cells. For
every i ∈ [n] where events S(A1 , pi ) and S(A2 , pi ) occur, we have
0
Pr [AD (q) = xi ] ≥ cx (i) − 2γ − 2ε.
q∼N (pi )
Proof. Note that cx (i) represents the probability mass of queries in the neighborhood of pi for
which the algorithm returns xi . We want to understand how much of that probability mass we
remove when we avoid probing the corrupted cells.
Since the dataset point pi is (K, γ)-weakly shattered by A1 and A2 , at most 2γ probability mass
of ci (x) will come from the slack of the shattering. In more detail, if q ∼ N (pi ), we have probability
ci (x) that the algorithm returns xi . If we query q ∼ N (pi ) \ β(pi ), in the worst case, every query
q ∈ β(pi ) returns xi ; thus, after removing β(pi ), we have removed at most 2γ probability mass over
queries that the algorithm returns correctly.
The remaining probability mass is distributed across various cells, where each cell has at most
mass for being probing in the first probe, and at most
1
K
1
K
mass for being probe in the second probe.
Therefore, if we remove εK cells, the first or second probe will probe those cells with probability
at most 2ε. If we avoid the εK corrupted cells, the algorithm has the same output as it did with
the uncorrupted data structure D. Therefore, the probability mass which returns xi on query q in
the corrupted data structure D0 is at least cx (i) − 2γ − 2ε.
Lemma 6.15. Fix γ > 0 to be a small enough constant. There exists a set S ⊂ [n] of size
|S| = Ω(n), such that whenever i ∈ S, we have that: events S(A1 , pi ) and S(A2 , pi ) occur, and
E [cx (i)] ≥
x∼X
where ν can be taken to be some small constant like
1
+ ν,
2
1
10 .
Proof. There is at most a 4γ-fraction of the dataset points which are not shattered. For simplifying
the notation, let P = Pri∈[n] [Ex∼X [cx (i)] ≥
1
2
+ ν, S(A1 , pi ) ∧ S(A2 , pi )]. We need to show that
23
P = Ω(1), since we will set S ⊂ [n] as
1
+ ν, S(A1 , pi ) ∧ S(A2 , pi ) .
2
S=
i ∈ [n] | E [cx (i)] ≥
x∼X
The argument is a straight-forward averaging argument.
2
− o(1) ≤
E [cx (i)]
3
x∼X ,i∈[n]
(52)
≤ 1 · 4γ + 1 · P +
1
1
− o(1) − 4γ − ν ≤ P ·
−ν .
6
2
1
+ν
2
· (1 − P)
(53)
(54)
We combine Lemma 6.14 and Lemma 6.15 to obtain the following condition on the dataset.
Lemma 6.16. Fix small enough γ > 0 and ε > 0. There exists a set S ⊂ [n] where |S| = Ω(n),
such that whenever i ∈ S,
1
Pr [A (q) = xi ] ≥ + η
2
q∼N (pi )
E
x∼X
D0
where η = ν − 2γ − 2ε and the algorithm probes a corrupted version of the data structure D.
Proof. Consider the set S ⊂ [n] satisfying the conditions of Lemma 6.15. Whenever i ∈ S, pi
gets (K, γ)-weakly shattered and on average over x, A will recover xi with probability
1
2
+ ν when
probing the data structure D on input q ∼ N (pi ), i.e
E
x∼X
1
Pr [A (q) = xi ] ≥ + ν.
2
q∼N (pi )
D
Therefore, from Lemma 6.14, if A probes D0 which is a corruption of D in any εK cells, A will
recover xi with probability at least
other words,
E
x∼X
1
2
+ ν − 2γ − 2ε averaged over all x ∼ X where q ∼ N (pi ). In
1
Pr [A (q) = xi ] ≥ + ν − 2γ − 2ε.
2
q∼N (pi )
D0
Theorem 6.17. There exists an algorithm A and a subset S ⊆ [n] of size S = Ω(n), where A
makes only 2 cell probes to D. Furthermore, for any corruption of D at εK cells, A can recover xi
with probability at least
1
2
+ η over the random choice of x ∼ X .
Proof. In order to extract xi , we generate a random query q ∼ N (pi ) and we probe the data
structure at the cells assuming the data structure is uncorrupted. From Lemma 6.16, there exists
24
a set S ⊂ [n] of size Ω(n) for which this algorithm recovers xi with probability at least
the probability is taken on average over all possible x ∈
1
2
+ η, where
{0, 1}n .
We fix the algorithm A and subset S ⊂ [n] satisfying the conditions of Theorem 6.17. Since we
fixed the dataset P = {pi }ni=1 satisfying the conditions of Lemma 6.11, we say that x ∈ {0, 1}n is
an input to algorithm A in order to initialize the data structure with dataset P = {pi }ni=1 and xi
is the bit associated with pi .
6.5
Decreasing the word size
We now reduce to the case when the word size is w = 1 bit.
Lemma 6.18. There exists a deterministic non-adaptive algorithm A0 which on input x ∈ {0, 1}n
builds a data structure D0 using m2w cells of width 1 bit. Any i ∈ S as well as any corruption C
to D0 in at most εK positions satisfies
E
x∈{0,1}n
η
1
Pr [A (q) = xi ] ≥ + 2w
2
2
q∼N (pi )
0C
Proof. Given algorithm A which constructs the data structure D : [m] → {0, 1}w on input x ∈
{0, 1}n , construct the following data structure D0 : [m · 2w ] → {0, 1}. For each cell Dj ∈ {0, 1}w ,
make 2w cells which contain all the parities of the w bits in Dj . This blows up the size of the data
structure by 2w .
Fix i ∈ S and q ∈ N (pi ) if algorithm A produces a function fq : {0, 1}w × {0, 1}w → {0, 1}
which succeeds with probability at least
1
2
+ ζ over x ∈ {0, 1}n , then there exists a signed parity on
some input bits which equals fq in at least
1
2
+
ζ
22w
inputs x ∈ {0, 1}n . Let Sj be the parity of the
bits of cell j and Sk be the parity of the bits of cell k. Let fq0 : {0, 1} × {0, 1} → {0, 1} denote the
parity or the negation of the parity which equals fq on
1
2
+
ζ
22w
possible input strings x ∈ {0, 1}n .
Algorithm A0 will evaluate fq0 at the cell containing the parity of the Sj bits in cell j and the
parity of Sk bits in cell k. Let ISj , ISk ∈ [m · 2w ] be the indices of these cells. Since we can find
such function for each fixed q ∈ N (pi ), any two cell probes to j, k ∈ [m], and any corrupted version
of D, the algorithm A0 satisfies
E
x∈{0,1}n
Pr [fq0 (CI0 S , CI0 S )
j
k
q∼N (pi )
= xi ] ≥
1
η
+
2 22w
whenever i ∈ S.
For the remainder of the section, we will prove a version of Theorem 6.1 for algorithms with
1-bit words. Given Lemma 6.18, we will modify the space to m · 2w and the probability to
1
2
+
η
22w
to obtain the answer. So for the remainder of the section, assume algorithm A has 1 bit words.
25
6.6
Connecting to Locally-Decodable Codes
To complete the proof of Theorem 6.1, it remains to prove the following lemma.
Lemma 6.19. Let A be a non-adaptive deterministic algorithm which makes 2 cell probes to a
data structure D of m cells of width 1 bit which can handle εK corruptions and recover xi with
probability
1
2
+ η on random input x ∈ {0, 1}n whenever i ∈ S for some fixed S of size Ω(n). Then
the following must hold
m log m
≥ Ω εKη 2 .
n
The proof of the lemma uses [KdW04] and relies heavily on notions from quantum computing,
in particular quantum information theory as applied to LDC lower bounds.
6.6.1
Crash Course in Quantum Computing
We introduce a few concepts from quantum computing that are necessary in our subsequent arguments. A qubit is a unit-length vector in C2 . We write a qubit as a linear combination of the basis
states (10 ) = |0i and (01 ) = |1i. The qubit α = (αα12 ) can be written
|αi = α1 |0i + α2 |1i
where we refer to α1 and α2 as amplitudes and |α1 |2 + |α2 |2 = 1. An m-qubit system is a vector in
the tensor product C2 ⊗ · · · ⊗ C2 of dimension 2m . The basis states correspond to all 2m bit-strings
of length m. For j ∈ [2m ], we write |ji as the basis state |j1 i ⊗ |j2 i ⊗ · · · ⊗ |jm i where j = j1 j2 . . . jm
is the binary representation of j. We will write the m-qubit quantum state |φi as unit-vector
P
given by linear combination over all 2m basis states. So |φi = j∈[2m ] φj |ji. As a shorthand, hφ|
corresponds to the conjugate transpose of a quantum state.
A mixed state {pi , |φi i} is a probability distribution over quantum states. In this case, we the
quantum system is in state |φi i with probability pi . We represent mixed states by a density matrix
P
pi |φi i hφi |.
A measurement is given by a family of positive semi-definite operators which sum to the identity
operator. Given a quantum state |φi and a measurement corresponding to the family of opera-
tors {Mi∗ Mi }i , the measurement yields outcome i with probability kMi |φi k2 and results in state
Mi |φi
,
kMi |φik2
where the norm k · k is the `2 norm. We say the measurement makes the observation Mi .
Finally, a quantum algorithm makes a query to some bit-string y ∈ {0, 1}m by starting with the
state |ci |ji and returning (−1)c·yj |ci |ji. One can think of c as the control qubit taking values 0 or
1; if c = 0, the state remains unchanged by the query, and if c = 1 the state receives a (−1)yj in its
P
amplitude. The queries may be made in superposition to a state, so the state c∈{0,1},j∈[m] αcj |ci |ji
P
becomes c∈{0,1},j∈[m] (−1)c·yj αcj |ci |ji.
26
6.6.2
Weak quantum random access codes from GNS algorithms
Definition 6.20. C : {0, 1}n → {0, 1}m is a (2, δ, η)-LDC if there exists a randomized decoding
algorithm making at most 2 queries to an m-bit string y non-adaptively, and for all x ∈ {0, 1}n ,
i ∈ [n], and y ∈ {0, 1}m where d(y, C(x)) ≤ δm, the algorithm can recover xi from the two queries
to y with probability at least
1
2
+ η.
In their paper, [KdW04] prove the following result about 2-query LDCs.
Theorem 6.21 (Theorem 4 in [KdW04]). If C : {0, 1}n → {0, 1}m is a (2, δ, η)-LDC, then m ≥
2Ω(δη
2 n)
.
The proof of Theorem 6.21 proceeds as follows. They show how to construct a 1-query quantumLDC from a classical 2-query LDC. From a 1-query quantum-LDC, [KdW04] constructs a quantum
random access code which encodes n-bit strings in O(log m) qubits. Then they apply a quantum
information theory lower bound due to Nayak [Nay99]:
Theorem 6.22 (Theorem 2 stated in [KdW04] from Nayak [Nay99]). For any encoding x → ρx
of n-bit strings into m-qubit states, such that a quantum algorithm, given query access to ρx , can
decode any fixed xi with probability at least 1/2 + η, it must hold that m ≥ (1 − H(1/2 + η))n.
Our proof will follow a pattern similar to the proof of Theorem 6.21. We assume the existence
of a GNS algorithm A which builds a data structure D : [m] → {0, 1}. We can think of D as a
length m binary string encoding x; in particular let Dj ∈ {0, 1} be the jth bit of D.
Our algorithm A from Theorem 6.17 does not satisfy the strong properties of an LDC, preventing
us from applying 6.21 directly. However, it does have some LDC-ish guarantees. In particular, we
can support εK corruptions to D. In the LDC language, this means that we can tolerate a noise
rate of δ =
εK
m .
Additionally, we cannot necessarily recover every coordinate xi , but we can recover
xi for i ∈ S, where |S| = Ω(n). Also, our success probability is
of i ∈ S and the random choice of the bit-string x ∈
arguments of [KdW04] to this weaker setting.
{0, 1}n .
1
2
+ η over the random choice
Our proof follows by adapting the
εK
and a ≤ 1 is a constant. Let D be the data structure
m
from above (i.e., satisfying the hypothesis of Lemma 6.19). Then there exists a quantum algorithm
Lemma 6.23. Let r =
2
δa2
where δ =
that, starting from the r(log m + 1)-qubit state with r copies of |U (x)i, where
|U (x)i = √
X
1
(−1)c·Dj |ci |ji
2m c∈{0,1},j∈[m]
can recover xi for any i ∈ S with probability
1
2
+ Ω(η) (over a random choice of x).
Assuming Lemma 6.23, we can complete the proof of Lemma 6.19.
27
Proof of Lemma 6.19. The proof is similar to the proof of Theorem 2 of [KdW04]. Let ρx represent
the s-qubit system consisting of the r copies of the state |U (x)i, where s = r(log m + 1); ρx is an
encoding of x. Using Lemma 6.23, we can assume we have a quantum algorithm that, given ρx ,
can recover xi for any i ∈ S with probability α =
1
2
+ Ω(η) over the random choice of x ∈ {0, 1}n .
We will let H(A) be the Von Neumann entropy of A, and H(A|B) be the conditional entropy
and H(A : B) the mutual information.
Let XM be the (n + s)-qubit system
1
2n
X
x∈{0,1}n
|xi hx| ⊗ ρx .
The system corresponds to the uniform superposition of all 2n strings concatenated with their
encoding ρx . Let X be the first subsystem corresponding to the first n qubits and M be the second
subsystem corresponding to the s qubits. We have
H(XM ) = n +
1
2n
X
x∈{0,1}n
H(ρx ) ≥ n = H(X)
H(M ) ≤ s,
(55)
(56)
since M has s qubits. Therefore, the mutual information H(X : M ) = H(X)+H(M )−H(XM ) ≤ s.
P
Note that H(X|M ) ≤ ni=1 H(Xi |M ). By Fano’s inequality, if i ∈ S,
H(Xi |M ) ≤ H(α)
where we are using the fact that Fano’s inequality works even if we can recover xi with probability
α averaged over all x’s. Additionally, if i ∈
/ S, H(Xi |M ) ≤ 1. Therefore,
s ≥ H(X : M ) = H(X) − H(X|M )
≥ H(X) −
n
X
i=1
H(Xi |M )
(57)
(58)
≥ n − |S|H(α) − (n − |S|)
(59)
= |S|(1 − H(α)).
(60)
Furthermore, 1 − H(α) ≥ Ω(η 2 ) since, and |S| = Ω(n), we have
2m
(log m + 1) ≥ Ω nη 2
2
a εK
m log m
≥ Ω εKη 2 .
n
28
(61)
(62)
It remains to prove Lemma 6.23, which we proceed to do in the rest of the section. We first
show that we can simulate our GNS algorithm with a 1-query quantum algorithm.
Lemma 6.24. Fix an x ∈ {0, 1}n and i ∈ [n]. Let D : [m] → {0, 1} be the data structure produced
by algorithm A on input x. Suppose Prq∼N (pi ) [AD (q) = xi ] =
1
2
+ b for b > 0. Then there exists a
quantum algorithm which makes one quantum query (to D) and succeeds with probability
1
2
+
4b
7
to
output xi .
Proof. We use the procedure in Lemma 1 of [KdW04] to determine the output algorithm A on
input x at index i. The procedure simulates two classical queries with one quantum query.
All quantum algorithms which make 1-query to D can be specified in the following manner:
there is a quantum state |Qi i, where
|Qi i =
X
c∈{0,1},j∈[m]
αcj |ci |ji
which queries D. After querying D, the resulting quantum state is |Qi (x)i, where
|Qi (x)i =
X
c∈{0,1},j∈[m]
(−1)c·Dj αcj |ci |ji .
There is also a quantum measurement {R, I − R} such that, after the algorithm obtains the state
|Qi (x)i, it performs the measurement {R, I − R}. If the algorithm observes R, it outputs 1 and if
the algorithm observes I − R, it outputs 0.
From Lemma 6.24, we know there must exist a state |Qi i and {R, I − R} where if algorithm A
succeeds with probability
probability
1
2
+
4η
7
1
2
+ η on random x ∼ {0, 1}n , then the quantum algorithm succeeds with
on random x ∼ {0, 1}n .
In order to simplify notation, we write p(φ) as the probability of making observation R from
state |φi. Since R is a positive semi-definite matrix, R = M ∗ M and so p(φ) = kM |φi k2 .
In exactly the same way as [KdW04], we can remove parts of the quantum state |Qi (x)i where
αcj >
√1
δm
=
√1 .
εK
If we let L = {(c, j) | αcj ≤
obtain the quantum state
1
a
|Ai (x)i, where
|Ai (x)i =
X
√1 },
εK
(−1)c·Dj αcj |ci |ji
after keeping only the amplitudes in L, we
a=
(c,j)∈L
s X
2
αcj
(c,j)∈L
Lemma 6.25. Fix i ∈ S. The quantum state |Ai (x)i satisfies
1
1
8η
p
A
(x)
|
x
=
1
−
p
A
(x)
|
x
=
0
≥ 2.
E
E
i
i
i
i
n
n
a
a
7a
x∈{0,1}
x∈{0,1}
29
Proof. Note that since |Qi (x)i and {R, I − R} simulate A and succeed with probability at least
1
2
+
4η
7
on a random x ∈ {0, 1}n , we have that
1
1
1 4η
E [p (Qi (x)) | xi = 1] +
E [1 − p (Qi (x)) | xi = 0] ≥ + ,
2 x∈{0,1}n
2 x∈{0,1}n
2
7
(63)
which we can simplify to say
E
x∈{0,1}n
[p (Qi (x)) | xi = 1] +
E
x∈{0,1}n
[p (Qi (x)) | xi = 0] ≥
8η
.
7
(64)
Since |Qi (x)i = |Ai (x)i + |Bi (x)i and |Bi (x)i contains at most εK parts, if all probes to D
in |Bi (x)i had corrupted values, the algorithm should still succeed with the same probability on
random inputs x. Therefore, the following two inequalities hold:
8η
7
x∈{0,1}n
x∈{0,1}n
8η
E [p (Ai (x) − B(x)) | xi = 1] + E [p (Ai (x) − B(x)) | xi = 0] ≥
n
n
7
x∈{0,1}
x∈{0,1}
E
[p (Ai (x) + B(x)) | xi = 1] +
[p (Ai (x) + B(x)) | xi = 0] ≥
E
Note that p(φ ± ψ) = p(φ) + p(ψ) ± (hφ| R |ψi + hψ| D |φi) and p( 1c φ) =
p(φ)
.
c2
(65)
(66)
One can verify by
averaging the two inequalities (65) and (66) that we get the desired expression.
Lemma 6.26. Fix i ∈ S. There exists a quantum algorithm that starting from the quantum state
1
a
|Ai (x)i, can recover the value of xi with probability
1
2
+
2η
7a2
over random x ∈ {0, 1}n .
Proof. The algorithm and argument are almost identical to Theorem 3 in [KdW04], we just check
that it works under the weaker assumptions. Let
1
q1 =
p
Ai (x) | xi = 1
E
a
x∈{0,1}n
From Lemma 6.25, we know q1 − q0 ≥
8η
.
7a2
on random x ∈ {0, 1}n . Assume q1 ≥
1
2
1
q0 =
p
Ai (x) | xi = 0 .
E
a
x∈{0,1}n
In order to simplify notation, let b =
a quantum algorithm which starting from state
1
a
4η
.
7a2
So we want
|Ai (x)i can recover xi with probability
+ b, since otherwise q0 ≤
1
2
1
2
+
b
2
− b and the same argument
will work for 0 and 1 flipped. Also, assume q1 + q0 ≥ 1, since otherwise simply outputting 1 on
observation R and 0 on observation I − R will work.
1
The algorithm works in the following way: it outputs 0 with probability 1 − q1 +q
and otherwise
0
makes the measurement {R, I − R} on state
1
a
|Ai (x)i. If the observation made is R, then the
algorithm outputs 1, otherwise, it outputs 0. The probability of success over random input x ∈
30
{0, 1}n is
E
x∈{0,1}n
[Pr[returns correctly]]
=
1
1
E [Pr[returns 1] | xi = 1] +
E [Pr[returns 0] | xi = 0] . (67)
2 x∈{0,1}n
2 x∈{0,1}n
When xi = 1, the probability the algorithm returns correctly is (1 − q)p
the probability the algorithm returns correctly is q + (1 − q)(1 −
1
a Ai (x)
p( a1 Ai (x))).
and when xi = 0,
So simplifying (67),
1
1
[Pr[returns correctly]] = (1 − q)q1 + (q + (1 − q)(1 − q0 ))
2
2
x∈{0,1}n
1 b
≥ + .
2 2
(68)
E
(69)
Now we can finally complete the proof of Lemma 6.23.
Proof of Lemma 6.23. Again, the proof is exactly the same as the finishing arguments of Theorem 3
in [KdW04], and we simply check the weaker conditions give the desired outcome. On input i ∈ [n]
and access to r copies of the state |U (x)i, the algorithm applies the measurement {Mi∗ Mi , I−Mi∗ Mi }
where
Mi =
√
εK
X
(c,j)∈L
αcj |c, ji hc, j| .
This measurement is designed in order to yield the state
makes the observation
{Mi∗ Mi , I
−
Mi∗ Mi }
Mi∗ Mi .
1
a
|Ai (x)i on |U (x)i if the measurement
The fact that the amplitudes of |Ai (x)i are not too large makes
a valid measurement.
The probability of observing Mi∗ Mi is hU (x)| Mi∗ Mi |U (x)i =
δa2
2 ,
where we used that δ =
εK
m .
So the algorithm repeatedly applies the measurement until observing outcome Mi∗ Mi . If it never
makes the observation, the algorithm outputs 0 or 1 uniformly at random. If the algorithm does
observe Mi∗ Mi , it runs the output of the algorithm of Lemma 6.26. The following simple calculation
(done in [KdW04]) gives the desired probability of success on random input,
E
x∈{0,1}n
2
r
[Pr[returns correctly]] ≥ 1 − (1 − δa /2)
≥
1
η
+ 2.
2 7a
31
1
2η
+ 2
2 7a
+ (1 − δa2 /2)r ·
1
2
(70)
(71)
6.6.3
On adaptivity
We can extend our lower bounds from the non-adaptive to the adaptive setting.
Lemma 6.27. If there exists a deterministic data structure which makes two queries adaptively
and succeeds with probability at least
1
2
+ η, there exists a deterministic data structure which makes
the two queries non-adaptively and succeeds with probability at least
1
2
+
η
2w .
Proof. The algorithm guesses the outcome of the first cell probe and simulates the adaptive algorithm with the guess. After knowing which two probes to make, we probe the data structure
non-adaptively. If the algorithm guessed the contents of the first cell-probe correctly, then we output the value of the non-adaptive algorithm. Otherwise, we output a random value. This algorithm
is non-adaptive and succeeds with probability at least 1 − 21w · 12 + 21w 12 + η = 12 + 2ηw .
Applying this theorem, from an adaptive algorithm succeeding with probability 23 , we obtain
1
2
a non-adaptive algorithm which succeeds with probability
the intended
2
3,
+ Ω(2−w ). This value is lower than
but we the reduction to a weak LDC still goes through when let γ = Θ(2−w ),
ε = Θ(2−w ). Another consequence is that |S| = Ω(2−w n).
One can easily verify that for small enough γ = Ω(2−w ),
m log m · 2Θ(w)
1
≥ Ω Φr
,γ
n
m
Which yields tight lower bounds (up to sub-polynomial factors) for the Hamming space when
w = o(log n).
In the case of the Hamming space, we can compute robust expansion in a similar fashion to
Theorem 1.2. In particular, for any p, q ∈ [1, ∞) where (p − 1)(q − 1) = σ 2 , we have
m log m · 2O(w)
≥ Ω(γ q m1+q/p−q )
n
mq−q/p+o(1) ≥ n1−o(1) γ q
m≥n
=n
Let p = 1 +
n → ∞.
wf (n)
log n
1−o(1)
q−q/p+o(1)
p
−o(1)
pq−q
γ
γ
q
q−q/p+o(1)
p
−o(1)
p−1
(72)
(73)
(74)
(75)
log n
and q = 1 + σ 2 wf
(n) where we require that wf (n) = o(log n) and f (n) → ∞ as
log n
1
m ≥ n σ2 −o(1) 2 log log n
≥n
1
−o(1)
σ2
32
(76)
(77)
7
Acknowledgments
We would like to thank Jop Briët for helping us to navigate literature about LDCs. We thank Omri
Weinstein for useful discussions.
References
[AC09]
Nir Ailon and Bernard Chazelle. The fast Johnson–Lindenstrauss transform and approximate
nearest neighbors. SIAM J. Comput., 39(1):302–322, 2009.
[ACP08]
Alexandr Andoni, Dorian Croitoru, and Mihai Pǎtraşcu. Hardness of nearest neighbor under
L-infinity. In Proceedings of the Symposium on Foundations of Computer Science (FOCS),
pages 424–433, 2008.
[ADI+ 06]
Alexandr Andoni, Mayur Datar, Nicole Immorlica, Piotr Indyk, and Vahab Mirrokni.
Locality-sensitive hashing scheme based on p-stable distributions. Nearest Neighbor Methods
for Learning and Vision: Theory and Practice, Neural Processing Information Series, MIT
Press, 2006.
[AI06]
Alexandr Andoni and Piotr Indyk. Near-optimal hashing algorithms for approximate nearest
neighbor in high dimensions. In Proceedings of the Symposium on Foundations of Computer
Science (FOCS), pages 459–468, 2006.
[AI08]
Alexandr Andoni and Piotr Indyk. Near-optimal hashing algorithms for approximate nearest
neighbor in high dimensions. Communications of the ACM, 51(1):117–122, 2008.
[AIL+ 15]
Alexandr Andoni, Piotr Indyk, Thijs Laarhoven, Ilya Razenshteyn, and Ludwig Schmidt.
Practical and optimal LSH for angular distance. In NIPS, 2015. Full version available at
http://arxiv.org/abs/1509.02897.
[AINR14]
Alexandr Andoni, Piotr Indyk, Huy L. Nguyen, and Ilya Razenshteyn. Beyond
locality-sensitive hashing. In Proceedings of the ACM-SIAM Symposium on Discrete
Algorithms (SODA), 2014. Full version at http://arxiv.org/abs/1306.1547.
[AIP06]
Alexandr Andoni, Piotr Indyk, and Mihai Pǎtraşcu. On the optimality of the dimensionality
reduction method. In Proceedings of the Symposium on Foundations of Computer Science
(FOCS), pages 449–458, 2006.
[And09]
Alexandr Andoni. Nearest Neighbor Search: the Old, the New, and the Impossible. PhD thesis,
MIT, 2009. Available at http://www.mit.edu/~andoni/thesis/main.pdf.
[AR15]
Alexandr Andoni and Ilya Razenshteyn. Optimal data-dependent hashing for approximate
near neighbors. In Proceedings of the Symposium on Theory of Computing (STOC), 2015. Full
version at http://arxiv.org/abs/1501.01062.
[AR16]
Alexandr Andoni and Ilya Razenshteyn. Tight lower bounds for data-dependent
locality-sensitive hashing. In Proceedings of the ACM Symposium on Computational Geometry
(SoCG), 2016. Available at http://arxiv.org/abs/1507.04299.
[AV15]
Amirali Abdullah and Suresh Venkatasubramanian. A directed isoperimetric inequality with
application to bregman near neighbor lower bounds. In Proceedings of the Forty-Seventh
Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA, June
14-17, 2015, pages 509–518, 2015.
[AW15]
Josh Alman and Ryan Williams. Probabilistic polynomials and hamming nearest neighbors. In
Proceedings of the Symposium on Foundations of Computer Science (FOCS), 2015.
33
[BDGL16] Anja Becker, Léo Ducas, Nicolas Gama, and Thijs Laarhoven. New directions in nearest
neighbor searching with applications to lattice sieving. In Proceedings of the ACM-SIAM
Symposium on Discrete Algorithms (SODA), 2016.
[BOR99]
Allan Borodin, Rafail Ostrovsky, and Yuval Rabani. Lower bounds for high dimensional
nearest neighbor search and related problems. Proceedings of the Symposium on Theory of
Computing, 1999.
[BR02]
Omer Barkol and Yuval Rabani. Tighter bounds for nearest neighbor search and related
problems in the cell probe model. J. Comput. Syst. Sci., 64(4):873–896, 2002. Previously
appeared in STOC’00.
[BRdW08] Avraham Ben-Aroya, Oded Regev, and Ronald de Wolf. A hypercontractive inequality for
matrix-valued functions with applications to quantum computing and ldcs. In 49th Annual
IEEE Symposium on Foundations of Computer Science, FOCS 2008, October 25-28, 2008,
Philadelphia, PA, USA, pages 477–486, 2008.
[CCGL99] Amit Chakrabarti, Bernard Chazelle, Benjamin Gum, and Alexey Lvov. A lower bound on the
complexity of approximate nearest-neighbor searching on the Hamming cube. Proceedings of
the Symposium on Theory of Computing (STOC), 1999.
[Cha02]
Moses Charikar. Similarity estimation techniques from rounding. In Proceedings of the
Symposium on Theory of Computing (STOC), pages 380–388, 2002.
[Cla88]
Ken Clarkson. A randomized algorithm for closest-point queries. SIAM Journal on
Computing, 17:830–847, 1988.
[CR04]
Amit Chakrabarti and Oded Regev. An optimal randomised cell probe lower bounds for
approximate nearest neighbor searching. Proceedings of the Symposium on Foundations of
Computer Science (FOCS), 2004.
[DG03]
Sanjoy Dasgupta and Anupam Gupta. An elementary proof of a theorem of Johnson and
Lindenstrauss. Random Structures Algorithms, 22(1):60–65, 2003.
[DG15]
Zeev Dvir and Sivakanth Gopi. 2-server PIR with sub-polynomial communication. In
Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC
2015, Portland, OR, USA, June 14-17, 2015, pages 577–584, 2015.
[DIIM04]
Mayur Datar, Nicole Immorlica, Piotr Indyk, and Vahab Mirrokni. Locality-sensitive hashing
scheme based on p-stable distributions. In Proceedings of the ACM Symposium on
Computational Geometry (SoCG), 2004.
[DRT11]
Inderjit S. Dhillon, Pradeep Ravikumar, and Ambuj Tewari. Nearest neighbor based greedy
coordinate descent. In Advances in Neural Information Processing Systems 24: 25th Annual
Conference on Neural Information Processing Systems 2011. Proceedings of a meeting held
12-14 December 2011, Granada, Spain., pages 2160–2168, 2011.
[GIM99]
Aristides Gionis, Piotr Indyk, and Rajeev Motwani. Similarity search in high dimensions via
hashing. Proceedings of the 25th International Conference on Very Large Data Bases (VLDB),
1999.
[HIM12]
Sariel Har-Peled, Piotr Indyk, and Rajeev Motwani. Approximate nearest neighbor: Towards
removing the curse of dimensionality. Theory of Computing, 1(8):321–350, 2012.
[HLM15]
Thomas Hofmann, Aurélien Lucchi, and Brian McWilliams. Neighborhood watch: Stochastic
gradient descent with neighbors. CoRR, abs/1506.03662, 2015.
[IM98]
Piotr Indyk and Rajeev Motwani. Approximate nearest neighbor: towards removing the curse
of dimensionality. Proceedings of the Symposium on Theory of Computing (STOC), pages
604–613, 1998.
34
[Ind01a]
Piotr Indyk. High-dimensional computational geometry. Ph.D. Thesis. Department of
Computer Science, Stanford University, 2001.
[Ind01b]
Piotr Indyk. On approximate nearest neighbors in `∞ norm. J. Comput. Syst. Sci.,
63(4):627–638, 2001. Preliminary version appeared in FOCS’98.
[JKKR04] T. S. Jayram, Subhash Khot, Ravi Kumar, and Yuval Rabani. Cell-probe lower bounds for the
partial match problem. Journal of Computer and Systems Sciences, 69(3):435–447, 2004. See
also STOC’03.
[JL84]
William B. Johnson and Joram Lindenstrauss. Extensions of lipshitz mapping into hilbert
space. Contemporary Mathematics, 26:189–206, 1984.
[Kap15]
Michael Kapralov. Smooth tradeoffs between insert and query complexity in nearest neighbor
search. In Proceedings of the ACM Symposium on Principles of Database Systems (PODS),
pages 329–342, New York, NY, USA, 2015. ACM.
[KdW04]
Iordanis Kerenidis and Ronald de Wolf. Exponential lower bound for 2-query locally decodable
codes via a quantum argument. Journal of Computer and System Sciences, 69(3):395–420,
2004.
[KKK16]
Matti Karppa, Petteri Kaski, and Jukka Kohonen. A faster subquadratic algorithm for finding
outlier correlations. In Proceedings of the ACM-SIAM Symposium on Discrete Algorithms
(SODA), 2016. Available at http://arxiv.org/abs/1510.03895.
[KKKÓ16] Matti Karppa, Petteri Kaski, Jukka Kohonen, and Padraig Ó Catháin. Explicit correlation
amplifiers for finding outlier correlations in deterministic subquadratic time. In Proceedings of
the 24th European Symposium Of Algorithms (ESA ’2016), 2016. To appear.
[KOR00]
Eyal Kushilevitz, Rafail Ostrovsky, and Yuval Rabani. Efficient search for approximate nearest
neighbor in high dimensional spaces. SIAM J. Comput., 30(2):457–474, 2000. Preliminary
version appeared in STOC’98.
[KP12]
Michael Kapralov and Rina Panigrahy. NNS lower bounds via metric expansion for `∞ and
EMD. In Proceedings of International Colloquium on Automata, Languages and Programming
(ICALP), pages 545–556, 2012.
[Laa15a]
Thijs Laarhoven. Search problems in cryptography: From fingerprinting to lattice sieving. PhD
thesis, Eindhoven University of Technology, 2015.
[Laa15b]
Thijs Laarhoven. Sieving for shortest vectors in lattices using angular locality-sensitive
hashing. In Advances in Cryptology - CRYPTO 2015 - 35th Annual Cryptology Conference,
Santa Barbara, CA, USA, August 16-20, 2015, Proceedings, Part I, pages 3–22, 2015.
[Laa15c]
Thijs Laarhoven. Tradeoffs for nearest neighbors on the sphere. CoRR, abs/1511.07527, 2015.
[Liu04]
Ding Liu. A strong lower bound for approximate nearest neighbor searching in the cell probe
model. Information Processing Letters, 92:23–29, 2004.
[LJW+ 07] Qin Lv, William Josephson, Zhe Wang, Moses Charikar, and Kai Li. Multi-probe LSH:
efficient indexing for high-dimensional similarity search. In VLDB, 2007.
[LPY16]
Mingmou Liu, Xiaoyin Pan, and Yitong Yin. Randomized approximate nearest neighbor
search with limited adaptivity. CoRR, abs/1602.04421, 2016.
[Mei93]
Stefan Meiser. Point location in arrangements of hyperplanes. Information and Computation,
106:286–303, 1993.
[Mil99]
Peter Bro Miltersen. Cell probe complexity-a survey. Proceedings of the 19th Conference on
the Foundations of Software Technology and Theoretical Computer Science, Advances in Data
Structures Workshop, page 2, 1999.
35
[MNP07]
Rajeev Motwani, Assaf Naor, and Rina Panigrahy. Lower bounds on locality sensitive hashing.
SIAM Journal on Discrete Mathematics, 21(4):930–935, 2007. Previously in SoCG’06.
[MNSW98] Peter B. Miltersen, Noam Nisan, Shmuel Safra, and Avi Wigderson. Data structures and
asymmetric communication complexity. Journal of Computer and System Sciences, 1998.
[Nay99]
Ashwin Nayak. Optimal lower bounds for quantum automata and random access codes. In
Foundations of Computer Science, 1999. 40th Annual Symposium on, pages 369–376. IEEE,
1999.
[Ngu14]
Huy L. Nguyên. Algorithms for High Dimensional Data. PhD thesis, Princeton University,
2014. Available at http://arks.princeton.edu/ark:/88435/dsp01b8515q61f.
[O’D14]
Ryan O’Donnell. Analysis of boolean functions. Cambridge University Press, 2014.
[OWZ14]
Ryan O’Donnell, Yi Wu, and Yuan Zhou. Optimal lower bounds for locality sensitive hashing
(except when q is tiny). Transactions on Computation Theory, 6(1):5, 2014. Previously in
ICS’11.
[Pag16]
Rasmus Pagh. Locality-sensitive hashing without false negatives. In Proceedings of the
ACM-SIAM Symposium on Discrete Algorithms (SODA), 2016. Available at
http://arxiv.org/abs/1507.03225.
[Pan06]
Rina Panigrahy. Entropy-based nearest neighbor algorithm in high dimensions. In Proceedings
of the ACM-SIAM Symposium on Discrete Algorithms (SODA), 2006.
[Pǎt11]
Mihai Pǎtraşcu. Unifying the landscape of cell-probe lower bounds. SIAM Journal on
Computing, 40(3):827–847, 2011. See also FOCS’08, arXiv:1010.3783.
[PP16]
Ninh Pham and Rasmus Pagh. Scalability and total recall with fast CoveringLSH. CoRR,
abs/1602.02620, 2016.
[PT06]
Mihai Pǎtraşcu and Mikkel Thorup. Higher lower bounds for near-neighbor and further rich
problems. Proceedings of the Symposium on Foundations of Computer Science (FOCS), 2006.
[PTW08]
Rina Panigrahy, Kunal Talwar, and Udi Wieder. A geometric approach to lower bounds for
approximate near-neighbor search and partial match. In Proceedings of the Symposium on
Foundations of Computer Science (FOCS), pages 414–423, 2008.
[PTW10]
Rina Panigrahy, Kunal Talwar, and Udi Wieder. Lower bounds on near neighbor search via
metric expansion. In Proceedings of the Symposium on Foundations of Computer Science
(FOCS), pages 805–814, 2010.
[Raz14]
Ilya Razenshteyn. Beyond Locality-Sensitive Hashing. Master’s thesis, MIT, 2014.
[SDI06]
Gregory Shakhnarovich, Trevor Darrell, and Piotr Indyk, editors. Nearest Neighbor Methods in
Learning and Vision. Neural Processing Information Series, MIT Press, 2006.
[TT07]
Tengo Terasawa and Yuzuru Tanaka. Spherical LSH for approximate nearest neighbor search
on unit hypersphere. Workshop on Algorithms and Data Structures, 2007.
[Val88]
Leslie G Valiant. Functionality in neural nets. In First Workshop on Computational Learning
Theory, pages 28–39, 1988.
[Val15]
Gregory Valiant. Finding correlations in subquadratic time, with applications to learning
parities and the closest pair problem. J. ACM, 62(2):13, 2015. Previously in FOCS’12.
[WLKC15] Jun Wang, Wei Liu, Sanjiv Kumar, and Shih-Fu Chang. Learning to hash for indexing big
data — a survey. Available at http://arxiv.org/abs/1509.05472, 2015.
[WSSJ14]
Jingdong Wang, Heng Tao Shen, Jingkuan Song, and Jianqiu Ji. Hashing for similarity search:
A survey. CoRR, abs/1408.2927, 2014.
36
[Yin16]
Yitong Yin. Simple average-case lower bounds for approximate near-neighbor from
isoperimetric inequalities. CoRR, abs/1602.05391, 2016.
[ZYS16]
Zeyuan Allen Zhu, Yang Yuan, and Karthik Sridharan. Exploiting the structure: Stochastic
gradient methods using raw clusters. CoRR, abs/1602.02151, 2016.
A
Random instances for `2
We first introduce the equivalent notion of the “random instance” (from Section 1.3) for `2 . This
instance is what lies at the core of the optimal data-dependent LSH from [AR15].
• All points and queries lie on a unit sphere S d−1 ⊂ Rd .
• The dataset P is generated by sampling n unit vectors in S d−1 independently and uniformly
at random.
• A query q is generated by first choosing a dataset point p ∈ P uniformly at
random, and then
√
choosing q uniformly at random from all points in S d−1 within distance
2
c
from p.
• The goal of the data structure is to preprocess P so given a query q generated as above, can
recover the corresponding data point p.
√
This instance must be handled by any data structure for c + o(1), c2 -ANN over `2 . In fact,
[AR15] show how to reduce any (c, r)-ANN instance into several (pseudo-)random instances from
above without increasing the time and space complexity by a polynomial factor. The resulting
instances are pseudo-random because they are not exactly the random instance described above,
but do have roughly the same distribution over distances from q to the data points.
Following the strategy from [AR15], we first analyze the random instance, and then reduce the
case for general subsets of Rd to pseudo-random instances.
B
Spherical case
We describe how to solve a random instance
of ANN on a unit sphere S d−1 ⊆ Rd , where near
√
neighbors are planted within distance
2
c
(as defined in Appendix A). We obtain the same time-
space tradeoff as in [Laa15c], namely (1). In Appendix C, we extend this algorithm to the entire
space Rd using the techniques from [AR15].
e
Below we assume that d = O(log
n) [JL84, DG03].
B.1
The data structure description
The data structure is a single rooted T -ary tree consisting of K + 1 levels. The zeroth level holds
the root r, and each node up to the K-th level has T children, so there are T K leaves. For every
37
node v, let Pv be the set of nodes on the path from v to the root except the root itself. Each node
v except the root holds a random Gaussian vector zv ∼ N (0, 1)d is stored . For each node v, we
define the subset of the dataset Pv ⊂ P :
Pv = p ∈ P | ∀v 0 ∈ Pv hzv0 , pi ≥ η ,
where η > 0 is a parameter to be chosen later. For instance, Pr = P , since Pr = ∅. Intuitively, each
set Pv corresponds to a subset of the dataset which lies in the intersection of sphere caps centered
around zv0 for all v 0 ∈ Pv . Every leaf v of the tree stores the subset Pv .
To process the query q ∈ S d−1 , we start with the root and make our way down the tree. We
consider all the children of the root v with hzv , qi ≥ η 0 , where η 0 > 0 is a parameter to be chosen
later, and recurse on them. If we end up in a leaf v, we try all the points from Pv until we find a
near neighbor. If we don’t end up in a leaf, or we do not find a neighbor, we fail.
B.2
Analysis
First, let us√ analyze the probability of success. Let q ∈ S d−1 and p ∈ P be the near neighbor
(kp − qk ≤
2
c ).
Lemma B.1. If
T ≥
100
,
Prz∼N (0,1)d [hz, pi ≥ η and hz, qi ≥ η 0 ]
then the probability of successfully finding p on query q is at least 0.9.
Proof. We prove this by induction. Suppose the querying algorithm is at node v, where p ∈ Pv .
We would like to prove that—if the conditions of the lemma are met—the probability of success is
0.9.
When v is a leaf, the statement is obvious. Suppose it is true for all the children of a node v,
then
Pr[failure] ≤
Y
v 0 child of v
1 − Przv0 hzv0 , pi ≥ η and hzv0 , qi ≥ η 0 · 0.9
T
= 1 − Prz∼N (0,1)d hz, pi ≥ η and hz, qi ≥ η 0 · 0.9 ≤ 0.1.
Now let us understand how much space the data structure occupies. In the lemma below,
u ∈ S d−1 is an arbitrary point.
Lemma B.2. The expected space consumption of the data structure is at most
no(1) · T K 1 + n · Prz∼N (0,1)d [hz, ui ≥ η]K .
38
Proof. The total space the tree nodes occupy is no(1) · T K .
At the same time, every point u participates on average in T K · Prz∼N (0,1)d [hz, ui ≥ η]K leaves,
hence the desired bound.
Finally, let us analyze the expected query time. As before, u ∈ S d−1 is an arbitrary point.
Lemma B.3. The expected query time is at most
K
no(1) · T K+1 · Prz∼N (0,1)d hz, ui ≥ η 0 · 1 + n · Prz∼N (0,1)d [hz, ui ≥ η]K .
Proof. First, a query touches at most T K · Prz∼N (0,1)d [hz, ui ≥ η 0 ]K tree nodes on average .
If a node is not a leaf, the time spent on it is at most no(1) · T .
For a fixed leaf and a fixed dataset point, the probability that they end up in the leaf together
with the query point is
K
Prz∼N (0,1)d hz, ui ≥ η 0 · Prz∼N (0,1)d [hz, ui ≥ η]K ,
hence we obtain the desired bound.
B.3
Setting parameters
First, we set K =
√
log n. Second, we set η > 0 such that
Prz∼N (0,1)d [hz, ui ≥ η] = n−1/K = 2−
√
log n
.
We can simply substitute the parameter setting of Lemma B.2 and Lemma B.3 This gives an
expected space of
no(1) · T K
,
and an expected query time
K
no(1) · T K+1 · Prz∼N (0,1)d hz, ui ≥ η 0 .
As discussed above in Lemma B.1, by setting
T ≥
100
,
Prz∼N (0,1)d [hz, pi ≥ η and hz, qi ≥ η 0 ]
the probability of success is 0.9.
In order to get the desired tradeoff, we can vary T and η 0 . Suppose we want space to be nρs +o(1)
39
for ρs ≥ 1. Then we let
T =n
ρs +o(1)
K
= 2(1+o(1))·ρs
√
log n
,
and η 0 > 0 to be the largest number such that for every p, q ∈ S d−1 with kp − qk ≤
√
2
c ,
we have
√
100
Prz∼N (0,1)d hz, pi ≥ η and hz, qi ≥ η 0 ≥
= 2−(1+o(1))·ρs log n .
T
Again, substituting in values of Lemma B.3, the query time is
K
√log n
no(1) · T K+1 · Prz∼N (0,1)d hz, ui ≥ η 0 = nρs +o(1) · Prz∼N (0,1)d hz, ui ≥ η 0
.
The trade-off between ρs and ρq follows from a standard computation of
Prz∼N (0,1)d hz, ui ≥ η 0
given that
√
Prz∼N (0,1)d [hz, ui ≥ η] = 2−
and
log n
√
Prz∼N (0,1)d hz, pi ≥ η and hz, qi ≥ η 0 ≥ 2−(1+o(1))·ρs log n .
The computation is relatively standard: see [AIL+ 15]. We verify next that the resulting tradeoff is the same as (1) obtained in [Laa15c].
√
2
c
Denote α, β to be real numbers such that k(1, 0) − (α, β)k2 =
√
α = 1 − c12 and β = 1 − α2 .
Lemma B.4. Suppose that η, η 0 > 0 are such that η, η 0 → ∞ and
every p, q ∈
S d−1
√
with kp − qk2 ≤
2
c
and k(α, β)k2 = 1. Namely,
η 2 +η 02 −2αηη 0
β2
→ ∞. Then, for
one has:
2
02
0
−(1+o(1))· η +η −2αηη
2β 2
Prz∼N (0,1)d hz, pi ≥ η and hz, qi ≥ η 0 = e
,
and,
Prz∼N (0,1)d [hz, pi ≥ η] = e−(1+o(1))·
η2
2
.
Proof. Using spherical symmetry of Gaussians, we can reduce the computation to computing the
Gaussian measure of the following two-dimensional set:
{(x, y) | x ≥ η 0 and αx + βy ≥ η}.
The squared distance from zero to the set is:
η 2 + η 02 − 2αηη 0
.
β2
40
The result follows from the Appendix A of [AIL+ 15].
From the discussion above, we conclude that one can achieve the following trade-off between
space nρs +o(1) and query time nρq +o(1) :
1 + α2 ρs − ρq − 2α
p
ρs − ρq = 0.
(78)
We now show that this is equivalent to the tradeoff of [Laa15c], i.e., (1), where ρs = 1 + ρu .
Indeed, squaring (78) and replacing ρs = 1 + ρu , we get:
(1 + α2 ) + α2 ρu − ρq
2
= 4α2 (1 + ρu − ρq ),
(79)
or
(1 + α2 )2 + α4 ρ2u + ρ2q + 2 · ((1 + α2 )α2 ρu − (1 + α2 )ρq − α2 ρu ρq ) = 4α2 + 4α2 ρu − 4α2 ρq .
(80)
Simplifying the equation, we get
(1 − α2 )2 + α4 ρ2u + ρ2q + 2 · ((α2 − 1)α2 ρu − (1 − α2 )ρq − α2 ρu ρq ) = 0.
Remember that we have α = 1 − 1/c2 , and hence α2 =
(c2 −1)2
c4
obtain:
(2c2 −1)2
c8
+
(c2 −1)4 2
ρu
c8
+ ρ2q − 2 (2c
2 −1)(c2 −1)2
c8
2
2c2 −1
.
c4
We further
ρu ρq = 0,
(82)
and 1 − α2 =
ρu − 2 2cc4−1 ρq − 2 (c
2 −1)2
c4
(81)
or, multiplying by c8 ,
(2c2 − 1)2 + (c2 − 1)4 ρ2u + c8 ρ2q − 2(2c2 − 1)(c2 − 1)2 ρu − 2(2c2 − 1)c4 ρq − 2(c2 − 1)2 c4 ρu ρq = 0. (83)
In a similar fashion, squaring (1), we obtain:
or equivalently,
√
c4 ρq + (c2 − 1)2 ρu + 2c2 (c2 − 1) ρq ρu = 2c2 − 1,
(84)
√
2c2 (c2 − 1) ρq ρu = 2c2 − 1 − c4 ρq − (c2 − 1)2 ρu .
(85)
Squaring again, we obtain
4c4 (c2 − 1)2 ρq ρu = (2c2 − 1)2 + c8 ρ2q + (c2 − 1)2 ρ2u
+ 2 · c4 (c2 − 1)2 ρq ρu − (2c2 − 1)c4 ρq − (2c2 − 1)(c2 − 1)2 ρu , (86)
41
or, simplifying,
(2c2 − 1)2 + c8 ρ2q + (c2 − 1)2 ρ2u − 2c4 (c2 − 1)2 ρq ρu − 2(2c2 − 1)c4 ρq − 2(2c2 − 1)(c2 − 1)2 ρu = 0 (87)
We now observe that we obtain the same equation as (83) and hence we are done proving
that (78) is equivalent to (1).
C
Upper Bound: General case
We show how to extend the result of [Laa15c] (and Appendix B) to the general case using the
techniques of [AR15]. In particular, we show how to reduce a worst-case instance to several instances
that are random-like. Overall the algorithm from below gives a data structure that solves the (c, r)ANN problem in the d-dimensional Euclidean space, using space O(n1+ρu +o(1) + dn), and query
time O(dnρq +o(1) ) for any ρu , ρq > 0 that satisfy:
p
√
√
c2 ρq + (c2 − 1) ρu = 2c2 − 1.
(88)
As in [AR15], our data structure is a decision tree. However, there are several notable differences
from [AR15]:
• The whole data structure is a single decision tree, while in [AR15] we consider a collection of
nΘ(1) trees.
• Instead of Spherical LSH used in [AR15], we use the partitioning procedure from Section B.
• In [AR15], one proceeds with partitioning a dataset until all parts contain less than no(1)
points. We change the stopping criterion slightly to ensure the number of “non-cluster”
nodes5 on any root-leaf branch is the same.
• Unlike [AR15], we do not use a “three-point property” of a random space partition in the
analysis. This is related to the fact that the probability success of a single tree is constant,
unlike [AR15], where it is polynomially small.
• In [AR15] we reduce the general case to the “bounded ball” case using LSH from [DIIM04].
Now we cannot quite do this, since we are aiming at getting a full time-space trade-off.
Instead, we use a standard trick of imposing a randomly shifted grid, which reduces an
√
e log n) [IM98]. Then, we invoke an upper
arbitrary dataset to a dataset of diameter O(
bound from [Laa15c] together with a reduction from [Val15], which for this case is enough to
proceed.
5
√
think K = O( log n) as in Section B.
42
√
( 2 − ε)R
(1 − Θ(ε2 ))R
√
Figure 1: Covering a spherical cap of radius ( 2 − ε)R
C.1
Overview
We start with a high-level overview. Consider a dataset P0 of n points. We can assume that
r = 1 by rescaling. We may also assume that the dataset lies in the Euclidean space of dimension
d = Θ(log n·log log n): one can always reduce the dimension to d by applying Johnson-Lindenstrauss
lemma [JL84, DG03] while incurring distortion at most 1 + 1/(log log n)Ω(1) with high probability.
For simplicity, suppose that the entire dataset P0 and a query lie on a sphere ∂B(0, R) of radius
√
R = Oc (1). If R ≤ c/ 2, we are done: this case corresponds to the “random instance” of points
and we can apply the data structure from Section B.
√
Now suppose that R > c/ 2. We split P0 into a number of disjoint components: l dense
components, termed C1 , C2 , . . . , Cl , and one pseudo-random component, termed Pe. The properties
of these components are as follows. For each dense component Ci we require that |Ci | ≥ τ n and
√
that Ci can be covered by a spherical cap of radius ( 2 − ε)R (see Fig. 1). Here τ, ε > 0 are
small quantities to be chosen later. The pseudo-random component Pe contains no more dense
components inside.
We proceed separately for each Ci and Pe. We enclose every dense component Ci in slightly
smaller ball Ei of radius (1 − Θ(ε2 ))R (see Figure 1). For simplicity, let us first ignore the fact that
Ci does not necessarily lie on the boundary ∂Ei . Once we enclose each dense cluster in a smaller
ball, we recurse on each resulting spherical instance of radius (1 − Θ(ε2 ))R. We treat the pseudorandom part Pe as described in Section B. we sample T Gaussian vectors z1 , z2 , . . . , zT ∼ N (0, 1)d ,
where T is a parameter to be chosen later (for each pseudo-random remainder separately), and
form T subsets of Pe as follows:
Pei = {p ∈ Pe | hzi , pi ≥ ηR},
where η > 0 is a parameter to be chosen later (for each pseudo-random remainder separately).
Then we recurse on each Pei . Note that after we recurse, there may appear new dense clusters in
some sets Pei (e.g., since it may become easier to satisfy the minimum size constraint).
43
Project(R1 , R2 , r)
r
R1
S1
R2
S2
Figure 2: The definition of Project
During the query procedure, we recursively query each Ci with the query point q. For the
pseudo-random component Pe, we identify all i’s such that hzi , qi ≥ η 0 R, and query all corresponding
children recursively. Here η 0 > 0 is a parameter to be chosen later (for each pseudo-random
remainder separately).
To analyze our algorithm, we show that we make progress in two ways. First, for dense clusters
we reduce the radius of a sphere by a factor of (1 − Θ(ε2 )). Hence, in Oc (1/ε2 ) iterations we
√
must arrive to the case of R ≤ c/ 2, which is easy (as argued above). Second, for the pseudo√
random component Pe, we argue that most points lie at a distance ≥ ( 2 − ε)R from each other.
√
In particular, the ratio of R to a typical inter-point distance is ≈ 1/ 2, exactly like in a random
case. This is the reason we call Pe pseudo-random. This setting is where the data structure from
Section B performs well.
We now address the issue deferred in the above high-level description: namely, that a dense
component Ci does not generally lie on ∂Ei , but rather can occupy the interior of Ei . In this case,
we partitioning Ei into very thin annuli of carefully chosen width δ and treat each annulus as a
sphere. This discretization of a ball adds to the complexity of the analysis, but is not fundamental
from the conceptual point of view.
C.2
Formal description
We are now ready to describe the data structure formally. It depends on the (small positive)
√
parameters τ , ε and δ, as well as an integer parameter K ∼ log n. We also need to choose
parameters T , η > 0, η 0 > 0 for each pseudo-random remainder separately.
Preprocessing. Our preprocessing algorithm consists of the following functions:
• ProcessSphere(P , r1 , r2 , o, R, k) builds the data structure for a dataset P that lies
on a sphere ∂B(o, R), assuming we need to solve ANN with distance thresholds r1 and r2 .
44
Moreover, we are guaranteed that queries will lie on ∂B(o, R). The parameter k is a counter
which, in some sense, measures how far are we from being done.
• ProcessBall(P , r1 , r2 , o, R, k) builds the data structure for a dataset P that lies inside
the ball B(o, R), assuming we need to solve ANN with distance thresholds r1 and r2 . Unlike
ProcessSphere, here queries can be arbitrary. The parameter k has the same meaning as
above.
• Process(P ) builds the data structure for a dataset P to solve the general (c, 1)-ANN;
• Project(R1 , R2 , r) is an auxiliary function computing the following projection. Suppose
we have two spheres S1 and S2 with a common center and radii R1 and R2 . Suppose there
are points p1 ∈ S1 and p2 ∈ S2 with kp1 − p2 k = r. Project(R1 , R2 , r) returns the distance
between p1 and the point pe2 that lies on S1 and is the closest to p2 (see Figure 2).
We now elaborate on algorithms in each of the above functions.
ProcessSphere. Function ProcessSphere follows the exposition from Section C.1. We consider
three base cases. First, if k = K, then we stop and store the whole P . Second, if r2 ≥ 2R, then the
goal can be achieved trivially, since any point from P works as an answer for any valid query. Third,
if an algorithm from Section B would give a desired point on the time-space trade-off (in particular,
√
if r2 ≥ 2R), then we just choose η, η 0 > 0 and T appropriately (in particular, we set η > 0 such that
for any u, v with ku−vk = r2 one has Prz∼N (0,1)d [hz, ui ≥ ηR and hz, vi ≥ ηR] = n−1/K = 2−
√
log n )
and make a single step .
√
Otherwise, we find dense clusters, i.e., non-trivially smaller balls, of radius ( 2 − ε)R, with
centers on ∂B(o, R) that contain many data points (at least τ |P |). These balls can be enclosed
e ≤ (1 − Ω(ε2 ))R. For these balls we invoke
into balls (with unconstrained center) of radius R
ProcessBall with the same k. Then, for the remaining points we perform a single step of the
algorithm from Section B with appropriate η, η 0 > 0 and T (in particular, we set η > 0 as above
√
for the distance 2R), and recurse on each part with k increased by 1.
ProcessBall.
First, we consider the following simple base case. If r1 + 2R ≤ r2 , then any point
from B(o, R) could serve as a valid answer to any query.
In general, we reduce to the spherical case via a discretization of the ball B(o, R). First, we
round all the distances to o up to a multiple of δ, which can change distance between any pair of
points by at most 2δ (by the triangle inequality). Then, for every possible distance δi from o to
a data point and every possible distance δj from o to a query (for admissible integers i, j), we build
a separate data structure via ProcessSphere (we also need to check that |δ(i − j)| ≤ r1 + 2δ to
ensure that the corresponding pair (i, j) does not yield a trivial instance). We compute the new
distance thresholds re1 and re2 for this data structure as follows. After rounding, the new thresholds
45
for the ball instance should be r1 + 2δ and r2 − 2δ, since distances can change by at most 2δ. To
compute the final thresholds (after projecting the query to the sphere of radius δi), we just invoke
Project (see the definition above).
Process. Process reduces the general case to the ball case. We proceed similarly to ProcessSphere, with two modifications. First, we apply a randomized partition using cubes with side
√
√
ec ( log n), and solve each part separately. Second, we seek to find dense clusters of raOc ( d) = O
dius Oc (1). After there are no such clusters, we apply the reduction to unit-norm case from [Val15,
Algorithm 25], and then (a single iteration of) the algorithm from Section B.
Project. This is implemented by a formula as in [AR15] (see Figure 2).
Overall, the preprocessing creates a decision tree, where the nodes correspond to procedures
ProcessSphere, ProcessBall, Process. We refer to the tree nodes correspondingly, using the
labels in the description of the query algorithm from below.
Query algorithm.
Consider a query point q ∈ Rd . We run the query on the decision tree,
starting with the root, and applying the following algorithms depending on the label of the nodes:
• In Process we first recursively query the data structures corresponding to the clusters.
Second, we locate q in the spherical caps, and query the data structure we built for the
corresponding subsets of P .
• In ProcessBall, we first consider the base case, where we just return the stored point if it
is close enough. In general, we check if kq − ok ≤ R + r1 . If not, we can return. Otherwise,
we round q so that the distance from o to q is a multiple of δ. Next, we enumerate the
distances from o to the potential near neighbor we are looking for, and query the corresponding
ProcessSphere children after projecting q on the sphere with a tentative near neighbor
(using, naturally, Project).
• In ProcessSphere, we proceed exactly the same way as Process modulo the base cases.
• In all the cases we try all the points if we store them explicitly (which happens when k = K).
C.3
How to set parameters
Here we briefly state how one sets the parameters of the data structure.
Recall that the dimension is d = Θ(log n · log log n). We set ε, δ, τ as follows:
• ε=
1
log log log n ;
• δ = exp −(log log log n)C ;
46
• τ = exp − log2/3 n ,
where C is a sufficiently large positive constant.
Now we need to specify how to set η, η 0 > 0 and T for each pseudo-random remainder. The
idea is to set η, η 0 and T such that
√
Prz∼N (0,1)d [hz, ui] = n−1/K = 2−
log n
while, at the same time, for every u and v at distance at most r1
T ∼
100
.
Prz∼N (0,1)d [hz, ui ≥ η, hz, vi ≥ η 0 ]
Finally, we choose T such that T K ∼ nρs +o(1) where ρs ≥ 1 is a parameter that governs the memory
consumption.
This gives us a unique value of η 0 > 0, which governs the query time.
√
A crucial relation between parameters is that τ should be much smaller than n−1/K = 2− log n .
√
This implies that the “large distance” is effectively equal to 2R, at least for the sake of a single
step of the random partition.
47
| 8cs.DS
|
Scalable Magnetic Field SLAM in 3D
Using Gaussian Process Maps
Manon Kok? and Arno Solin†
?
University of Cambridge, United Kingdom, e-mail: [email protected]
†
Aalto University, Finland, e-mail: [email protected]
arXiv:1804.01926v1 [cs.RO] 5 Apr 2018
Abstract
We present a method for scalable and fully 3D magnetic field simultaneous localisation and mapping
(SLAM) using local anomalies in the magnetic field as a source of position information. These
anomalies are due to the presence of ferromagnetic material in the structure of buildings and in
objects such as furniture. We represent the magnetic field map using a Gaussian process model and
take well-known physical properties of the magnetic field into account. We build local magnetic
field maps using three-dimensional hexagonal block tiling. To make our approach computationally
tractable we use reduced-rank Gaussian process regression in combination with a Rao–Blackwellised
particle filter. We show that it is possible to obtain accurate position and orientation estimates
using measurements from a smartphone, and that our approach provides a scalable magnetic SLAM
algorithm in terms of both computational complexity and map storage.
1
Introduction
The use of the magnetic field as a source of position information for indoor navigation is a promising
novel approach that has gained interest in recent years [1–8]. It relies upon the spatial variation of the
ambient magnetic field which is typically due to ferromagnetic material in the structures of buildings
and to a lesser extent due to the presence of e.g. furniture. The advantage of using the magnetic field
for positioning is that it can be measured by a small device, without additional infrastructure and
without line-of-sight requirements. Furthermore, magnetometers are nowadays present in (almost) any
smartphone. Crucial to this approach is the ability to build maps of the magnetic field that can be
used for interpolation and extrapolation and the ability to localise inside this map. In this paper, we
present a scalable and fully three-dimensional simultaneous localisation and mapping (SLAM) approach
that builds a map of the magnetic field while simultaneously localising the sensor in the map.
Extending previous work [9], we build the magnetic field map using Gaussian process regression [10]
and incorporate physical knowledge about the magnetic field known from Maxwell’s equations in the
Gaussian process prior. To overcome issues with computational complexity, we use the rank-reduced
approach first presented in [11]. This results in a representation that fits perfectly in a Rao–Blackwellised
particle filter [12] which can be used both for building the map and for localisation in the map, resulting
in a tractable localisation algorithm. Example results can be found in Fig. 1.
The approach presented in [11] relies on a basis function expansion on a specific domain where the
number of basis functions that is needed scales with the size of the domain. To allow for mapping large
3D areas, we propose a map representation using hexagonal block tiling, which aims at providing a very
compact (in terms of required storage per volume) representation of the magnetic field map (all three
vector field components and associated uncertainties). Using this approach, hexagonal tiles are created
whenever particles enter unexplored areas, resulting in a growing number of local Gaussian process maps.
We present experimental results where we obtain magnetometer measurements as well as odometry
from a smartphone. This odometry gives us information about the change in position and orientation but
the resulting pose estimates drift over time. This drift is corrected by making use of the magnetometer
measurements and the magnetic field map. We show accurate 6D pose and 3D map estimates can be
obtained in large scale and 3D experiments, as illustrated in Fig. 1.
Previous work on magnetic field SLAM typically assumes that the motion is in 2D or 2.5D [13–16].
To the best of the authors’ knowledge, this is the first fully 3D approach to magnetic field SLAM. It is
also the first time the reduced-rank approach from [11] has been used for SLAM and that the domain in
this approach has been chosen to be hexagonal.
1
Figure 1: Illustration of the magnetic field map built by our SLAM approch. Each hexagonal block (with
radius of 5 m and height of 4 m) can cover a Gaussian process magnetic map volume of ∼260 m3 in a
state dimension of 256. A more detailed map provided in Fig. 5.
2
Problem formulation
We are interested both in estimating the position and the orientation of the sensor and in building a map
of the magnetic field. We denote the position of the sensor at time instance t by pw
t . The superscript w
refers to the world frame. In our SLAM formulation, it is always possible to move or rotate the estimated
map while simultaneously moving and rotating the position and orientation estimates. In other words,
the absolute position and orientation are not observable. Because of this, without loss of generality we
choose the origin of the world frame to be equal to the initial position of the sensor and its axes to be
aligned with the coordinate system defined by the initial orientation.
The orientation of the sensor is denoted by qwb
t , where b refers to the body frame. Its origin lies in
the centre of the magnetometer triad and its axes are aligned with the axes of the magnetometer. We
encode the orientation using a unit quaternion. There is a direct mapping from the quaternion qwb
to
t
and
because
of
this
we
use
these
interchangeably.
The
double
superscript
wb
the rotation matrix Rwb
t
denotes the rotation from body frame to world frame. The reverse rotation is denoted by bw and hence,
bw T
Rwb
.
t = Rt
We denote our estimate of the magnetic field map by mt . More details about the exact form of mt
can be found in Section 3.1. Note that we assume that the map of the magnetic field does not change
over time. The subscript t refers to the fact that our belief of the map changes over time because more
data results in more information about the map.
The resulting state vector xt is given by
T
xt = (pw
t )
T
(qwb
t )
mT
t
T
.
(1)
The dynamics of the position and orientation are modelled in terms of the change in position ∆pt and
orientation ∆qt obtained from the odometry information. We consider these to be inputs to our dynamic
model. Using the assumption that the magnetic field does not vary over time, our dynamic model can
be written as
pw
t + ∆pt + εp,t
expq (εq,t ) .
xt+1 = ∆qt qwb
(2)
t
mt
We assume that the noise on the position (εp,t ) and on the orientation (εq,t ) is Gaussian with εp,t ∼
N (0, ∆T σp2 I3 ) and εq,t ∼ N (0, ∆T σq2 I3 ). Here, ∆T is the (possibly time-varying) time between subsequent samples. In (2), denotes the quaternion multiplication and expq (εq,t ) refers to the representation
of the noise vector εq,t ∈ R3 as a unit quaternion. The dynamics of the orientation qwb
t can be interpreted
2
as the rotation equivalent of the dynamics of the position pw
t . For more information about quaternion
algebra and representations of orientation, see e.g.[17].
We use the magnetometer measurements to provide information about the magnetic field map as well
as the position and orientation of the sensor. This can be modelled as
w
ym,t = Rbw
t ∇Φt (pt )mt + εm,t .
(3)
w
Here, ∇Φt (pw
t )mt is the prediction of the magnetic field at the location pt . More details about the
w
specific form of ∇Φt (pt ) can be found in Section 3.1. Note that the magnetic field map is represented
in world frame. Hence, the magnetometer measurements are modelled as the prediction ∇Φt (pw
t )mt ,
rotated to body frame.
Together with a prior on the state at the first time instance, the models (2) and (3) constitute the
state space model that we use in the RBPF to estimate the state defined in (1). Because the initial
wb
position and orientation fix the world coordinate frame w, pw
are set to the initial position
1 and q1
and orientation with zero uncertainty. The magnetic field map at the first time instance is set to be
equal to the Gaussian process prior. More details about the Gaussian process map will be discussed in
Section 3.1. The RBPF implementation will subsequently be discussed in Section 3.2.
3
3.1
Methods
Gaussian process magnetic field map
Building on previous work [9], we build a Gaussian process magnetic field map in which we encode physical
knowledge about the magnetic field. First of all, we know that the measured magnetic field consists of
the Earth’s magnetic field and the magnetic field due to anomalies induced by small-scale variations
and building structures. Furthermore, using Maxwell’s equations and classical electromagnetism [18], we
know that the magnetic field due to the anomalies can be modelled using a latent scalar potential field
ϕ(p). Here, ϕ : R3 → R, where p ∈ R3 is the spatial coordinate in world frame and the magnetometer
measures the derivative of this scalar potential. We assume the scalar potential field to be a realisation
of a Gaussian process prior and the magnetometer measurements to be corrupted by Gaussian noise,
resulting in the following model
ϕ(p) ∼ GP(0, κlin. (p, p0 ) + κSE (p, p0 )),
ym,t = −∇ϕ(p)
p=pw
t
+ εm,t ,
(4)
2
I3 ), for each observation t = 1, 2, . . . , NT . The local Earth’s magnetic field
where εm,t ∼ N (0, σm
contributes linearly to the scalar potential as
2
κlin. (p, p0 ) = σlin.
pT p,
(5)
2
where σlin.
is the magnitude scale hyperparameter. The magnetic field anomalies are captured by the
squared exponential covariance function
kp − p0 k2
0
2
,
(6)
κSE (p, p ) = σSE exp −
2`2
2
where σSE
is a magnitude and ` the characteristic length-scale hyperparameter. Including the additional
prior information introduced by the scalar potential in (4) results in a coupling between the three
components of the magnetic field according to physical laws. This has been shown in [9] to improve
prediction accuracy over simply modelling each magnetic field component separately.
Using the model (4), it is possible to predict the magnetic field at previously unseen locations.
In practice, however, this quickly becomes computationally intractable because of the large amount
of magnetometer measurements we collect and the fact that the computational complexity scales as
O(n3 ). A computationally tractable solution was presented in [9] where the problem was projected on
the eigenbasis of the (negative) Laplace operator in a confined domain Ω ⊂ R3 . In this domain, the
eigendecomposition of the Laplace operator subject to Dirichlet boundary conditions can be solved as
(
−∇2 φj (x) = λ2j φj (x), x ∈ Ω,
(7)
φj (x) = 0,
x ∈ ∂Ω.
3
Using this eigendecomposition, it is possible to approximate the covariance function in (4) as
κ(p, p0 ) = κlin. (p, p0 ) + κSE (p, p0 )
m
X
≈ κlin. (p, p0 ) +
SSE (λj ) φj (p) φj (p0 ),
(8)
j=1
where κSE (p, p0 ) is approximated using m basis functions φj and their corresponding eigenvalues λj . The
exact form of φj and λj depends on the shape of the domain Ω. Note that SSE (·) denotes the spectral
density function of the squared exponential kernel SSE (·) for which a closed form solution exists, see [9].
Using the approximation (8) with the boundary conditions defined in (7) implies that our model of the
magnetic field reverts back to the Earth’s magnetic field at the boundary of the domain.
A fixed domain Ω is not desirable in the SLAM problem, as we do not know a priori the spatial
extent of the final magnetic field map. Furthermore, the number of basis functions that is needed for
a good approximation scales with the size of the domain and for large scale SLAM problems will hence
become intractable. We therefore propose a scalable representation where we split the magnetic map into
a three-dimensional grid of hexagonal blocks with radius r and height 2Lz , such that each subdomain d
is given by
(d)
(d)
(d)
Ω(d) = {p | (p1 , p2 ) ∈ hexagon(r, p1,2 ), p3 ∈ [p3 − Lz , p3 + Lz ]}.
(9)
Here, p(d) is the center point of the dth hexagon cell in the grid representation. The hexagonal grid
provides an efficient way of maximising the area per number of basis functions for representing the
magnetic field map, thus requiring less memory for storing the map representation.
To compute the basis functions φj and their corresponding eigenvalues λj , we need to solve the
eigendecomposition of the Laplace operator (7) in a two-dimensional hexagon. This can not be solved in
closed form, but instead it can be solved numerically. We set up the eigenvalue problem by composing the
sparse stencil matrix corresponding to the Laplacian using a nine-point finite difference approximation.
The solution itself is given by the Lanczos algorithm (see, e.g., [19], callable in Matlab through eigs) that
2
we use for solving the m largest real eigenvalues [λhex
n ] —together with the corresponding eigenfunctions
hex
φn (p). Fig. 2 shows the 40 first eigenfunctions of the negative Laplacian in a unit hexagonal domain
with respect to Dirichlet boundary conditions.
The eigendecomposition is extended to cover the three-dimensional hexagonal grid cell Ω(d) by considering the numerically solved horizontal eigendecomposition and a closed-form solution to the separable
vertical dimension. The final eigenbasis then becomes the combination:
πnj,2 (p3 + Lz )
1
√
sin
,
(10)
φj (p) = φhex
(p
)
1,2
nj,1
2Lz
Lz
2
πnj,2
2
hex 2
,
(11)
λj = [λnj,1 ] +
2Ld
where the matrix n ∈ Rm×2 consists of an index set of permutations of integers {1, 2, . . . , m} (i.e.,
{(1, 1), (2, 1), . . . , (1, 2), . . . , (2, 2), . . .}) corresponding to the m largest eigenvalues.
The numerical solution for the two-dimensional hexagon is fast and takes less than a second even
with a tight discretisation on an average laptop computer. Furthermore, because the basis functions (10)
are independent of the hyperparameters and the input locations, they only need to be evaluated once,
not causing any computational overhead in our SLAM algorithm.
The basis function expansion described above allows us to write the prior model for each hexagonal
tile at time t = 0 in terms of a mean m0 = 0m+3 and a covariance
2
2
2
P0 = diag σlin.
, σlin.
, σlin.
, SSE (λ1 ), SSE (λ2 ), . . . , SSE (λm ) .
(12)
The posterior can be updated sequentially [20] as new magnetometer data arrives at t = 1, 2, . . . , NT
and can be written as
2
St = ∇Φt Pt−1 [∇Φt ]T + σm
I3 ,
Kt = Pt−1 [∇Φt ]T S−1
t ,
mt = mt−1 + Kt (ym,t − Rbw
t ∇Φt mt−1 ),
Pt = Pt−1 − Kt St KT
t,
4
(13)
Figure 2: The first 40 eigenfunctions of the negative Laplace operator in a hexagonal domain subject to
Dirichlet boundary conditions. The eigenbasis can be effectively solved by a numerical solver, and the
calculations are only required once.
with
∇Φt = ∇pT
t , ∇φ1 (pt ), ∇φ2 (pt ), . . . , ∇φm (pt ) .
(14)
Note that (13) is a Kalman filter measurement update of the state mt representing the magnetic field
map in a specific tile. This fact will be exploited in our magnetic field SLAM algorithm presented in
Section 3.2.
3.2
Magnetic field SLAM
As discussed in Section 2, we are interested in estimating the position and the orientation of the sensor
while building a map of the magnetic field. We represent the map in terms of a varying number of
hexagonal block tiles, each using the basis function expansion presented in Section 3.1. We use a
nonlinear filtering technique to estimate the position, the orientation and the map of each tile that
has been visited. The state vector (1) has a relatively large dimension due to the representation of the
magnetic field map mt . For instance, in Section 4 we set the dimension of the state mt to 256 per
hexagonal block. However, mt enters conditionally linearly into the state space model. It is possible to
exploit this by using a Rao–Blackwellised particle filter (RBPF) [12] which uses a Kalman filter for the
conditionally linear states xlt and a particle filter xnt for the nonlinear states where
T
xnt = (pw
t )
T
(qwb
,
t )
D,t
xlt = {mt }d=1
.
N
(15)
Here, ND,t is the number of hexagonal block tiles that have been created up to time t.
w,(i)
wb,(i)
In the RBPF, i = 1, . . . , NP particles having a position pt
and an orientation qt
are used to
(i)
represent the state at time t. Each of the particles also contains an estimate of the map mt and its
(i)
(i)
covariance Pt . Note that mt consists of the magnetic field map of each hexagonal block that particle i
has visited. We assume that each tile is independent. The covariance of the linear state can therefore be
ND,t
described as Pt = {Pt }d=1
.
The dynamics of the state is chosen as in (2) and the measurement model is given in (3). An important
thing to note is that, as discussed in Section 2, the absolute position and orientation are not observable.
w,(i)
wb,(i)
and orientation qt
of all particles i = 1, . . . , NP
Because of this, we set the initial position p0
wb,(i)
equal to an initial position and orientation pw,(i) = 03 , q0
= qwb
.
We
initialise the magnetic field
0
according to the Gaussian process prior (see also Section 3.1).
The state space model presented in Section 2 results in a fairly straightforward RBPF implementation.
The major difference from a standard RBPF is that each particle only updates the local magnetic field
map at its current location. Note that the computational complexity of the algorithm scales linearly
with the number of particles and linearly with the number of hexagonal tiles per particle.
5
Alg 1: Magnetic field SLAM
NT
wb
Input: {ym,t , ∆pw
t , ∆qt }t=1 , q0 , Σp , Σq , m.
w
wb
Output: For each t, pt , qt , mt of the highest-weight particle at t.
w,(i)
wb,(i)
(i)
(i)
1: Initialisation: Initialise the particles as p0
= 03 , q0
= qwb
0 , set m0 = 0m+3 , P0 as given
(i)
in (12) and initialise the weights of the particles as w0 = N1P , for i = 1, . . . NP .
2: For t = 1, . . . , NT ,
w,(i)
3: Create new hexagonal block tiles: If pt
is in a new hexagonal block tile d, create this tile for
(i)
(i)
particle i and initialise md,t = 0m+3 and Pd,t as given in (12).
(i)
Evaluate the importance weights: For i = 1, . . . , NP , evaluate the importance weights wt based
on the measurement model (3) and normalise.
5: Resampling: If 90% of the particles is in a tile that they are revisiting, resample those particles
with replacement.
(i)
6: Kalman filter measurement update: For i = 1, . . . , NP , update the magnetic field map md,t and
4:
(i)
its covariance Pd,t of the hexagon d that the particle is in and its close neighbours using the
w,(i)
wb,(i)
nonlinear states pt
and Rt
according to (13).
7: Obtain a filtered estimate: Output the position, orientation, and estimated map of the particle
with the highest weight.
n,(i)
8: Particle filter time update: For i = 1, . . . , NP , predict new particles xt+1 as
!
w,(i)
w,(i)
(i)
pt
+ ∆pt
+ εp,t
n,(i)
,
(16)
xt+1 =
wb,(i)
(i)
∆qt qt
expq (εq,t )
(i)
(i)
where εp,t and εq,t are drawn from N (0, Σp ) and N (0, Σq ), respectively.
9: Set t := t + 1.
In a practical magnetic field SLAM implementation care must be taken of the following in terms of
updating the map:
• To avoid boundary effects due to the Dirichlet boundary conditions, the actual tile extends slightly
outside of the hexagonal block domain for updating.
• To allow for a smooth transition from one tile to the next, particles close to the border of a tile
also update the map of the neighbouring tiles.
During the “exploration phase” when particles start building a map of previously unseen locations, very
little information is available to distinguish between the particles. Because of this, we would like to
ensure that the particle cloud keeps spreading during this phase. To this end, we adapt the RBPF
implementation in two different ways:
• To maintain the particle spread during the exploration phase, we delay the updating of the map
by one lengthscale `.
• To not collapse the particle cloud too quickly, we only resample the particles when they revisit a
tile. Resampling is further delayed until at least 90% of the particles has arrived in a tile that they
revisit.
4
Results
In the empirical experiments, we use data captured by a smartphone for simultaneously building the
magnetic field map and localising on it. We use an Apple iPhone 6s (published in September 2015), a
previous flagship product of Apple which by current standards represents more of a standard smartphone.
The interest in this paper is purely on the map building and localisation, so we choose to leverage on
the recent developments in visual-inertial pedestrian dead-reckoning (PDR) methods and use the built-in
ARKit PDR provided by Apple. ARKit (released September 2017, used version running in iOS 11.2.5) is
using the IMU and camera of the phone to provide a 6-DoF (position and orientation) PDR movement
6
5 meters
40
(a) At t = 10 s
0
20
(b) At t = 25 s
−20
0
(c) At t = 48 s
−60
(d) At t = 90 s
(e) At t = 120 s
60
(f) At t = 260 s
−40
(g) Magnetic x (µT) (h) Magnetic y (µT) (i) Magnetic z (µT)
(j) Odometry
(k) Tilted 3D view
Figure 3: The SLAM solution for walking in a square for several loops for around 250 meters. Subfigures
(a)–(f) show the norm of the magnetic field map for the highest weight particle. The transparency of
the maps scales with the marginal variance of the GP map. Subfigures (g)–(i) show the corresponding
vector field components for the final map in (f). The odometry trajectory is visualised in (j), and (k)
shows the 3D hexagonal tiles with a cross-section of the magnetic field maps.
trajectory. ARKit does visual relocalisation under the hood, which aims to correct for long-term drift,
resulting in discontinuities in the PDR track. These discontinuities turned out to be more harmful than
useful for our approach, and we implemented a heuristic filtering approach for removing the relocalisation
jumps. The PDR—albeit being surprisingly robust—suffers from long-term drift both in position and
orientation (see Fig. 3).
The data was captured by a customised data capture application running on the phone. It collects the
IMU (at 100 Hz) and three-axis magnetometer (100 Hz) data time-locked to the ARKit PDR estimates
(60 Hz). Furthermore, the captured data was complemented with video frames of the back-facing phone
camera (resolution 1280×720 at 60 fps), the barometric air pressure (10 Hz), and GNSS locations captured by the phone, none of which were used in the SLAM implementation, but captured for validation
and reference. All data was stored on the device, downsampled to 10 Hz before use, and the SLAM
implementation was run off-line in a Mathworks Matlab implementation.
For simplicity, we use the same model parameters in all our experiments.
• We set the size of the hexagonal blocks to r = 5 m, Lz = 2 m and the number of basis functions
per tile to m = 256. The radius and height used in solving the eigenbasis of the Laplacian extends
one meter outside the hexagon.
2
2
• We set the Gaussian process hyperparameters defined in Section 3.1 to σlin.
= 650, σSE
= 200,
` = 1.3 and σn2 = 10. This assumes that the lengthscale of the magnetic field anomalies is in the
order of 1.3 m. The measurement noise σn is rather high to also account for model discrepancies.
• We set the number of particles in the RBPF to NP = 100.
• Neighbouring tiles are updated if samples are closer than 10 cm to the hexagonal block boundary.
• The process noise parameters were set to Σp = diag(0.12 , 0.12 , 0.022 ) (values specified in meters
drift per second) and Σq = diag(0.012 , 0.012 , 0.242 ) (values specified in degrees drift per second).
7
In terms of memory usage, the choices mean that for each initialised map tile, the particle state dimension
grows by 256. For fast look-up, our implementation uses a hashmap (per particle) for storing and keeping
track of the map tiles. Each tile stores the GP mean and covariance matrix.
4.1
Illustrating magnetic SLAM in 2D
In a first experiment, a data set is collected in the DPO at the Engineering Department of the University
of Cambridge. Significant excitation in the magnetic field is present for instance due the presence of
radiators and a large number of computers. The trajectory is around 250 meters long and covers an area
of around 20 × 5 m. The same circular trajectory is traversed several times. Because of this, it is easy to
visually inspect the quality of the estimates from our SLAM algorithm which are shown in Fig. 3. The
different subfigures illustrate the workings of our algorithm.
Figs. 3(a) – 3(f) show how the estimates progress over time. For visualisation purposes, the norm of
the magnetic field is predicted by the map of the highest weight particle at a large number of locations
in each hexagon. The transparency in the map visualises the uncertainty of the map. It can be seen
that the uncertainty is large in unexplored regions (no map is actually visible) and that the uncertainty
of the map decreases over time in the re-visited regions.
The black line displays the trajectory of the highest weight particle. In Figs. 3(a) and (b), the
deliberate slight delay in updating the map can clearly be seen. As discussed in Section 3.2, we wait
with resampling until at least 90% of the particles is in a tile that they have previously updated and
left. Because of this, no resampling is done in the first loop, until around 50 seconds and the spread in
the particles in Fig. 3(c) is therefore fairly large. This allows the algorithm to properly close the loop as
visualised in Fig. 3(d). Another aspect of our algorithm to note is that in Fig. 3(c), most particles are
on the border between the tile in which the trajectory was started and a new tile below. To allow for
smooth transitions between these two tiles, many of the particles will in this case therefore update the
local magnetic field maps in both tiles.
The map that is visualised in Figs. 3(a) – (f), shows the norm of the predicted magnetic field.
Our model (4), however, explicitly models the magnetic field as a three-dimensional vector. Instead of
visualising the norm of the predicted magnetic field, it is therefore also possible to visualise the predicted
magnetic field in three directions. This is shown in Figs. 3(g) – (i). Because the Earth’s magnetic field is
fairly large in Cambridge, UK, the z-component of the magnetic field is of significantly larger magnitude
than the x- and y-components.
Based on Fig. 3(f) it can be concluded that our SLAM algorithm estimates the trajectory with high
accuracy since the same trajectory can be seen to be traversed multiple times. For comparison, we also
show the odometry in Fig. 3(j) which clearly shows a drift in the position. Our magnetic SLAM approach
is able to correct for this drift. To illustrate the fact that even though the movement for this data set is
in a 2D plane the estimation is still done in 3D, the tilted 3D view is shown in Fig. 3(k).
4.2
Fully 3D magnetic SLAM
In Section 4.1, the movement was close to two-dimensional and hexagonal tiles with their own local
magnetic field map were created whenever the particles were at the border of a tile. Exactly the same
procedure also applies to movement and tile block structures in three dimensional space. In Fig. 4 we
show the estimated trajectory and the magnetic field map for an empirical experiment in which we first
walked up and subsequently down a set of stairs at the Engineering Department of the University of
Cambridge.
Again, each hexagonal block has its own local magnetic field map. To visualise these, the norm of the
predicted magnetic field for a hexagonal slice in the middle of each block is shown. The transparency
visualises the uncertainty of the map. As in the results presented in Section 4.1, our algorithm only
starts to resample particles on the way down the stairs. As can be seen, this results in an estimated path
that is fairly similar in both directions.
4.3
Large-scale magnetic SLAM
In a third empirical data set we walk around St. John’s college in Cambridge. The college consists of
several courts and even though the data set is captured outdoors, the close proximity of the buildings
results in small magnetic field anomalies that can be used for localisation. The SLAM results were
previously presented in Fig. 1 and are presented in more detail in Fig. 5. The map extends over 27 tiles
covering the courtyard. For many algorithms the extent of the data set would therefore be prohibitively
8
Figure 4: Example showing the three-dimensional nature of the map-building. The path covers a flight
of stairs in the University of Cambridge Engineering Department. Complete path length 125 meters.
The cross-sections of the tiles show the magnetic map norm (color scale same as in subfigure (f) in the
previous figure).
50 meters
30
40
50
Figure 5: The SLAM map solution for a large-scale outdoor–indoor example. The visualised color field
is the norm of the magnetic field, and the opacity follows the uncertainty. The layout corresponds to
that of Fig. 1.
large. However, due to our computationally efficient model in terms of hexagonal tiling, our approximate
Gaussian process model and the RBPF algorithm, the computations remain feasible. It can be seen from
Fig. 5 that our approach scales to large-scale magnetic SLAM.
5
Conclusions and future work
We have presented a scalable algorithm for magnetic field SLAM in 3D. It builds local Gaussian process
maps where each map is an approximation to the local full GP solution in terms of a basis function
expansion. This approximation allows us to map a magnetic map volume of ∼ 260 m3 in a state
dimension of only 256. Each local map is modelled in terms of a hexagonal block tile, which minimises
9
the number of tiles that are needed to cover the total mapping volume. The formulation fits perfectly
in an RBPF resulting in a computationally tractable algorithm. The computational complexity scales
linearly both with the number of particles in the particle filter and with the number of hexagonal tiles
of each particle. The algorithm is shown to perform well on three challenging empirical data sets.
One of the challenges in the data we used for testing, is that the odometry we obtain from the ARKit
does not always obey our motion model. More specifically, for straight paths the odometry shows almost
no drift, while fast turns can introduce sudden errors. This does not follow our motion model in which
we model a constant process noise. In future work we would like to focus on using inertial PDR as
odometry instead. Apart from avoiding these issues, this would also result in a positioning algorithm
only based on widely available inertial and magnetometer sensors. Other directions of future work are
in the direction of using a Rao–Blackwellised Particle Smoother and further reducing the computational
complexity of the algorithm.
Acknowledgments
This research was financially supported by the EPSRC grant Autonomous behaviour and learning in an
uncertain world (Grant number: EP/J012300/1) and the Academy of Finland grant Sequential inference
for real-time probabilistic modelling (Grant number: 308640).
References
[1] Janne Haverinen and Anssi Kemppainen. Global indoor self-localization based on the ambient
magnetic field. Robotics and Autonomous Systems, 57(10):1028–1035, 2009.
[2] Binghao Li, Thomas Gallagher, Andrew G. Dempster, and Chris Rizos. How feasible is the use
of magnetic field alone for indoor positioning? In Proceedings of the International Conference on
Indoor Positioning and Indoor Navigation (IPIN), pages 1–9, 2012.
[3] Michael Angermann, Martin Frassl, Marek Doniec, Brian J. Julian, and Patrick Robertson. Characterization of the indoor magnetic field for applications in localization and mapping. In Proceedings
of the International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–9,
2012.
[4] Etienne Le Grand and Sebastian Thrun. 3-axis magnetic field mapping and fusion for indoor
localization. In IEEE Conference on Multisensor Fusion and Integration for Intelligent Systems
(MFI), pages 358–364, 2012.
[5] A. Solin, S. Särkkä, J. Kannala, and E. Rahtu. Terrain navigation in the magnetic landscape:
Particle filtering for indoor positioning. In Proceedings of the European Navigation Conference,
2016.
[6] David Hanley, Alexander B. Faustino, Scott D. Zelman, David A. Degenhardt, and Timothy Bretl.
MagPIE: A dataset for indoor positioning with magnetic anomalies. In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8, 2017.
[7] Chao Gao and Robert Harle. MSGD: Scalable back-end for indoor magnetic field-based GraphSLAM. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA),
pages 3855–3862, 2017.
[8] Joaquín Torres-Sospedra, David Rambla, Raul Montoliu, Oscar Belmonte, and Joaquín Huerta.
UJIIndoorLoc-Mag: A new database for magnetic field-based localization problems. In Proceedings
of the International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–10,
2015.
[9] A. Solin, M. Kok, N. Wahlström, T. B. Schön, and S. Särkkä. Modeling and interpolation of
the ambient magnetic field by Gaussian processes. IEEE Transactions on Robotics (accepted for
publication), 2018. arXiv:1509.04634.
[10] Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning.
MIT Press, 2006.
10
[11] Arno Solin and Simo Särkkä. Hilbert space methods for reduced-rank Gaussian process regression.
ArXiv preprint arXiv:1401.5508, 2014.
[12] Thomas B. Schön, Fredrik Gustafsson, and Per-Johan Nordlund. Marginalized particle filters for
mixed linear/nonlinear state-space models. IEEE Transactions on Signal Processing, 53(7):2279–
2289, 2005.
[13] Patrick Robertson, Martin Frassl, Michael Angermann, Marek Doniec, Brian J. Julian, Maria Garcia Puyol, Mohammed Khider, Michael Lichtenstern, and Luigi Bruno. Simultaneous localization
and mapping for pedestrians using distortions of the local magnetic field intensity in large indoor
environments. In Proceedings of the International Conference on Indoor Positioning and Indoor
Navigation (IPIN), pages 1–10. IEEE, 2013.
[14] Martin Frassl, Michael Angermann, Michael Lichtenstern, Patrick Robertson, Brian J. Julian, and
Marek Doniec. Magnetic maps of indoor environments for precise localization of legged and nonlegged locomotion. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS), pages 913–920, 2013.
[15] Ilari Vallivaara, Janne Haverinen, Anssi Kemppainen, and Juha Röning. Simultaneous localization
and mapping using ambient magnetic field. In Proceedings of the IEEE Conference on Multisensor
Fusion and Integration for Intelligent Systems (MFI), pages 14–19, 2010.
[16] Ilari Vallivaara, Janne Haverinen, Anssi Kemppainen, and Juha Röning. Magnetic field-based SLAM
method for solving the localization problem in mobile robot floor-cleaning task. In Proceedings of
the 15th International Conference on Advanced Robotics (ICAR), pages 198–203, 2011.
[17] M. Kok, J. D. Hol, and T. B. Schön. Using inertial sensors for position and orientation estimation.
Foundations and Trends on Signal Processing, 11(1–2):1–153, 2017.
[18] David Jeffrey Griffiths and Reed College. Introduction to Electrodynamics. Prentice hall, Upper
Saddle River, NJ, 1999.
[19] Gene H Golub and Charles F Van Loan. Matrix Computations. Johns Hopkins University Press, 3
edition, 2012.
[20] Simo Särkkä. Bayesian Filtering and Smoothing. Cambridge University Press, 2013.
11
| 3cs.SY
|
On Reduced Input-Output Dynamic Mode Decomposition
∗
Christian Himpe†
Tim Mitchell
‡
ep
rin
t
Abstract
The identification of reduced-order models from high-dimensional data
is a challenging task, and even more so if the identified system should not
only be suitable for a certain data set, but generally approximate the
input-output behavior of the data source. In this work, we consider the
input-output dynamic mode decomposition method for system identification. We compare excitation approaches for the data-driven identification
process and describe an optimization-based stabilization strategy for the
identified systems.
Keywords: Dynamic Mode Decomposition, Model Reduction,
System Identification, Cross Gramian, Optimization
MSC: 93B30, 90C99
Pr
arXiv:1712.08447v1 [cs.SY] 22 Dec 2017
Peter Benner
∗ Computational Methods in Systems and Control Theory, Max Planck Institute for Dynamics of Complex Technical Systems, Sandtorstr. 1, 39106 Magdeburg, Germany
ORCID: 0000-0003-3362-4103, [email protected]
† Computational Methods in Systems and Control Theory, Max Planck Institute for Dynamics of Complex Technical Systems, Sandtorstr. 1, 39106 Magdeburg, Germany
ORCID: 0000-0003-2194-6754, [email protected]
‡ Computational Methods in Systems and Control Theory, Max Planck Institute for Dynamics of Complex Technical Systems, Sandtorstr. 1, 39106 Magdeburg, Germany
ORCID: 0000-0002-8426-0242, [email protected]
1
1
Introduction
Pr
ep
rin
t
In various applications, it can be difficult (sometimes even impossible) to derive
models from first principles. However, the input-output response of a system
and data of the inner state may still be available; we refer to such a setup as
a graybox. For example, a natural process whose underlying action is not well
understood can be considered a graybox since we may only be able to measure its behavior. In applications, such as manufacturing and design, it may
be necessary to model a third-party provided subcomponent whose exact or
full specifications may not be obtainable due to it containing proprietary information. In order to gain insight into such natural or technical processes, and
derive models that simulate and/or predict their behaviors, it is often beneficial
and perhaps necessary to create models using measured or generated data. The
discipline of system identification investigates methods for the task of obtaining
such models from data. One class of methods for data-driven identification is
dynamic mode decomposition (DMD), which also provides a modal analysis of
the resulting systems. In this work, we investigate variants of DMD for the class
of input-output systems and compare data sampling strategies.
DMD has its roots in the modal decomposition of the Koopman operator
[15, 26], which has been recently rediscovered for the spectral analysis of fluid
dynamics [19]. The basic DMD method was introduced in [27], and various
extensions have been added, such as Exact DMD [28] or Extended DMD [9].
Furthermore, DMD can also be used as a tool for model order reduction: [25]
proposed using DMD for flow analysis and control while DMD has also been
combined with Galerkin-projection techniques to model nonlinear systems [1].
For a comprehensive survey of DMD and its variants, see [16].
Our work here builds upon two specific variants of DMD. The first is Dynamic Mode Decomposition with Control (DMDc) [23], and thus by relation,
also Koopman with Inputs and Control (KIC) [24]. The second is Input-Output
Dynamic Mode Decomposition (ioDMD) [3, 4], which itself is closely related
to the Direct Numerical Algorithm for Sub-Space State System Identification
(N4SID) [31]. DMDc extends DMD to scenarios where the input of the discrete system approximation is given by a functional while ioDMD additionally
handles the case when the system’s output is specified and also a functional.
To generically identify a system without prior knowledge on the relevant input functions, techniques such as persistent excitation [6] have been well known
for some time now. We propose an extension to the ioDMD method of [3] with
a new excitation variant related to the cross Gramian matrix [11]. Additionally,
since DMD-based identification does not guarantee that its resulting models
are stable, we propose a post-processing procedure to stabilize ioDMD-derived
models, by employing the nonsmooth constrained optimization method of [10]
to solve a corresponding stabilization problem.
This document is structured as follows. In Section 2, an overview of the
prerequisite dynamic mode decomposition method and its relevant variants is
given. Section 3 describes the new excitation strategy while our optimizationbased stabilization procedure is discussed in Section 4. Finally, two numerical
experiments are conducted in Section 5.
2
2
Dynamic Mode Decomposition
Consider the time-invariant ordinary differential equation (ODE):
ẋ(t) = f (x(t)),
(1)
with state x(t) ∈ RN and vector field f : RN → RN . Furthermore, consider for
now that (1) is sampled at uniform intervals for times t0 , . . . , tK . The most basic
version of DMD [27, 16] aims to approximate (1) by constructing a discrete-time
linear system
xk+1 = Axk ,
(2)
ep
rin
t
with a linear operator A ∈ RN ×N , such that if xk = x(tk ), then xk+1 ≈ x(tk+1 )
should also hold, for all k = 0, . . . , K − 1.
Starting at an initial state x0 , the sequence defined by (2) corresponds to a
trajectory of the state vectors xk . The corresponding data matrix X ∈ RN ×K
is simply the in-order concatenation of these state vectors, that is,
X = x0 x1 . . . x K .
By forming the following two partial trajectories:
X0 = x0 x1 . . . xK−1
X1 = x1 x2 . . . xK ,
and
where X0 ∈ RN ×K−1 is just the data matrix X with its last data point removed
while X1 ∈ RN ×K−1 is just X with its first data point removed, the idea behind
(plain) DMD [27] is to then solve the linear system of equations:
X1 = AX0 ,
Pr
which is in fact just (2), where xk+1 and xk have been replaced by X1 and X0 ,
respectively. A best-fit solution to this problem is given by the pseudoinverse:
A ≈ X1 X0+ ,
which is also the solution to minimizing the least-squares error in the Frobenius
norm (k · kF ):
A = arg min (kX1 − ÃX0 kF ).
Ã∈RN ×N
The DMD modes of (1) are given by the eigenvectors of the matrix A:
AΛ = ΛV.
Beyond just using a single uniformly-sampled time series, DMD has also
been generalized to a method called Exact DMD [28], which can additionally
support the concatenation of multiple different time series and/or non-uniform
time steppings. This generalization of DMD is made possible by reducing the
requirements of X0 and X1 to the lesser condition that their columns need only
be composed in pairs of data such that X1 (:, k) = f (X0 (:, k)) is fulfilled.
3
2.1
Practical Computation
We now give a high-level algorithmic description of DMD identification. The
pseudoinverse of the data matrix can be computed using the (rank-revealing)
singular value decomposition (SVD), X0 = UΣV ∗ , as follows:
X0+ = VΣ−1 U ∗ .
ep
rin
t
However, inverting tiny but nonzero singular values in the computed diagonal
matrix Σ poses a numerical problem, as these small singular values may be
inaccurate. Applying Σ−1 could overamplify components of X0 , in particular,
the less important ones. To counteract this effect, computing the pseudoinverse
via the SVD is done in practice by truncating any singular values that are
smaller than some fixed ε ∈ R+ , which is equivalent to adding a regularization
term to the least-squares solution for A:
A = arg min kX1 − ÃX0 k2F + βkÃk2F ,
Ã∈RN ×N
for some value of the regularization parameter β ∈ R+ . Following [28], the
algorithm for the DMD-based computation of the system matrix A, given a
state-space trajectory X and a lower bound ε > 0 for discarding tiny singular
values, is as follows:
1. Sample the state-space trajectory and form data matrix X = x0 . . . xK .
2. Assemble the shifted partitions X0 and X1 .
3. Compute the SVD of X0 = UΣV ∗ .
e ii = Σii if Σii ≥ ε; 0 otherwise.
4. Truncate Σ: Σ
e −1
5. Identify the approximate discrete system matrix: A := U ∗ X1 V Σ
Pr
In this DMD variant, the order (dimension) of the computed matrix A is equal
to the number of retained (nontruncated) singular values, but this truncation is
done solely for numerical accuracy; the intention is to keep as many components
of the dynamics as possible. In contrast, model order reduction typically aims
to significantly reduce the system down to just a small set of essential dynamics,
and accomplishing this goal will be the focus of Section 2.4.
2.2
Dynamic Mode Decomposition with Control
Considering systems whose vector field f depends not just on the state but also
on an input function u : R → RM :
ẋ(t) = f (x(t), u(t)),
(3)
has led to a DMD variant called Dynamic Mode Decomposition with Control
(DMDc) [23]. We focus on a specific DMDc variant [23, Sec. 3.3] that aims to
approximate (3) by a linear discrete-time control system:
xk+1 = Axk + Buk ,
4
(4)
where B ∈ RN ×M (called the input operator) must also be identified in addition
to A and input uk = u(tk ) is a discretely-sampled version of the continuous input
function u(t) for some sampling times given by tk . In addition to forming X
and X0 as in plain DMD (Section 2), an analogous construction of matrices for
input data is also done. An in-order concatenation of the series of input data
uk vectors is done to obtain matrix U ∈ RM ×K while the partial data matrix
U0 ∈ RM ×K−1 is simply U without its last column (the last input sample):
U = u0 u1 . . . uK−1 uk
and
U0 = u0 u1 . . . uK−1 .
ep
rin
t
Then, A and B can be obtained by computing the approximate solutions to the
linear system of equations given by:
X0
X1 = AX0 + BU0 = A B
,
U0
which is (4) with uk replaced by U , xk by X0 , and xk+1 by X1 , and has solution:
+
X0
A B = X1
.
U0
As mentioned in Section 1, DMDc is actually a special case of the KIC
method [24]. For KIC, the state of the system is also augmented with the discretized input uk , which leads the resulting augmented system to have additional
operators:
xk+1
A B
xk
=
,
uk+1
Bu Au uk
Pr
where Bu ∈ RM ×N and Au ∈ RM ×M . Of course, these two additional operators
must now also be identified along with matrices A and B. However, if one
assumes that the input is known and no identification of the associated operators
is required, then Bu and Au are just zero matrices, and it is thus clear that KIC
is a generalization of DMDc.
2.3
Input-Output Dynamic Mode Decomposition
Extending the underlying system once more to now also include an output
function y : R → RQ , with an associated output functional g : RN × RM → RQ
that also depends on the state x and the input u, yields the following system:
ẋ(t) = f (x(t), u(t)),
y(t) = g(x(t), u(t)).
(5)
Data-based modeling of systems of the form given by (5) has given rise to a class
of DMD methods called Input-Output Dynamic Mode Decomposition (ioDMD)
[3]. Similar to previously discussed DMD variants, the ioDMD method also
approximates the original system by a discrete-time linear system, but now
with input and output:
xk+1 = Axk + Buk ,
yk+1 = Cxk + Duk ,
5
(6)
where C ∈ RQ×N and D ∈ RQ×M are the output and feed-through operators,
respectively. Since this approximation includes output data, the discrete output
instances yk = y(tk ) ∈ RQ are also correspondingly arranged into a matrix
Y ∈ RQ×K by in-order concatenation while Y0 ∈ RQ×K−1 simply omits the last
column (output sample) of Y :
Y = y0 y1 . . . yK−1 yK
and
Y0 = y0 y1 . . . yK−1 .
B
D
X0
,
U0
(7)
ep
rin
t
Matrices A, B, C, and D are then approximated by solving:
(
X1 = AX0 + BU0
X1
A
or equivalently
=
Y
C
0
Y0 = CX0 + DU0
which is (6) but where uk , xk , xk+1 , and yk+1 have been replaced by U0 , X0 ,
X1 and Y0 , respectively, and has the solution:
A
C
+
B
X1 X0
=
.
D
Y0 U0
(8)
This procedure is equivalent to the Direct N4SID algorithm [30, 14, Ch. 6.6]
mentioned in Section 1.
Note that all the DMD variants discussed so far identify the original continuoustime systems by linear discrete-time models. However, one can create a continuousb B,
b C,
b D},
b that is a first-order approximation to the
time model given by {A,
discrete-time model obtained by ioDMD, using for example, the standard firstorder Euler method:
Pr
b k + Bu
b k)
xk+1 = xk + h(Ax
b k + hBu
b k
⇒ Axk + Buk = xk + hAx
i
h
b B
b = h(I − A) hB .
⇒ A
The output and feed-through operators for the continuous-time model remain
b = C,
the same as in the discrete-time model produced by ioDMD, that is, C
b = D. Finally, it is important to note that (io)DMD derived models are
D
generally only valid for the time horizon over which the data has been gathered.
2.4
Reduced Order DMD
To accelerate the computation of ioDMD-derived models, we follow [7, 4] and
reduce the order of the possibly high-dimensional state trajectories using a
projection-based approach. The data matrices X0 and X1 are compressed using
a truncated (Galerkin) projection Q ∈ RN ×n , n N , Q∗ Q = I:
∗
Q X1
Ar Br Q∗ X0
=
Y0
Cr Dr
U0
∗ ∗ +
Q X1 Q X0
Ar B r
=
.
Y0
U0
Cr Dr
6
The order of the identified system is thus determined by the rank of the projection.
A popular method to compute such a truncating projection Q is proper
orthogonal decomposition (POD) [12], which is practically obtained as the left
singular vectors (POD modes) of the data matrix X:
SVD
X = U DV ∗ → Xr = U1| X ≈ X.
3
ep
rin
t
The number of resulting computed
modes n depends on the desired proPPOD
n
2 1/2
jection error kX − U U ∗ XkF ≤ ( i=1 Dii
) , which is a consequence of the
Schmidt-Eckhard-Young-Mirsky Lemma (see for example [5]).
Note again that this data compression scheme has a very different motivation
compared to that of the aforementioned dimensionality-reduction done when
computing the pseudoinverse via the truncated SVD (discussed in Section 2.1).
The latter truncation, based on the magnitude of the singular values, is done for
reasons of numerical accuracy when computing the pseudoinverse (and applying
it in subsequent computations). The projection-based truncation discussed here,
using the sum of the singular values squared, allows for the possibility of a
much less onerous computational burden, as the state space of the models can
often be greatly reduced by discarding all but a handful of essential modes in
order to obtain a desired approximation error. Other projection-based datadriven model reduction techniques for the compression of the state trajectory
are similarly applicable, for example empirical balanced truncation [17].
Training Data and Generic Identification
Pr
DMD is a data-driven method, hence, the source of the data used for the system
identification needs to be considered. Usually it is possible to identify an inputoutput system (for provided discrete input, state, and output functions) so that
the identified system produces similar outputs given the input used for the
identification. To identify a model from data, the associated system needs to
produce outputs approximating the data source, not only for specific data sets,
but for a whole class of relevant input functions and perhaps even arbitrarily
admissible ones. For such generic linear system identification, the matrices A, B,
C, D have to be computed in such a manner that (a) does not require a specific
data set and (b) allows behaviors of the system being modeled to be predicted
for a given initial condition and input function. This can be accomplished, for
example, by persistent excitation (PE) [6, 23], which utilizes signals such as step
functions or Gaussian noise as training input data. Here, we propose a related
approach that also exploits random (Gaussian) sampling, yet is based on the
cross operator.
3.1
Cross Excitation
The cross operator [13] WX : RN → RN is a tool for balancing-type model reduction and encodes the input-output coherence of an associated system, which for
linear time-invariant systems is the so-called cross Gramian matrix [11]. This operator is defined as the composition of the controllability operator C : L2 → RN
7
with the observability operator O : RN → L2 :
WX = C ◦ O.
Thus, for a square system with the same input and output space dimension, WX
maps a given initial state x0 via the observability operator to an output function,
and is in turn passed to the controllability operator as an input function resulting
in a final state1 :
O
C
x0 7→ y 7→ x(T ).
ep
rin
t
To generate trajectory data, we modify the cross operator by replacing the
controllability operator with the input-to-state map ξ : L2 → L2 [5, Ch. 4.3].
This yields an operator WX : RN → L2 :
ξ
O
x0 7→ y 7→ x,
which maps, as before, an initial state to an output function, but then maps this
output (as an input) function to a state trajectory (instead of to a final state).
Compared to PE, the cross(-operator-based) excitation (CE) is a two-stage procedure using perturbations of the initial state to generate the excitation, as
opposed to perturbing the input directly.
The cross excitation is related to indirect identification of closed-loop systems
[29, 22], which is also a two-stage process. First, an intermediary system (openloop) system is identified, which then is used in the second step to generate
a signal that acts as an excitation for the identification of the actual closedloop system. A distinct difference of CE compared to the indirect closed-loop
identification approach is that the latter exclusively uses input-output data while
the former also uses state trajectory data in addition to the input-output data.
4
Stabilization
Pr
As DMD and its variants are time-domain methods (ioDMD included), it is
typically desired to preserve stability in the (reduced) identified discrete-time
systems. However, models derived by ioDMD are not guaranteed to be stable.
To enforce stability, an additional post-processing step is required. For example,
[2] proposed stabilizing models derived using Petrov-Galerkin projections by
solving a sequence of semidefinite programs. In this paper, we take a much
more direct approach.
A square matrix A is discrete-time stable if its spectral radius is less than
one, that is, ρ(A) < 1, where
ρ(A) := max{|λ| : λ ∈ Λ(A)}.
Although the spectral radius is nonconvex, it is a continuous function with
respect to the matrix A and furthermore, it is continuously differentiable almost
everywhere (in the mathematical sense). In other words, the set of points where
the spectral radius is nonsmooth only has measure 0 and so it holds that points
chosen randomly will, with probability 1, be outside of this set. Hence, despite
1 This
is also illustrated in [13, Fig. 1]
8
the nonsmoothness of the spectral radius, it is still possible to attain a wealth of
information from its gradient, since it is defined on all but a subset of measure
0 in the full space. Thus if the matrix A from the ioDMD-derived model,
that is, from (8), is not stable, we could consider employing a gradient-based
optimization method to stabilize it, while hopefully ensuring that the resulting
stabilized version of the ioDMD solution still models the original large-scale
system. In order to meet these two objectives, we consider solving the following
constrained optimization problem:
2
As Bs
X1
à B̃ X0
−
= arg min
s.t. |λ(Ã)| < 1 − τ, (9)
Cs Ds
Y0
C̃ D̃ U0 F
Ã,B̃,C̃,D̃
ep
rin
t
where à ∈ Rr×r , B̃ ∈ Rr×m , C̃ ∈ Rp×r , D̃ ∈ Rp×m , and τ ≥ 0 is a margin for
the stability tolerance. As the unstabilized ioDMD model is already a solution
to (7), solving (9) should promote solutions that are still close to the original
ioDMD model while simultaneously enforcing the requirement that these models
must now be stable, due to the presence of the stability radius in the inequality
constraint. Furthermore, the unstabilized ioDMD model should make a good
point to start the optimization method at.
There are however some difficulties with solving (9) iteratively via optimization techniques. The first is that the objective function of (9) is typically underdetermined in DMD settings, which can adversely impact a method’s usual rate
of convergence, as minimizers are no longer locally unique. However, as our goal
is mainly to stabilize the ioDMD model, without changing its other properties
too much, we do not necessarily need to solve (9) exactly. A few iterations may
be all that is needed to accomplish this task.
As an alternative, one could consider solving
2
ADMD BDMD
à B̃
s.t. |λ(Ã)| < 1 − τ
min
−
CDMD DDMD
C̃ D̃ F
Ã,B̃,C̃,D̃
Pr
in lieu of (9), where ADMD , BDMD , CDMD , and DDMD are the matrices produced
by ioDMD. On the upside, this helps to avoid the problem of underdeterminedness arising in (9) and encourages that a stable solution close to the original
ioDMD-derived system is found. However, this modified objective no longer
takes any observed data into account. We did evaluate this alternate optimization problem in our experiments, but the performance of the models it produced
was sometimes worse. As such, we will only report results for our experiments
done using (9) in Section 5.
The second issue for trying to solve (9) is that the spectral radius can be
a rather difficult function to optimize, relatively speaking. First, despite being
continuously differentiable almost everywhere, the spectral radius is still a nonsmooth function, specifically at matrices which have multiple eigenvalues that
attain the maximum modulus, that is, the value of the spectral radius. Typically, minimizers of the spectral radius will be at such matrices (for example,
see the plots of spectral configurations post optimization in [10, Section 4.1
and Appendix 9.1]), so optimizing the spectral radius often means that an optimization method must try to converge to a nonsmooth minimizer, a difficult
prospect. Worse still is that the spectral radius is also non-locally Lipschitz at
matrices where these multiple eigenvalues attaining the value of the spectral radius coincide (see [8]). Many of the available continuous optimization methods
9
5
ep
rin
t
are designed under the assumption that functions they optimize are smooth or
if not, at least locally Lipschitz. If the functions being optimized do not meet
these criteria, these methods typically break down. Furthermore, the nonconvexity of the spectral radius means that optimization codes may get stuck in
local minima that may or may not provide sufficiently acceptable solutions.
Although the inclusion of the spectral radius constraint makes (9) a nonsmooth, nonconvex optimization problem, with a non-locally-Lipschitz constraint
function, again we do not necessarily need to solve it exactly (though this will
remain to be seen in our experiments). Furthermore, while much of the literature on nonsmooth optimization has historically focused on unconstrained problems, there has also been recent interest in addressing problems with nonsmooth
constraints. For example, a new algorithm combining quasi-Newton BFGS
(Broyden-Fletcher-Goldfarb-Shanno) updating and SQP (sequential quadratic
programming)2 was recently proposed for general nonsmooth, nonconvex, constrained optimization problems [10], where no special knowledge or structure is
assumed about the objective or constraint functions. Particularly relevant to
our approach here, this BFGS-SQP method was evaluated on 100 different spectral radius constrained optimization problems, with promising results relative
to other solvers [10, Section 6]. This indicates that it may also be a good solver
for our nonsmooth constrained optimization problem and thus we propose using it to solve (9). Specifically, we use GRANSO: GRadient-based Algorithm
Non-Smooth Optimization [20], an open-source MATLAB code implementing
the aforementioned BFGS-SQP method of [10].
Numerical Results
Pr
We implemented our new ioDMD variant using both PE and CE sampling
strategies (presented in Section 3.1) to collect observations of the original system’s behaviors. Furthermore, our software can also optionally stabilize the
resulting ioDMD-derived models by using GRANSO [10] on our associated nonsmooth constrained optimization problem.
As discussed in Section 4, (9) is an underdetermined optimization problem in DMD settings. Since such problems are in a sense quite flat, the norm
of the gradient merely being small can be a poor measure of when to terminate; correspondingly, we tightened GRANSO’s default termination tolerance
of 10−6 to 10−8 (i.e. opts.opt_tol = 1e-8). Relatedly, convergence can also
be slow so the choice of a starting point can also be critical. As our goal,
specified by (9), is to stabilize a model while minimizing the tradeoff in any
increased approximation error (that may occur due to stabilization), we simply
used the unstable ioDMD-derived models as starting points for GRANSO and
used GRANSO’s custom termination feature to halt optimization once a model
had been found that was both stable (for (9) we used τ := 0) and had an objective value that was less than 1000 times the objective function evaluated at the
original ioDMD-derived unstable model. We found that this loose limit on how
much the objective function was allowed to increase was more than adequate to
retain good output errors. We ran GRANSO using its full-memory BFGS mode
2 For a good introductory reference on many of the optimization techniques referred to in
this paper, see [21].
10
(its default behavior and the recommended choice for nonsmooth problems) and
kept all other GRANSO options at their default values as well.
The numerical experiments were implemented in the Matlab language and
were run under MATLAB R2017a on a workstation computer with an Intel Core
i7-6700 (4 Cores @ 3.4 GHz) and 8 GB memory.
5.1
Excitation and Stabilization Evaluation
ep
rin
t
We demonstrate the effects of different types of excitation used for the ioDMDbased system identification by a numerical example. Specifically, for a given
target data set, we identify a discrete-time linear system first using the target
data itself, second by persistent excitation (PE) and third by utilizing the herein
proposed cross excitation (CE) from Section 3.1.
The considered input-output system is based on the transport equation, with
the left boundary of the domain selected as input and the right boundary as
output:
∂
∂
z(x, t) = a z(x, t),
∂t
∂x
z(0, t) = u(t),
x ∈ [0, 1],
z(x, 0) = 0,
y(t) = z(1, t).
The partial differential equation is discretized in space using the first-order
1
finite-difference upwind scheme, with a spatial resolution of ∆x = 1000
and
a = 1.3. The resulting ODE input-output system is then given by:
ẋ(t) = Ax(t) + Bu(t),
y(t) = Cx(t).
(10)
Pr
The target data is given by discrete input, state and output functions from a
simulation, which is performed by a first order implicit Runge-Kutta method.
1
We used a time-step width of ∆t = 1000
and a time horizon of T = 1, with a
1
1 2
Gaussian-bell type input function û(t) = e− 1000 (t− 10 ) .
For the comparison, a discrete-time linear system was first identified from the
1000 snapshots generated by a simulation of (10) excited by input û, to obtain a
baseline for the modeling performance of ioDMD. The PE variant was sampled
using zero-mean, unit-covariance Gaussian noise and a unit step function input
was used for the ioDMD-based identification. Finally, our CE variant had the
initial state sampled from a unit Gaussian distribution and a (component-wise)
shifted initial state x0,i = 1 was tested. All methods were tested with ioDMD
only and then also with stabilization.
5.1.1
ioDMD without Stabilization
Fig. 1 depicts the relative output error for simulations of systems identified
using data associated to the target input û, PE, and CE for increasingly accurate state-space data compression (that is, for increasingly smaller amounts
of compression). The state-space data dimensionality was reduced using the
POD method with prescribed projection errors of 10−1 , . . . , 10−8 . In this set of
experiments, we did not use stabilization.
11
10 0
10 -1
Output Error
Output Error
10 0
10 -2
Target
PE
CE
0.01
0.0001
1e-06
Projection Error
1e-08
(a) Non-regularized identification by
noise input or randomly sampled initial
state.
10 -1
10 -2
0.01
1e-08
10 0
Target
PE
CE
10 -3
0.0001
1e-06
Projection Error
(b) Non-regularized identification by step
input or shifted initial state.
Output Error
Output Error
10 0
Target
PE
CE
10 -10
ep
rin
t
10 -3
10 -5
0.01
0.0001 1e-06
Projection Error
Target
PE
CE
10 -10
1e-08
(c) Regularized identification by noise input or randomly sampled initial state.
10 -5
0.01
0.0001
1e-06
Projection Error
1e-08
(d) Regularized identification by step input or shifted initial state.
Pr
Figure 1: First numerical experiment (Section 5.1.1): Non-stabilized ioDMDbased system identification. The plots show the identified (reduced order) system’s output error compared to the original system’s output for varying accuracies of the POD projection-based model reduction.
In Figs. 1a and 1b, the ioDMD procedure was not regularized by truncating
singular values, while regularization Σii < 10−5 was used in Figs. 1c and 1d. In
Figs. 1a and 1c, system identification was performed using zero-mean Gaussian
noise for the PE and an initial state sampled from a zero-mean Gaussian distribution for CE, respectively. In Figs. 1b and 1d, respectively, the identification
was driven by a step input for the PE and a shifted initial state x0,i = 1 for CE.
For this set of experiments, using the target data only produced stable systems for large values of acceptable projection error while the PE-derived models
were always unstable (and thus had very poor performance regardless of the projection error). In contrast, the CE method produced stable systems for all levels
of projection error tested. Performance-wise, when comparing to the few targetdata-derived models that also happened to be stable, the CE-derived models had
12
errors that were less than one order of magnitude higher (see Figs. 1a and 1c).
On the other hand, when using the step input or shifted initial state, both PE
and CE produced models with increasing accuracy as the level of acceptable projection error of the data was decreased, as seen in Figs. 1b and 1d. In Fig. 1d,
we see that the regularization limited the attainable accuracy for both PE and
CE. The target-data-derived system had a constant error independent from the
projection error of the data.
5.1.2
ioDMD with Stabilization
Pr
ep
rin
t
In this second set of experiments, Fig. 2 still shows the relative output error
for simulations of systems identified using the target data, PE, and CE for
increasingly accurate state-space data compression, but now the systems have
been post-processed using our optimization-based approach to enforce stability,
as necessary. The state-space data dimensionality was again reduced using
the POD method, with prescribed projection errors of 10−1 , . . . , 10−8 . The
subfigures are arranged as they were in Fig. 1.
The step function PE and shifted initial state CE are unaffected by our
stabilization post-processing phase, as these systems were already stable; thus
their plots are the same in Figs. 2b and 2d as they were in Figs. 1b and 1d. In the
case of using Gaussian noise or randomly sampled initial state (Figs. 2a and 2c),
which had not yielded stable systems for the target data or PE (either with or
without regularization), our optimization-based post-processing procedure now
enforced stability.
For the ioDMD-derived models that were unstable, GRANSO was able to
stabilize all 24 of them. The average number of iterations needed to find the first
stabilized version was 13.5 while the maximum was just 61. Furthermore, for
12 of the problems, our full termination criteria were met in less than 20 iterations while the average and maximum iteration counts over all 24 problems were
respectively 84.2 and 329. This demonstrates that our optimization-based approach is indeed able to stabilize such models reliably and efficiently. Solving (9)
via GRANSO also met our secondary goal, that stabilization is achieved without
deviating too significantly from the original unstable models. The largest observed relative change between an initial unstable model and its corresponding
GRANSO-derived stabilized version was just 1.44% while the average observed
relative change was merely 0.231%; the relative differences were calculated by
comparing vec[As , Bs ; Cs , Ds ], where the matrices are GRANSO’s computed stabilized solution to (9), to a similar vec of the original (unstable) ioDMD-derived
model.
5.1.3
Reduced Orders and Runtimes
We now compare the order of the identified systems. The order of the identified
system is determined by the projection error selected for the state-space compression in the POD. For each data set (Target, Gaussian noise PE, Gaussian
sample CE, Step input PE, shifted initial state CE), the resulting reduced order
is plotted for the different prescribed projection errors in Fig. 3a. As we can see,
the step-input PE and shifted-initial-state CE method behave similar in terms
of system dimension while for the Gaussian-noise PE and Gaussian-sampled
initial-state CE, the CE variant resulted in smaller system dimensions.
13
10 0
10 -1
Output Error
Output Error
10 0
10 -2
Target
PE
CE
0.01
0.0001
1e-06
Projection Error
1e-08
(a) Non-regularized identification by
noise input or randomly sampled initial
state.
10 -1
10 -2
0.01
1e-08
10 0
Target
PE
CE
10 -3
0.0001
1e-06
Projection Error
(b) Non-regularized identification by step
input or shifted initial state.
Output Error
Output Error
10 0
Target
PE
CE
10 -10
ep
rin
t
10 -3
10 -5
0.01
0.0001
1e-06
Projection Error
Target
PE
CE
10 -10
1e-08
(c) Regularized identification by noise input or randomly sampled initial state.
10 -5
0.01
0.0001
1e-06
Projection Error
1e-08
(d) Regularized identification by step input or shifted initial state.
Pr
Figure 2: Second numerical experiment (Section 5.1.2): Stabilized ioDMD-based
system identification. The plots show the identified (reduced order) system’s
output error compared to the original system’s output for varying accuracies of
the POD projection-based model reduction.
In terms of computational cost, only the Target and Gaussian-noise PE
variants required stabilization and as such, it is only for these that we see
increased runtimes, as shown by the red and green plots (respectively) in Fig. 3b.
Otherwise, the runtimes were mostly identical for the other variants in the
comparison.
5.2
Limited-Memory BFGS for Stabilization?
One potential downside to our optimization-based stabilization procedure is
that full-memory BFGS inverse Hessian updating (GRANSO’s recommended
setting for nonsmooth problems) requires per-iteration work and storage that is
quadratic in the number of optimization variables. As the number of optimiza-
14
10 2
100
System Dimension
90
80
Stabilization Runtime
Target
PE (Gauss)
CE (Gauss)
PE (Step)
CE (Shift)
70
60
50
40
10 1
10 0
10 -1
0.01
0.0001 1e-06
Projection Error
1e-08
0.01
0.0001
1e-06
Projection Error
1e-08
ep
rin
t
30
Target
PE (Gauss)
CE (Gauss)
PE (Step)
CE (Shift)
(a) Comparison of the identified system’s
reduced order for different pre-selected
mean L2 projection errors used in both
experiments (see Section 5.1.1).
(b) Comparison of runtimes in seconds for
the stabilization procedure in the second
experiment (Section 5.1.2).
Figure 3: Comparison of reduced identified system orders and stabilization runtimes.
Pr
tion variables is (r + m) × (r + p), the running time required to solve (9) could
become unacceptably long as r, the reduced order of the model, is increased.
Thus, we now also consider whether or not limited-memory BFGS updating can
also be effective for solving (9).
Using limited-memory BFGS has the benefit that the per-iteration work
and storage is reduced to a linear amount (again in the number of optimization
variables). However, one of the tradeoffs is that convergence can be quite slow
in practice. For smooth problems, full-memory BFGS converges superlinearly
while limited-memory BFGS only does so linearly; on nonsmooth problems,
linear convergence is the best one can typically expect. Another potential issue
is that while there has been much evidence supporting that full-memory BFGS
is very reliable for nonsmooth optimization, the case for using limited-memory
BFGS on nonsmooth problems is much less clear; for a good overview of the
literature on this topic, see [18, Section 1].
To investigate this question, we reran the experiments from Section 5.1 a
second time but where GRANSO’s limited-memory BFGS mode was now enabled. Specifically, we configured GRANSO to approximate the inverse Hessian
at each iteration using only the 10 most recently computed gradients, accomplished by setting opts.limited_mem_size = 10. All other parameters of our
experimental setup were kept as they were described earlier.
In this limited-memory configuration, GRANSO often required significantly
more iterations, as one might expect. The average and max number of iterations
to find the first stable version of a model were respectively 73.6 and 474, about
an order of magnitude more iterations than incurred when using full-memory
BFGS. On the other hand, for 19 of the 24 problems, stable models were encountered within the first 20 iterations. To meet our full termination criteria, the
average and max number of iterations were respectively 222.0 and 811, roughly
about two and a half times more than incurred when using full-memory BFGS.
15
6
ep
rin
t
Nevertheless, 12 of the 24 problems were still satisfactorily solved in less than
20 iterations, matching the earlier result when using full-memory BFGS. Despite the large increases in iteration counts, GRANSO’s overall runtime was on
average 3.83 times faster when enabling limited-memory BFGS.
With respect to output error in this limited-memory evaluation, the resulting
stabilized models still essentially matched the earlier results using full-memory
BFGS. There was one notable exception, for Target data using the smallest
projection error of 10−8 , where GRANSO remarkably found a better-performing
model when using limited-memory BFGS. However, we did observe that the
quality of the stabilized models appeared to be much more sensitive to changing
GRANSO’s parameters than they were when using full-memory BFGS. As a
consequence, we still advocate that solving (9) with GRANSO is generally best
done using its default full-memory BFGS updating. Nonetheless, if this is simply
not feasible computationally, one may still be able to obtain good results using
limited-memory BFGS but perhaps not as reliably or consistently.
As a final clarifying remark on this topic, we note that one cannot necessarily expect good performance on nonsmooth problems when using just any
BFGS-based optimization code and that generally it is critical that the choice
of software is one specifically designed for nonsmooth optimization. Indeed, this
is highlighted in the evaluation done in [10, Section 6], where off-the-shelf quasiNewton-based codes built for smooth optimization perform much worse on a test
set of nonsmooth optimization problems compared to the quasi-Newton-based
codes specifically built with nonsmooth optimization in mind.
Conclusion
Pr
In this work, we evaluated the approximation quality of ioDMD system identification using a novel excitation scheme and a new optimization-based, postprocessing procedure to ensure stability of the identified systems. Our new
cross excitation strategy, particularly when used with random sampling, often
produces better results than when using persistent excitation, and our experiments indicate that both excitation schemes are useful for efficiently obtaining
good models for approximating the target data. Furthermore, we show that
directly solving a nonsmooth constrained optimization problem can indeed be
a viable approach for stabilizing ioDMD-derived systems while retaining the
salient properties for approximating the output response.
Code Availability
The source code of the presented numerical examples can be obtained from:
http://runmycode.org/companion/view/2902
and is authored by: Christian Himpe and Tim Mitchell.
16
References
[1] A. Alla and J. N. Kutz. Nonlinear model order reduction via dynamic
model decomposition. SIAM J. Sci. Comput., 39(5):B778–B796, 2017. doi:
10.1137/16M1059308.
[2] D. Amsallem and C. Farhat. Stabilization of projection-based reducedorder models. Numerical Methods in Engineering, 91(4):358–377, 2012.
doi:10.1002/nme.4274.
ep
rin
t
[3] J. Annoni, P. Gebraad, and P. Seiler. Wind farm flow modeling using inputoutput dynamic mode decomposition. In American Control Conference
(ACC), pages 506–512, 2016. doi:10.1109/ACC.2016.7524964.
[4] J. Annoni and P. Seiler. A method to construct reduced-order parametervarying models. International Journal of Robust and Nonlinear Control,
27(4):582–597, 2017. doi:10.1002/rnc.3586.
[5] A. C. Antoulas. Approximation of Large-Scale Dynamical Systems, volume 6 of Adv. Des. Control. SIAM Publications, Philadelphia, PA, 2005.
doi:10.1137/1.9780898718713.
[6] K. J. Åström and P. Eykhoff. System identification – a survey. Automatica,
7(2):123–162, 1971. doi:10.1016/0005-1098(71)90059-8.
[7] B. W. Brunton, L. A. Johnson, J. G. Ojemann, and J. N. Kutz. Extracting
spatial-temporal coherent patterns in large-scale neural recordings using
dynamic mode decomposition. Journal of Neuroscience Methods, 258:1–
15, 2016. doi:10.1016/j.jneumeth.2015.10.010.
[8] J. V. Burke and M. L. Overton. Variational analysis of non-Lipschitz
spectral functions. Math. Program., 90(2, Ser. A):317–352, 2001. doi:
10.1007/s102080010008.
Pr
[9] K. K. Chen, J. H. Tu, and R. W. Rowley. Variants of dynamic mode decomposition: Boundary condition, Koopman, and Fourier analyses. Nonlinear
Science, 22(6):887–915, 2012. doi:10.1007/s00332-012-9130-9.
[10] F. E. Curtis, T. Mitchell, and M. L. Overton. A BFGS-SQP method for
nonsmooth, nonconvex, constrained optimization and its evaluation using
relative minimization profiles. Optim. Methods Softw., 32(1):148–181, 2017.
doi:10.1080/10556788.2016.1208749.
[11] K. V. Fernando and H. Nicholson. On the structure of balanced and other
principal representations of SISO systems. IEEE Trans. Autom. Control,
28(2):228–231, 1983. doi:10.1109/TAC.1983.1103195.
[12] P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley. Turbulence,
coherent structures, dynamical systems and symmetry. Cambridge Monographs on Mechanics. Cambridge University Press, Cambridge, 2012. doi:
10.1017/CBO9780511919701.
17
[13] T. C. Ionescu, K. Fujimoto, and J. M. A. Scherpen. The cross operator
and the singular value analysis for nonlinear symmetric systems. In Proc.
Eur. Control Conf., pages 1565–1570, 2009. URL: http://hdl.handle.
net/11370/1bcfd3be-c006-49e0-bbbe-5c09ca055d87.
[14] K. Katayama. Subspace Methods for System Identification. Communications and Control Engineering. Springer London, 2005. doi:10.1007/
1-84628-158-X.
[15] B. O. Koopman. Hamiltonian systems and transformation in Hilbert space.
Proceedings of the National Academy of Sciences, 17(5):315–381, 1931.
URL: http://www.pnas.org/content/17/5/315.full.pdf.
ep
rin
t
[16] J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor. Dynamic
Mode Decomposition: Data-Driven Modeling of Complex Systems. Society
of Industrial and Applied Mathematics, Philadelphia, USA, 2016. doi:
10.1137/1.9781611974508.
[17] S. Lall, J. E. Marsden, and S. Glavaški. Empirical model reduction of controlled nonlinear systems. In Proc. of the IFAC World Congress, volume F,
pages 473–478, 1999. URL: resolver.caltech.edu/CaltechAUTHORS:
20101007-154754737.
[18] A. S. Lewis and M. L. Overton. Nonsmooth optimization via quasi-Newton
methods. Math. Program., 141(1–2, Ser. A):135–163, 2013. doi:10.1007/
s10107-012-0514-2.
[19] I. Mezic. Spectral properties of dynamical systems, model reduction and
decompositions. Nonlinear Dynamics, 41(1):309–325, 2005. doi:10.1007/
s11071-005-2824-x.
[20] T. Mitchell. GRANSO: GRadient-based Algorithm for Non-Smooth Optimization. http://timmitchell.com/software/GRANSO. See also [10].
Pr
[21] J. Nocedal and S. J. Wright. Numerical Optimization. Springer, New York,
1999. doi:10.1007/b98874.
[22] H. Oku and T. Fujii. Direct subspace model identification of LTI systems
operating in closed-loop. In 43rd IEEE Conference on Decision and Control, pages 2219–2224, 2004. doi:100.1109/CDC.2004.143378.
[23] J. L. Proctor, S. L. Brunton, and J. N. Kutz. Dynamic mode decomposition
with control. SIAM J. Applied Dynamical Systems, 15(1):142–161, 2016.
doi:10.1137/15M1013857.
[24] J. L. Proctor, S. L. Brunton, and J. N. Kutz. Generalizing Koopman
theory to allow for inputs and control. arXiv e-prints 1602.07647, Cornell
University, 2016. math.OC. URL: https://arxiv.org/abs/1602.07647.
[25] C. W. Rowley and S. T. M. Dawson. Model reduction for flow analysis
and control. Annual Review of Fluid Mechanics, 49:387–417, 2017. doi:
10.1146/annurev-fluid-010816-060042.
18
[26] C. W. Rowley, I. Mezic, S. Bagheri, P. Schlatter, and D. S. Henningson.
Spectral analysis of nonlinear flows. Journal of Fluid Mechanics, 641:115–
1127, 2009. doi:10.1017/S0022112009992059.
[27] P. J. Schmid. Dynamic mode decomposition of numerical and experimental
data. J. Fluid Mech., 656:5–28, 2010. doi:10.1017/S0022112010001217.
[28] J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, and J. N.
Kutz. On dynamic mode decomposition: Theory and applications. Journal
of Computational Dynamics, 1(2):391–421, 2014. doi:10.3934/jcd.2014.
1.391.
ep
rin
t
[29] P. M. J. Van Den Hof and R. J. P. Schrama. An indirect method for transfer
function estimation from closed loop data. Automatica, 29(6):1523–1527,
1993. doi:10.1016/0005-1098(93)90015-L.
[30] P. Van Overschee and B. de Moor. N4SID: numerical algorithms for state
space subspace system identification. In IFAC Proceedings Volumes, volume 26, pages 55–58, 1993. doi:10.1016/S1474-6670(17)48221-8.
Pr
[31] M. Viberg. Subspace-based methods for the identification of linear timeinvariant systems. Automatica, 31(12):1835–1851, 1995. doi:10.1016/
0005-1098(95)00107-5.
19
| 3cs.SY
|
Message-passing algorithms for synchronization problems
over compact groups
Amelia Perry∗†1 , Alexander S. Wein∗‡1 , Afonso S. Bandeira§3 , and Ankur Moitra¶1,2
arXiv:1610.04583v1 [cs.IT] 14 Oct 2016
1
2
Department of Mathematics, Massachusetts Institute of Technology
Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology
3
Department of Mathematics and Center for Data Science, Courant Institute of
Mathematical Sciences, New York University, NY, USA
October 17, 2016
Abstract
Various alignment problems arising in cryo-electron microscopy, community detection, time synchronization, computer vision, and other fields fall into a common framework of synchronization problems over
compact groups such as Z/L, U (1), or SO(3). The goal of such problems is to estimate an unknown vector
of group elements given noisy relative observations. We present an efficient iterative algorithm to solve a
large class of these problems, allowing for any compact group, with measurements on multiple ‘frequency
channels’ (Fourier modes, or more generally, irreducible representations of the group). Our algorithm is
a highly efficient iterative method following the blueprint of approximate message passing (AMP), which
has recently arisen as a central technique for inference problems such as structured low-rank estimation
and compressed sensing. We augment the standard ideas of AMP with ideas from representation theory
so that the algorithm can work with distributions over compact groups. Using standard but non-rigorous
methods from statistical physics we analyze the behavior of our algorithm on a Gaussian noise model,
identifying phases where the problem is easy, (computationally) hard, and (statistically) impossible. In
particular, such evidence predicts that our algorithm is information-theoretically optimal in many cases,
and that the remaining cases show evidence of statistical-to-computational gaps.
1
Introduction
Among the most common data problems in the sciences is that of recovering low-rank signal present in a
noisy matrix. The standard tool for such problems is principal component analysis (PCA), which estimates
the signal by the top eigenvectors. One example out of many is in macroeconomics, where large, noisy
correlation matrices reveal useful volatility and yield predictions in their top eigenvectors [LS91, ELW10].
However, many particular applications involve extra structure such as sparsity in the signal, and this structure
is ignored by conventional PCA, leading to sub-optimal estimates. Thus, a major topic of recent interest in
∗ The
first two authors contributed equally.
[email protected]. This work is supported in part by NSF CAREER Award CCF-1453261 and a grant from
the MIT NEC Corporation.
‡ Email: [email protected]. This research was conducted with Government support under and awarded by DoD, Air Force
Office of Scientific Research, National Defense Science and Engineering Graduate (NDSEG) Fellowship, 32 CFR 168a.
§ Email: [email protected]. A.S.B. was supported by NSF Grant DMS-1317308. Part of this work was done while
A.S.B. was with the Department of Mathematics at the Massachusetts Institute of Technology.
¶ Email: [email protected]. This work is supported in part by NSF CAREER Award CCF-1453261, a grant from the MIT
NEC Corporation and a Google Faculty Research Award.
† Email:
1
machine learning is to devise efficient algorithms for sparse PCA [AW08, BR13b], non-negative PCA [MR16],
general Bayesian PCA with a prior [Bis99], and other variants. These problems pose a major computational
challenge. While significant advances have appeared, it is also expected that there are fundamental gaps
between what is statistically possible and what can be done efficiently [BR13a, LKZ15a, LKZ15b, MW15],
and thus carried out in practice on very large datasets that are now prevalent.
A number of low-rank recovery problems involve a significant amount of symmetry and group structure.
A compelling example is the orientation problem in cryo-electron microscopy (cryo-EM), where one is given
many noisy 2D images of copies of an unknown molecule, each in different unknown 3D orientations. The
goal is to estimate the orientations, in order to assemble the images into an estimate of the molecule structure
[SS11]. Thus, one is tasked with learning elements gu of SO(3), one for each image u, based on some loss
function derived from the observed images. Moreover, this loss function has a symmetry: it depends only
on the relative alignments gu gv−1 , as there is no a priori reference frame. One previous approach to this
problem, due to [CSSS10, SS11], produces a matrix of pairwise image comparisons, and then attempts to
extract the rotations gu from the top eigenvectors of this matrix. However, it is reasonable to imagine that
this approach could be significantly sub-optimal: PCA does not exploit the significant group structure of
the signal.
Many problems with similar patterns of group symmetry have been previously studied under the general
heading of synchronization problems. In general, a synchronization problem asks us to recover a vector of
group elements given noisy pairwise measurements of the relative group elements gu gv−1 , and PCA-based
‘spectral methods’ are among the most common techniques used for such problems. Singer [Sin11] introduced
a PCA approach for angular synchronization, a 2D analogue of the problem above with symmetry over SO(2),
in which one estimates the orientations of noisy, randomly rotated copies of an unknown image. Cucuringu
et al. [CLS12] applied a similar approach to a sensor localization problem with synchronization structure over
the Euclidean group Euc(2). The problem of detecting two subcommunities in a random graph can be viewed
as synchronization over Z/2 [Ban15], and spectral methods have a long history of use in such community
detection and minimum cut problems (e.g. [McS01]). Further instances of synchronization appear in time
synchronization in networks [GK06], computer vision [ARC06], optics [RW01], and alignment in signals
processing [BCSZ14].
Further work on synchronization has focused on ways to better exploit the group structure. One method
used in practice for cryo-EM and related problems is alternating minimization, which alternates between
estimating the rotations by aligning the images with a previous guess of the molecule structure, and then
estimating the molecule structure from the images using these rotations. This method only appears to succeed
given a strong initial guess of the molecule structure, and then it is unclear whether the final estimate mainly
reflects the observations or simply the initial guess, leading to a problem of model bias; see e.g. [Coh13]. In
this paper we are interested in de novo estimation without a substantial initial guess, steering clear of this
pitfall.
Convex relaxations of maximum likelihood estimators have shown promise, such semidefinite relaxation
for angular synchronization introduced by Singer [Sin11] and proven tight by Bandeira et al. [BBS14],
or in semidefinite programs for community detection and correlation clustering problems that constitute
Z/2-synchronization (e.g. [GW95, ABH16, HWX16, ABBS14, MS16]). Indeed, a very general semidefinite
relaxation for synchronization problems was introduced by Bandeira et al. [BCS15], but its performance
remains unclear even empirically: while this convex program can be solved in polynomial time, it is large
enough to make experiments or application difficult.
An alternate approach is an iterative method due to Boumal [Bou16], for the following Gaussian model
of angular synchronization: one wants to estimate a vector x ∈ Cn whose entries are unit-norm complex
numbers (standing in for 2D rotations), given a matrix
Y =
1
λ ∗
xx + √ W,
n
n
where λ is a signal-to-noise ratio (SNR) parameter, W is a GUE matrix (independent complex Gaussians up
to Hermitian symmetry), and ∗ denotes conjugate transpose. One could perform ordinary PCA by initializing
2
with a small random guess v and repeatedly assigning v ← Y v; this is the method of power iteration. Instead,
Boumal proposes1 to iterate v ← f (Y v), where f divides each entry by its norm, thus projecting to the
unit circle. This method is highly efficient, and is moreover observed to produce a better estimate than
PCA once the signal-to-noise parameter λ is sufficiently large. However, while PCA produces a nontrivial
estimate for all λ > 1 (e.g. [FP07, BGN11]), this projected power method does not appear to produce a
meaningful √
estimate until λ is somewhat larger. (In fact, a heuristic analysis similar to Section 6.3 suggests
that λ > 2/ π ≈ 1.128 is required.) This behavior suggests that some iterative method combining the best
features of PCA and the projected power method might outperform both statistically, while remaining very
efficient. More importantly, we are motivated to find analogous iterative methods for groups other than
U (1), and for more complicated observation models.
One such observation model is as follows. Instead of observing only the matrix Y as above, suppose we
are given matrices corresponding to different Fourier modes:
Y1 =
Y2 =
λ1 ∗
1
xx + √ W1 ,
n
n
1
λ2 2 2 ∗
x (x ) + √ W2 ,
n
n
..
.
YK =
λK K K ∗
1
x (x ) + √ WK ,
n
n
where xk denotes entrywise power, and Wk are independent GUE matrices. With a PCA-based approach,
it is not clear how to effectively couple the information from these matrices to give a substantially better
estimate than could be derived from only one. The “non-unique games” semidefinite program of [BCS15] is
able to use data from an observation model such as this, but it is not yet empirically clear how it performs.
Can we hope for some very efficient iterative algorithm to strongly leverage data from multiple ‘frequencies’
or ‘channels’ such as this?
This question has more than abstract relevance: due to Fourier theory, a very large class of measurement
models for U (1)-synchronization decomposes into matrix-based observations on different frequencies, in a
manner resembling the model above. Moreover, an analogous statement holds over SO(3), and over other
compact groups, with Fourier theory replaced by the noncommutative setting of representation theory.
Thus, the aforementioned spectral approach to cryo-EM applies PCA to only the lowest frequency part
of the observations; an algorithm that can use all frequencies effectively might demonstrate dramatically
improved statistical performance on cryo-EM.
In this paper we present an iterative algorithm to meet the challenges above. Our algorithm aims to
solve a general formulation of the synchronization problem: it can apply to multiple-frequency problems for
a large class of observation models, with symmetry over any compact group. Our approach is statistically
powerful, empirically providing a better estimate than both PCA and the projected power method on U (1)synchronization, and leveraging multiple frequencies to give several orders of magnitude improvement in
estimation error in experiments (see Figures 3 and 4). Indeed, we conjecture based on ideas from statistical
physics that in many regimes our algorithm is statistically optimal, providing a minimum mean square error
(MMSE) estimator asymptotically as the matrix dimensions become infinite (see Section 8). Finally, our
approach is highly efficient, with each iteration taking time linear in the (matrix) input, and with roughly
15 iterations sufficing for convergence in experiments.
Our algorithm follows the framework of approximate message passing (AMP), based on belief propagation
on graphical models [Pea86] and the related cavity method in statistical physics [MPV86]. Following a general
blueprint, AMP algorithms have previously been derived and analyzed for compressed sensing [DMM09,
DMM10, BM11, JM13], sparse PCA [DM14], non-negative PCA [MR16], cone-constrained PCA [DMR14],
1 Projected
power methods have appeared earlier in the literature, for instance, in [MR16].
3
planted clique [DM15] and general structured PCA [RF12]. In fact, AMP has already been derived for Z/2synchronization under a Gaussian observation model [DAM16], and our algorithm will generalize this one to
all compact groups. A striking feature of AMP is that its asymptotic performance can be captured exactly
by a particular fixed-point equation called state evolution, which has enabled the rigorous understanding
of its performance on some problems [BM11, JM13]. AMP is provably statistically optimal in many cases,
including Gaussian Z/2 synchronization (modulo a technicality whereby the proof supposes a small warmstart) [DAM16].
AMP algorithms frequently take a form similar to the projected power method of Boumal described
above, alternating between a matrix–vector product with the observations and an entrywise nonlinear transformation, together with an extra ‘Onsager’ correction term. In the case of Z/2- or U (1)-synchronization,
we will see that the AMP derivation reproduces Boumal’s algorithm, except with the projection onto the
unit circle replaced by a soft, sigmoid-shaped projection function to the unit disk (see Figure 1), with the
magnitude maintaining a quantitative measure of confidence. Integrating the usual AMP blueprint with
the representation theory of compact groups, we obtain a broad generalization of this method, to synchronization problems with multiple frequencies and noncommutative groups such as SO(3). In full generality,
the nonlinear transformation has a simple interpretation through representation theory and the exponential
function.
One drawback of our approach is that although we allow for a very general observation model, we do insist
that the noise on each pairwise measurement is independent. This fails to capture certain synchronization
models such as multireference alignment [BCSZ14] and cryo-EM that have noise on each group element rather
than on each pair. For instance, the noise in cryo-EM occurs on each image rather than independently on
each pairwise comparison. Adapting AMP to these more general models is left for future work.
This paper is organized as follows. We begin in Section 2 with an outline of our methods in the simplified
cases of synchronization over Z/2 and U (1), motivating our approach from a detailed discussion of prior work
and its shortfalls. In Section 3 we provide our general algorithm and the general problem model for which
it is designed. Several experiments on this Gaussian model and other models are presented in Section 4,
demonstrating strong empirical performance. We then offer two separate derivations of our AMP algorithm:
in Section 5, we derive our algorithm as a simplification of belief propagation, and then in Section 6 we give
an alternative self-contained derivation of the nonlinear update step and use this to provide a non-rigorous
analysis of AMP (based on standard assumptions from statistical physics). In particular, we derive the
state evolution equations that govern the behavior of AMP, and use these to identify the threshold above
which AMP achieves non-trivial reconstruction. Namely, we see that AMP has the same threshold as PCA
(requiring the SNR λ to exceed 1 on at least one frequency), but AMP achieves better recovery error above
the threshold. In Section 7 we argue for the correctness of the above non-rigorous analysis, providing both
numerical and mathematical evidence. It is known that inefficient estimators can beat the λ = 1 threshold
[PWBM16] but we conjecture that no efficient algorithm is able to break this barrier, thus concluding in
Section 8 with an exploration of statistical-to-computational gaps that we expect to exist in synchronization
problems, driven by ideas from statistical physics.
2
Intuition: iterative methods for Z/2 and U (1) synchronization
We begin with a discussion of synchronization methods over the cyclic group Z/2 and the group of unit
complex numbers (or 2D rotations) U (1). These examples will suffice to establish intuition and describe much
of the novelty of our approach, while avoiding the conceptual and notational complication of representation
theory present in the general case. Sections 2.1, 2.2, and some of 2.3 discuss prior work on these problems
in more depth, while Sections 2.3 and 2.4 develop a special case of our algorithm.
4
2.1
Z/2 synchronization
The problem of Gaussian Z/2 synchronization is to estimate a uniformly drawn signal x ∈ {±1}n given the
matrix
λ
1
Y = xx⊤ + √ W,
n
n
where W is a symmetric matrix whose entries are distributed independently (up to symmetry) as N (0, 1),
and λ > 0 is a signal-to-noise parameter. With this scaling, the signal and noise are of comparable size in
spectral norm; we can not hope to recover x exactly, but we can hope to produce an estimate x
b ∈ {±1} that
bi2 > ε with
is correlated nontrivially with x, i.e. there exists ε > 0 (not depending on n) such that n12 hx, x
⊤
⊤
probability 1 − o(1) as n → ∞. As xx = (−x)(−x) , we can only hope to estimate x up to sign; thus we
aim to achieve a large value of hx, x
bi2 . We now review three algorithmic methods for this problem.
Spectral methods. With the scaling above, the spectral norm of the signal nλ xx⊤ is λ, while that of
the noise √1n W is 2. By taking the top eigenvector of Y , x may be estimated with significant correlation
provided that λ is a large enough constant.
Specifically, the generative model for Y above is a special case of the spiked Wigner model, and the
eigenvalues and eigenvectors of such spiked models are among the main objects of study in random matrix
theory. When λ > 1, the (unit norm) top eigenvector vmax (Y ) correlates nontrivially with x; more specifically,
as n → ∞, we have n1 hx, vmax (Y )i2 → 1 − 1/λ2 in probability [FP07, BGN11]. When λ < 1, this squared
correlation tends to zero; in fact, this is known to be true of all estimators [DAM16, PWBM16], reflecting a
sharp statistical phase transition.
Note that a top eigenvector may be computed through power iteration as follows: an initial guess v 0 is
drawn randomly, and then we iteratively compute v (t) = Y v (t−1) , rescaling the result as appropriate. Thus
P
(t−1)
(t−1)
(t)
each entry is computed as vu = w Yuw vw ; we can imagine that each entry w sends a ‘message’ Yuw vw
to each entry u – the ‘vote’ of entry w as to the identity of entry u – and then each entry sums the incoming
votes to determine its new value. The result has both a sign, reflecting the weighted majority opinion as to
whether that entry should ultimately be +1 or −1, and also a magnitude, reflecting a confidence and serving
as the weight in the next iteration. Thus we can envision the spectral method as a basic “message-passing
algorithm”.
While this approach is effective as quantified above, it would seem to suffer from two drawbacks:
• the spectral method is basis-independent, and thus cannot exploit the entrywise ±1 structure of the
signal;
• the vertex weights can grow without bound, potentially causing a few vertices to exert undue influence.
Indeed, these drawbacks cause major issues in the stochastic block model, a variant of the model above
with the Gaussian observations replaced by low-probability Bernoulli observations, usually envisioned as the
adjacency matrix of a random graph. Here a few sporadically high-degree vertices can dominate the spectral
method, causing asymptotically significant losses in the statistical power of this approach.
Projected power iteration. Our next stepping-stone toward AMP is the projected power method studied
by [Bou16, CC16], a variant of power iteration that exploits entry-wise structure. Here each iteration takes
the form v (t) = sgn(Y v (t−1) ), where the sign function sgn : R → {±1} applies entrywise. Thus each iteration
is a majority vote that is weighted only by the magnitudes of the entries of Y ; the weights do not become
more unbalanced with further iterations. This algorithm is also basis-dependent in a way that plausibly
exploits the ±1 structure of the entries.
Empirically, this algorithm obtains better correlation with the truth, on average, when λ > 2.4 approximately; see Figure 2. However, for very noisy models with 1 < λ < 2.4, this method appears weaker than
the spectral method. The natural explanation for this weakness is that this projected power method forgets
the distinction between a 51% vote and a 99% vote, and thus is over-influenced by weak entries. This is
5
particularly problematic for low signal-to-noise ratios λ, for which 51% votes are common. In fact, a heuristic
analysis similar to Section 6.3 suggests that this method
does not achieve the correct threshold for λ, failing
√
to produce nontrivial correlation whenever λ < 2/ π ≈ 1.128.
Soft-threshold power iteration. A natural next step is to consider iterative algorithms of the form
v (t) = f (Y v (t−1) ), where f applies some function R → [−1, 1] entrywise (by abuse of notation, we will also
denote the entrywise function by f ). Instead of the identity function, as in the spectral method, or the
sign function, as in the projected power method, we might imagine that some continuous, sigmoid-shaped
function performs best, retaining some sense of the confidence of the vote without allowing the resulting
weights to grow without bound. It is natural to ask what the optimal function for this purpose is, and
whether the resulting weights have any precise meaning.
Given the restriction to the interval [−1, 1], one can imagine treating each entry as a sign with confidence
in a more precise way, as the expectation of a distribution over {±1}. At each iteration, each entry u might
(t−1)
then obtain the messages Yuw vw
from all others, and compute the posterior distribution, summarized as
(t)
an expectation vu . As one can compute, this corresponds to the choice of transformation f (t) = tanh(λt)
where λ is the signal-to-noise parameter from above (see Figure 1).
1
f (t)
0.5
0
-0.5
-1
-5
0
5
t
Figure 1: Soft threshold functions used by AMP. The solid line is f (t) = tanh(t),
used for Z/2 synchronization. The dashed line is f (t) = I1 (2t)/I0 (2t) (modified Bessel
functions of the first kind), used for U (1) synchronization with one frequency.
2.2
Belief propagation and approximate message passing
The soft-projection algorithm above may remind the reader of belief propagation, due to [Pea86] in the
context of inference, and to [MPV86] as the cavity method in the context of statistical physics. We may
envision the problem of estimating x as probabilistic inference over a graphical model. The vertices of the
model represent the unknown entries of x, and every pair of vertices u, w participates in an edge interaction
based on the matrix entry Yuw = Ywu . Specifically, it may be computed that the posterior distribution for
x ∈ {±1}n after observing Y is given by
Y
Pr(x) ∝
exp(λYuw xu xw ),
u<w
which is precisely the factorization property that a graphical model captures.
Given such a model, belief propagation proceeds in a fashion reminiscent of the previous algorithm: each
vertex w sends a message to each neighbor u encoding the posterior distribution of xu based the previous
6
distribution of xw and the direct interaction λYuw . Each vertex u then consolidates all incoming messages
into a new ‘posterior’ distribution on xu given these messages, computed as if the messages were independent.
However, belief propagation introduces a correction to this approach: rather than letting information passed
from w to u propagate back to w on the next iteration, belief propagation is designed to pass information
along only those paths that do not immediately backtrack. Specifically, at each iteration, the message from
w to u is based on only the synthesis of messages from all vertices except u from the previous iteration.
This algorithm differs from the iterative methods presented above, both in this non-backtracking behavior,
and in the fact that the transformation from the distribution at w to the message w → u is not necessarily
linear (as in the multiplication Yuw vw above). Both of these differences can be reduced by passing to the
framework of approximate message passing [DMM09], which simplifies belief propagation in dense models
with weak interactions, through the following two observations (inspired by [TAP77] in the theory of spin
glasses):
√
• As the interaction λYuw is small, scaling as O(1/ n) as n → ∞, we may pass to an expansion in
small Yuw when computing the message w → u from the mean w. In this example, we find that the
(t−1)
(t−1)
2
is the mean
+ O(Yuw
), where vw
message w → u should be Rademacher with mean λYuw vw
of the distribution for xw in the previous iteration. This linear expansion ensures that the main
message-passing step can be expressed as a matrix–vector product.
• Rather than explicitly computing non-backtracking messages, which is computationally more involved,
we may propagate the more naı̈ve backtracking messages and then subtract the bias due to this simplification, which concentrates well. This correction term is called an Onsager correction. If vertex w
passes messages to all neighbors based on its belief at iteration t − 2, and then all of these neighbors
send return messages based on their new beliefs at time t − 1, then when updating the belief for vertex
w at time t, one can explicitly subtract off the ‘reflected’ influence of the previous belief at time t − 2.
It turns out that this is the only correction necessary: all other error contributions (e.g. 3-cycles) are
o(1).
Following these simplifications, one can arrive at an approximate message passing (AMP) algorithm for
Z/2-synchronization:
Algorithm 2.1 (AMP for Z/2 synchronization [DAM16]). —
• Initialize v 0 to small (close to zero) random values in [−1, 1].
• Iterate for 1 ≤ t ≤ T :
– Set c(t) = λY v (t−1) − λ2 (1 − h(v (t−1) )2 i)v (t−2) , the Onsager-corrected sum of incoming messages.
(t)
(t)
– Set vu = tanh(cu ) for each vertex u, the new estimated posterior mean.
• Return x
b = v (T ) (or the approximate MAP estimate x
b = sgn(v (T ) ) if a proper estimate in {±1}n is
desired).
Here h(v (t−1) )2 i denotes the average of the squared entries of v (t−1) . Detailed derivations of this algorithm
appear in Sections 5 and 6 in much higher generality.
In the setting of Z/2 synchronization, an algorithm equivalent to the above approach appears in [DAM16],
where a statistical optimality property is proven: if AMP is warm-started with a state v 0 with nontrivial
correlation with the truth, then it converges to an estimate of x that achieves minimum mean-squared error
(MMSE) asymptotically as n → ∞. The warm-start requirement
√ is technical and likely removable: if AMP
is initialized to small randomness, with trivial correlation O(1/ n) with the truth, then its early iterations
resemble PCA and should produce nontrivial correlation in O(log n) iterations. The statistical strength of
AMP is confirmed empirically, as it appears to produce a better estimate than either PCA or the projected
power method, for every λ > 1; see Figure 2.
7
2.3
AMP for Gaussian U(1) synchronization with one frequency
As a first step toward higher generality, consider the following Gaussian synchronization model over the unit
complex numbers U (1). The goal is to estimate a uniformly drawn signal x ∈ U (1) given the matrix
Y =
λ ∗
1
xx + √ W,
n
n
where W is a Hermitian matrix whose entries are distributed independently as CN (0, 1), the complex
normal distribution given by N (0, 1/2) + N (0, 1/2)i, and where λ > 0 is a signal-to-noise parameter. As xx∗
is invariant under a global phase shift of x, we can only hope to estimate up to the same ambiguity, and so
we would like an estimator x
b that maximizes |hx, x
bi|2 , where the inner product is conjugated in the second
variable. Many of the previously discussed iterative techniques adapt to this new case.
Spectral methods. The same analysis of the spectral method holds in this case; thus when λ > 1, the top
eigenvector achieves nontrivial correlation with x, while for λ < 1, the spectral method fails and nontrivial
estimation is provably impossible [PWBM16].
Projected power method. After each matrix–vector product, we can project v (t) entrywise onto the
unit circle, preserving the phase of each entry while setting the magnitude to 1. This algorithm is analyzed
in [Bou16] in a lower-noise setting, where it is shown to converge to the maximum likelihood estimator.
Soft-threshold power method. One might imagine applying some entrywise function after each matrix–
vector product, which preserves the phase of each entry while mapping the magnitude to [0, 1]. Thus the
vector entries vu live in the unit disk, the convex hull of the unit circle; these might be envisioned as estimates
of the posterior expectation of xu .
Belief propagation & AMP. Belief propagation is somewhat problematic in this setting: all messages
should express a distribution over U (1), and it is not a priori clear how this should be expressed in finite
space. However, under the simplifications of approximate message passing, the linearity of the messagepassing stage enables a small summary of this distribution to suffice: we need only store the expectation of
each distribution, a single value in the unit disk. Approximate message passing takes the following form:
Algorithm 2.2 (AMP for U (1) synchronization with one frequency). —
• Initialize v 0 to small random values in the unit disk conv(U (1)).
• Iterate for 1 ≤ t ≤ T :
– Set c(t) = λY v (t−1) − λ2 (1 − h|v (t−1) |2 i)v (t−2) , the Onsager-corrected sum of incoming messages.
(t)
(t)
– Set vu = f (cu ) for each vertex u, the new estimated posterior mean. Here f applies the function
f (t) = I1 (2t)/I0 (2t) to the magnitude, leaving the phase unchanged.
• Return x
b = v (T ) (or the approximate MAP estimate x
b = phase(v (T ) ) if a proper estimate in U (1)n is
desired).
Here Ik denotes the modified Bessel functions of the first kind. The function f is depicted in Figure 1.
Detailed derivations of this algorithm appear in Sections 5 and 6 in much higher generality.
8
2.4
AMP for Gaussian U(1) synchronization with multiple frequencies
Consider now the following more elaborate synchronization problem. The goal is to estimate a spike x ∈
U (1)n from the observations
λ1 ∗
1
Y1 =
xx + √ W1 ,
n
n
Y2 =
1
λ2 2 2 ∗
x (x ) + √ W2 ,
n
n
..
.
YK =
1
λK K K ∗
x (x ) + √ WK ,
n
n
where the Wk are independent Hermitian matrices whose entries are distributed independently (up to symmetry) as CN (0, 1), the λk > 0 are signal-to-noise parameters, and xk denotes the entrywise kth power of
x.
Thus we are given K independent noisy matrix-valued observations of x; we can imagine these observa√
tions as targeting different frequencies or Fourier modes. Given two independent draws of λ xx∗ /n + W/ n
as in the previous section,
the spectral method applied to their average will produce a nontrivial estimate of
√
x as soon as λ > 1/ 2. However, under the multiple frequencies model above, with K = 2 and λ1 = λ2 = λ,
nontrivial estimation is provably impossible for λ < 0.937 [PWBM16]; we present non-rigorous evidence in
Section 8 that the true statistical threshold should in fact remain λ = 1. Thus the multiple frequencies model
would seem to confound attempts to exploit the multiple observations together. However, we will discuss
how AMP enables us to obtain a much better estimate when λ > 1 than is possible with one matrix alone.
Let us return to the issue of belief propagation over U (1), and of how to represent distributions. One
crude approach might be to discretize U (1) and express the density on a finite subset of points; however,
this is somewhat messy (e.g. the discretization is not preserved under rotation) and only becomes worse for
more elaborate groups such as SO(3) (here one can not even find arbitrarily fine discretizations on which
the group acts transitively).
Instead, we could exploit the rich structure of Fourier theory, and express a distribution on U (1) by the
(t)
Fourier series of its density2 . Thus, if µw is the belief distribution at vertex w and time t, we can express:
(t)
X
dµw
=
vw,k eikθ ,
dθ/2π
k∈Z
with vw,0 = 1 and vw,−k = vw,k . Computing the distributional message mw→u from w to u, we obtain
dmt+1
w→u
=1+
dθ/2π
X
λk (Yk )uw vw,k eikθ + O((Y• )2uw ),
X
λk (Yk )uw vw,k eikθ + O((Y• )2uw ).
1≤|k|≤K
√
where we take Y−k = Yk . As (Yk )uw is order 1/ n in probability, this approximation will be asymptotically
accurate. Thus it suffices to represent distributions by the coefficients vw,k with |k| ≤ K. By conjugate
symmetry, the coefficients with 1 ≤ k ≤ K suffice. The sufficiency of this finite description of each belief
distribution is a key insight to our approach.
The other crucial observation concerns the remaining BP step of consolidating all incoming messages into
a new belief distribution. As each incoming message is a small perturbation of the uniform distribution, the
approximation log(1 + x) ≈ x allows us to express the log-density of the message distribution:
log
dmt+1
w→u
=
dθ/2π
1≤|k|≤K
2 A dense subset of distributions satisfies appropriate continuity assumptions to discuss their densities with respect to uniform
measure, a Fourier series, etc., and we will not address these analytic technicalities further.
9
We now add these log-densities to obtain the log-density of the new belief distribution, up to normalization:
t+1
X
X
dµu
+ const. =
λk (Yk )uw vw,k eikθ + O((Y• )2uw ).
log
dθ/2π
1≤|k|≤K
w6=u
We thus obtain the Fourier coefficients of the log-density of the new belief from the Fourier coefficients of the
density of the old belief, by matrix–vector products. Remarkably, this tells us that the correct per-vertex
nonlinear transformation to apply at each iteration is the transformation from Fourier coefficients of the
log-density to those of the density! (In section 6 we will see an alternative interpretation of this nonlinear
transformation as an MMSE estimator.)
The only constraints on a valid log-density are those of conjugate symmetry on Fourier coefficients; thus
log-densities form an entire linear space. By contrast, densities are subject to non-negativity constraints,
and form a nontrivial convex body in RK . The latter space is the analogue of the unit disk or the interval
[−1, 1] in the preceding examples, and this transformation from the Fourier series of a function to those of
its exponential (together with normalization) forms the analogue of the preceding soft-projection functions.
We thus arrive at an AMP algorithm for the multiple-frequency problem:
Algorithm 2.3 (AMP for U (1) synchronization with multiple frequencies). —
0
• For each 1 ≤ k ≤ K and each vertex u, initialize vu,k
to small random values in C.
• Iterate for 1 ≤ t ≤ T :
(t)
(t−1)
(t−1)
(t−2)
1. For each 1 ≤ k ≤ K, set ck = λk Yk vk
− λ2k (1 − h(vk
)2 i)vk
, the vector of kth Fourier
components of the estimated posterior log-densities, with Onsager correction.
(t)
2. Compute vk , the vector of kth Fourier components of the estimated posterior densities.
(T )
• Return x
b = v1 (or some rounding if a proper estimate in U (1)n is desired, or even the entire per-vertex
posteriors represented by c(T ) ).
Again, a more detailed derivation can be found in Sections 5 and 6.
It is worth emphasizing that only the expansion
log
(t)
X (t)
dµu
+ const. = 2Re
ck,w eikθ
dθ/2π
1≤k≤K
is an accurate expansion of the estimated vertex posteriors. While this log-density is band-limited, this
still allows for the density to be a very spiked, concentrated function, without suffering effects such as the
Gibbs phenomenon. By contrast, the finitely many v coefficients that this algorithm computes do not suffice
to express the Fourier expansion of the density, and a truncated expansion based on only the computed
coefficients might even become negative.
We conclude this section by noting that nothing in our derivation depended crucially on the Gaussian
observation model. The choice of model tells us how to propagate beliefs along an edge according to a
matrix–vector product, but we could carry this out for a larger class of graphical models. The essential
properties of a model, that enables this approach to adapt, are:
• The model can be expressed as a graphical model with only pairwise interactions:
Y
Pr(x) ∝
Luw (xu , xw ).
u<w
√
• The interaction graph is dense, with all pair potentials individually weak (here, 1 + O(1/ n)).
10
• The pair potentials Luw (xu , xw ) depend only on the group ratio xu x−1
w . (This is the core property of
a synchronization problem.)
• The pair potentials Luw are band-limited as a function of xu x−1
w . This assumption (or approximation)
allows the algorithm to track only finitely many Fourier coefficients.
The formulation of AMP for general models of this form is discussed in the next section.
3
AMP over general compact groups
The approach discussed above for U (1)-synchronization with multiple frequencies readily generalizes to the
setting of an arbitrary compact group G, with Fourier theory generalized to the representation theory of
G. Just as the Fourier characters are precisely the irreducible representations of U (1), we will represent
distributions over G by an expansion in terms of irreducible representations, as described by the Peter–Weyl
theorem. Under the assumption (or approximation) of band-limited pairwise observations, it will suffice
to store a finite number of coefficients of this expansion. (Note that finite groups have a finite number of
irreducible representations and so the band-limited requirement poses no restriction in this case.)
A geometric view on this is as follows. Belief propagation ideally sends messages in the space of distributions on G; this is a form of formal convex hull on G, and is illustrated in the case of Z/2-synchronization by
sending messages valued in [−1, 1]. When G is infinite, however, this space is infinite-dimensional and thus
intractable. We could instead ask whether the convex hull of G taken in some finite-dimensional embedding
is a sufficient domain for messages. The key to our approach is the observation that, when observations are
band-limited, it suffices to take an embedding of G described by a sum of irreducible representations.
This section will be devoted to presenting our AMP algorithm in full generality, along with the synchronization model that it applies to. In particular, the algorithm can run on the general graphical model
formulation of Section 3.2, but when we analyze its performance we will restrict to the Gaussian observation
model of Section 3.4.
3.1
3.1.1
Representation theory preliminaries
Haar measure
A crucial property of compact groups is the existence of a (normalized) Haar measure, a positive measure
µ on the group that is invariant under left and right translation by any group element, normalized such
that µ(G) = 1 [BtD13]. This measure amounts to a concept of ‘uniform distribution’ on such a group, and
specializes to the ordinary uniform distribution on a finite group. Throughout this paper, integrals of the
form
Z
f (g) dg,
G
are understood to be taken with respect to Haar measure.
3.1.2
Peter–Weyl decomposition
Fix a compact group G. We will be working with the density functions of distributions over G. In order
to succinctly describe these, we use the representation-theoretic analogue of Fourier series: the Peter–Weyl
decomposition. The Peter–Weyl theorem asserts that L2 (G) (the space of square-integrable, complex scalar
functions on G) is the closure of the span (with coefficients from C) of the following basis, which is furthermore
orthonormal with respect to the Hermitian inner product on L2 (G) [BtD13]:
p
Rρab (g) = dρ ρ(g)ab ,
indexed over all complex irreducible representations ρ of G, and all 1 ≤ a ≤ dρ , 1 ≤ b ≤ dρ where dρ = dim ρ.
The representations are assumed unitary (without loss of generality). The inner product is taken to be
conjugate-linear in the second input.
11
Since we want our algorithm to be able to store the description of a function using finite space, we fix
a finite list P of irreducible representations to use. From now on, all Peter–Weyl decompositions will be
assumed to only use representations from P; we describe functions of this form as band-limited. We exclude
the trivial representation from this list because we will only need to describe functions up to an additive
constant. Given a real-valued function f : G → R, we will often write its Peter–Weyl expansion in the form
X
f (g) =
hfbρ , Rρ (g)i,
ρ
p
where fbρ and Rρ (g) = dρ ρ(g) are dρ × dρ complex matrices. Here ρ ranges over the irreducibles in P; we
assume that the functions f we are working with can be expanded in terms of only these representations. The
matrix inner product used here is defined by hA, Bi = Tr(AB ∗ ). The Peter–Weyl coefficients of a function
can be extracted by integration against the appropriate basis functions:
Z
Rρ (g)f (g) dg.
fbρ =
G
By analogy to Fourier theory, we will sometimes refer to the coefficients fbρ as Fourier coefficients, and refer
to the irreducible representations as frequencies.
3.1.3
Representations of real, complex, quaternionic type
Every irreducible complex representation of a compact group G over C is of one of three types: real type,
complex type, or quaternionic type [BtD13]. We will need to deal with each of these slightly differently. In
particular, for each type we are interested in the properties of the Peter–Weyl coefficients that correspond
to the represented function being real-valued.
A complex representation ρ is of real type if it can be defined over the reals, i.e. it is isomorphic to a realvalued representation. Thus we assume without loss of generality that we are working with a real-valued
ρ. In this case it is clear that if f is a real-valued function then (by integrating against Rρ ) fbρ is real.
Conversely, if fbρ is real then the term hfbρ , Rρ (g)i (from the Peter–Weyl expansion) is real.
A representation ρ is of complex type if ρ is not isomorphic to its conjugate representation ρ, which is the
irreducible representation defined by ρ(g) = ρ(g). We will assume that the complex-type representations in
our list P come in pairs, i.e. if ρ is on the list then so is ρ. If f is real-valued, we see (by integrating against
Rρ and Rρ ) that fbρ = fbρ . Conversely, if fbρ = fbρ holds then hfbρ , Rρ (g)i + hfbρ , Rρ (g)i is real.
Finally, a representation ρ is of quaternionic type if it can be defined over the quaternions in the following
sense: dρ is even and ρ(g) is comprised of 2 × 2 blocks, each of which encodes a quaternion by the following
relation:
a + bi c + di
a + bi + cj + dk ↔
.
−c + di a − bi
Note that this relation respects quaternion addition and multiplication. Furthermore, quaternion conjugation
(negate b, c, d) corresponds to matrix conjugate transpose. If a matrix is comprised of 2×2 blocks of this form,
we will call it block-quaternion. Now let ρ be of quaternionic type (and assume without loss of generality
that ρ takes the above block-quaternion form), and let f be a real function G → R. By integrating against
Rρ we see that fbρ must also be block-quaternion. Conversely, if fbρ is block-quaternion then hfbρ , Rρ (g)i is
real; to see this, write hfbρ , Rρ (g)i = Tr(fbρ Rρ (g)∗ ), note that fbρ Rρ (g)∗ is block-quaternion, and note that the
trace of any quaternion block is real.
3.2
Graphical model formulation
As in Section 2.2, we take the standpoint of probabilistic inference over a graphical model. Thus we consider
the task of estimating g ∈ Gn from observations that induce a posterior probability factoring into pairwise
12
likelihoods:
Pr(g) ∝
Y
{u,w}
Luv (gu , gw )
(1)
where {u, w} ranges over all (undirected) edges (without self-loops). We assume that the pair interactions
−1
Luw (gu , gw ) are in fact a function of gu gw
∈ G, depending only on the relative orientation of the group
elements. Without loss of generality, we can take Luu = 0 for all u. This factorization property amounts to
a graphical model for g, with each entry gu ∈ G corresponding to a vertex u, and each pair interaction Luw
represented by an edge of the model.
−1
Taking a Peter–Weyl decomposition of Luw as a function of gu gw
allows us to write:
X
ρ
−1
Yuw
, ρ(gu gw
) ,
Luw (gu , gw ) = exp
ρ
ρ
where ρ runs over all irreducible representations of G. We require coefficients Yuv
∈ Cdρ ×dρ for which this
−1
expansion is real-valued (for all gu gv ). We also require the symmetry Luv (gu , gv ) = Lvu (gv , gu ), which
ρ
ρ ∗
ρ
means Yuv
= (Yvu
) . Let Yρ be the ndρ × ndρ matrix with blocks Yuv
.
The input to our synchronization problem will simply be the coefficients Yρ . These define a posterior
distribution µ on the latent vector g of group elements, and our goal is to approximately recover g up to a
global right-multiplication by some group element.
We suppose that the observations are band-limited : Yρ = 0 except on a finite set P of irreducible
representations. This will allow us to reduce all Peter–Weyl decompositions to a finite amount of relevant
information. We will always exclude the trivial representation from P: this representation can only contribute
a constant factor to each pair likelihood, which then disappears in the normalization, so that without loss
of generality we can assume the coefficient of the trivial representation to always be zero.
One might also formulate a version of this model that allows node potentials:
!
Y
Y
Luv (gu , gw )
Pr(g) ∝
Lu (gu ) ,
u
{u,w}
expressing a nontrivial prior or observation on each group element. Although this model is compatible with
our methods (so long as the node potentials are also band-limited), we suppress this generality for the sake
of readability.
Many synchronization problems (for instance, sensor localization) have noise on each pairwise measurement, and fit this graphical model formulation well. Other synchronization problems (for instance, cryo-EM)
are based on per-vertex measurements with independent randomness; one can derive pairwise information
by comparing these measurements, but these pairwise measurements do not have independent noise and
do not strictly fit the model described above. We note that prior work has often run into the same issue
and achieved strong results nonetheless. Indeed, this model is closely related to the “non-unique games”
model of [BCS15] for which the application to cryo-EM is discussed; this model suggests minimizing an
objective of the form (1) without interpreting it as a posterior likelihood. We defer a close examination of
message-passing algorithms for synchronization problems with per-vertex noise to future work.
3.3
AMP algorithm
We now state our AMP algorithm. The algorithm takes as input the log-likelihood coefficients Yρ ∈ Cndρ ×ndρ ,
for each ρ in a finite list P of irreducibles (which must not contain the trivial representation; also for each
complex-type representation ρ in the list, ρ must also appear in the list). The algorithm’s state at time t is
(t)
comprised of the Fourier coefficients Cρ ∈ Cndρ ×dρ , which are updated as follows.
Algorithm 3.1 (AMP for synchronization over compact groups). —
(0)
• For each ρ ∈ P and each vertex u, initialize Cu,ρ to small random values in C.
13
• Iterate for 1 ≤ t ≤ T :
– For each ρ ∈ P, set
(t)
Cu,ρ
= d−1
ρ
X
w6=u
ρ 2 (t−2)
ρ
(t−1)
Yuw
Vw,ρ
− d−2
ρ |Ytyp | Vu,ρ
X
w
(t−1) ∗ (t−1)
,
dρ I − (Vw,ρ
) Vw,ρ
the Fourier coefficients of the estimated posterior log-densities, with Onsager correction. Here
ρ 2
|Ytyp
| denotes the average squared-norm of the entries of Y ρ .
(t)
(t)
– For each u and each ρ ∈ P, set Vu,ρ = Eρ (Cu ), where as in Section 2.4, the nonlinear transformation
Z
.Z
X
X
Rρ (g) exp
Eρ (C) =
hCρ′ , Rρ′ (g)i dg
exp
hCρ′ , Rρ′ (g)i dg
(2)
G
G
ρ′
ρ′
takes the Fourier coefficients for a function f on G and returns those of exp ◦f , re-normalized
(t)
to have integral 1. These Vu,ρ are the Fourier coefficients of the estimated posterior densities,
truncated to the contribution from irreducibles P, which suffice for the next iteration.
(T )
• Return the posteriors represented by Cu,ρ , or some rounding of these (e.g. the per-vertex MAP
estimate).
This algorithm follows the intuition of Section 2, and derivations can be found in Sections 5 and 6.
Note that each iteration runs in time O(n2 ), which is linear in the input matrices. This runtime is due
to the matrix–vector products; the rest of the iteration takes O(n) time. We expect O(log n) iterations to
suffice, resulting in a nearly-linear-time algorithm with respect to the matrix inputs. Some applications may
derive from per-vertex observations that are pairwise compared to produce edge observations, hacked into
this framework by an abuse of probability; our algorithm will then takes nearly-quadratic time with respect
to the vertex observations. However, some such per-vertex applications produce matrices with a low-rank
factorization Yρ = Uρ Uρ⊤ , for which the matrix–vector product can be performed in O(n) time.
3.4
Gaussian observation model
Our AMP algorithm handles the general graphical model formulation above, but we will be able to analyze its
performance in more detail when restricted to the following concrete Gaussian observation model (which we
first introduced in [PWBM16]), generalizing the Gaussian models of Section 2. First, latent group elements
gu are drawn independently and uniformly from G (from Haar measure). Then for each representation ρ in
P, we observe the ndρ × ndρ matrix
Mρ =
1
λρ
Xρ Xρ∗ + p
Wρ .
n
ndρ
Here Xρ is the ndρ × dρ matrix formed by vertically stacking the dρ × dρ matrices ρ(gu ) for all vertices u. λρ
is a signal-to-noise parameter for the frequency ρ. The noise Wρ is a Gaussian random matrix drawn from
the GOE (Gaussian orthogonal ensemble), GUE, or GSE, depending on whether ρ is of real type, complex
type, or quaternionic type, respectively. In any case, Wρ is normalized so that each off-diagonal entry has
expected squared-norm 1. To be concrete, in the real case the entries are N (0, 1) and in the complex case,
the real and imaginary parts of each entry are N (0, 1/2). For the quaternionic case, each 2 × 2 block encodes
a quaternion value a + bi + cj + dk in the usual way (see Section 3.1.3) where a, b, c, d are N (0, 1/2). The
noise matrices Wρ are independent across representations except when we have a conjugate pair of complextype representations we draw Mρ randomly as above and define Mρ = Mρ and λρ = λρ . Note that the
normalization is such that the signal term has spectral norm λρ and the noise term has spectral norm 2.
14
Special cases of this model have appeared previously: Z/2 with one frequency [DAM16, JMR16] and
U (1) with one frequency [JMR16]. In fact, [DAM16] derives AMP for the Z/2 case and proves that it is
information-theoretically optimal. We introduced the general model in [PWBM16], which presents some
statistical lower bounds.
In Appendix A, we show how the Gaussian observation model fits into the graphical model formulation
by deriving the corresponding coefficient matrices Yρ . In particular, we show that Yρ = dρ λρ Mρ , a scalar
multiple of the observed Gaussian matrix.
3.5
Representation theory of some common examples
In this section we discuss the representation theory of a few central examples, namely Z/L, U (1), and SO(3),
and connect the general formalism back to the examples of Section 2.
Representations of Z/L and U (1). The irreducible representations of these groups are one-dimensional,
described by the discrete Fourier transform and the Fourier series, respectively. U (1) has frequencies indexed
by k ∈ Z, given by ρ(g) = g k where g ∈ U (1) (i.e. a unit-norm complex number). All of these representations
are of complex type. We will say “U (1) with K frequencies” to refer to the frequencies 1, . . . , K along with
their conjugates, the frequencies −1, . . . , −K. Similarly, if we identify Z/L with the complex Lth roots of
unity, we have frequencies defined the same way as above, except to avoid redundancy we restrict the range
of k as follows. If L is odd, we allow k ∈ {1, 2, . . . , (L − 1)/2} along with their conjugates (negations). If L
is even, we have complex-type representations k ∈ {1, 2, . . . , L/2 − 1} (along with their conjugates), plus an
additional real-type representation k = L/2. Again, “Z/L with K frequencies” means we take frequencies
1, . . . , K along with their conjugates (when applicable).
For the case of Z/2 we can now see how the tanh function from the AMP algorithm of Section 2.2 arises
as a special case of the nonlinear transformation E occurring in AMP. The only nontrivial representation of
Z/2 is the ‘parity’ representation in which −1 acts as −1. In this context, E will first input the Fourier series
of a log-density with respect to the uniform measure on {±1}:
dµu
+ const. = cx
dx
and evaluate this at ±1 to obtain the values ±c. We then compute the exponential of this at each point to
obtain the un-normalized density of e−c at −1 and ec at 1. Normalizing, the density values are e−c /(e−c +ec )
and ec /(e−c + ec ), so that the new parity coefficient is
log
Eparity (c) =
ec − e−c
= tanh c.
ec + e−c
Representations of SO(3). This group has one irreducible representation ρk of each odd dimension dk =
2k + 1; thus the k = 0 representation is the trivial representation ρ0 (g) = 1, and the k = 1 representation is
the standard representation of SO(3) as rotations of three-dimensional space. All of these representations are
of real type, and may be described as the action of rotations on the 2k + 1-dimensional space of homogeneous
degree k spherical harmonics. Frequently in the literature (for instance in molecular chemistry), a complex
basis for the spherical harmonics is given, and the representation matrices are the complex-valued Wigner Dmatrices; however, the representation can be defined over the reals, as is demonstrated by any real orthogonal
basis for the spherical harmonics. See e.g. Section II.5 of [BtD13] for a more detailed account. As in the
cases above, we will often refer to synchronization problems over “SO(3) with K frequencies”, in which the
observations are assumed to be band-limited to the first K nontrivial irreducibles with 1 ≤ k ≤ K.
4
Experimental results
We present a brief empirical exploration of the statistical performance of AMP in various settings, and as
compared to other algorithms.
15
Figure 2: Comparison of iterative algorithms for Gaussian Z/2 synchronization. The
horizontal axis represents the signal-to-noise ratio λ, and the vertical axis depicts the
log-error ln(1 − |hx, x
bi/n|) where x ∈ {±1}n is the ground truth and x
b ∈ {±1}n is
the (rounded) output of the algorithm. From top to bottom, as measured on the left
side: projected power iteration (green), soft-threshold power iteration (red), spectral
method (blue), and AMP (cyan). Each data point is an average of 200 trials with
n = 2000 vertices.
In Figure 2 we compare the performance of the spectral method, projected power iteration, soft-threshold
power iteration without an Onsager correction, and full AMP (see Sections 2.1 and 2.2) for Gaussian Z/2
synchronization. The spectral method achieves the optimal threshold of λ = 1 as to when nontrivial recovery
is possible, but does not achieve the optimal correlation afterwards. The projected power method appears
to asymptotically achieve the optimal correlation as λ → ∞, but performs worse than the spectral method
for small λ. Soft-thresholding offers a reasonable improvement on this, but the full AMP algorithm strictly
outperforms all other methods. This reflects the optimality result of [DAM16] and highlights the necessity
for the Onsager term. The gains are fairly modest in this setting, but increase with more complicated
synchronization problems.
Figures 3 and 4 compare the performance of AMP on Gaussian U (1) synchronization with multiple
frequencies; see Section 2.4 for the model. In sharp contrast to spectral methods, which offer no reasonable
way to couple the frequencies together, AMP produces an estimate that is orders of magnitude more accurate
than what is possible with a single frequency.
In Figures 5 and 6, we see similar results over SO(3), under the Gaussian model of Section 3.4. This
also demonstrates that AMP is an effective synchronization algorithm for more complicated, non-abelian Lie
groups.
This ability to exploit multiple frequencies represents a promising step toward improved algorithms for
cryo-electron microscopy, which may be viewed as a synchronization problem over SO(3). Some previous
approaches to this problem effectively band-limit the observations to a single frequency and then apply
a spectral method [SS11], and the experiments in Figures 3–6 demonstrate that our algorithm stands a
compelling chance of achieving a higher-quality reconstruction.
We remark that some numerical issues arise when computing the nonlinear transformation E in our
AMP algorithm, which involves integration over the group. Our implementation of E for U (1) and SO(3) is
based on evaluating each log-density on a discretization of the group, taking the pointwise exponential, and
thus approximating each integral by a discrete sum. This approach is somewhat crude but appears to work
16
Figure 3: Gaussian U (1) synchronization with K
frequencies; from bottom to top, K = 1, . . . , 6.
The signal-to-noise ratios λk are all equal, with the
common value given by the horizontal axis. Each
curve depicts the correlation |hx, x
bi/n| between the
ground truth and the AMP estimate. Each data
point is an average of 50 trials with n = 1000 vertices.
Figure 4: Here the vertical axis depicts the logerror ln(1 − |hx, x
bi/n|). From top to bottom: K =
1, . . . , 6.
Figure 5: Gaussian SO(3) synchronization with K
frequencies; from bottom to top, K = 1, 2, 3. The
signal-to-noise ratios λk are all equal, with the common value given by the horizontal axis. Each curve
√
b F /(n 3)
depicts the squared correlation kX ⊤Xk
between the ground truth and the AMP estimate.
b are 3n × n matrices where each
Here X and X
3 × 3 block encodes an element of SO(3) via the
standard representation (rotation matrices). Each
data point is an average of 5 trials with n = 100
vertices.
Figure 6: Here the vertical
axis depicts the log√
b F /(n 3)). From top to bottom:
error ln(1−kX ⊤Xk
K = 1, 2, 3.
17
adequately in our experiments; there is undoubtedly room for this numerical procedure to be improved. More
sophisticated methods may be necessary to obtain adequate results on any higher-dimensional Lie groups.
Note also that if the vertex posterior in question is extremely concentrated near a point, the numerical value
of each integral will depend significantly on whether this spike lies near a discretization point; however, this
should affect both the numerator and denominator integrals in (2) by approximately equal factors, so as to
have a minimal effect on the normalized value of Eρ .
5
Derivation of AMP from belief propagation
In this section we derive the general AMP algorithm of Section 3 starting from belief propagation, similarly
(t)
to [DMM10]. We begin with the belief propagation update step (see e.g. [MM09]), writing messages µu→v
as densities with respect to Haar measure:
(t)
(t−1)
Y Z
dµu→v
1
dµw→u
= (t)
dgw .
Luw (gu , gw )
dgu
dgw
Zu→v w6=u,v G
(t)
Here t denotes the timestep and Zu→v is the appropriate normalization constant. Expand this (positive)
probability density as the exponential of an L2 function, expressed as a Peter–Weyl expansion:
(t)
X (t)
dµu→v
= exp
Cu→v,ρab Rρab (gu ).
dgu
ρab
(t)
We can extract these Fourier coefficients Cu→v,ρab by integrating against the basis functions above.
Assume that ρ is not the trivial representation; then:
(t)
Cu→v,ρab =
Z
G
(t−1)
Rρab (gu ) log
dµu→v
dgu
dgu
Y Z
1
(t−1)
dµw→u
Rρab (gu ) log
=
dgw dgu
Luw (gu , gw )
(t−1)
dgw
G
Zu→v w6=u,v G
Z
Z
E dµ(t−1)
X
XD ′
w→u
ρ
−1
dgw dgu .
Rρab (gu )
log exp
=
Yuw
, ρ′ (gu gw
)
dgw
G
G
′
Z
ρ
w6=u,v
′
ρ
As the Yuw
are small, we can pass to a linear expansion about these, incurring o(1) error as n → ∞:
Z
Z
Z D
Z
(t−1)
(t−1)
E dµ(t−1)
X
X
′
dµ
dµ
w→u
w→u
w→u
ρ
−1
log
dgw +
dgw
dgw dgu
≈
Rρab (gu )
Yuw
, ρ′ (gu gw
)
dgw
dg
dg
w
w
G
G
G
G
′
w6=u,v
=
ρ
X XZ Z
w6=u,v ρ′
G
D ′
E dµ(t−1)
w→u
ρ
−1
dgw dgu .
Rρab (gu ) Yuw
, ρ′ (gu gw
)
dgw
G
To progress further, we will expand the middle factor of the integrand:
E
D ′
E D ′
ρ
−1
ρ
Yuw
, ρ′ (gu gw
) = Yuw
, ρ′ (gu )ρ′ (gw )∗
X ρ′
=
Yuw,a′ b′ ρ′ (gu )a′ c′ ρ′ (gw )b′ c′ .
a′ b′ c′
18
Returning to the previous derivation:
(t)
Cu→v,ρab =
X X X Z Z
w6=u,v ρ′ a′ b′ c′
=
X X X
G
G
Z
′
ρ
Yuw,a
′ b′
G
w6=u,v ρ′ a′ b′ c′
= d−1
ρ′
X X X
Rρab (gu )ρ′ (gu )a′ c′ dgu ·
Z
′
ρ
Yuw,a
′ b′
G
w6=u,v ρ′ a′ b′ c′
= d−1
ρ′
X X X
(t−1)
′
ρ
′
′
Rρab (gu )Yuw,a
′ b′ ρ (gu )a′ c′ ρ (gw )b′ c′
X X
′
ρ
Yuw,a
′ b′ δρ,ρ′ δa,a′ δb,c′
ρ
Yuw,ab
′
w6=u,v b′
Z
G
(t)
Cu→v,ρ
= d−1
ρ
Cu→v,ρ of log
dρ ×dρ
ρC
(t)
dµu→v
to
dgu
→
L
ρ
Z
Z
(t−1)
Rρ′ b′ c′ (gw )
G
dµw→u
dgw
dgw
(t−1)
Rρ′ b′ c′ (gw )
G
dµw→u
dgw
dgw
dµw→u
dgw .
dgw
X
ρ
Yuw
w6=u,v
L
G
dµw→u
dgw
dgw
(t−1)
Rρb′ b (gw )
In matrix form,
Let E :
(t−1)
ρ′ (gw )b′ c′
Rρab (gu )Rρ′ a′ c′ (gu ) dgu ·
w6=u,v ρ′ a′ b′ c′
= d−1
ρ
Z
dµw→u
dgw dgu
dgw
Z
(t−1)
Rρ (gw )
G
dµw→u
dgw .
dgw
Cdρ ×dρ denote the transformation from the nontrivial Fourier coefficients
the Fourier coefficients of
dµ(t)
u→v
dgu .
(t)
Cu→v,ρ
= d−1
ρ
X
w6=u,v
Then we have
ρ
(t−1)
Yuw
Eρ (Cw→u
).
The map E amounts to exponentiation in the evaluation basis, except that the trivial Fourier coefficient is
missing from the input, causing an unknown additive shift. This corresponds to an unknown multiplicative
dµ(t)
should normalize to 1. Thus E amounts to
shift in the output, which we correct for by noting that dgu→v
u
exponentiation followed by normalization.
Explicitly, we can let
Z
X
Iρab (C) =
Rρab (g) exp
Cρ′ a′ b′ Rρ′ a′ b′ (g) dg.
G
ρ′ a′ b′
Then Eρab (C) = Iρab (C)/Itriv (C) where triv denotes the trivial representation Rtriv (g) = 1.
5.1
Onsager correction
In this section we complete the derivation of AMP by replacing the non-backtracking nature by an Onsager correction term, reducing the number of messages from n2 to n. This is similar to the derivation in
Appendix A of [BM11].
In order to remove the non-backtracking nature of the AMP recurrence, let us define
(t)
Cu,ρ
= d−1
ρ
X
w6=u
=
(t)
Cu→v,ρ
ρ
(t−1)
Yuw
Eρ (Cw→u
)
(t)
+ δu→v,ρ
,
19
(t−1)
(t−1)
(t−1)
(t)
ρ
where δu→v,ρ = d−1
ρ Yuv Eρ (Cv→u ). Then, substituting Cw→u = Cw
(t)
Cu,ρ
= d−1
ρ
X
w6=u
≈
d−1
ρ
X
w6=u
(t−1)
(t−2)
ρ′
ρ
Yuw
Eρ {Cw,ρ′ − d−1
ρ′ Ywu Eρ′ (Cu→w )}ρ′
ρ
(t−1)
Yuw
Eρ (Cw
) − d−1
ρ
X
(t−1)
− δw→u , we have
ρ
Yuw
DEρ
(t−1)
Cw
i
h
(t−2)
ρ′
′
′ (C
{d−1
)}
Y
E
′
ρ
ρ
u→w
wu
ρ
ρ
Yuw
DEρ
(t−1)
Cw
i
h
(t−2)
ρ′
′ .
′ (C
{d−1
)}
Y
E
′
ρ
ρ
u
wu
ρ
w6=u
where D denotes the total derivative
≈ d−1
ρ
X
w6=u
ρ
(t−1)
Yuw
Eρ (Cw
) − d−1
ρ
X
w6=u
Under the assumption that Y consists of per-edge O(n−1/2 ) noise and O(n−1 ) signal, the error incurred
in these two steps should be o(1). We thus reach an entirely non-backtracking recurrence where the first term
is a message-passing step and the second term is the so-called Onsager correction. It remains to simplify
this. We focus on a single matrix coefficient of the correction:
h
i
XX
(t)
(t−2)
ρ′
ρ
)}
Onsu,ρab = d−1
Yuw,ac
DEρ C (t−1) {d−1
ρ
ρ′ Ywu Eρ′ (Cu
w
cb
w6=u c
= d−1
ρ
X X
ρ
Yuw,ac
w6=u ρ′ cef
= d−1
ρ
X X
∂Eρcb
∂Cρ′ ef
ρ
Yuw,ac
w6=u ρ′ cef h
(ρ)
′
(t−1)
Cw
∂Eρcb
∂Cρ′ ef
(t−2)
ρ
(d−1
))ef
ρ′ Ywu Eρ′ (Cu
′
(t−1)
Cw
ρ
(t−2)
d−1
).
ρ′ Ywu,eh Eρ′ hf (Cu
(ρ)
Note that Ywu = (Yuw )∗ , and that other than this relation, the entries of Y consist of O(n−1/2 )-size
independent noise and lower-order signal. Hence, each of the O(n) terms contributing to this sum is O(n−1 ),
ρ′
ρ
and is only of this order if Yuw,ac
and Ywu,eh
are dependent, i.e. if a = h, c = e, and ρ = ρ′ . Thus:
(t)
Onsu,ρab = d−2
ρ
X
cf
Eρaf (Cu(t−2) )
X
w6=u
ρ
|Yuw,ac
|2
∂Eρcb
∂Cρ′ cf
(t−1)
Cw
.
ρ
Similarly to the derivation of [BM11], suppose that |Yuw,ac
| depends sufficiently little on
we can treat
ρ
|Yuw,ac
|
as a constant
ρ 2
|Ytyp
| .
(t)
ρ 2
Onsu,ρab = dρ−2 |Ytyp
|
Then:
X
f
Eρaf (Cu(t−2) )
X X ∂Eρcb
∂Cρ′ cf
c
(t−1)
Cw
.
w6=u
An interlude, understanding derivatives of E:
Z
X
∂Iρab
Rρab (g)Rρ′ cd (g) exp
=
Cρ′′ a′ b′ Rρ′′ a′ b′ (g) dg.
∂Cρ′ cd
G
′′ ′ ′
ρ ab
In particular,
∂Itriv
=
∂Cρab
Z
Rρab (g) exp
G
= Iρab .
X
ρ′ a′ b′
20
Cρ′ a′ b′ Rρ′ a′ b′ (g) dg
∂Eρcb
(t−1)
∂Cρ′ cf Cw
that
Note the following convenient identity:
Z
X ∂Iρcb
=
∂Cρcf
G
c
= dρ
X
Rρcb (g)Rρcf (g) exp
c
Z
X
G
= dρ
!
Z
G
= dρ δbf
ρ(g)cb ρ(g
−1
)f c
c
ρ(g −1 g)f b exp
Z
X
!
X
Cρ′ a′ b′ Rρ′ a′ b′ (g) dg
ρ′ a′ b′
exp
X
Cρ′ a′ b′ Rρ′ a′ b′ (g) dg
ρ′ a′ b′
Cρ′ a′ b′ Rρ′ a′ b′ (g) dg
ρ′ a′ b′
exp
G
X
Cρ′ a′ b′ Rρ′ a′ b′ (g) dg
ρ′ a′ b′
= dρ δbf Itriv (C).
Recalling that Eρab (C) = Iρab (C)/Itriv (C), we have
X ∂Eρcb
Itriv
=
∂Cρcf
c
P
∂Iρcb
c ∂Cρcf −
2
Itriv
= dρ δbf −
X
c
P
c Iρcb Iρcf
Eρcb (C)Eρcf (C)
= (dρ I − Eρ (C)∗ Eρ (C))f b .
Thus we obtain the following form for the Onsager correction:
X
ρ 2
−2
(t−2)
(t−1) ∗
(t−1)
Ons(t)
)Mρ(t) , Mρ(t) =
dρ I − Eρ (Cw
) Eρ (Cw
),
u,ρ = dρ |Ytyp | Eρ (Cu
w
with each AMP iteration reading as
(t)
Cu,ρ
= d−1
ρ
X
w6=u
6
ρ
(t−1)
Yuw
Eρ (Cw
) − Ons(t)
u,ρ .
MMSE derivation and state evolution
The goal of this section is to derive the state evolution equations that govern the behavior of AMP on
the Gaussian synchronization model of Section 3.4 (in the large n limit). Along the way, we will give an
alternative derivation of the algorithm (excluding the Onsager term) which shows that the nonlinear function
E has an interpretation as an MMSE (minimum mean squared error) estimator. This derivation is similar to
[DAM16] and based on ideas first introduced by [DMM09]. We do not give a proof that the state evolution
equations derived here are correct (i.e. that AMP obeys them) but we will argue for their correctness in
Section 7.
6.1
MMSE estimator
We begin by defining a ‘scalar’ problem: a simplification of the Gaussian synchronization model where we
attempt to recover a single group element from noisy measurements. We will be able to analyze the Gaussian
synchronization model by connection to this simpler model. (This is the idea of single letterization from
information theory.) Suppose there is an unknown group element g drawn uniformly from G (Haar measure)
and for each irreducible representation ρ in our list P we are given a measurement uρ = µρ ρ(g) + σρ zρ
(for some constants µρ , σρ ). Here zρ is a dρ × dρ non-symmetric matrix of Gaussian entries (real, complex,
or block-quaternion, depending on the type of ρ) with all entries (or blocks) independent and each entry
21
normalized to have expected squared-norm 1. (Note that zρ is the same as an off-diagonal block of the
matrix Wρ from Section 3.4.) For ρ of complex type, we only get a measurement uρ for one representation
in each conjugate pair, and define uρ = uρ . The MMSE estimator for ρ(g) (minimizing the matrix mean
d − ρ(g)k2 ) is simply the conditional expectation
squared error Ekρ(g)
F
h
E ρ(g) {uq }q
i
! Z
!
.
X 1
X 1
ρ(h) exp −
exp −
=
kuq − µq q(h)k2F
kuq − µq q(h)k2F
2
2
2σ
2σ
h∈G
h∈G
q
q
q
q
! Z
!
Z
.
X µq
X µq
ρ(h) exp
huq , q(h)i
huq , q(h)i
=
exp
σq2
σq2
h∈G
h∈G
q
q
!
µq
≡ Fρ
uq
σq2
q
Z
where
Fρ ({wq }q ) =
Z
ρ(h) exp
h∈G
X
q
!
hwq , q(h)i
.Z
exp
h∈G
X
q
!
hwq , q(h)i .
Here q ranges over irreducible representations in our list P (which includes both q and q for representations of
complex type). The likelihoods used in the above computation are derived similarly to those in Appendix A.
We recognize F as a rescaling of the function E from the AMP update step.
6.2
AMP update step
Consider the Gaussian observation model Mρ =
λρ
∗
√1
n Xρ Xρ + ndρ Wρ
from Section 3.4. Similarly to [DAM16],
the MMSE-AMP update step (without Onsager term) is
Uρt+1
= M ρ Fρ
µtq
Ut
(σqt )2 q
!
q
where t indicates the timestep and µtρ , σρt will be defined based on state evolution below. Here the AMP
state Uρt is ndρ × dρ with a dρ × dρ block for each vertex. Fρ is applied to each of these blocks separately.
We will motivate this AMP update step below, but notice its similarity to the MMSE estimator above.
6.3
State evolution
The idea of state evolution is that the AMP iterates can be approximately modeled as ‘signal’ plus ‘noise’
[DMM09]. Namely, we postulate that Uρt = µtρ Xρ + σρt Zρ for some constants µt , σt , where Zρ is a ndρ × dρ
Gaussian noise matrix with each dρ × dρ block independently distributed like zρ (from the scalar model)
with Zρ = Zρ for conjugate pairs. Recall Xρ has blocks ρ(gu ), the ground truth. Note that this sheds light
on the AMP update step above: at each iteration we are given Uqt , a noisy copy of the ground truth; the
first thing we do is to apply the MMSE estimator entrywise.
We will derive a recurrence for how the parameters µρ and σρ change after one iteration. To do this,
we assume that the noise Wρ is independent from Zρ at each timestep. This assumption is far from true;
however, it turns out that AMP’s Onsager term corrects for this (e.g. [BM11]). In other words, we derive
state evolution by omitting the Onsager term and assuming independent noise at each timestep. Then if
we run AMP (with the Onsager term and the same noise at each timestep), it behaves according to state
evolution. We now derive state evolution:
22
Uρt+1
= M ρ Fρ
=
=
where γqt =
µtq
σqt
2
µtq
Ut
(σqt )2 q
!
q
!
µtq
µt Xq + σqt Zq
Fρ
(σqt )2 q
q
!
n
q
o
γqt Xq + γqt Zq
Wρ Fρ
1
λρ
Xρ Xρ∗ + p
Wρ
n
ndρ
λρ
1
Xρ Xρ∗ + p
n
ndρ
!
q
n
n
q
q
o
o
1
λρ
∗
t
t
t
t
+p
.
Xρ Xρ Fρ
γq X q + γq Z q
=
Wρ Fρ
γq X q + γq Z q
n
q
q
ndρ
First focus on the signal term:
n
n
q
q
o
o
λρ
γqt q(g) + γqt zq
≈ λρ Xρ Eg,zq ρ(g)∗ Fρ
γqt Xq + γqt Zq
Xρ Xρ∗ Fρ
n
q
q
where g is drawn from Haar measure on G, and zq is a non-symmetric Gaussian matrix of the appropriate
type (as in Section 6.1). Define Atρ ∈ Cdρ ×dρ to be the second matrix in the expression above:
n
q
o
γqt q(g) + γqt zq
Atρ ≡ Eg,zq ρ(g)∗ Fρ
.
q
We will see shortly that Atρ is a multiple atρ ∈ R of the identity and so we can now write the signal term as
λρ atρ Xρ . Therefore our new signal parameter is µt+1
= λρ atρ .
ρ
We take a short detour to state some properties of Atρ , which we prove in Appendix B.
Lemma 6.1. Atρ is a real multiple of the identity: Atρ = atρ Idρ for some atρ ∈ R. Furthermore, we have the
following equivalent formulas for atρ :
i
h
p
(i) Eg,zq ρ(g)∗ Fρ γqt q(g) + γqt zq q
∗
(ii) Eg,zq Fρ (· · · ) Fρ (· · · )
h
p
(iii) Ezq Fρ γqt Idq + γqt zq
∗
(iv) Ezq Fρ (· · · ) Fρ (· · · )
q
i
where · · · denotes the argument to Fρ from the previous line.
Returning to state evolution, we now focus on the noise term:
n
q
o
1
t
t
p
γq X q + γq Z q
.
Wρ Fρ
q
ndρ
23
Each entry of this ndρ × dρ matrix is Gaussian. The variance (expected squared-norm) of entry (i, j) is
(approximately)
n
ndρ
q
o
1 X
Fρ
γqt Xq + γqt Zq
ndρ
q k,j
k=1
2
n
dρ
q
o
X
1
Fρ
γqt q(g) + γqt zq
≈ Eg,zq
dρ
q k,j
k=1
1
= E [Fρ (· · · )∗ Fρ (· · · )]jj
dρ
1
= (Atρ )jj
dρ
1
= atρ .
dρ
We therefore have the new noise parameter (σρt+1 )2 =
atρ
dρ .
To summarize, we now have the state evolution recurrence µt+1
= λρ atρ and (σρt+1 )2 =
ρ
6.4
2
atρ
dρ .
Simplified AMP update step
Note that the state evolution recurrence implies the relation
µt+1
ρ
(σρt+1 )2
= dρ λρ .
Provided our initial values of µρ , σρ satisfy this relation (which can always be arranged by scaling the initial
Uρ appropriately), our AMP update step (without Onsager term) becomes
Uρt+1 = Mρ Fρ dρ λρ Uqt q .
This is convenient because we can implement AMP without keeping track of the state evolution parameters
p
µtρ , σρt . Also note that this variant of AMP matches the original derivation after the rescaling Cρt = dρ λρ Uρt
(and excluding the Onsager term).
6.5
Reduction to single parameter (per frequency)
We will rewrite the state evolution recurrence in terms of a single parameter per frequency. This parameter
µt 2
will be γρt , which was introduced earlier: γρt = σρt . Recall the state evolution recurrence µt+1
= λρ atρ
ρ
ρ
and
(σρt+1 )2
=
atρ
dρ .
We therefore have the update step
γρt+1 =
µt+1
ρ
σρt+1
!2
=
(λρ atρ )2
= dρ λ2ρ atρ .
atρ /dρ
Using part (iii) of Lemma 6.1 we can write this as:
γρt+1 = λ2ρ Ezq Tr Fρ
n
γqt Idq +
q
o
γqt zq
.
q
(3)
This is the final form of our state evolution recurrence. The relation between µρ , σρ , γρ can be summarized
as γρ = dρ λρ µρ = d2ρ λ2ρ σρ2 .
We expect that the state evolution recurrence (3) exactly governs the behavior of AMP in the large n
limit. Although the derivation above was heuristic, we discuss its correctness in Section 7. There is a caveat
24
regarding how it should be initialized (see Section 7) but in practice we can imagine the initial γ value
is a small random vector. (Note that the initialization γ = ~0 is problematic because state evolution will
never leave zero.) We expect that state evolution converges to some fixed point of the recurrence. Some
complications arise if there are multiple fixed points (see Section 8) but we expect there to be a unique fixed
point that is reached from any small initialization. This fixed point γ ∗ describes the output of AMP in the
sense that (following the postulate of state evolution) the final AMP iterate p
is approximately distributed
as Uρ ≈ µ∗ρ Xρ + σρ∗ Zρ , which in terms of γ ∗ is (up to scaling) Uρ ≈ γρ∗ Xρ + γρ∗ Zρ . (See [BM11] for the
precise sense in which we expect this to be true.) Note that one can use this to translate a γ ∗ value into any
measure of performance, such as MSE. This gives an exact asymptotic characterization of the performance
of AMP for any set of λρ values. The most prominent feature of AMP’s performance is the threshold at
λ = 1, which we derive in the next section.
One can check that our state evolution recurrence matches the Bayes-optimal cavity and replica predictions of [JMR16] for Z/2 and U (1) with one frequency. Indeed, we expect AMP to be statistically optimal in
these settings (and many others too; see Section 8), and this has been proven rigorously for Z/2 [DAM16].
6.6
Threshold at λ = 1
In this section we use the state evolution occurrence to derive the threshold above which AMP achieves
nontrivial recovery. In particular, if λρ < 1 for all frequencies ρ then the AMP fixed point γ ∗ is equal to
the zero vector and so AMP gives trivial performance (random guessing) in the large n limit. On the other
hand, if λρ > 1 for at least one frequency ρ then γ ∗ is nonzero and AMP achieves nontrivial recovery.
The zero vector is always a fixed point of state evolution. Whether or not AMP achieves nontrivial
performance depends on whether the zero vector is a stable or unstable fixed point. Therefore we consider
the regime where γρ is small for all ρ. When the input {wq }q to Fρ is small, we can approximate Fρ by its
linearization.
"
# Z
Z
X
X
ρ(h) 1 +
hwq , q(h)i =
ρ(h)
hwq , q(h)i
Fρ {wq }q ≈
h
h
q
q
and so
Fρ {wq }q
ab
≈
=
Z
ρ(h)ab
h
X
X
wqcd q(h)cd
qcd
wqcd
Z
ρ(h)ab q(h)cd
h
qcd
=
X
wqcd
qcd
1
δρab,qcd
dρ
wρab
=
dρ
which means Fρ {wq }q ≈
wρ
dρ .
Now the state evolution update step becomes
γρt+1 = λ2ρ Ezq Tr Fρ
n
q
o
γqt Idq + γqt zq
q
1 t
γρ Idρ + γρt zq
≈ λ2ρ Ezq Tr
dρ
q
= λ2ρ γρt .
This means that when γ is small (but nonzero), γρ shrinks towards zero if λρ < 1 and grows in magnitude
if λρ > 1. We conclude the threshold at λ = 1.
25
7
Correctness of state evolution?
In this section we justify the heuristic derivation of state evolution in the previous section and argue for its
correctness. We first discuss prior work that provides a rigorous foundation for the methods we used, in
related settings. We then show numerically that our AMP algorithm obeys the state evolution equations.
7.1
Rigorous work on state evolution
State evolution was introduced along with AMP by [DMM09]. It was later proven rigorously that AMP
obeys state evolution in the large n limit (in a particular formal sense) for certain forms of the AMP iteration
[BM11, JM13]. In particular, Z/2 synchronization with Gaussian noise (a special case of our model) falls
into this framework and thus admits a rigorous analysis [DAM16]. Although the proofs of [BM11, JM13]
only consider the case of real-valued AMP, it has been stated [MAYB13] that the proof extends to the
complex-valued case. This covers our synchronization model over U (1) with one frequency. In order to cover
our general formulation of AMP over any group with any number of frequencies, one needs to replace the
complex numbers by a different real algebra (namely a product of matrix algebras). We expect that this
generalization should follow from the existing methods.
There is, however, an additional caveat involving the initialization of state evolution. In practice, we
initialize AMP to small random values. Recall that we only need to recover the group elements up to a
global right-multiplication and so there exists a favorable global right-multiplication so that our random
initialization has some correlation with the truth. However, this correlation is o(1) and corresponds to γ = ~0
in the large n limit. This means that technically, the formal proof of state evolution (say for Z/2) tells
us that for any fixed t, AMP achieves γ = ~0 after t iterations in the large n limit. Instead we would like
to show that after ω(1) iterations we achieve a nonzero γ. It appears that proving this would require a
non-asymptotic analysis of AMP, such as [RV16]. It may appear that this initialization issue can be fixed by
initializing AMP with a spectral method, which achieves Ω(1) correlation with the truth; however this does
not appear to easily work due to subtle issue about correlation between the noise and iterates. In practice,
the initialization issue is actually not an issue at all: with a small random initialization, AMP consistently
escapes from the trivial fixed point (provided some λ exceeds 1). One way to explain this is that when the
AMP messages are small, the nonlinear function F is essentially the identity (see Section 6.6) and so AMP
is essentially just the power method; this roughly means that AMP automatically initializes itself to the
output of the spectral method.
7.2
Experiments on state evolution
We now present experimental evidence that AMP obeys the state evolution equations. In Figure 7 we show
two experiments, one with U (1) and one with SO(3). In both cases we see that the performance of AMP
closely matches the state evolution prediction. We see some discrepancy near the λ = 1 threshold, which
can be attributed to the fact that here we are running AMP with finite n whereas state evolution describes
the n → ∞ behavior.
8
Statistical-to-computational gaps
In various settings it has been shown, using standard but non-rigorous methods from statistical physics, that
the analysis of AMP and state evolution yields a complete picture of the various “phase transitions” that occur
in a computational problem (e.g. [LKZ15b, LKZ15a]). In some settings, certain features of these predictions
have been confirmed rigorously (e.g. [KXZ16, BDM+ 16]). In this section we will use these methods to give
non-rigorous predictions about statistical-to-computational gaps in the Gaussian synchronization model.
In Section 6.6 we have seen that (in the large n limit) AMP achieves nontrivial recovery if and only if
λ > 1 on at least one frequency. In this section, we will see that it is sometimes statistically possible to
succeed below this threshold, although no known efficient algorithm achieves this. A rigorous analysis of an
26
−0.5
−1
−1.0
log error
log error
−2
−3
−4
−1.5
−2.0
−2.5
−5
−3.0
−6
−3.5
0.5
1.0
1.5
λ
2.0
2.5
3.0
0.5
1.0
1.5
λ
2.0
2.5
3.0
Figure 7: AMP compared to the state evolution equations experimentally. Left: U (1) with K frequencies,
for K = 1, 2, 3, 4 (from top to bottom) with n = 100. The solid line is AMP and the dotted line is the
state evolution prediction. The horizontal axis is the signal-to-noise ratio λ, which we take to be equal on all
frequencies. The vertical axis is the natural logarithm of error, which is defined as error = 1−|hx, x
bi/n| ∈ [0, 1]
where x ∈ U (1)n is the truth and x
b ∈ U (1)n is the (rounded) output of AMP. In particular, a log error value
of zero (top of the figure) indicates trivial recovery (random guessing), and lower values are better. Right:
b F ∈ [0, 1] where
SO(3) with one frequency, with n = 50. Now error is measured as error = 1 − √13n kX ⊤ Xk
b are 3n × n matrices whose 3 × 3 blocks encode elements of SO(3) via the standard representation (3D
X, X
rotation matrices).
inefficient estimator has indeed confirmed that the λ = 1 threshold can be beaten in some cases [PWBM16];
the non-rigorous computations in this section give sharp predictions for exactly when this is possible.
8.1
Free energy
Recall the parameter γ = {γρ }ρ from the state evolution recurrence (3); γ captures the amount of information
that AMP’s current state has about each frequency, with γρ = 0 indicating no information and γρ → ∞
indicating complete knowledge.
An important quantity is the Bethe free energy per variable (also called the replica symmetric potential
function) of a state γ, which for the Gaussian synchronization model is given (up to constants) by
!
X
1X 2 2 1X
1 X γρ2
√
f (γ) = −
− Ez log Eg exp
d λ +
dρ γρ +
hρ(g), γρ Idρ + γρ zρ i
4 ρ ρ ρ 2 ρ
4 ρ λ2ρ
ρ
where zρ is a dρ × dρ matrix of i.i.d. standard Gaussians (of the appropriate type: real, complex, or quaternionic, depending on ρ), and g is drawn from Haar measure on the group. We do not include the derivation
of this expression, but it can be computed from belief propagation (as in [LKZ15a]) or from the replica
calculation (as in [JMR16]).
Roughly speaking, the interpretation of the Bethe free energy is that it is the objective value that AMP is
trying to minimize. AMP can be thought of as starting from the origin γ = 0 and performing naı̈ve gradient
descent in the free energy landscape until it reaches a local minimum; the value of γ at this minimum
describes the final state of AMP. (It can be shown that the fixed points of the state evolution recurrence
(3) are precisely the stationary points of the Bethe free energy.) As is standard for these types of problems,
we conjecture that AMP is optimal among all polynomial-time algorithms. However, with no restriction on
efficiency, the information-theoretically optimal estimator is given by the global minimum of the free energy.
(This has been shown rigorously for the related problem of rank-one matrix estimation [BDM+ 16].) The
27
intuition here is that the optimal estimator should use exhaustive search to enumerate all fixed points of
AMP and return the one of lowest Bethe free energy. Note that just because we can compute the γ value
that minimizes the Bethe free energy it does not mean we can achieve this γ with an efficient algorithm; γ
represents correlation between the AMP iterates and the ground truth, and since the truth is unknown it is
hard to find iterates that have a prescribed γ.
8.2
Examples
We now examine the Bethe free energy landscapes of some specific synchronization problems at various
values of λ, and discuss the implications. Our primary examples will be U (1) and Z/L with various numbers
of frequencies, as discussed in Section 3.5. Recall that references to U (1) or Z/L “with K frequencies” means
that observations are band-limited to the Fourier modes eikθ with |k| ≤ K.
Our first example is U (1) with a single frequency, shown in Figure 8. Here we see that the problem
transitions from (statistically) ‘impossible’ to ‘easy’ (AMP achieves nontrivial recovery) at λ = 1, with no
(computationally) ‘hard’ regime. In particular, AMP is statistically optimal for every value of λ.
0.2
−0.62
0.1
−0.64
0.0
−0.66
f
f
−0.1
−0.68
−0.2
−0.70
−0.3
−0.72
−0.4
−0.5
0.0
0.2
0.4
0.6
γ
0.8
1.0
−0.74
0.0
1.2
0.2
0.4
0.6
γ
0.8
1.0
1.2
Figure 8: Free energy landscape for U (1) with 1 frequency. Left: λ < 1. The global minimum of free
energy occurs at γ = 0, indicating that AMP or any other estimator achieves zero correlation with the truth.
Right: λ > 1. Now the global minimum occurs at nonzero γ, and this is achieves by AMP. Therefore AMP
achieves the statistically optimal MSE (mean squared error). This MSE departs continuously from zero at
the λ = 1 threshold.
Our next example is a single-frequency problem that exhibits a computational gap (a ‘hard’ phase). In
Figure 9 we take the alternating group A4 with its irreducible 3-dimensional representation as the rotational
symmetries of a tetrahedron. When λ > 1, AMP achieves statistically optimal performance but when λ is
below 1 but sufficiently large, AMP gives trivial performance while the statistically optimal estimator gives
nontrivial performance. This means we have a computational gap, i.e. there are values of λ below the AMP
threshold (λ = 1) where nontrivial recovery is statistically possible.
Next we move on to some 2-frequency problems, where γ is now a 2-dimensional vector. In Figure 10 we
see an example with no computational gap, and an example with a computational gap. Note that the free
energy landscape at the AMP threshold λ = (1, . . . , 1) reveals whether or not a computational gap exists:
there is a gap if and only if the global minimum of free energy does not occur at the origin.
We now state some experimental results regarding which synchronization problems have computational
gaps. For U (1) with (the first) K frequencies, there is a gap iff K ≥ 3. For Z/L with K frequencies, there
is a gap for K ≥ 3 and no gap for K = 1; when K = 2 there is only a gap for L = 5. For SO(3) with K
frequencies, there is a gap iff K ≥ 2.
In [PWBM16] we gave some rigorous lower bounds for Gaussian synchronization problems, showing for
28
−0.2
−1.835
−0.4
−1.840
−0.6
−1.845
f
f
−0.8
−1.850
−1.0
−1.855
−1.2
−1.860
−1.4
−1.6
0.0
0.5
1.0
1.5
2.0
γ
2.5
3.0
0.0
(a) λ = 0.8, impossible
0.2
0.4
0.6
0.8
γ
1.0
1.2
1.4
1.6
(b) λ = 0.91, impossible
−1.894
−2.6
−1.896
−2.7
−1.898
−2.8
f
f
−1.900
−1.902
−2.9
−1.904
−3.0
−1.906
−3.1
−1.908
−1.910
0.0
0.2
0.4
0.6
0.8
γ
1.0
1.2
1.4
1.6
−3.2
1.8
(c) λ = 0.92, hard
0
1
2
γ
3
4
5
(d) λ = 1.1, easy
Figure 9: Free energy landscape for A4 with 1 frequency: the standard 3-dimensional representation (rigid
motions of a tetrahedron). (a) λ = 0.8. The global minimizer is γ = 0 so no estimator achieves nontrivial
recovery. (b) λ = 9.1. A new local minimum in the free energy has appeared, but the global minimum
is still at γ = 0 and so nontrivial recovery remains impossible. (c) λ = 9.2. AMP is stuck at γ = 0 but
the (inefficient) statistically optimal estimator achieves a nontrivial γ (the global minimum). AMP is not
statistically optimal. This computational gap appears at λ ≈ 0.913, at which point the global minimizer
transitions discontinuously from γ = 0 to some positive value. (d) λ = 1.1. AMP achieves optimal recovery.
The AMP γ value transitions discontinuously from zero to optimal at λ = 1.
instance that U (1) with one frequency is statistically impossible below λ = 1. The non-rigorous results above
predict further results that we were unable to show rigorously, e.g. U (1) with two frequencies and Z/3 (with
one frequency) are statistically impossible below the λ = 1 threshold.
In the examples above we saw that when every λ is below 1, AMP gives trivial performance, and when
some λ exceeds 1, AMP gives statistically optimal performance. However, the behavior can be more complicated, namely AMP can exhibit nontrivial but sub-optimal performance. In Figure 11 we show such an
example: Z/25 with 9 frequencies.
29
0.35
0.30
0.30
0.25
0.25
0.20
0.20
γ2
γ2
0.35
0.15
0.15
0.10
0.10
0.05
0.05
0.00
0.00
0.05
0.10
0.15
0.20
γ1
0.25
0.30
0.00
0.00
0.35
0.05
0.10
0.15
0.20
γ1
0.25
0.30
0.35
Figure 10: Free energy landscape for 2-frequency problems at the critical value λ = (1, 1). Darker colors
indicate lower free energy. Left: Z/6 with 2 frequencies. Here the origin is the global minimizer of free
energy and so there is no computational gap, i.e. nontrivial recovery is statistically impossible when both λ1
and λ2 are below 1. Right: Z/5 with 2 frequencies. Here the global minimizer (marked with an X) does
not lie at the origin and so there is a computational gap, i.e. there is a regime where nontrivial recovery is
statistically possible yet AMP fails.
1.2
1
γ
0.8
0.6
0.4
0.2
0
0
5
10
15
20
25
30
35
40
t
Figure 11: An example where AMP gives nontrivial but sub-optimal performance. Here we take Z/25 with 9
frequencies. Set λk = 0.8 for k = 1, . . . , 8 and λ9 = 1.1. Since we cannot visualize the free energy landscape
in 9 dimensions, we instead plot the state evolution recurrence as it evolves over time (number of iterations
t) from two different starting points. The bottom two curves correspond to AMP’s performance, where we
initialize γ to be small: γ = (0.05, 0.05). The solid line is γ9 and the dashed line is γ1 (which is representative
of γ2 , . . . , γ8 ). The top two curves correspond to a “warm start” γ = (0.7, 0.7). We see that with the warm
start, state evolution converges to a different fixed point with larger γ values, and thus better correlation with
the truth. Furthermore, this fixed point has lower free energy (not shown) than the lower one, indicating
that the information-theoretically optimal estimator outperforms AMP.
Acknowledgements
The authors would like to thank Amit Singer, Roy Lederman, Yutong Chen, Nicholas Boumal, others from
Amit Singer’s group, and Yash Deshpande, for helpful discussions.
30
References
[ABBS14]
Emmanuel Abbe, Afonso S Bandeira, Annina Bracher, and Amit Singer. Decoding binary
node labels from censored edge measurements: Phase transition and efficient recovery. IEEE
Transactions on Network Science and Engineering, 1(1):10–22, 2014.
[ABH16]
Emmanuel Abbe, Afonso S Bandeira, and Georgina Hall. Exact recovery in the stochastic block
model. IEEE Transactions on Information Theory, 62(1):471–487, 2016.
[ARC06]
Amit Agrawal, Ramesh Raskar, and Rama Chellappa. What is the range of surface reconstructions from a gradient field? In European Conference on Computer Vision, pages 578–591.
Springer, 2006.
[AW08]
Arash A Amini and Martin J Wainwright. High-dimensional analysis of semidefinite relaxations
for sparse principal components. In 2008 IEEE International Symposium on Information Theory,
pages 2454–2458. IEEE, 2008.
[Ban15]
Afonso S Bandeira. Convex Relaxations for Certain Inverse Problems on Graphs. PhD thesis,
Princeton University, June 2015.
[BBS14]
Afonso S Bandeira, Nicolas Boumal, and Amit Singer. Tightness of the maximum likelihood
semidefinite relaxation for angular synchronization. arXiv:1411.3272, 2014.
[BCS15]
Afonso S. Bandeira, Yutong Chen, and Amit Singer. Non-unique games over compact groups
and orientation estimation in cryo-EM. arXiv:1505.03840, May 2015.
[BCSZ14]
Afonso S Bandeira, Moses Charikar, Amit Singer, and Andy Zhu. Multireference alignment using
semidefinite programming. In Proceedings of the 5th conference on Innovations in theoretical
computer science, pages 459–470. ACM, 2014.
[BDM+ 16] Jean Barbier, Mohamad Dia, Nicolas Macris, Florent Krzakala, Thibault Lesieur, and Lenka
Zdeborova. Mutual information for symmetric rank-one matrix estimation: A proof of the
replica formula. arXiv:1606.04142, 2016.
[BGN11]
Florent Benaych-Georges and Raj Rao Nadakuditi. The eigenvalues and eigenvectors of finite,
low rank perturbations of large random matrices. Advances in Mathematics, 227(1):494–521,
2011.
[Bis99]
Christopher M Bishop. Bayesian PCA. Advances in neural information processing systems,
pages 382–388, 1999.
[BM11]
Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense graphs, with
applications to compressed sensing. IEEE Transactions on Information Theory, 57(2):764–785,
2011.
[Bou16]
Nicolas Boumal. Nonconvex phase synchronization. arXiv:1601.06114, 2016.
[BR13a]
Quentin Berthet and Philippe Rigollet. Complexity theoretic lower bounds for sparse principal
component detection. In COLT, pages 1046–1066, 2013.
[BR13b]
Quentin Berthet and Philippe Rigollet. Optimal detection of sparse principal components in
high dimension. The Annals of Statistics, 41(4):1780–1815, 2013.
[BtD13]
Theodor Bröcker and Tammo tom Dieck. Representations of compact Lie groups, volume 98.
Springer Science & Business Media, 2013.
[CC16]
Yuxin Chen and Emmanuel Candès. The projected power method: An efficient algorithm for
joint alignment from pairwise differences. arXiv:1609.05820, 2016.
31
[CLS12]
Mihai Cucuringu, Yaron Lipman, and Amit Singer. Sensor network localization by eigenvector
synchronization over the euclidean group. ACM Transactions on Sensor Networks (TOSN),
8(3):19, 2012.
[Coh13]
Jon Cohen. Is high-tech view of HIV too good to be true? Science, 341(6145):443–444, 2013.
[CSSS10]
Ronald R Coifman, Yoel Shkolnisky, Fred J Sigworth, and Amit Singer. Reference free structure
determination through eigenvectors of center of mass operators. Applied and Computational
Harmonic Analysis, 28(3):296–312, 2010.
[DAM16]
Yash Deshpande, Emmanuel Abbe, and Andrea Montanari. Asymptotic mutual information
for the binary stochastic block model. In 2016 IEEE International Symposium on Information
Theory (ISIT), pages 185–189. IEEE, 2016.
[DM14]
Yash Deshpande and Andrea Montanari. Information-theoretically optimal sparse pca. In 2014
IEEE International Symposium on Information Theory, pages 2197–2201. IEEE, 2014.
p
Yash Deshpande and Andrea Montanari. Finding hidden cliques of size N/e in nearly linear
time. Foundations of Computational Mathematics, 15(4):1069–1128, 2015.
[DM15]
[DMM09]
David L Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for compressed sensing. Proceedings of the National Academy of Sciences, 106(45):18914–18919, 2009.
[DMM10]
David L Donoho, Arian Maleki, and Andrea Montanari. Message passing algorithms for compressed sensing: I. motivation and construction. IEEE Information Theory Workshop (ITW),
pages 115–144, 2010.
[DMR14]
Yash Deshpande, Andrea Montanari, and Emile Richard. Cone-constrained principal component
analysis. In Advances in Neural Information Processing Systems, pages 2717–2725, 2014.
[ELW10]
Daniel Egloff, Markus Leippold, and Liuren Wu. The term structure of variance swap rates and
optimal variance swap investments. Journal of Financial and Quantitative Analysis, 45(5):1279,
2010.
[FP07]
Delphine Féral and Sandrine Péché. The largest eigenvalue of rank one deformation of large
Wigner matrices. Communications in Mathematical Physics, 272(1):185–228, 2007.
[GK06]
Arvind Giridhar and PR Kumar. Distributed clock synchronization over wireless networks:
Algorithms and analysis. In Proceedings of the 45th IEEE Conference on Decision and Control,
pages 4915–4920. IEEE, 2006.
[GW95]
Michel X Goemans and David P Williamson. Improved approximation algorithms for maximum
cut and satisfiability problems using semidefinite programming. Journal of the ACM (JACM),
42(6):1115–1145, 1995.
[HWX16]
Bruce Hajek, Yihong Wu, and Jiaming Xu. Achieving exact cluster recovery threshold via
semidefinite programming. IEEE Transactions on Information Theory, 62(5):2788–2797, 2016.
[JM13]
Adel Javanmard and Andrea Montanari. State evolution for general approximate message passing algorithms, with applications to spatial coupling. Information and Inference, 2(2):115–144,
2013.
[JMR16]
Adel Javanmard, Andrea Montanari, and Federico Ricci-Tersenghi. Phase transitions in semidefinite relaxations. Proceedings of the National Academy of Sciences, 113(16):E2218–E2223, 2016.
[KXZ16]
Florent Krzakala, Jiaming Xu, and Lenka Zdeborová. Mutual information in rank-one matrix
estimation. arXiv:1603.08447, 2016.
32
[LKZ15a]
Thibault Lesieur, Florent Krzakala, and Lenka Zdeborov. Mmse of probabilistic low-rank matrix
estimation: Universality with respect to the output channel. In 2015 53rd Annual Allerton
Conference on Communication, Control, and Computing (Allerton), pages 680–687. IEEE, 2015.
[LKZ15b]
Thibault Lesieur, Florent Krzakala, and Lenka Zdeborová. Phase transitions in sparse PCA. In
2015 IEEE International Symposium on Information Theory (ISIT), pages 1635–1639. IEEE,
2015.
[LS91]
Robert B Litterman and Jose Scheinkman. Common factors affecting bond returns. The Journal
of Fixed Income, 1(1):54–61, 1991.
[MAYB13] Arian Maleki, Laura Anitori, Zai Yang, and Richard G Baraniuk. Asymptotic analysis of complex
LASSO via complex approximate message passing (CAMP). IEEE Transactions on Information
Theory, 59(7):4290–4308, 2013.
[McS01]
Frank McSherry. Spectral partitioning of random graphs. In Foundations of Computer Science,
2001. Proceedings. 42nd IEEE Symposium on, pages 529–537. IEEE, 2001.
[MM09]
Marc Mezard and Andrea Montanari. Information, physics, and computation. Oxford University
Press, 2009.
[MPV86]
Marc Mézard, Giorgio Parisi, and MA Virasoro. SK model: The replica solution without replicas.
Europhys. Lett, 1(2):77–82, 1986.
[MR16]
Andrea Montanari and Emile Richard. Non-negative principal component analysis: Message passing algorithms and sharp asymptotics. IEEE Transactions on Information Theory,
62(3):1458–1484, 2016.
[MS16]
Andrea Montanari and Subhabrata Sen. Semidefinite programs on sparse random graphs and
their application to community detection. In Proceedings of the 48th Annual ACM SIGACT
Symposium on Theory of Computing, pages 814–827. ACM, 2016.
[MW15]
Zongming Ma and Yihong Wu. Computational barriers in minimax submatrix detection. The
Annals of Statistics, 43(3):1089–1116, 2015.
[Pea86]
Judea Pearl. Fusion, propagation, and structuring in belief networks. Artificial intelligence,
29(3):241–288, 1986.
[PWBM16] Amelia Perry, Alexander S Wein, Afonso S Bandeira, and Ankur Moitra. Optimality and suboptimality of PCA for spiked random matrices and synchronization. arXiv:1609.05573, 2016.
[RF12]
Sundeep Rangan and Alyson K Fletcher. Iterative estimation of constrained rank-one matrices
in noise. In IEEE International Symposium on Information Theory (ISIT), pages 1246–1250.
IEEE, 2012.
[RV16]
Cynthia Rush and Ramji Venkataramanan. Finite sample analysis of approximate message
passing. arXiv:1606.01800, 2016.
[RW01]
J Rubinstein and G Wolansky. Reconstruction of optical surfaces from ray data. Optical review,
8(4):281–283, 2001.
[Sin11]
Amit Singer. Angular synchronization by eigenvectors and semidefinite programming. Applied
and computational harmonic analysis, 30(1):20–36, 2011.
[SS11]
Amit Singer and Yoel Shkolnisky. Three-dimensional structure determination from common
lines in cryo-EM by eigenvectors and semidefinite programming. SIAM Journal on Imaging
Sciences, 4(2):543–572, 2011.
[TAP77]
David J Thouless, Philip W Anderson, and Robert G Palmer. Solution of ‘Solvable model of a
spin glass’. Philosophical Magazine, 35(3):593–601, 1977.
33
A
Log-likelihood expansion for the Gaussian observation model
In this section we show how the Gaussian observation model fits into the graphical model formulation by
deriving the corresponding coefficient matrices Yρ . In particular, we show that Yρ = dρ λρ Mρ , a scalar
multiple of the observed Gaussian P
matrix.
We can write log Luv (gu , gv ) = ρ log Lρuv (gu , gv ) and consider each representation separately. There are
three cases for the three types of representations (see Section 3.1.3).
For convenience we recall the Gaussian observation model:
Mρ =
Restricting to the u, v submatrix:
ρ
Muv
=
Real type.
λρ
1
Xρ Xρ∗ + p
Wρ .
n
ndρ
1
λρ
ρ
ρ(gu gv−1 ) + p
Wuv
.
n
ndρ
ρ
Let ρ be of real type. Recall that in this case, each entry of Wuv
is N (0, 1). We have
log Lρuv (gu , gv ) =
λρ
−ndρ
ρ
ρ(gu gv−1 )
Muv
−
2
n
2
F
ρ
= dρ λρ Muv
, ρ(gu gv−1 ) + const.
ρ
Here k · kF denotes the Frobenius norm. The additive constant in the last step depends on Muv
but not on
ρ
ρ
gu , gv . Thus the log-likelihood coefficients are Yuv
= dρ λρ Muv
and so Yρ = dρ λρ Mρ .
Complex type. Now consider a representation ρ of complex type, along with its conjugate ρ. Recall that
ρ
in this case, each entry of Wuv
has independent real and imaginary parts drawn from N (0, 1/2). We have
2
λρ
ρ(gu gv−1 )
n
E
D F
ρ
ρ
−1
, ρ(gu gv−1 ) + const.
= dρ λρ Muv , ρ(gu gv ) + dρ λρ Muv
ρ
log Lρuv (gu , gv ) = −ndρ Muv
−
Therefore we have Yρ = dρ λρ Mρ and Yρ = dρ λρ Mρ = dρ λρ Mρ .
ρ
Quaternionic type. Now consider a representation ρ of quaternionic type. Recall that in this case, Wuv
is
block-quaternion where each 2 × 2 block encodes a quaternion value whose 4 entries are drawn independently
from N (0, 1/2). Note the following relation between the norm of a quaternion and its corresponding 2 × 2
matrix:
2
1
a + bi c + di
.
ka + bi + cj + dkk2 ≡ a2 + b2 + c2 + d2 =
2 −c + di a − bi F
We have
2
1
λρ
ρ
ρ(gu gv−1 )
Muv
−
2
n
F
ρ
−1
= dρ λρ Re Muv , ρ(gu gv ) + const
log Lρuv (gu , gv ) = −ndρ ·
ρ
, ρ(gu gv−1 ) + const
= dρ λρ Muv
ρ
where Re denotes real part. In the last step we used the fact that Muv
and ρ(gu gv−1 ) are block-quaternion
and so their inner product is real (see Section 3.1.3). Therefore Yρ = dρ λρ Mρ .
34
B
Proof of Lemma 6.1
To see that (i) and (ii) are equal,
recall the
p interpretation of Fρ as a conditional expectation: Fρ (· · · ) =
E[ρ(g)| · · · ] where · · · stands for γqt q(g) + γqt zq q . (This is related to the Nishimori identities in statistical
physics.)
We have the following symmetry properties of Fρ .
Lemma B.1. (1) For any γqt ∈ R, zq ∈ Cdρ ×dρ , and g, h ∈ G, we have
Fρ
and
n
γqt q(hg) +
n
q
q
o
o
γqt q(g) + γqt q(h−1 )zq
= ρ(h)Fρ
γqt zq
q
q
n
n
q
q
o
o
t
t
−1
t
t
γq q(g) + γq zq q(h )
γq q(gh) + γq zq
= Fρ
Fρ
ρ(h).
q
(2) Therefore, if we define
fρ (g) ≡ Ezq Fρ
q
n
q
o
γqt q(g) + γqt zq
q
we have fρ (hg) = ρ(h)fρ (g) and fρ (gh) = fρ (g)ρ(h).
Proof. Part (1) is a straightforward computation using the definition of Fρ . Part (2) follows from part (1)
because zq has the same distribution as q(h−1 )zq and zq q(h−1 ).
We now return to the proof of Lemma 6.1. The equality of (i) and (iii) follows from part (2) of Lemma B.1.
The equality of (ii) and (iv) follows from part (1) of Lemma B.1. Combining this with the equality of (i)
and (ii) from above, we have now shown equality of (i),(ii),(iii),(iv). It remains to show that Atρ is a real
multiple of the identity.
Letting e ∈ G be the identity, we have
fρ (e)ρ(g) = fρ (eg) = fρ (ge) = ρ(g)fρ (e)
and so by Schur’s lemma, this means fρ (e) is a (possibly complex) multiple of the identity. But fρ (e) is just
(iii), so we are done. To see that the multiple atρ is real, note that the trace of (ii) is real.
35
| 8cs.DS
|
arXiv:cs/0312031v1 [cs.DC] 16 Dec 2003
Theory and Practice of Logic Programming 1(3) 251–282, May 2001.
1
Distributed WWW Programming using
(Ciao-)Prolog and the PiLLoW library∗
DANIEL CABEZA and MANUEL HERMENEGILDO
CLIP Group
http://www.clip.dia.fi.upm.es – http://www.cliplab.org
Facultad de Informática
Universidad Politécnica de Madrid (UPM)
28660-Boadilla del Monte, Madrid, Spain
(e-mail: {dcabeza,herme}@fi.upm.es)
Abstract
We discuss from a practical point of view a number of issues involved in writing distributed
Internet and WWW applications using LP/CLP systems. We describe PiLLoW, a publicdomain Internet and WWW programming library for LP/CLP systems that we have
designed in order to simplify the process of writing such applications. PiLLoW provides
facilities for accessing documents and code on the WWW; parsing, manipulating and
generating HTML and XML structured documents and data; producing HTML forms;
writing form handlers and CGI-scripts; and processing HTML/XML templates. An important contribution of PiLLoW is to model HTML/XML code (and, thus, the content
of WWW pages) as terms. The PiLLoW library has been developed in the context of the
Ciao Prolog system, but it has been adapted to a number of popular LP/CLP systems,
supporting most of its functionality. We also describe the use of concurrency and a highlevel model of client-server interaction, Ciao Prolog’s active modules, in the context of
WWW programming. We propose a solution for client-side downloading and execution of
Prolog code, using generic browsers. Finally, we also provide an overview of related work
on the topic.
Keywords: WWW, HTML, XML, CGI, HTTP, Distributed Execution, (Constraint)
Logic Programming.
1 Introduction
The wide diffusion of the Internet and the popularity of the “World Wide
Web” (Berners-Lee et al., 1994) –WWW– protocols are effectively providing a
novel platform that facilitates the development of new classes of portable and
user-friendly distributed applications. Good support for network connectivity and
the protocols and communication architectures of this novel platform are obviously requirements for any programming tool to be useful in this arena. However, this alone may not be enough. It seems natural that significant parts
0
This paper is an expanded and improved version of (Cabeza & Hermenegildo, 1996a;
Cabeza & Hermenegildo, 1996b; Cabeza & Hermenegildo, 1997; Cabeza et al., 1996).
2
D. Cabeza and M. Hermenegildo
of network applications will require symbolic and numeric capabilities which
are not necessarily related with distribution. Important such capabilities are,
for example, high-level symbolic information processing, dealing with combinatorial problems, and natural language processing in general. Logic Programming (LP) (Kowalski, 1974; Colmerauer, 1975) and Constraint Logic Programming
(CLP) systems (Jaffar & Lassez, 1987; Van Hentenryck, 1989; Colmerauer, 1990;
M. Dincbas & Van Hentenryck, 1990; ECRC, 1993) have been shown particularly
successful at tackling these issues (see, for example, the proceedings of recent conferences on the “Practical Applications of Prolog” and “Practical Applications of
Constraint Technology”). It seems natural to study how LP/CLP technology fares
in developing applications which have to operate over the Internet.
In fact, Prolog, its concurrent and constraint based extensions, and logic programming languages in general have many characteristics which appear to set
them particularly well placed for making an impact on the development of practical networked applications, ranging from the simple to the quite sophisticated.
Notably, LP/CLP systems share many characteristics with other recently proposed
network programming tools, such as Java, including dynamic memory management,
well-behaved structure and pointer manipulation, robustness, and compilation to
architecture-independent bytecode. Furthermore, and unlike the scripting or application languages currently being proposed (e.g., shell scripts, Perl, Java, etc.),
LP/CLP systems offer a quite unique set of additional features including dynamic
databases, search facilities, grammars, sophisticated meta-programming, and well
understood semantics.
In addition, most LP/CLP systems also already offer some kind of low level support for remote communication using Internet protocols. This generally involves
providing a sockets (ports) interface whereby it is possible to make remote data connections via the Internet’s native protocol, TCP/IP. A few systems support higherlevel communication layers on top of this interface including linda-style blackboards (e.g., SICStus Prolog (Carlsson, 1988), Ciao (Carro & Hermenegildo, 1999;
Cabeza & Hermenegildo, 1995;
Hermenegildo & CLIP Group, 1994;
Hermenegildo et al., 1995a;
Hermenegildo et al., 1999a;
Hermenegildo et al., 1999b;
Bueno et al., 1997),
BinProlog/µ2Prolog (Tarau, 1996; De Bosschere, 1989), etc.) or shared variable-based communication (e.g., KL1 (Chikayama et al., 1994), AKL (Janson & Haridi, 1991), Oz
(Smolka, 1994), Ciao (Hermenegildo et al., 1995b; Cabeza & Hermenegildo, 1995),
etc.). In some cases, this functionality is provided via libraries, building on top of
the basic TCP/IP primitives. This is the case, for example, of the SICStus and Ciao
distributed linda-style interfaces. In fact, as we have shown in previous work, sharedvariable based communication can also be implemented in conventional systems
via library predicates, by using attributed variables (Hermenegildo et al., 1995b;
Cabeza & Hermenegildo, 1995). In addition to these communication primitives,
several systems offer concurrency and even higher-level abstractions (distributed
objects, mobile code, ...) which are very useful for developing general-purpose distributed applications.
Our concrete interest here is WWW applications. These applications generally
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
3
use specific high-level protocols (such as HTTP or FTP), data formats (such as HTML
or XML), and application architectures (e.g., the CGI interface) which are different from, e.g., the shared-variable or linda-based protocols typically used in other
types of distributed applications. In this paper we study how good support for
these WWW-related protocols, data formats, and architectures can be provided for
LP/CLP systems, building on the widely available interfaces to the basic TCP/IP
protocols. Our aim is to discuss from a practical point of view a number of the new
issues involved in writing WWW applications using LP/CLP systems, as well as
the architecture of some typical solutions. In the process, we will describe PiLLoW
(“Programming in Logic Languages on the Web”), a public domain Internet/WWW
programming library for LP/CLP systems which, we argue, significantly simplifies
the process of writing such applications. PiLLoW provides facilities for generating
HTML/XML structured documents by handling them as Herbrand terms, producing HTML forms, writing form handlers, processing HTML/XML templates,
accessing and parsing WWW documents, etc. We also describe the architecture of
some relatively sophisticated application classes, using a high-level model of clientserver interaction, active modules (Cabeza & Hermenegildo, 1995). Finally we describe an architecture for automatic LP/CLP code downloading for local execution,
using just the library and generic browsers.
Apart from the tutorial value of the paper, we present a number of technical
contributions which include the idea of representing HTML and XML code (and
structured text in general) as Prolog terms, the use of the logical variable in such
terms leading to a model of an “HTML template” (a pair comprising a term with
free variables and a dictionary associating names to those variables), the notion
of “active logic modules” and its application to solving efficiency issues in CGI
interaction in a very simple way, the idea of “Prolog scripts” and its application
to CGIs, and the identification of a number of features that should be added to
existing systems in order to facilitate the programming of WWW applications –
mainly concurrency.
The argument throughout the paper is that, with only very small limitations in
functionality (which disappear when concurrency is added, as in systems such as
BinProlog/µ2-Prolog, AKL, Oz, KL1, and Ciao Prolog), it is possible to add an
extremely useful Internet/WWW programming layer to any LP/CLP system without making any significant changes in the implementation. We argue that this layer
can simplify the generation of applications in LP/CLP systems including active
WWW pages, search tools, content analyzers, indexers, software demonstrators,
collaborative work systems, MUDs and MOOs, code distributors, etc.
The purpose of the paper is also to serve as a tutorial, containing sufficient
information for developing relatively complex WWW applications in Prolog and
other LP and CLP languages using the PiLLoW library. The PiLLoW library has
been developed in the context of the Ciao Prolog system, but it has been adapted to
a number of other popular LP/CLP systems, supporting most of its functionality.
The Ciao Prolog system and the PiLLoW library can be freely downloaded from
http://www.clip.dia.fi.upm.es and http://www.cliplab.org.
4
D. Cabeza and M. Hermenegildo
www.xxx.yyy
HTTP
Server
1
WWW
Browser
std. doc. request
4
output
2
3
hello_world
Fig. 1. The CGI interface
2 Writing basic cgi-bin applications
The simplest way of writing WWW applications is through the use of the “Common
Gateway Interface” (CGI). A CGI executable is a standard executable file but such
that the HTTP server (the program that responds to HTTP requests in a machine
which serves a WWW site) can tell that it in fact contains a program that is to be
run, rather than a document text that is to be sent to the client (the browser) as
usual. The file can be distinguished by belonging to a special directory, commonly
named cgi-bin, or by a special filename ending, such as .cgi. This is normally set
during configuration of the HTTP server. The basic idea behind the CGI interface
is illustrated in Figure 1. When the user selects an address of a CGI executable
in a document, such as http://www.xxx.yyy/cgi-bin/hello_world (or perhaps
http://www.xxx.yyy/foo/hello_world.cgi) the browser issues a standard document request (1). The HTTP server, recognizing that it is a CGI executable rather
than a document, starts the executable (2), and during such execution stores the
output of the executable in a buffer (3). Upon termination of the executable, the
contents of the buffer (which should be in a format that the browser can handle,
such as HTML) are returned to the browser as if a normal page with that content
had been accessed (4).
The following is an example of how a very simple such executable can be written
in an LP/CLP language. The source might be as follows:1
main :write(’Content-type: text/html’), nl, nl,
write(’<HTML>’),
write(’Hello world.’),
write(’</HTML>’).
And the actual executable could be generated as usual, for example in the Ciao
1
Note that in the examples presented, and in order to shorten them, the HTML code may
be slightly simplified, and as a result of this it may not be completely standard-conforming.
However, the examples can be used as is with all popular browsers.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
5
system, using the standalone compiler, by writing at a UNIX shell “ciaoc -o
hello_world.cgi hello_world”.2 The executable then has to be placed in an
appropriate place (accessible via an HTTP address by a browser) and have the
right permissions for being executed by the server (for example, in some systems
this means being executable by the user “nobody”).
In systems which make executables through saved states (which usually have the
disadvantage of their generally large size), at the system prompt one could create
the executable by writing something like:
?:- compile(’hello_world.pl’),
save(’hello_world.cgi’), main.
3 LP/CLP Scripts for CGI Applications
CGI executables are often small- to medium-sized programs that perform relatively
simple tasks. This, added to the slow speed of the network connection in comparison with that of executing a program (which makes program execution speed less
important) has made scripting languages (such as shell scripts or Perl) very popular
for writing these programs. The popularity is due to the fact that no compilation
is necessary (extensive string handling capabilities also play an important role in
the case of Perl), and thus changes and updates to the program imply only editing
the source file.
Logic languages are, a priori, excellent candidates to be used as scripting languages.3 However, the relative complication in making executables (needing in some
systems to start the top-level, compile or consult the file, and make a saved state)
and the often large size of the resulting executables may deter CGI application
programmers. It appears convenient to provide a means for LP/CLP programs to
be executable as scripts, even if with reduced performance.
It is generally relatively easy to support scripts with the same functionality in
most LP/CLP systems. In Ciao, the program ciao-shell –which has also been
adapted to SICStus (Hermenegildo, 1996)– accomplishes this task, by first loading
the file given to it as the first argument (but skipping the first lines and avoiding
loading messages) and then starting execution at main/1 (the argument provides
the list of command line options). Then, for example, in a Unix system, the following
program can be run directly as a script without any need for compilation:
#!/usr/local/bin/ciao-shell
main(_) :write(’Content-type: text/html’), nl, nl,
write(’<HTML>’),
2
3
It often convenient to use options (such as ciaoc’s -s or -S) which will generate a standalone
executable which is independent of any libraries.
For example, the built-in grammars and databases greatly simplify many typical script-based
applications.
6
D. Cabeza and M. Hermenegildo
www.xxx.yyy
<form action="http://www.xxx.yyy/handler.cgi">
2
WWW
Browser
1
HTTP
Server
form data
6
3
form
reply
4
5
handler.cgi
Fig. 2. The Forms interface
write(’Hello world.’),
write(’</HTML>’).
Note that in some UNIX versions either the program ciao-shell must be included in the /etc/shells listing or the first line should be replaced by these
two:
#!/bin/sh
exec ciao-shell $0 "$@"
The execution of Prolog scripts may be optimized in some systems. For example,
in Ciao, the first time a script is run it is also compiled and its bytecode is saved
to a file. At subsequent times, if the script has not changed, the object code is
retrieved from that file, avoiding compilation or interpretation overhead.
4 Form Handling in HTTP
So far we have shown CGI executables which produce output, but this output is not
a function of input coming from the request, which is obviously of limited interest.
CGI executables become most useful when combined with HTML forms. HTML
forms are HTML documents (or parts of HTML documents) which include special
fields such as text areas, menus, radio buttons, etc. which allow providing input to
CGI executables. The steps involved in the handling of the input contained in a
form are illustrated in Figure 2. When a document containing a form is accessed
via a form-capable browser (Mosaic, Netscape, Lynx, etc.), the browser displays the
input fields, buttons, menus, etc. indicated in the document, and locally allows the
user to perform input by modifying such fields. However, this input is not ultimately
handled by the browser. Instead, it will be sent to a “handler” CGI program, which
can be anywhere on the net, and whose address must be given in the form itself
(1). Forms generally have a “submit” button such that, when pressed, the input
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
7
provided through the menus, text areas, etc. is sent by the browser to the HTTP
server corresponding to the handler (2). Two methods for sending this input exist:
“GET” and “POST”. In the meantime, the sending browser waits for a response
from that program, which should come in the form of a new HTML document.
The handler program is invoked in much the same way as a cgi-bin application (3),
except that the information from the form is supplied to the handler (in different
ways depending on the system, the method of invocation, and the content type)
(4). This information is encoded in a predefined format, which relates each piece of
information to the corresponding field in the form, by means of a keyword associated
with each field. The handler then identifies the information corresponding to each
field in the original form, processes it, and then responds by writing an HTML
document to its standard output (5), which is forwarded by the server to the waiting
browser when the handler terminates (6). An important point to be noted is that,
as with simple cgi-bin applications, the handler is started and should terminate for
each transaction. The reader is referred, for example, to (Grobe & Naseer, 1998)
for a more complete introduction to CGI scripts and HTML forms.
5 Writing Form Handlers with Pi LLoW
The only complication in writing form handlers compared to writing simple CGI
applications is the need to capture and parse the form data. As we said before, this
data can be provided in several ways, depending on the system and the method
used to invoke the form, and is encoded with escape sequences. It is relatively easy
to write a Prolog program to parse such input (using, for example, definite clause
grammars –DCGs). The PiLLoW library provides some predicates which do this
and simplify the whole task, hiding the low-level protocol behind. The principal
predicates provided include:
get_form_input(Dic ) Translates input from the form (with either the POST or
GET methods, and even with CONTENT_TYPE multipart/form-data) to a dictionary Dic of attribute=value pairs. It translates empty values (which indicate
only the presence of an attribute) to the atom ’$empty’, values with more than
one line (from text areas or files) to a list of lines as strings, the rest to atoms or
numbers (using name/2). This is implemented using DCG parsers.
get_form_value(Dic,Var,Val ) Gets value Val for attribute Var in dictionary
Dic. Does not fail: value is ’’ if not found (this simplifies merging form producers
and form handlers, see later).
form_empty_value(V ) Useful to check that a value V from a text area is empty
(filters spaces, newlines, linefeeds, etc.).
form_default(Val,Default,NewVal ) Useful when a form is only partially filled
(and also in the first invocation of a combined form handler/producer – see
Section 7). If the value of Val is empty then NewVal=Default, else NewVal=Val.
my_url(URL ) Returns in URL the Uniform Resource Locator (WWW address) of
this cgi executable.
form_request_method(Method ) Returns in Method the method of invocation of
the form handler (“GET” or “POST”).
8
D. Cabeza and M. Hermenegildo
For example, suppose we want to make a handler which implements a database
of telephone numbers and is queried by a form including a single entry field with
name person_name. The handler might be coded as follows:
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
write(’Content-type: text/html’), nl, nl,
write(’<HTML><TITLE>Telephone database</TITLE>’), nl,
write(’<IMG SRC="phone.gif">’),
write(’<H2>Telephone database</H2><HR>’),
write_info(Name),
write(’</HTML>’).
write_info(Name) :form_empty_value(Name) ->
write(’You have to provide a name.’)
; phone(Name, Phone) ->
write(’Telephone number of <B>’),
write(Name),
write(’</B>: ’),
write(Phone)
; write(’No telephone number available for <B>’),
write(Name),
write(’</B>.’).
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
The code above is quite simple. On the other hand, the interspersion throughout
the text of calls to write with HTML markup inside makes the code somewhat
inelegant. Also, there is no separation between computation and input/output, as
is normally desirable. It would be much preferable to have an encoding of HTML
code as Prolog terms, which could then be manipulated easily in a more elegant way,
and a predicate to translate such terms to HTML for output. This functionality,
provided by the PiLLoW library, is presented in the next section.
6 Handling HTML as Prolog Terms
Since LP/CLP systems perform symbolic processing using Herbrand terms, it seems
natural to be able to handle HTML code directly as terms. Then, such structures
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
9
only need to be translated by appropriate predicates to HTML code when they
need to be output. In general, this relationship between HTML code and Prolog
terms allows viewing any WWW page as a Herbrand term. The predicates which
provide this functionality in PiLLoW are:
output_html(F ) Accepts in F an HTML term (or a list of HTML terms) and
sends to the standard output the text which is the rendering of the term(s) in
HTML format.
html2terms(Chars,Terms ) (also, xml2terms/2) Relates a list of HTML
(resp. XML) terms and a list of ASCII characters which are the rendering of
the terms in HTML format. This predicate is reversible (but it normalizes in
the reverse direction –see later). output_html/2 uses this predicate to transform
HTML terms in characters. Again, this is implemented via DCG parsing.
In an HTML term certain atoms and structures represent special functionality
at the HTML level. An HTML term can be recursively a list of HTML terms. The
following are legal HTML terms:
hello
[hello, world]
["This is an ", em(’HTML’), " term"]
When converting HTML terms to characters, html2terms/2 translates special
structures into the corresponding format in HTML, applying itself recursively to
their arguments. Strings are always left unchanged. HTML terms may contain logic
variables, provided they are instantiated before the term is translated or output.
This allows creating documents piecemeal, back-patching of references in documents, etc.
In the following sections we list the meaning of the principal Prolog structures
that represent special functionality at the HTML level. Only special atoms are
translated, the rest are assumed to be normal text and will be passed through to
the HTML document.
6.1 General Structures
Basically, HTML has two kinds of components: HTML elements and HTML environments. An HTML element has the form “<NAME Attributes >” were NAME
is the name of the element and Attributes is a (possibly empty) sequence of attributes, each of them being either an attribute name or an attribute assignment
as name="Value".
An HTML environment has the form “<NAME Attributes > Text </NAME>”
were NAME is the name of the environment an Attributes has the same form as
before.
The general Prolog structures that represent these two HTML constructions are:
Name $Atts (‘$/2’ is defined as an infix, binary operator.) Represents an HTML
element of name Name and attributes Atts, were Atts is a (possibly empty) list
10
D. Cabeza and M. Hermenegildo
of attributes, each of them being either an atom or a structure name =value . For
example, the term
img$[src=’images/map.gif’,alt="A map",ismap]
is translated into the HTML source
<img src="images/map.gif" alt="A map" ismap>
Note that HTML is not case-sensitive, so we can use lower-case atoms.
name (Text ) (A term with functor name/1 and argument Text) Represents an
HTML environment of name name and included text Text. For example, the
term
address(’[email protected]’)
is translated into the HTML source
<address>[email protected]</address>
name (Atts,Text ) (This is a term with functor name/2 and arguments Atts and
Text) Represents an HTML environment of name name, attributes Atts and
included text Text. For example, the term
a([href=’http://www.clip.dia.fi.upm.es/’],"Clip home")
represents the HTML source
<a href="http://www.clip.dia.fi.upm.es/">Clip home</a>
env(Name,Atts,Text ) Equivalent to Name (Atts,Text ).
begin(Name,Atts ) It translates to the start of an HTML environment of name
Name and attributes Atts. There exists also a begin(Name ) structure. Useful,
in conjunction with the next structure, when including in a document output
generated by an existing piece of code (e.g. Name = pre). Its use is otherwise
discouraged.
end(Name ) Translates to the end of an HTML environment of name Name.
Now we can rewrite the previous example as follows (note how the use of the
logic variable Response allows injecting the result of the call to response/1 into
the output term, using unification):
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
’Content-type: text/html\n\n’,
html([title(’Telephone database’),
img$[src=’phone.gif’],
h2(’Telephone database’),
hr$[],
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 11
Response])]). %% Using the logic variable.
response(Name, Response) :form_empty_value(Name) ->
Response = ’You have to provide a name.’
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone]
; Response = [’No telephone number available for ’,b(Name),’.’].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
Any HTML construction can be represented with these structures (except comments and declarations, which could be included as atoms or strings), but the
PiLLoW library provides additional, specific structures to simplify HTML creation.
6.2 Specific Structures
In this section we will list some special structures for HTML which PiLLoW understands. While in many cases using the general structures (with the native HTML
names) is probably good practice, using specific structures such as these can sometimes be convenient. Also, some of these structures have special functionality (e.g.,
prolog_term/1). A predicate html_expansion/2 is provided which allows defining new structures (tables, layers, etc). Specific structures include (the reader is
referred to the PiLLoW manual for a full listing):
start Used at the beginning of a document (translates to <html>).
end Used at the end of a document (translates to </html>).
-- Produces a horizontal rule (translates to <hr>).
\\ Produces a line break (translates to <br>).
$ Produces a paragraph break (translates to <p>).
comment(Comment ) Used to insert an HTML comment (translates to
<!-- Comment -->).
declare(Decl ) Used to insert an HTML declaration – seldom used (translates to
<!Decl >).
image(Addr ) Used to include an image of address (URL) Addr (translates to an
<img> element).
image(Addr,Atts ) As above with the list of attributes Atts.
ref(Addr,Text ) Produces a hypertext link, Addr is the URL of the
referenced resource, Text is the text of the reference (translates to
<a href="Addr">Text</a>).
label(Label,Text ) Labels Text as a target destination with label Label (translates to <a name="Label">Text</a>).
heading(N,Text ) Produces a heading of level N (1 ≤ N ≤ 6), Text is the text to
be used as heading – useful when one wants a heading level relative to another
heading (translates to a <hN > environment).
12
D. Cabeza and M. Hermenegildo
itemize(Items ) Produces a list of bulleted items, Items is a list of corresponding
HTML terms (translates to a <ul> environment).
enumerate(Items ) Produces a list of numbered items, Items is a list of corresponding HTML terms (translates to an <ol> environment).
description(Defs ) Produces a list of defined items, Defs is a list whose elements
are definitions, each of them being a Prolog sequence (composed by ’,’/2 operators). The last element of the sequence is the definition, the other (if any) are
the defined terms (translates to an <dl> environment).
nice_itemize(Img,Items ) Produces a list of bulleted items, using the image Img
as bullet. The predicate icon_address/2 provides a colored bullet.
preformatted(Text ) Used to include preformatted text, Text is a list of HTML
terms, each element of the list being a line of the resulting document (translates
to a <pre> environment).
verbatim(Text ) Used to include text verbatim, special HTML characters
(<,>,&,") are translated into its quoted HTML equivalent.
prolog_term(Term ) Includes any prolog term Term, represented in functional
notation. Variables are output as _.
nl Used to include a newline in the HTML source (just to improve human readability).
entity(Name ) Includes the entity of name Name (ISO-8859-1 special character).
cgi_reply This is not HTML, rather, the CGI protocol requires this content
descriptor to be used by CGI executables (including form handlers) when replying
(translates to “Content-type: text/html”).
pr Includes in the page a graphical logo with the message “Developed using the
PiLLoW Web programming library”, which points to the manual and library
source.
With these additional structures, we can rewrite the previous example as follows
(note that in this example the use of heading/2 or h2/1 is equally suitable):
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
cgi_reply,
start,
title(’Telephone database’),
image(’phone.gif’),
heading(2,’Telephone database’),
--,
Response,
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 13
end]).
response(Name, Response) :form_empty_value(Name) ->
Response = ’You have to provide a name.’
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone]
; Response = [’No telephone number available for ’,b(Name),’.’].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
We have not included above the specific structures for creating forms. They are
included and explained in the following section.
6.3 Specific Structures for Forms
In this section we explain the structures which represent the various elements related
to forms:
start_form(Addr [,Atts ]) Specifies the beginning of a form. Addr is
the address (URL) of the program that will handle the form, and
Atts other attributes of the form, as the method used to invoke it.
If Atts is not present the method defaults to POST. (Translates to
<form action="Addr " Atts >.)
start_form Specifies the beginning of a form without assigning address to the
handler, so that the form handler will be the cgi-bin executable producing the
form.
end_form Specifies the end of a form (translates to </form>).
checkbox(Name,State ) Specifies an input of type checkbox with name Name,
State=on if the checkbox is initially checked (translates to an <input> element).
radio(Name,Value,Selected ) Specifies an input of type radio with name Name
(several radio buttons which are interlocked must share their name), Value is the
the value returned by the button, if Selected=Value the button is initially checked
(translates to an <input> element).
input(Type,Atts ) Specifies an input of type Type with a list of attributes Atts.
Possible values of Type are text, hidden, submit, reset, . . . (translates to an
<input> element).
textinput(Name,Atts,Text ) Specifies an input text area of name Name. Text
provides the default text to be shown in the area, Atts a list of attributes (translates to a <textarea> environment).
option(Name,Val,Options ) Specifies a simple option selector of name Name,
Options is the list of available options and Val is the initial selected option
(if Val is not in Options the first item is selected) (translates to a <select>
environment).
14
D. Cabeza and M. Hermenegildo
menu(Name,Atts,Items ) Specifies a menu of name Name, list of attributes Atts
and list of options Items. The elements of the list Items are marked with the
prefix operator ‘$’ to indicate that they are selected (translates to a <select>
environment).
For example, in order to generate a form suitable for sending input to the previously described phone database handler one could execute the following goal:
output_html([
start,
title(’Telephone database’),
heading(2,’Telephone database’),
$,
start_form(’http://www.clip.dia.fi.upm.es/cgi-bin/phone_db.pl’),
’Click here, enter name of clip member, and press Return:’,
\\,
input(text,[name=person_name,size=20]),
end_form,
end]).
Of course, one could have also simply written directly the resulting HTML document:
<html>
<title>Telephone database</title>
<h2>Telephone database</h2>
<p>
<form method="POST"
action="http://www.clip.dia.fi.upm.es/cgi-bin/phone_db.pl">
Click here, enter name of clip member, and press Return:
<br>
<input type="text" name="person_name" size="20">
</form>
</html>
7 Merging the Form Producer and the Handler
An interesting practice when producing HTML forms and handlers is to merge the
operation of the form producer and the handler into the same program. The idea is
to produce a generalized handler which receives the form input, parses it, computes
the answer, and produces a new document which contains the answer to the input,
as well as a new form. A special case must be made for the first invocation, in
which the input would be empty, and then only the form should be generated. The
following is an example which merges the producer and the handler for the phones
database:4
4
Notice that when only one text field exists in a form, the form can be submitted by simply
pressing “Return” inside the text field.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 15
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
cgi_reply,
start,
title(’Telephone database’),
image(’phone.gif’),
heading(2,’Telephone database’),
--,
Response,
start_form,
’Click here, enter name of clip member, and press Return:’,
\\,
input(text,[name=person_name,size=20]),
end_form,
end]).
response(Name, Response) :form_empty_value(Name) ->
Response = []
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone,$]
; Response = [’No telephone number available for ’,b(Name),’.’,$].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
This combination of the form producer and the handler allows producing applications that give the impression of being interactive, even if each step involves starting
and running the handler to completion. Note that forms can contain fields which
are not displayed and are passed as input to the next invocation of the handler.
This allows passing state from one invocation of the handler to the next one.
Finally, a note about testing and debugging CGI scripts: this is unfortunately
not as straightforward as it could be. Useful techniques include carefully checking
permissions, looking at the data logs of the server, replacing predicates such as
get_form_... with versions that print what is really being received, etc.
16
D. Cabeza and M. Hermenegildo
8 Templates
A problem in the previous programs is that the layout of the output page is not
easily configurable –it is hard-coded in the source and can only be changed by
modifying the program. This is something that a normal user (or even an expert
programmer if the size of the program is large) may not want to do. In order to
address this, PiLLoW provides a facility for reading in “HTML templates” (also
XML templates), and converting them into a term format in which it is very natural
to manipulate them. An HTML template is a file which contains standard HTML
code, but in which “slots” can be defined and given an identifier by means of a
special tag. These slots represent parts of the HTML code in which other HTML
code can be inserted. Once the HTML template is read by PiLLoW, such slots
appear as free logic variables in the corresponding PiLLoW terms. In this way, the
user can define a layout with an HTML editor of choice, taking care of marking the
“left out” parts with given names. These parts will then be filled appropriately by
the program. The functionality associated with parsing such terms is encapsulated
in the following predicate:
html_template(Chars, Terms, Dict ) Parses the string Chars as the contents
of an HTML template and unifies Terms with the list of HTML terms comprised
in the template, substituting occurrences of the special tag <V>name</V> with
prolog variables. Dict is instantiated to the dictionary of such substitutions, as a
list of name=Variable pairs.
In the following example a template file called TlfDB.html is assumed to hold the
formatting of the output page, defining an HTML variable called “response” which
will be substituted by the response of the CGI program. Note that the predicate
file_to_string/2 (defined in Ciao library file_utils) reads a file and returns
in its second argument the contents of the file as a list of character codes. Note also
that calling html_template/3 with the third argument instantiated to [response
= Response] has the effect of instantiating the “slot” in HTML_terms to the contents
of Response (this makes use of the fact that there is only one slot on the template;
normally, a call to member/2 is used to locate the appropriate name=Variable pair).
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
:- use_module(library(file_utils)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
file_to_string(’TlfDB.html’, Contents),
html_template(Contents, HTML_terms, [response = Response]),
output_html([cgi_reply|HTML_terms]).
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 17
response(Name, Response) :form_empty_value(Name) ->
Response = []
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone,$]
; Response = [’No telephone number available for ’,b(Name),’.’,$].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
An example of the contents of the template file could be:
<HTML><HEAD><TITLE>Telephone database</TITLE></HEAD>
<BODY background="bg.gif">
<IMG src="phone.gif">
<H2>Telephone database</H2>
<HR>
<V>response</V>
<FORM method="POST">
Click here, enter name of clip member, and press Return:<BR>
<INPUT type="text" name="person_name" size="20"></FORM>
</BODY>
</HTML>
9 Accessing WWW documents
The facilities presented in the previous sections allow generating HTML documents,
including forms, and handling the input coming from forms. In many applications
such as search tools, content analyzers, etc., it is also desirable to be able to access
documents on the Internet. Such access is generally accomplished through protocols
such as FTP and HTTP which are built on top of TCP/IP. In LP/CLP systems which
have TCP/IP connectivity (i.e., a sockets/ports interface) the required protocols
can be easily coded in the source language using such facilities and DCG parsers.
At present, only the HTTP protocol is supported by PiLLoW. As with HTML code,
the library uses an internal representation of Uniform Resource Locators (URLs), to
be able to manipulate them easily, and provides predicates which translate between
the internal representation and the textual form. The facilities provided by PiLLoW
for accessing WWW documents include the following predicates:
url_info(URL,Info ) Translates a URL URL to an internal structure Info which
details its various components and vice-versa. For now non-HTTP URLs make
the predicate fail. E.g.
url_info(’http://www.foo.com/bar/scooby.txt’,Info)
gives Info = http(’www.foo.com’,80,"/bar/scooby.txt"),
url_info(URL, http(’www.foo.com’,2000,"/bar/scooby.txt")
gives URL = "http://www.foo.com:2000/bar/scooby.txt" (a string).
18
D. Cabeza and M. Hermenegildo
url_info_relative(URL,BaseInfo,Info ) Translates a relative URL URL which
appears in the HTML page referred to by BaseInfo (given as an url_info structure) to a complete url_info structure Info. Absolute URLs are translated as
with the previous predicate. E.g.
url_info_relative("/guu/intro.html",
http(’www.foo.com’,80,"/bar/scoob.html"), Info)
gives Info = http(’www.foo.com’,80,"/guu/intro.html")
url_info_relative("dadu.html",
http(’www.foo.com’,80,"/bar/scoob.html"), Info)
gives Info = http(’www.foo.com’,80,"/bar/dadu.html").
url_query(Dic,Args ) Translates a list of attribute=value pairs Dic (in the same
form as the dictionary returned by get_form_input/1) to a string Args for appending to a URL pointing to a form handler.
fetch_url(URL,Request,Response ) Fetches a document from the Internet. URL
is the Uniform Resource Locator of the document, given as a url_info structure.
Request is a list of options which specify the parameters of the request, Response
is a list which includes the parameters of the response. The request parameters
available include:
head To specify that we are only interested in the header.
timeout(Time ) Time specifies the maximum period of time (in seconds) to wait
for a response. The predicate fails on timeout.
if_modified_since(Date ) Get
document
only if newer than Date. An example of a structure that represents a date
is date(’Tuesday’,15,’January’,1985,’06:14:02’).
user_agent(Name ) Provide a user-agent field.
authorization(Scheme,Params ) Provides an authentication field when accessing restricted sites.
name (Param ) Any other functor translates to a field of the same name (e.g.
from(’user@machine’)).
The parameters which can be returned in the response list include (see the
HTTP/1.0 definition for more information):
content(Content ) Returns in Content the actual document text, as a list of
characters.
status(Type,Code,Phrase ) Gives the status of the response. Type can be any
of informational, success, redirection, request_error, server_error or
extension_code, Code is the status code and Phrase is a textual explanation
of the status.
pragma(Data ) Miscellaneous data.
message_date(Date ) The time at which the message was sent.
location(URL ) The document has moved to this URL.
http_server(Server ) Identifies the server responding.
allow(methods ) List of methods allowed by the server.
last_modified(Date ) Date/time at which the sender believes the resource was
last modified.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 19
expires(Date ) Date/time after which the entity should be considered stale.
content_type(Type,Subtype,Params ) Returns the MIME type/subtype of the
document.
content_encoding(Type ) Encoding of the document (if any).
content_length(Length ) Length is the size of the document, in bytes.
authenticate(Challenges ) Request for authentication.
html2terms(Chars,Terms ) We have already explained how this predicate transforms HTML terms to HTML format. Used the other way around it can parse
HTML code, for example retrieved by fetch_url. The resulting list of HTML
terms Terms is normalized: it contains only comment/1, declare/1, env/3 and
$/2 structures.
For example, a simple fetch of a document can be done as follows:
url_info(’http://www.foo.com’,UI), fetch_url(UI,[],R),
member(content(C),R), html2terms(C, HTML_Terms).
Note that if an error occurs (the document does not exist or has moved, for example)
this will simply fail. The following call retrieves a document if it has been modified
since October 6, 1999:
fetch_url(http(’www.foo.com’,80,"/doc.html"),
[if_modified_since(’Wednesday’,6,’October’,1999,’00:00:00’)],
R).
This last one retrieves the header of a document (with a timeout of 10 seconds) to
get its last modified date:
fetch_url(http(’www.foo.com’,80,"/last_news.html"),
[head,timeout(10)],R),
member(last_modified(Date),R).
The following is a simple application illustrating the use of fetch_url and
html2terms. The example defines check_links(URL,BadLinks ). The predicate
fetches the HTML document pointed to by URL and scours it to check for links
which produce errors when followed. The list BadLinks contains all the bad links
found, stored as compound terms of the form: badlink(Link,Error) where Link
is the problematic link and Error is the error explanation given by the server.
check_links(URL,BadLinks) :url_info(URL,URLInfo),
fetch_url(URLInfo,[],Response),
member(content_type(text,html,_),Response),
member(content(Content),Response),
html2terms(Content,Terms),
check_source_links(Terms,URLInfo,[],BadLinks).
check_source_links([],_,BL,BL).
check_source_links([E|Es],BaseURL,BL0,BL) :-
20
D. Cabeza and M. Hermenegildo
check_source_links1(E,BaseURL,BL0,BL1),
check_source_links(Es,BaseURL,BL1,BL).
check_source_links1(env(a,AnchorAtts,_),BaseURL,BL0,BL) :member((href=URL),AnchorAtts), !,
check_link(URL,BaseURL,BL0,BL).
check_source_links1(env(_Name,_Atts,Env_html),BaseURL,BL0,BL) :- !,
check_source_links(Env_html,BaseURL,BL0,BL).
check_source_links1(_,_,BL,BL).
check_link(URL,BaseURL,BL0,BL) :url_info_relative(URL,BaseURL,URLInfo), !,
fetch_url_status(URLInfo,Status,Phrase),
( Status \== success ->
name(P,Phrase),
name(U,URL),
BL = [badlink(U,P)|BL0]
; BL = BL0
).
check_link(_,_,BL,BL).
fetch_url_status(URL,Status,Phrase) :fetch_url(URL,[head,timeout(20)],Response), !,
member(status(Status,_,Phrase),Response).
fetch_url_status(_,timeout,"Timeout").
10 Providing Code Through the WWW
A facility which can be easily built on top of the primitives presented so far is that
of “remote WWW modules,” i.e., program modules which reside on the net at a
particular HTTP address in the same way that normal program modules reside
in a particular location in the local file system. This allows for example always
fetching the most recent version of a given library (e.g., PiLLoW) when a program
is compiled. For example, the form handler of Section 6.1, if rewritten as
#!/usr/local/bin/ciao-shell
:- use_module(’http://www.clip.dia.fi.upm.es/lib/pillow.pl’).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
...
would load the current version of the library each time it is executed. This generalized module declaration is just syntactic sugar, using expand_term, for a document
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 21
www.xxx.yyy
<form action="http://www.xxx.yyy/am_inter.cgi">
HTTP
Server
2
1
WWW
Browser
form data
8
3
form
reply
4
7
5
Active
Module
am_inter
6
predicate(Arg1,Arg2,...,ArgN)
Fig. 3. The Forms Interface using Active Modules
fetch, using fetch_url, followed by a standard use_module declaration. It is obviously interesting to combine this facility with caching strategies. An interesting
(and straightforward to implement) additional feature is to fetch remote byte-code
(as generally done by use_module), if available, but this is only possible if the two
systems use the same byte-code (this can normally be checked easily in the bytecode itself). Also, it may be interesting to combine this type of code downloading
with WWW document accesses, so that code is downloaded automatically when a
particular document is fetched. This issue is addressed in Section 12. Finally, there
are obvious security issues related to downloading code in general, which can be
addressed with standard techniques such as security signatures.
11 A High-Level Model of Client-Server Interaction: Active Modules
Despite its power, the cgi-bin interface also has some shortcomings. The most serious is perhaps the fact that the handler is started and expected to terminate for
each interaction. This has two disadvantages. First, no state is preserved from one
query to the next. However, as mentioned before, this can be fixed by passing the
state through the form (using hidden fields), by saving it in a temporary file at the
server side, by using “cookies”, etc. Second, and more importantly, starting and
stopping the application may be inefficient. For example, if the idea is to query
a large database or a natural language understanding system, it may take a long
time to start and stop the system. In order to avoid this we propose an alternative
architecture for cgi-bin applications (a similar idea, although not based on the idea
of active modules, has been proposed independently by Ken Bowen (Bowen, 1996)).
The basic idea is illustrated in Figure 3. The operation is identical to that of
standard form handlers, as illustrated in Figure 2, up to step 3. In this step, the
handler started is not the application itself, but rather an interface to the actual
application, which is running continuously and thus contains state. Thus, only the
22
D. Cabeza and M. Hermenegildo
interface is started and stopped with every transaction. The interface simply passes
the form input received from the server (4) to the running application (5) and then
forwards the output from the application (6) to the server before terminating, while
the application itself continues running. Both the interface and the application can
be written in LP/CLP, using the predicates presented. The interface can be a simple
script, while the application itself will be typically compiled.
An interesting issue is that of communication between interface and application.
This can of course be done through sockets. However, as a cleaner and much simpler alternative, the concept of active modules (Cabeza & Hermenegildo, 1995) can
be used to advantage in this application. An active module (or an active object,
if modularity is implemented via objects) is an ordinary module to which computational resources are attached (for example, a process on a UNIX machine), and
which resides at a given (socket) address on the network.5 Compiling an active
module produces an executable which, when running, acts as a server for a number of relations, which are the predicates exported by the module. The relations
exported by the active module can be accessed by any program on the network
by simply “loading” the module and thus importing such “remote relations.” The
idea is that the process of loading an active module does not involve transferring
any code, but rather setting up things so that calls in the local module are executed as remote procedure calls to the active module, possibly over the network.
Except for compiling it in a special way, an active module is identical from the
programmer point of view to an ordinary module. Also, a program using an active
module imports it and uses it in the same way as any other module, except that it
uses “use_active_module” rather than “use_module” (see below). Also, an active
module has an address (network address) which must be known in order to use it.
The address can be announced by the active module when it is started via a file or
a name server (which would be itself another active module with a fixed address).
We now present the constructs related to active modules in Ciao:
:- use_active_module(Module,Predicates ) A declaration used to import the
predicates in the list Predicates from the active module Module . From this
point on, the code should be written as if a standard use_module/2 declaration
had been used. The declaration needs the following predicate to be accessible
from the module.
module_address(Module,Address ) This predicate must return in Address the
address of Module , for any active module imported in the code. There are a
number of standard libraries defining versions of this predicate.
save_addr_actmod(Address ) This predicate should define a way to publish
Address , to be used in active modules (the name of the active module is taken
as the name of the current executable). There are a number of standard libraries
defining versions of this predicate, which are in correspondence with the libraries
which define versions of the previous predicate.
5
It is also possible to provide active modules via a WWW address. However, we find it more
straightforward to simply use socket addresses. In any case, this is generally hidden inside the
access method and can be thus made transparent to the user.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 23
make_actmod(ModuleFile,PublishModule ) Makes an active module executable
from the module residing in ModuleFile , using address publish module of
name PublishModule . When the executable is run (for example, at the operating system level by “Module &”), a socket is created and the hook predicate save_addr_actmod/1 mentioned above (which is supposed to be defined in
PublishModule ) is called in order to export the active module address as required. Then, a standard driver is run to attend network requests for the module
exported predicates. Note that the code of ModuleFile does not need to be
written in any special way.
This scheme is very flexible, allowing to completely configure the way active
modules are located. This is accomplished by writing a pair of libraries, one defining the way an active module address is published, and a second defining the way
the address of a given active module is found. For example, the Ciao standard libraries include as an example an implementation (libraries filebased_publish and
filebased_locate) which uses a directory accessible by all the involved machines
(via NFS) to store the addresses of the active modules, and the module_address/2
predicate examines this directory to find the required data. Other solutions provided
as examples include posting the address at a WWW address (webbased_publish
/ webbased_locate), and an implementation of a name server, that is, another
active module (this one with a known, fixed address) that records the addresses of
active modules and supplies this data to the modules that import it, serving as a
contact agency between servers and clients.
From the implementation point of view, active modules are essentially daemons:
Prolog executables which are started as independent processes at the operating
system level. In the Ciao system library, communication with active modules is
implemented using sockets (thus, the address of an active module is a UNIX socket
in a machine). Requests to execute goals in the module are sent through the socket
by remote programs. When such a request arrives, the process running the active
module takes it and executes it, returning through the socket the computed results.
These results are then taken by the remote processes.
Thus, when the compiler finds a use_active_module declaration, it defines the
imported predicates as remote calls to the active module. For example, if the predicate P is imported from the active module M , the predicate would be defined as
P :- module_address(M,A), remote_call(A,P )
Compiling the following code as an active module, by writing at the Ciao
toplevel “make_actmod(phone_db, ’actmods/filebased_publish’)” (or, using
the standalone compiler, by executing “ciaoc -a ’actmods/filebased_publish’
phone_db”), creates an executable phone_db which, when started as a process (for
example, by typing “phone_db &” at a UNIX shell prompt) saves its address (i.e.,
that of its socket) in file phone_db.addr and waits for queries from any module
which “imports” this module (it also provides a predicate to dynamically add information to the database):
:- module(phone_db,[response/2,add_phone/2]).
24
D. Cabeza and M. Hermenegildo
response(Name, Response) :form_empty_value(Name) ->
Response = ’You have to provide a name.’
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone]
; Response = [’No telephone number available for ’,b(Name),’.’].
add_phone(Name, Phone) :assert(phone(Name, Phone)).
:- dynamic phone/2.
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
The following simple script can be used as a cgi-bin executable which will be the
active module interface for the previous active module. When started, it will process
the form input, issue a call to response/2 (which will be automatically handled
by the phone_db active module), and produce a new form before terminating. It
will locate the address of the phone_db active module via the module_address/2
predicate defined in library ’actmods/filebased_locate’.
#!/usr/local/bin/ciao-shell
:- use_active_module(phone_db,[response/2]).
:- use_module(library(’actmods/filebased_locate’)).
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
cgi_reply,
start,
title(’Telephone database’),
image(’phone.gif’),
heading(2,’Telephone database’),
--,
Response,
$,
start_form,
’Click here, enter name of clip member, and press Return:’,
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 25
\\,
input(text,[name=person_name,size=20]),
end_form,
end]).
There are many enhancements to this simple schema which, for brevity, are only
sketched here. One is to add concurrency to the active module (or whatever means of
handling the client-server interaction is being used), in order to handle queries from
different clients concurrently. This is easy to do in systems that support concurrency natively, such as Ciao, BinProlog/µ2-Prolog, AKL, Oz, and KL1. We feel that
Ciao can offer advantages in this area because it offers compatibility with Prolog
and CLP systems while at the same time efficiently supporting concurrent execution of clause goals via local or distributed threads (Carro & Hermenegildo, 1999).
Such goals can communicate at different levels of abstraction: sockets/ports, the
shared fact database (similarly to a blackboard), or shared variables. BinProlog/µ2Prolog also supports threads, with somewhat different communication mechanisms (Tarau, 1996; De Bosschere, 1989). Finally, as shown in (Szeredi et al., 1996),
it is also possible to exploit the concurrency present in or-parallel Prolog systems
such as Aurora for implementing a multitasking server.
It is also interesting to set up things so that a single active module can handle
different forms. This can be done even dynamically (i.e., the capabilities of the active
module are augmented on the fly, being able to handle a new form), by designating
a directory in which code to be loaded by the active module would be put, the
active module consulting the directory periodically to increase its functionalities.
Finally, another important issue that has not been addressed is that of providing
security, i.e., ensuring that only allowed clients connect to the active module. As
in the case of remote code downloading, standard forms of authentication based on
codes can be used.
12 Automatic Code Downloading and Local Execution
In this section we describe an architecture which, using only the facilities we have
presented in previous sections, allows the downloading and local execution of Prolog
(or other LP/CLP) code by accessing a WWW address, without requiring a special
browser. This is a complementary approach to giving WWW access to an active
module in the sense that it provides code which will be executed in the client
machine (à la Java). More concretely, the functionality that we desire is that by
simply clicking on a WWW pointer, and transparently for the user, remote Prolog
code is automatically downloaded in such a way that it can be queried via forms
and all the processing is done locally.
To allow this, the HTTP server on the server machine is configured to give a
specific mime.type (for example application/x-prolog) to the files which will
hold WWW-downloadable Prolog code (for example those with a special suffix, like
.wpl). On the other side, the browser is configured to start the wpl_handler helper
application when receiving data of type application/x-prolog. This wpl_handler
26
D. Cabeza and M. Hermenegildo
<form action="http://localhost/wpl_questioner.cgi">
www.xxx.yyy
WWW
Browser
http://www.xxx.yyy/app.wpl
form data
/tmp/x1349
form reply
1
app.wpl
2
11
wpl_handler /tmp/x1349
7
8
wpl_handler
loaded
wpl_questioner
3
?
6
4
10
9
5
FormReply
Active
Module
code
FormData
answerform(FormData, FormReply)
loadcode(’/tmp/x1349’)
Fig. 4. Automatic code downloading architecture
application is the interface to a Prolog engine which will execute the WWW downloaded code, acting as an active module. We now sketch the procedure (see figure
4):
1. The form that will be used to query the downloaded code (and which we
assume already loaded on the browser) contains a link which points to a
WWW-aware Prolog code file. Clicking on this link produces the download
as explained below. Note that for browsers that can handle multipart/mixed
mime types (such as most modern browsers), the form and the code file could
alternatively be combined in the same document. However, for brevity, we
will only describe the case when they are separate. The handler for the form
is specified as the local cgi-bin executable wpl_questioner.cgi.
2. As the server of the file tells the browser that this page is of type
application/x-prolog, the browser starts a wpl_handler and passes the
file to it (in this example by saving the file in a temporal directory and passing its name).
3. The wpl_handler process checks whether a Prolog engine is currently running
for this browser and, if necessary, starts one. This Prolog engine is configured
as an active module.
4. Then, through a call to a predicate of the active module “loadcode(File )”
the handler asks the active module to read the code.
5. The active module reads the code and compiles it.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 27
6. wpl_handler waits for the active module to complete the compilation and
writes a “done” message to the browser.
7. The browser receives the “done” message.
8. Now, when the “submit” button in the form is pressed, and following the
standard procedure for forms, the browser starts a wpl_questioner process,
sending it the form data.
9. The wpl_questioner process gets this form data, translates it to a dictionary
FormData and passes it to the active module through a call to its exported
predicate answerform(FormData,FormReply ).
10. The active module processes this request, and returns in FormReply a WWW
page (as a term) which contains the answer to it (and possibly a new form).
11. The wpl_questioner process translates FormReply to raw HTML and gives
it back to the browser, dying afterwards. Subsequent queries to the active
module can be accomplished either by going back to the previous page (using
the “back” button present in many browsers) or, if the answer page contains
a new query form, by using it. In any case, the procedure continues at 8.
The net effect of the approach is that by simply clicking on a WWW pointer,
remote Prolog code is automatically downloaded to a local Prolog engine. Queries
posed via the form are answered locally by the Prolog engine.
There are obvious security issues that need to be taken care of in this architecture. Again, standard authentication techniques can be used. However, since source
code is being passed around, it is comparatively easy to verify that no dangerous
predicates (for example, perhaps those that can access files) are executed. Note
again that it is also possible to download bytecode, since this is supported by most
current LP/CLP systems, using a similar approach.
13 Related Work
Previous general purpose work on WWW programming using computational logic
systems includes, to the best of our knowledge, the publicly available html.pl
library (Cabeza & Hermenegildo, 1996a) and manual, and the LogicWeb system
(Loke & Davison, 1996) (the PiLLoW library was also described previously in
(Cabeza et al., 1996)). The html.pl library was built by D. Cabeza and M.
Hermenegildo, using input from L. Naish’s forms code for NU-Prolog and M.
Hermenegildo and F. Bueno’s experiments building a WWW interface to the CHAT80 (Warren & Pereira, 1982) program. It was released as a publicly available WWW
library for LP/CLP systems and announced, among other places, in the Internet
comp.lang.prolog newsgroup (Cabeza & Hermenegildo, 1996b). The library has
since been ported to a large number of systems and adapted by several Prolog vendors, as well as used by different programmers in various institutions. In particular,
Ken Bowen has ported the library to ALS Prolog and extended it to provide group
processing of forms and an alternative to our use of active modules (Bowen, 1996).
The present work is essentially a significant extension of the html.pl library.
The main other previous body of work related to general-purpose interfacing of
28
D. Cabeza and M. Hermenegildo
logic programming and the WWW that we have knowledge of is the LogicWeb
(Loke & Davison, 1996) system, by S.W. Loke and A. Davison. The aim of LogicWeb is to use logic programming to extend the concept of WWW pages, incorporating in them programmable behavior and state. In this, it shares goals with Java.
It also offers rich primitives for accessing code in remote pages and module structuring. The aims of LogicWeb are different from those of html.pl/PiLLoW. LogicWeb
is presented as a system itself, and its implementation is done through a tight integration with the Mosaic browser, making use of special features of this browser.
In contrast, html.pl/PiLLoW is a general purpose library, meant to be used by a
general computational logic systems and is browser-independent. html.pl/PiLLoW
offers a wide range of functionalities, such as syntax conversion between HTML and
logic terms, access predicates for WWW pages, predicates for handling forms, etc.,
which are generally at a somewhat lower level of abstraction than those of LogicWeb. We believe that using PiLLoW and the ideas sketched in this paper it is
possible to add the quite interesting functionality offered by LogicWeb to standard
LP and CLP systems. We have shown some examples including access to passive
remote code (modules with an ftp or http address) from programs and automatic
remote code access and querying using standard browsers and forms. In addition,
we have discussed active remote code, where the functionality, rather than the code
itself, is exported.
More recently, a larger body of work on the topic was presented at the workshop held on the topic of Logic Programming and the Internet at the 1996
Joint International Conference and Symposium on Logic Programming (where
also a previous version of this paper was presented). The work presented in
(S.W. Loke & Sterling, 1996) is based on LogicWeb, and aims to provide distributed lightweight databases on the WWW. As with the basic LogicWeb system,
we believe that the PiLLoW library can be used to implement in other systems
the interesting ideas proposed therein. As briefly mentioned before, the work in
(Szeredi et al., 1996) proposes an architecture similar to that of our active modules in order to handle form requests. In this solution the handling multiple requests is performed by using or-parallelism. While we feel that and-parallelism (as
in &-Prolog’s or Ciao’s threads) is more natural for modeling this kind of concurrency, the ideas proposed are quite interesting. The ECLiPSe HTTP-library
(Ph. Bonnet & Thomsen, 1996), aimed at implementing INTERNET agents, offers
functionality that is in part similar to that of the Ciao html.pl/PiLLoW libraries,
including facilities that are similar to our active modules. The approach is different,
however, in several respects. The ECLiPSe library implements special HTTP servers
and clients. In contrast, PiLLoW uses standard HTTP servers and interfaces. Using special purpose servers may be interesting because the approach possibly allows
greater functionality. On the other hand this approach in general requires either
the substitution of the standard server on a given machine or setting the special
server at a different socket address from the standard one. The ECLiPSe library
also contains functionality that is related to our active modules, although the interface provided is at a lower level. Finally, other papers describing very interesting
WWW applications are being presented regularly, which underline the suitability of
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 29
computational logic systems for the task. We believe that the Ciao PiLLoW library
can contribute to making it even easier to develop such applications in the future.
Additional work on the topic of Logic Programming and the Internet can
be found in the proceedings of the workshop sponsored by the CompulogNet research network. The reader is referred to the tutorials and papers presented in these two workshops for more information on a number of applications, other libraries, and topics such as interfacing and compilation from
computational logic systems to Java. Examples of Prolog systems interfaced
with Java are BinProlog (see http://clement.info.umoncton.ca/BinProlog),
Ciao (Bueno et al., 1997), and others (Calejo, 1999). Experimental Prolog to
Java compilers have been built both in academia (see for example jProlog at
http://www.cs.kuleuven.ac.be/~bmd/PrologInJava/) and Commercially (see
for example the IF Prolog tools http://www.ifcomputer.com). This approach is
quite attractive, although the results cannot compete in performance with conventional Prolog compilers (it is open for research whether improvements in Java
performance or improved Prolog-to-Java compilation technology can bridge the
gap). Other commercial work on the topic of interfacing Prolog and the WWW (in
addition to that done on the ALS system mentioned above) include the Amzi! Prolog WebLS System (http://www.amzi.com/share.htm ) and the LPA PrologWeb
System (http://www.lpa.co.uk ).
Recent work using PiLLoW includes the “Web Integrator”(Davulcu et al., 1999)
–a webbase system that integrates data from various Web sources, and allows
users to query these Web sources as if they were a single database– and WebDB
(Cabeza & Hermenegildo, 1998) –a WWW-based database management interface.
Also, within the RadioWeb project (Partners, 1997), we have developed (in collaboration with the group of M. Codish at Ben Gurion University) a constraint-based
language for describing WWW page layout and style rules and an engine which,
by interpreting these rules, can generate WWW sites which dynamically adapt to
parameters such as user characteristics (Cederberg & CLIP Group, 1999).
Additional applications developed with the PiLLoW library can be accessed
from the PiLLoW WWW site (see later). A page with pointers to the proceedings of the previously mentioned workshops, as well as other information (including technical reports and tutorial) regarding the topic of Logic
Programming, Constraint Programming, and the Internet is maintained at
http://www.clip.dia.fi.upm.es/lpnet/.
14 Conclusions and Future Work
We have discussed from a practical point of view a number of issues involved in
writing Internet and WWW applications using LP/CLP systems. In doing so, we
have described PiLLoW, an Internet/WWW programming library for LP/CLP
systems. PiLLoW provides facilities for generating HTML/XML structured documents, producing HTML forms, writing form handlers, processing HTML/XML
templates, accessing and parsing WWW documents, and accessing code posted
at HTTP addresses. We have also described the architecture of some application
30
D. Cabeza and M. Hermenegildo
classes, including automatic code downloading, using a high-level model of clientserver interaction, active modules. Finally we have also described an architecture for
automatic LP/CLP code downloading for local execution, using generic browsers.
We believe that the Ciao PiLLoW library can ease substantially the process of
developing WWW applications using computational logic systems.
We have recently developed several extensions to the library (for example, for
setting and getting “cookies”), and sample applications which make extensive use
of concurrency (on those LP/CLP systems that support it) to overlap network
requests. We have also developed a complementary library for interfacing Prolog
with the Virtual Reality Modeling Language VRML (Smedbäck et al., 1999).
In addition to being included as part of the Ciao system, the PiLLoW library is
provided as a standard, standalone public domain library for SICStus Prolog and
other Prolog and CLP systems, supporting most of its functionality. Please contact
the authors or consult our WWW site http://www.clip.dia.fi.upm.es and the
PiLLoW
page
at http://www.clip.dia.fi.upm.es/Software/pillow/pillow.html for download details and an up-to-date online version of the PiLLoW manual. The Ciao
Prolog system is also freely available from http://www.clip.dia.fi.upm.es and
http://www.ciaoprolog.org.
Acknowledgments
The authors would like to thank Lee Naish, Mats Carlsson, Tony Beaumont, Ken
Bowen, Michael Codish, Markus Fromherz, Paul Tarau, Andrew Davison, and Koen
De Bosschere for useful feedback on previous versions of this document and the
PiLLoW code. The first versions of the Ciao system and the html.pl library were
developed under partial support from the ACCLAIM ESPRIT project. Subsequent
development has occurred in the context of MCYT projects “ELLA” and “EDIPIA”
(MCYT TIC99-1151), ESPRIT project RADIOWEB, and NSF/CICYT collaboration “ECCOSIC” (Fulbright 98059).
References
Berners-Lee, T., Cailliau, R., Luotonen, A., Nielsen, H.F., & Secret, A. (1994). The
World-Wide Web. Communications of the ACM, 37(8), 76–82.
Bowen, K. 1996 (March).
Personal communication.
Available from
http://www.als.com/als/html_pl.html.
Bueno, F., Cabeza, D., Carro, M., Hermenegildo, M., López-Garcı́a, P., & Puebla, G. 1997
(August). The Ciao Prolog System. Reference Manual. The Ciao System Documentation
Series–TR CLIP3/97.1. School of Computer Science, Technical University of Madrid
(UPM).
Cabeza, D., & Hermenegildo, M. (1995). Distributed Concurrent Constraint Execution
in the CIAO System. Proc. of the 1995 compulog-net workshop on parallelism and
implementation technologies. Utrecht, NL: U. Utrecht / T.U. Madrid. Available from
http://www.clip.dia.fi.upm.es/.
Cabeza, D., & Hermenegildo, M. 1996a (March). html.pl: An HTML package for (C)LP
systems. Spain. Available from http://www.clip.dia.fi.upm.es/miscdocs/.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 31
Cabeza, D., & Hermenegildo, M. 1996b (February).
LP/CLP HTML and
WWW interface publicly available. Posting in comp.lang.prolog. Available from
http://www.clip.dia.fi.upm.es/.
Cabeza, D., & Hermenegildo, M. 1997 (April). WWW Programming using Computational
Logic Systems (and the PiLLoW/Ciao Library). Proceedings of the workshop on logic
programming and the WWW at WWW6.
Cabeza, D., & Hermenegildo, M. 1998 (June). A WWW Database Management Interface for Prolog. Technical Report CLIP2/98.0. School of Computer Science, Technical
University of Madrid (UPM), Facultad Informática UPM, 28660-Boadilla del Monte,
Madrid-Spain.
Cabeza, D., Hermenegildo, M., & Varma, S. 1996 (September). The PiLLoW/Ciao Library
for INTERNET/WWW Programming using Computational Logic Systems. Proceedings
of the 1st workshop on logic programming tools for internet applications. Available from
http://clement.info.umoncton.ca/~lpnet.
Calejo, M. (1999). Java+prolog: A land of opportunities. Pages 1–2 of: Proceedings of the
first international conference on the practical application of constraint technologies and
logic programming. The Practical Application Company Ltd. Also available at
http://dev.servisoft.pt/interprolog/paclp99/default.htm.
Carlsson, M. 1988 (February). Sicstus Prolog User’s Manual. Po Box 1263, S-16313
Spanga, Sweden.
Carro, M., & Hermenegildo, M. (1999). Concurrency in Prolog Using Threads and a Shared
Database. Pages 320–334 of: 1999 international conference on logic programming. MIT
Press, Cambridge, MA, U.S.A.
Cederberg, Per, & CLIP Group, The. 1999 (June). Flexible Layout and Styling – The LaSt
Language. Technical Report D2.2.M3 CLIP 3/99.0. RADIOWEB Project.
Chikayama, T., Fujise, T., & Sekita, D. (1994). A Portable and Efficient Implementation
of KL1. Tick, Evan (ed), Proc. of the 1994 icot/nsf workshop on parallel and concurrent
programming. U. of Oregon.
Colmerauer, A. (1975). Les gramaire de metamorphose. Tech. rept. Univ. D’aix-Marseille,
Groupe De Ia.
Colmerauer, A. (1990). An Introduction to Prolog III. Communications of the ACM,
28(4), 412–418.
Davulcu, Hasan, Freire, Juliana, Kifer, Michael, & Ramakrishnan, I.V.
1999 (June).
A layered architecture for querying dynamic web content.
ACM SIGMOD international conference on management of data.
URL:
http://www.acm.org/sigmod/sigmod99/eproceedings/.
De Bosschere, K. (1989). Multi–Prolog, Another Approach for Parallelizing Prolog. Pages
443–448 of: Proceedings of parallel computing. Elsevier, North Holland.
ECRC. (1993). Eclipse user’s guide. European Computer Research Center.
Grobe, M., & Naseer, Hasan. 1998 (July).
An instantaneous introduction
to
CGI
scripts
and
HTML
forms.
Available
from
http://www.cc.ukans.edu/~acs/docs/other/forms-intro.shtml.
Hermenegildo, M. 1996 (April). Writing “Shell Scripts” in SICStus Prolog. Posting in
comp.lang.prolog. Available from http://www.clip.dia.fi.upm.es/.
Hermenegildo, M., & CLIP Group, The. (1994). Some Methodological Issues in the Design of CIAO - A Generic, Parallel, Concurrent Constraint System. Pages 123–133 of:
Principles and practice of constraint programming. LNCS, no. 874. Springer-Verlag.
Hermenegildo, M., Bueno, F., Garcı́a de la Banda, M., & Puebla, G. 1995a (December).
The CIAO Multi-Dialect Compiler and System: An Experimentation Workbench for
32
D. Cabeza and M. Hermenegildo
Future (C)LP Systems. Proceedings of the ilps’95 workshop on visions for the future of
logic programming. Available from http://www.clip.dia.fi.upm.es/.
Hermenegildo, M., Cabeza, D., & Carro, M. (1995b). Using Attributed Variables in the
Implementation of Concurrent and Parallel Logic Programming Systems. Pages 631–
645 of: Proc. of the twelfth international conference on logic programming. MIT Press.
Hermenegildo, M., Bueno, F., Cabeza, D., Carro, M., M. Garcı́a de la Banda, LópezGarcı́a, P., & Puebla, G. (1999a). The CIAO Multi-Dialect Compiler and System: An
Experimentation Workbench for Future (C)LP Systems. Pages 65–85 of: Parallelism
and Implementation of Logic and Constraint Logic Programming. Commack, NY, USA:
Nova Science.
Hermenegildo, M., Puebla, G., & Bueno, F. (1999b). Using Global Analysis, Partial Specifications, and an Extensible Assertion Language for Program Validation and Debugging.
Pages 161–192 of: Apt, K. R., Marek, V., Truszczynski, M., & Warren, D. S. (eds), The
Logic Programming Paradigm: a 25–Year Perspective. Springer-Verlag.
Jaffar, Joxan, & Lassez, Jean-Louis. (1987). Constraint Logic Programming. Pages 111–
119 of: ACM symposium on principles of programming languages. ACM.
Janson, S., & Haridi, S. (1991). Programming Paradigms of the Andorra Kernel Language.
Pages 167–183 of: 1991 international logic programming symposium. MIT Press.
Kowalski, R. A. (1974). Predicate Logic as a Programming Language. Pages 569–574 of:
Proceedings ifips.
Loke, S.W., & Davison, A. (1996). Logic programming with the World-Wide Web.
Pages 235–245 of: 7th. ACM Conference on Hypertext. ACM Press. Available from
http://www.cs.unc.edu/~barman/HT96/P14/lpwww.html.
M. Dincbas, H. Simonis, & Van Hentenryck, P. (1990). Solving Large Combinatorial
Problems in Logic Programming. Journal of logic programming, 8(1 & 2), 72–93.
Partners, The RADIOWEB Project. 1997 (July). RADIOWEB EP25562: Automatic Generation of Web Sites for the Radio Brodcasting Industry – Project Description / Technical Annex. Technical Report. RADIOWEB Project.
Ph. Bonnet, S. Bressan, L. Leth, & Thomsen, B. 1996 (September).
Towards ECLiPSe Agents on the INTERNET.
Proceedings of the 1st workshop on logic programming tools for internet applications.
Available from
http://clement.info.umoncton.ca/~lpnet/lpnet2.html.
Smedbäck, G., Carro, M., & Hermenegildo, M. (1999). Interfacing Prolog and VRML and
its Application to Constraint Visualization. Pages 453–471 of: The practical application
of constraint technologies and logic programming. The Practical Application Company.
Smolka, G. 1994 (November). The Definition of Kernel Oz. DFKI Oz documentation
series. German Research Center for Artificial Intelligence (DFKI).
S.W. Loke, A. Davison, & Sterling, L. 1996 (September).
Lightweight Deductive Databases on the World-Wide Web.
Proceedings of the 1st workshop on logic programming tools for internet applications.
Available from
http://clement.info.umoncton.ca/~lpnet/lpnet10.html.
Szeredi, Péter, Molnár, Katalin, & Scott, Rob. 1996 (September). Serving Multiple HTML Clients from a Prolog Application.
Proceedings of the 1st workshop on logic programming tools for internet applications.
Available from
http://clement.info.umoncton.ca/~lpnet/lpnet9.html.
Tarau, P. 1996 (April). Binprolog 5.00. Posting in comp.lang.prolog. Available from
http://clement.info.umoncton.ca/~tarau.
Van Hentenryck, P. (1989). Constraint Satisfaction in Logic Programming. MIT Press.
Warren, D.H.D., & Pereira, F. C. N. (1982). An Efficient, Easily Adaptable System For
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 33
Interpreting Natural Language Queries. American journal of computational linguistics,
8(3-4), 110–122.
| 2cs.AI
|
Porcellio scaber algorithm (PSA) for solving constrained optimization problems
Yinyan Zhang1 ,⋆ , Shuai Li1 ,⋆⋆ , and Hongliang Guo2 ,⋆⋆⋆
arXiv:1710.04036v1 [cs.NE] 11 Oct 2017
1
Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China
Center for Robotics, School of Automation Engineering, University of Electronic Science and Technology of China, Chengdu,
China
2
Abstract. In this paper, we extend a bio-inspired algorithm called the porcellio scaber algorithm (PSA) to solve
constrained optimization problems, including a constrained mixed discrete-continuous nonlinear optimization
problem. Our extensive experiment results based on benchmark optimization problems show that the PSA has a
better performance than many existing methods or algorithms. The results indicate that the PSA is a promising
algorithm for constrained optimization.
1 Introduction
Modern optimization algorithms may be roughly classified
into deterministic optimization algorithms and stochastic
ones. The former is theoretically sound for well-posed
problems but not efficient for complicated problems. For
example, when it comes to nonconvex or large-scale optimization problems, deterministic algorithms may not be
a good tool to obtain a globally optimal solution within a
reasonable time due to the high complexity of the problem.
Meanwhile, while stochastic ones may not have a strong
theoretical basis, they are efficient in engineering applications and have become popular in recent years due to
their capability of efficiently solving complex optimization
problems, including NP-hard problems such as the travelling salesman problem. Bio-inspired algorithms take an
important role in stochastic algorithms for optimization.
These algorithms are designed based on the observations
of animal behaviors. For example, one of the well known
bio-inspired algorithm called particle swarm optimization
initially proposed by Kennedy and Eberhart [1] is inspired
by the social foraging behavior of some animals such as
the flocking behavior of birds.
There are some widely used benchmark problems in
the field of stochastic optimization. The pressure vessel
design optimization problem is an important benchmark
problem in structural engineering optimization [2]. The
problem is a constrained mixed discrete-continuous nonlinear optimization problem. In recent years, many bioinspired algorithms have been proposed to solve the problem [3–6]. The widely used benchmark problems also include a nonlinear optimization problem proposed by Himmelblau [7].
⋆ e-mail: [email protected]
⋆⋆ e-mail: [email protected]
⋆⋆⋆ e-mail: [email protected]
Recently, a novel bio-inspired algorithm called the
porcellio scaber algorithm (PSA) has been proposed by
Zhang and Li [8], which is inspired by two behaviors of
porcellio scaber. In this paper, we extend the result in [8]
to solve constrained optimization problems. As the original algorithm proposed in [8] deals with the case without
constraints, we provide some improvements for the original PSA so as to make it capable of solving constrained optimization problems. Then, we compare the corresponding
experiment results with reported ones for the aforementioned benchmark problems as case studies. Our extensive
experiment results show that the PSA has much better performance in solving optimization problems than many existing algorithms. Before ending this introductory section,
the main contributions of this paper are listed as follows:
1)We extend the PSA to solve constrained optimization
problems, including the constrained mixed discretecontinuous nonlinear optimization problem.
2)We show that the PSA is better than many other existing algorithms in solving constrained optimization problems by extensive numerical experiments.
2 Problem Formulation
The constrained optimization problem (COP) considered
in this paper is presented as follows:
minimize fˇ(x),
subject to g j (x) ≤ 0,
li ≤ xi ≤ ui ,
(1)
with i = 1, 2, · · · , d and j = 1, 2, · · · , m, where x =
[x1 , x2 , · · · , xd ]T is a d-dimension decision vector; li and ui
are the corresponding lower bound and upper bound of the
ith decision variable; fˇ(x) : Rd → R is the cost function
Algorithm 1 Original PSA
Cost function f (x), x = [x1 , x2 , · · · , xd ]T
Generate initial position of porcellio scaber x0i (i =
1, 2, · · · , N)
Environment condition Ex at position x is determined by
f (x)
Set weighted parameter λ for decision based on aggregation and the propensity to explore novel environments
Initialize f∗ to an extremely large value
Initialize each element of vector x∗ ∈ Rd to an arbitrary
value
while k < MaxS tep do
Get the position with the best environment condition,
i.e., xb = arg minxkj { f (xkj )} at the current time among
the group of porcellio scaber
if minxkj { f (xkj )} < f∗ then
x∗ = xb
f∗ = minxkj { f (xkj )}
end if
Randomly chose a direction τ = [τ1 , τ2 , · · · , τd ]T to
detect
Detect the best environment condition min{Ex } and
worst environment condition max{Ex } at position xki +
τ for i = 1 : N all N porcellio scaber
for i = 1 : N all N porcellio scaber do
Determine the difference with respect to the position to aggregate i.e., xki − arg minxkj { f (xkj )})
Determine where to explore, i.e., pτ
Move to a new position according to (2)
end for
end while
Output x∗ and the corresponding function value f∗
Visualization
to be minimized. For the case that the problem is convex,
there are many standard algorithms to solve the problem.
However, for the case that the problem is not convex, the
problem is difficult to solve.
3 Algorithm Design
where λ ∈ (0, 1), τ is a vector with each element being a
random number, and p is defined as follows:
p=
f (xki + τ) − min{ f (xki + τ)}
max{ f (xki + τ)} − min{ f (xki + τ)}
Evidently, the original PSA does not take constraints into
consideration. Thus, it cannot be directly used to solve
COPs.
3.2 Inequality constraint conversion
In this subsection, we provide some improvements for the
original PSA and make it capable of solving COPs. As the
original PSA focuses on solving unconstrained problem,
we first incorporate the inequality constraints g j (x) ≤ 0
( j = 1, 2, · · · , m) into the cost function. To this end, the
penalty method is used, and a new cost function is obtained as follows:
fˇ(x) = f (x) + γ
m
X
g2i (x)h(gi(x)),
3.1 Original PSA
For the sake of understanding, the original PSA is given
in algorithm 1 [8], which aims at solving unconstrained
optimization problems of the following form:
minimize f (x),
where x is the decision vector and f is the cost function
to be minimized. The main formula of the original PSA is
given as follows [8]:
xik+1 = xki − (1 − λ)(xki − arg min{ f (xkj )}) + λpτ,
xkj
(2)
(3)
i=1
where h(gi (x)) is defined as
1, if gi (x) > 0,
h(gi (x)) =
0, if gi (x) ≤ 0,
and γ ≫ 1 is the penalty parameter. By using a large
enough value of γ (e.g., 1012 ), unless all the inequality constraints gi (x) ≤ 0 (i = 1, 2, · · · , m) are satisfied,
PN 2
the term γ i=1
gi (x)h(gi(x)) takes a dominant role in the
cost function. On the other hand, when all the inequality constraints gi (x) ≤ 0 (i = 1, 2, · · · , m) are satisfied,
h(gi (x)) = 0, ∀i, and thus fˇ(x) = f (x).
3.3 Addressing simple bounds
In terms of the simple bounds l j ≤ x j ≤ u j with j =
1, 2, · · · , d, they are handled via two methods. Firstly, to
satisfy the simple bounds, the initial position of each porcellio scaber is set via the following formula:
x0i, j = l j + (u j − l j ) × rand(0, 1)
In this section, we modify the original PSA [8] and provide
an improved PSA for solving COPs.
.
(4)
where x0i, j denotes the initial value of the jth variable of
the position vector of the ith (with i = 1, 2, · · · , N) porcellio scaber; rand(0, 1) denotes a random number in the
region (0, 1), which can be realized by using the rand function in Matlab. The formula (4) guarantees that the initial
positions of all the porcellio scaber satisfy the the simple
bounds l j ≤ x j ≤ u j with j = 1, 2, · · · , d.
Secondly, if the positions of all the porcellio scaber are
updated according to (2) by replacing f (x) with fˇ(x) defined in (3) for the constrained optimization problem (1),
then the updated values of the position vector xki may violate the simple bound constraints. To handle this issue,
based on (2), a modified evolution rule is proposed as follows:
xik+1 = PΩ (xki − (1 − λ)(xki − arg min{ fˇ(xkj )}) − λpτ), (5)
xkj
Algorithm 2 Algorithm for the evaluation of PΩ (x) with
x = [x1 , x2 , · · · , xd ]T
for i = 1 : d do
if x1 < li then
xi = li
end if
if xi > ui then
xi = u i
end if
end for
return y = [x1 , x2 , · · · , xd ]T
Figure 1. A diagram showing the design parameters of a pressure
vessel [9].
where λ ∈ (0, 1), τ is a vector with each element being a
random number, and
p=
fˇ(xki + τ) − min{ fˇ(xki + τ)}
.
max{ fˇ(xk + τ)} − min{ fˇ(xk + τ)}
i
i
Besides, PΩ is a projection function and make the updated
position satisfy the simple bound constraints, where Ω =
{x ∈ Rd |li ≤ xi ≤ ui , i = 1, 2 · · · , d}. The mathematical
definition of PΩ (x) is PΩ (x) = arg miny∈Ω ky − xk2 with
k · k2 denoting the Euclidean norm. The algorithm for the
evaluation of PΩ (x) is given in Algorithm 2.
Algorithm 3 PSA for COPs
Cost function fˇ(x) as defined in (3),
x =
[x1 , x2 , · · · , xd ]T
Generate initial position of porcellio scaber x0i (i =
1, 2, · · · , N) according to (4)
Environment condition Ex at position x is determined by
fˇ(x)
Set weighted parameter λ for decision based on aggregation and the propensity to explore novel environments
Set penalty parameter γ in fˇ(x) to a large enough value
Initialize f∗ to an extremely large value
Initialize each element of vector x∗ ∈ Rd to an arbitrary
value
while k < MaxS tep do
Get the position with the best environment condition,
i.e., xb = arg minxkj { f (xkj )} at the current time among
the group of porcellio scaber
if minxkj { f (xkj )} < f∗ then
x∗ = xb
f∗ = minxkj { f (xkj )}
end if
Randomly chose a direction τ = [τ1 , τ2 , · · · , τd ]T to
detect
Detect the best environment condition min{Ex } and
worst environment condition max{Ex } at position xki +
τ for i = 1 : N all N porcellio scaber
for i = 1 : N all N porcellio scaber do
Determine the difference with respect to the position to aggregate i.e., xki − arg minxkj { f (xkj )})
Determine where to explore, i.e., pτ
Move to a new position according to (5) where
PΩ (x) is evaluated via Algorithm 2
end for
end while
Output x∗ and the corresponding function value f∗
Visualization
3.4 PSA for COPs
Based on the above modifications, the resultant PSA for
solving COPs is given in Algorithm 3. In the following
section, we will use some benchmark problems to test the
performance of the PSA in solving COPs.
4 Case Studies
In this section, we present experiment results regarding using the PSA for solving COPs.
4.1 Case I: Pressure vessel problem
In this subsection, the pressure vessel problem is considered. The pressure vessel problem is to find a set of
four design parameters, which are demonstrated in Fig. 1,
to minimize the total cost of a pressure vessel considering
the cost of material, forming and welding [1]. The four
design parameters are the inner radius R, and the length
L of the cylindrical section, the thickness T h of the head,
the thickness T s of the body. Note that, T s and T h are integer multiples of 0.0625 in., and R and L are continuous
variables.
Table 1. Comparisons of best results for the pressure vessel problem
[4]
[9]
[3]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
PSA
x1
x2
x3
x4
g1 (x)
g2 (x)
g3 (x)
g4 (x)
f (x)
0.8125
0.7782
0.8125
1.125
1.125
1.125
0.8125
0.9375
1.125
0.8125
1.000
0.8125
1
0.8125
0.8125
0.4375
0.3846
0.4375
0.625
0.625
0.625
0.4375
0.5000
0.625
0.4375
0.625
0.4375
0.625
0.4375
0.4375
42.0984
40.3196
42.0984
58.2789
48.97
58.1978
40.3239
48.3290
58.291
41.9768
51.000
42.0870
51.2519
42.0984
42.0952
176.6366
200.000
176.6366
43.7549
106.72
44.2930
200.0000
112.6790
43.690
182.2845
91.000
176.7791
90.9913
176.6378
176.8095
8.00e-11†
-3.172e-5
8.00e-11†
-0.0002
-0.1799
-0.00178
-0.034324
-0.0048
0.000016
-0.0023
-0.0157
-2.210e-4
-1.011
-8.8000e-7
-6.2625e-5
-0.0359
4.8984e-5†
-0.0359
-0.06902
-0.1578
-0.06979
-0.05285
-0.0389
-0.0689
-0.0370
-0.1385
-0.03599
-0.136
-0.0359
-0.0359
-2.724e-4
1.3312†
-2.724e-4
-3.71629
97.760
-974.3
-27.10585
-3652.877
-21.2201
-22888.07
-3233.916
-3.51084
-18759.75
-3.5586
-738.7348
-63.3634
-40
-63.3634
-196.245
-132.28
-195.707
-40.0000
-127.3210
-196.3100
-57.7155
-149
-63.2208
-149.009
-63.3622
-63.1905
6059.7143
5885.33
6059.7143
7198.433
7980.894
7207.494
6288.7445
6410.3811
7198.0428
6171.000
7079.037
6061.1229
7172.300
6059.7258
6063.2118
† means that the corresponding constraint is violated.
Algorithm 4 Algorithm for PΩ (x = [x1 , x2 , x3 , x4 ]T ) in the
pressure vessel problem
y1 = round(x1 /0.0625) × 0.0625
if y1 < 0.0625 then
y1 = 0.0625
end if
if y1 > 99 × 0.0625 then
y1 = 99 × 0.0625
end if
y2 = round(x2 /0.0625) × 0.0625
if y2 < 0.0625 then
y2 = 0.0625
end if
if y2 > 99 × 0.0625 then
y2 = 99 × 0.0625
end if
if x3 < 10 then
y3 = 10
end if
if x3 > 200 then
y3 = 200
end if
if x4 < 10 then
y4 = 10
end if
if x4 > 200 then
y4 = 200
end if
return y = [y1 , y2 , y3 , y4 ]T
Let x = [x1 , x2 , x3 , x4 ]T = [T s , T h , R, L]T . The pressure
vessel problem can be formulated as follows [9]:
minimize f (x) = 0.6224x1 x3 x4 + 1.7781x2 x23
+ 3.1661x21 x4 + 19.84x21 x3 ,
subject to g1 (x) = −x1 + 0.0193x3 ≤ 0,
g2 (x) = −x2 + 0.00954x3 ≤ 0,
4
g3 (x) = −πx23 x4 − πx33 + 1296000 ≤ 0,
3
g4 (x) = x4 − 240 ≤ 0,
Evidently, this problem has a nonlinear cost function, three
linear and one nonlinear inequality constraints. Besides,
there are two discrete and two continuous design variables.
Thus, the problem is relatively complicated. As this problem is a mixed discrete-continuous optimization, the projection function PΩ (x) is slightly modified and presented
in Algorithm 4. Besides, the initialization of the initial
positions of porcellio scaber is modified as follows:
x0i,1 = 0.0625 + f loor((99 − 1) × rand) × 0.0625,
x0i,2 = 0.0625 + f loor((99 − 1) × rand) × 0.0625,
x0i,3 = 10 + f loor(200 − 10) × rand,
x0i,4 = 10 + f loor(200 − 10) × rand,
where x0i, j denotes the jth variable of the position vector
of the ith porcellio scaber; f loor(y) = argminx∈{0,1,2,··· } {x +
1 > y}, i.e., the f loor function obtains the integer part
of a real number; rand denotes a random number in the
region (0, 1). The functions f loor and rand are available
at Matlab.
The best result we obtained using the PSA in 1000
instances of executions and those by using various existing algorithms or methods for solving this problem are
listed in Table 1. Note that, in the experiments, 40 porcellio scaber are used, the parameter λ is set to 0.6, and
the MaxS tep is set to 100000 with τ being a zero-mean
random number with the standard deviation being 0.1. As
seen from Table 1, the best result obtained by using the
PSA is better than most of the existing results. Besides,
the difference between the best function value among all
the ones in the table and the best function value obtained
via using the PSA is quite small.
4.2 Case II: Himmelblau’s nonlinear optimization
problem
In this subsection, we consider a nonlinear optimization
problem proposed by Himmelblau [7]. This problem is
Table 2. Comparisons of best results for Himmelblau’s nonlinear optimization problem
[21]
[22]
[23]
[7]
[24]
PSA
x1
x2
x3
x4
x5
g1 (x)
g2 (x)
g3 (x)
f (x)
78.0
78.00
81.4900
78.6200
78.00
79.9377
33.0
33.00
34.0900
33.4400
33.00
33.8881
27.07997
29.995
31.2400
31.0700
29.995256
28.5029
45.0
45.00
42.2000
44.1800
45.00
41.3052
44.9692
36.776
34.3700
35.2200
36.775813
41.7704
92.0000
90.7147
90.5225
90.5208
92
91.6157
100.4048
98.8405
99.3188
98.8929
98.8405
100.4943
20.0000
19.9999†
20.0604
20.1316
20
20.0055
-31025.5602
-30665.6088
-30183.576
-30373.949
-30665.54
-30667.8113
† means that the corresponding constraint is violated.
also one of the well known benchmark problems for bioinspired algorithms. The problem is formally described as
follows [7]:
minimize f (x) =5.3578547x23 + 0.8356891x1 x5
+ 37.29329x1 − 40792.141,
subject to g1 (x) =85.334407 + 0.0056858x2 x5
+ 0.00026x1 x4 − 0.0022053x3 x5 ,
g2 (x) =80.51249 + 0.0071317x2 x5
+ 0.0029955x1 x2 + 0.0021813x23,
g3 (x) =9.300961 + 0.0047026x3 x5
+ 0.0012547x1 x3 + 0.0019085x3 x4 ,
0 ≤ g1 (x) ≤ 92,
90 ≤ g2 (x) ≤ 110,
20 ≤ g3 (x) ≤ 25,
78 ≤ x1 ≤ 102,
33 ≤ x2 ≤ 45,
27 ≤ x3 ≤ 45,
27 ≤ x4 ≤ 45,
27 ≤ x5 ≤ 45,
with x = [x1 , x2 , x3 , x4 , x5 ]T being the decision vector. In
this problem, each double-side nonlinear inequality can be
represented by two single-side nonlinear inequality constraints. For example, the constraint 90 ≤ g2 (x) ≤ 110 can
be replaced by the following two constraints:
−g2 (x) ≤ −90,
g2 (x) ≤ 110.
Thus, this problem can also be solved by the PSA proposed in this paper.
The best result we obtained via using the PSA in 1000
instances of executions, together with the result obtained
by other algorithms or methods, is listed in Table 2. In the
experiments, 40 porcellio scaber are used, the parameter λ
is set to 0.6, and the MaxS tep is set to 100000 with τ being
a zero-mean random number with the standard deviation
being 0.1. Evidently, the best result generated by the PSA
is ranked No. 2 among all the results in Table 2.
By the above results, we conclude that the PSA is a
relatively promising algorithm for solving constrained optimization problems. The quite smalle performance difference between the PSA and the best one may be the result
of the usage of the penalty method with a constant penalty
parameter.
5 Conclusions
In this paper, the bio-inspired algorithm PSA has been extended to solve nonlinear constrained optimization problems by using the penalty method. Case studies have validated the efficacy and superiority of the resultant PSA. The
results have indicated that the PSA is a promising algorithm for solving constraint optimization problems. There
are several issues that requires further investigation, e.g.,
how to select a best penalty parameter that not only guarantees the compliance with constraints but also the optimality of the obtained solution. Besides, how to enhance
the efficiency of the PSA is also worth investigating.
Acknowledgement
This work is supported by the National Natural Science
Foundation of China (with numbers 91646114, 61370150,
and 61401385), by Hong Kong Research Grants Council
Early Career Scheme (with number 25214015), and also
by Departmental General Research Fund of Hong Kong
Polytechnic University (with number G.61.37.UA7L).
References
[1] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc. IEEE Int. Conf. Neural Netw., 1995,
1942–1948.
[2] A. H. Gandomi, X. Y. Yang, “Benchmark problems
in structural optimization,” in Computational Optimization Methods And Algorithms, Berlin: Springer-Verlag,
2011, pp. 259–281.
[3] S. He , E. Prempain, and Q. H. Wu, “An improved particle swarm optimizer for mechanical design optimization problems,” Eng. Opt., vol. 36, no. 5, pp. 585–605,
2004.
[4] A. H. Gandomi, X. S. Yang, and A. H. Alavi, “Cuckoo
search algorithm: A metaheuristic approach to solve
structural optimization problems,” Eng. Comput., vol.
29, pp. 17–35, 2013.
[5] O. Pauline, H. C. Sin, D. D. C. V. Sheng, S. C. Kiong,
and O. K. M., “Design optimization of structural engineering problems using adaptive cuckoo search algorithm,” in Proc. 3rd Int. Conf. Control Autom. Robot.,
2017, pp. 745–748.
[6] V. Garg and K. Deep, “Effectiveness of constrained
laplacian biogeography based optimization for solving
structural engineering design problems,” in Proc. 6th
Int. Conf. Soft. Comput. Problem Solv., 2017, pp. 206–
219.
[7] D. Himmelblau, Applied Nonlinear Programming,
New York: McGraw-Hill, 1972.
[8] Y. Zhang and S. Li, “PSA: A novel optimization algorithm based on survival rules of porcellio scaber,”
Available at https://arxiv.org/abs/1709.09840.
[9] X. S. Yang, “Firefly algorithm, stochastic test functions and design optimisation,” Int. J. Bio-Inspired
Comp., vol. 2, no. 2, pp. 2, pp. 78–84, 2010.
[10] K. S. Lee and Z. W. Geem, “A new meta-heuristic algorithm for continuous engineering optimization: Harmony search theory and practice,” Comput. Methods
Appl. Mech. Engrg., vol. 194, pp. 3902–3933, 2005.
[11] E. Sandgren, “Nonlinear integer and discrete programming in mechanical design optimization,” J. Mech.
Des. ASME, vol. 112, 223–229, 1990.
[12] S. J. Wu and P. T. Chow, “Genetic algorithms for
nonlinear mixed discrete-integer optimization problems
via meta-genetic parameter optimization,” Engrg. Optim., vol. 24, 137–159, 1995.
[13] C. A. C. Coello, “Use of a self-adaptive penalty approach for engineering optimization problems,” Comput. Ind., vol. 41, no. 2, pp. 113–127, 2000.
[14] K. Deb, “GeneAS: a robust optimal design technique
for mechanical component design”, in: D. Dasgupta,
Z. Michalewicz Eds., Evolutionary Algorithms in Engineering Applications, Springer-Verlag, Berlin, 1997,
pp. 497–514.
[15] B. K. Kannan and S. N. Kramer, “An augmented Lagrange multiplier based method for mixed integer discrete continuous optimization and its applications to
mechanical design,” J. Mecha. Design, Trans. ASME,
vol. 116, pp. 318–320, 1994.
[16] S. Akhtar, K. Tai, T. Ray, “A socio-behavioural simulation model for engineering design optimization,” Eng.
Optmiz., vol. 34, no. 4, pp. 341–354, 2002.
[17] J. F. Tsai, H. L. Li, and N. Z. Hu, “Global optimization for signomial discrete programming problems in
engineering design,” Eng. Optmiz. no. 34, no. 6, pp.
613–622, 2002.
[18] C. A. C. Coello and N. C. Corteś, “Hybridizing a
genetic algorithm with an artificial immune system for
global optimization,” Eng. Optmiz., vol. 36, no. 5, pp.
607–634, 2004.
[19] H. L. Li and C. T. Chou, “A global approach for nonlinear mixed discrete programming in design optimization,” Eng. Optmiz. vol. 22, pp. 109–122 1994.
[20] A. Kaveh and S. Talatahari, “An improved ant colony
optimization for constrained engineering design problems,” Eng. Comput. vol. 27, no. 1, pp. 155–182, 2010.
[21] G. H. M. Omran and A. Salman, “Constrained optimization using CODEQ,” Chaos Soliton. Fract., vol.
42, pp. 662–668, 2009.
[22] A. Homaifar, S. Lai, X. Qi, “Constrained optimization via genetic algorithms,” Simulation, vol. 62, no. 4,
pp. 242–253, 1994.
[23] M. Gen and R. Cheng, Genetic Algorithms and Engineering Design, Wiley, New York, 1997.
[24] G. G. Dimopoulos, “Mixed-variable engineering optimization based on evolutionary and social
metaphors,” Comput. Method. Appl. Mech. Eng., vol.
196, pp. 803–817, 2007.
| 9cs.NE
|
Logical Methods in Computer Science
Vol. 2 (3:3) 2006, pp. 1–22
www.lmcs-online.org
Submitted
Published
Dec. 16, 2005
Jul. 27, 2006
RELATIONAL PARAMETRICITY AND CONTROL
MASAHITO HASEGAWA
Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-8502 Japan, and
PRESTO, Japan Science and Technology Agency
e-mail address: [email protected]
Abstract. We study the equational theory of Parigot’s second-order λµ-calculus in connection with a call-by-name continuation-passing style (CPS) translation into a fragment
of the second-order λ-calculus. It is observed that the relational parametricity on the target calculus induces a natural notion of equivalence on the λµ-terms. On the other hand,
the unconstrained relational parametricity on the λµ-calculus turns out to be inconsistent. Following these facts, we propose to formulate the relational parametricity on the
λµ-calculus in a constrained way, which might be called “focal parametricity”.
Dedicated to Prof. Gordon Plotkin on the occasion of his sixtieth birthday
1. Introduction
The λµ-calculus, introduced by Parigot [26], has been one of the representative term
calculi for classical natural deduction, and widely studied from various aspects. Although
it still is an active research subject, it can be said that we have some reasonable understanding of the first-order propositional λµ-calculus: we have good reduction theories,
well-established CPS semantics and the corresponding operational semantics, and also some
canonical equational theories enjoying semantic completeness [16, 24, 25, 36, 39]. The last
point cannot be overlooked, as such complete axiomatizations provide deep understanding of equivalences between proofs and also of the semantic structure behind the syntactic
presentation.
The second-order λµ-calculus (λµ2), again due to Parigot [27], has been studied in depth
as a calculus for second-order classical natural deduction. In particular, strong normalization results of λµ2 [27, 23] and its extensions, e.g. with inductive types [21], have been a
central research topic, because of the proof-theoretical importance of strong normalization.
However, for λµ2, it seems that there are few attempts of giving an equational theory supported by some fine semantic structure. This situation is rather frustrating, since without
such equational and semantic accounts, we cannot discuss e.g. the correctness of the impredicative encoding of the datatypes in λµ2. For the second-order λ-calculus λ2 (system
2000 ACM Subject Classification: F.3.2.
Key words and phrases: polymorphism, parametricity, continuations.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-2 (3:3) 2006
CC
Masahito Hasegawa
Creative Commons
2
MASAHITO HASEGAWA
F) [9, 33], a subsystem of λµ2, there are several beautiful results on the relational parametricity [34] and the universal properties of impredicative constructions [1, 15, 31, 40, 44],
e.g. that µX.σ = ∀X.(σ → X) → X (where σ covariant in X) gives an initial algebra of
the functor ΛX.σ in a suitable sense. We certainly wish to have such a story for λµ2 too.
This work is an attempt to identify such an equational theory which is backed up by
certain semantic structures. Specifically, we propose a relational parametricity principle
which is sound and sufficiently powerful for deriving such equivalences on the λµ-terms.
1.1. Parametric CPS semantics. We first consider the semantics of λµ2 given by a
CPS-translation into a fragment of λ2 — that of the second-order existential types ∃X.τ ,
conjunction types τ1 ∧ τ2 , and arrow types τ → R into a distinguished type R (this choice
of the target calculus is due to a recent work of Fujita [7]). The translation (−)◦ sends
a type variable X to X, arrow type σ1 → σ2 to (σ1◦ → R) ∧ σ2◦ , and the universal type
∀X.σ to ∃X.σ ◦ — while a term M : σ is sent to [[M ]] : σ ◦ → R. It can be considered as a
natural extension of Streicher’s call-by-name CPS translation [36, 39, 43]. It follows that
this translation already gives a reasonable equational theory on λµ2, in that it validates the
standard βη-equalities1 . In fact, this is a consequence of a fibred version of the ”category
of continuations” construction [16, 36, 39].
However, this is just a starting point; we observe that, if some of the impredicative
constructions in the target calculus satisfy certain universal properties (e.g. ∃X.X is a
terminal object) which follow from the relational parametricity, then so do the impredicative
constructions in the source λµ2-calculus — but not quite in the way that we first might
expect. For instance, the type ⊥ = ∀X.X does not give an initial object (cf. [37]) — instead
it plays the role of the falsity type (or the “answer type”); in fact, we have a double-negation
elimination from (σ → ⊥) → ⊥ to σ for any σ which actually is an algebra of the doublenegation monad ((−) → ⊥) → ⊥. As another major example, ∀X.(σ → X) → X does
not give an initial algebra of ΛX.σ; it gives an initial algebra of ΛX.(σ → ⊥) → ⊥ — not
with respect to all terms but to a certain class of terms (the “focal terms”, to be mentioned
below). In particular, if X is not free in σ, ∀X.(σ → X) → X is isomorphic not to σ but
to (σ → ⊥) → ⊥. In short, impredicative encodings in λµ2 get extra double negations,
and the relational parametricity of λ2 is not consistent with the equational theory of λµ2
induced by the CPS semantics. As a consequence, we cannot encode cartesian products in
λµ2, though they can be added easily. Also we cannot express the classical disjunctions
[36], though they can be added without changing the target of the CPS translation.
1.2. Focal parametricity. These results suggest that the CPS translation into parametric
target calculus gives a reasonable semantic foundation and equational theory for λµ2, which
is sufficient for obtaining various interesting results. However, here the parametricity is
used rather indirectly, via the CPS translation; we also wish to have a decent notion of
parametricity directly within λµ2. To figure out what sort of parametricity principle can be
expected for λµ2, recall the following fact on λ2 with parametricity: given a polymorphic
term M : ∀X.F [X] → G[X] (with X covariant in F and G) and types σ1 , σ2 , the instances
1We can say more – we can show that this CPS-semantics is sound and complete with respect to the
βη-theory of λµ2. This result, together with further syntactic analysis of this CPS translation, will appear
in a forthcoming paper with Ken-etsu Fujita.
RELATIONAL PARAMETRICITY AND CONTROL
3
M σ1 : F [σ1 ] → G[σ1 ] and M σ2 : F [σ2 ] → G[σ2 ] obey the naturality, in that the following
diagram
F [f ]
- F [σ2 ]
F [σ1 ]
M σ1
M σ2
?
G[σ1 ]
?
- G[σ2 ]
G[f ]
commutes for any f : σ1 → σ2 . This is no longer true for λµ2. For example, let F [σ] = (σ →
⊥) → ⊥, G[σ] = σ and M be the double-negation elimination (which does not exist in λ2);
then the naturality for arbitrary maps implies inconsistency — we get σ ≃ (σ → ⊥) → ⊥
for every σ by letting f be the obvious map from σ to (σ → ⊥) → ⊥, which is enough to
kill the theory [20]. Similar result can be observed for other “classical” proofs, e.g. of the
Peirce law.
To this end, we look at the focus [36] (centre [32, 42], C-maps [16]) of λµ2; a focal map
is no other than an algebra morphism between the the double-negation monad mentioned
above, i.e., a map making the naturality diagram for the double-negation elimination commute. It follows that a notion of relational parametricity on λµ2 in which the construction
of the graph relations is allowed only for focal maps is consistent, as there are nontrivial
models. Together with the definability (fullness) of the CPS translation, we see that it is at
least as powerful as the parametricity on the CPS target calculus which we have mentioned
above, thus gives a powerful principle for deriving the equivalences of terms in λµ2. (We
actually conjecture that these two notions of parametricity do agree, but it is open as of
writing this article.) This principle, which we shall call focal parametricity, should be a
natural notion of parametricity for λµ2. We will sketch some use of focal parametricity for
deriving “free theorems” for λµ2 syntactically.
1.3. Towards parametricity for computational effects. At the conceptual and abstract level, this story closely resembles to the study of linear parametricity and recursion
[4, 30]. In the case of linear parametricity, the graph relations are allowed to be constructed
only from the linear maps, and a linear map is an algebra map w.r.t. the lifting monad.
We claim that, just like the linear parametricity gives a solution of accommodating nontermination and recursion in the polymorphic setting (as advocated by Plotkin [30]), the
focal parametricity provides a way of accommodating control features in the polymorphic
setting. In short:
focal parametricity
linear parametricity
=
non-termination
first-class control
As future work, it would be an interesting challenge to find a unifying framework
of linear parametricity and focal parametricity; it should be useful to have parametric
polymorphism, recursion, and control at once, as in the realistic programming languages (cf.
[14, 18, 19]). More ambitiously, we are keen to see an adequate notion of parametricity for
fairly general “effectful” settings. Possible starting points for this direction might include
the “parametricity graphs” approach [5] which allows us to deal with parametricity at
a general level (including the linear parametricity as an instance), and the “category of
4
MASAHITO HASEGAWA
linear continuations construction” [12] which induces both the CPS translation and Girard
translation as special cases. See Section 7 for further discussions related to this issue.
1.4. Construction of this paper. The rest of this paper is organised as follows. In section
2 and 3 we introduce the calculi and CPS-translation which are the subject of this study.
In section 4 we consider the implications of the relational parametricity on the CPS-target
calculus. The focal parametricity is introduced in section 5, followed by examples in section
6, including focally initial algebras and the type of Church numerals. Section 7 gives an
alternative characterisation of focus, which suggests a generalisation of this work to a theory
of parametricity for general effects. We then give some concluding remarks in Section 8.
2. The calculi
2.1. The second-order λµcalculus. The second-order λµ-calculus, λµ2, is given as follows. We essentially follow Parigot’s formulation [27] (with some flavour from Selinger’s
[36]). The types are the same as those of the second-order λ-calculus λ2:
σ ::= X | σ → σ | ∀X.σ
In a typing judgement Γ ⊢ M : σ | ∆, Γ stands for the typing context of variables, while ∆
for the context of names (continuation variables).
Γ, x : σ, Γ′ ⊢ x : σ | ∆
Γ, x : σ1 ⊢ M : σ2 | ∆
Γ ⊢ λxσ1 .M : σ1 → σ2 | ∆
Γ ⊢ M : σ1 → σ2 | ∆ Γ ⊢ N : σ1 | ∆
Γ ⊢ M N : σ2 | ∆
Γ ⊢ M : σ | ∆ (X 6∈ F T V (Γ, ∆))
Γ ⊢ ΛX.M : ∀X.σ | ∆
Γ ⊢ M : ∀X.σ1 | ∆
Γ ⊢ M σ2 : σ1 [σ2 /X] | ∆
Γ ⊢ M : σ2 | α : σ1 , ∆ (β : σ2 ∈ α : σ1 , ∆)
Γ ⊢ µασ1 .[β]M : σ1 | ∆
RELATIONAL PARAMETRICITY AND CONTROL
5
The axioms for the equational theory are again the standard ones — note that we consider
the extensional theory, i.e. with the η-axioms.
(λxσ .M ) N
λxσ .M x
(ΛX.M ) σ
ΛX.M X
=
=
=
=
M [N/x]
M (x 6∈ F V (M ))
M [σ/X]
M (X 6∈ F T V (M ))
µα.[β](µγ.M )
µασ .[α]M
σ
→σ
1
2 .M ) N
(µα
∀X.σ
1 .M ) σ
(µα
2
=
=
=
=
µα.M [β/γ]
M (α 6∈ F N (M ))
µβ σ2 .M [[β](− N )/[α](−)]
µβ σ1 [σ2 /X] .M [[β](− σ2 )/[α](−)]
In the last two axioms, we make uses of so-called ”mixed substitution”; for instance,
M [[β](− N )/[α](−)] means replacing occurances of the form [α]L in M by [β](L N ) recursively.
In the sequel, we frequently use the following syntactic sugar. First, we let ⊥ be the
type ∀X.X — the type of falsity. We may also write ¬σ for σ → ⊥. Using ⊥, we define the
“named term”
[β]M ≡ µα⊥ .[β]M : ⊥
(where M : σ, β : σ, with α fresh) and the µ-abstraction
µασ .M : σ ≡ µασ .[α](M σ)
for M : ⊥. It follows that µασ .[β]M = µασ .[β]M holds.
With this ⊥, we can express the double-negation elimination in λµ2 by making use of
both the polymorphic and classical features:
Cσ = λm¬¬σ .µασ .m (λxσ .[α]x) : ¬¬σ → σ
As expected, we have Cσ (λk¬σ .k M ) = M . The properties of ⊥ and Cσ will be further
studied later under parametricity assumptions.
2.2. Target: the {∃, ∧, ¬}calculus. In tthe literature, the second-order λ-calculus (λ2)
is often taken as the target of the CPS translation for λµ2. Fujita observed that it actually
suffices to consider a fragment of λ2 with negations, conjunctions and existential types as
a target [7]. In this paper we follow this insight.
τ
::= X | R | ¬τ | τ ∧ τ | ∃X.τ
¬τ can be considered as a shorthand of τ → R. The type R can be replaced by ∃X.¬X ∧ X,
but for simplicity we keep R as a type constant. The syntax of terms is a fairly standard one,
though for conjunctions we employ a slightly less familiar elimination rule (with let-binding)
so that it parallels that of the existential types.
Γ, x : τ, Γ′ ⊢ x : τ
Γ, x : τ ⊢ M : R
Γ ⊢ λxτ .M : ¬τ
Γ ⊢ M : ¬τ Γ ⊢ N : τ
Γ⊢MN :R
6
MASAHITO HASEGAWA
Γ ⊢ M : τ1 Γ ⊢ N : τ2
Γ ⊢ hM, N i : τ1 ∧ τ2
Γ ⊢ M : τ1 ∧ τ2 Γ, x : τ1 , y : τ2 ⊢ N : τ3
Γ ⊢ let hxτ1 , y τ2 i be M in N : τ3
Γ ⊢ M : τ1 [τ2 /X]
Γ ⊢ hτ2 , M i : ∃X.τ1
Γ ⊢ M : ∃X.τ1 Γ, x : τ1 ⊢ N : τ2 X 6∈ F T V (Γ, τ2 )
Γ ⊢ let hX, xτ1 i be M in N : τ2
Again, we employ the standard βη-axioms.
(λxσ .M ) N
λxσ .M x
let hx, yi be hL, M i in N
let hx, yi be M in N [hx, yi/z]
let hX, xi be hτ, M i in N
let hX, xi be M in N [hX, xi/z]
=
=
=
=
=
=
M [N/x]
M (x 6∈ F V (M ))
N [L/x, M/y]
N [M/z]
N [τ /X, M/x]
N [M/z]
3. CPS translation
3.1. The CPS translation. We present a call-by-name CPS translation which can be
considered as an extension of that introduced by Streicher [16, 39, 36] (rather than the
translations by Plotkin [29], Parigot [27] or Fujita [7] which introduce extra negations and
do not respect extensionality).
X◦ = X
(σ1 → σ2 )◦ = ¬σ1◦ ∧ σ2◦
(∀X.σ)◦ = ∃X.σ ◦
[[xσ ]]
[[λxσ1 .M σ2 ]]
[[M σ1 →σ2 N σ1 ]]
[[ΛX.M σ ]]
[[M ∀X.σ1 σ2 ]]
σ
[[µα 1 .[β σ2 ]M σ2 ]]
where
λhxσ , y τ i.M
λhX, y τ i.M
=
=
=
=
=
=
◦
x¬σ
◦
◦
λhx¬σ1 , kσ2 i.[[M ]] k
◦
λkσ2 .[[M ]] h[[N ]], ki
◦
λhX, kσ i.[[M ]] k
◦
λkσ1 [σ2 /X] .[[M ]] hσ2◦ , ki
◦
λασ1 .[[M ]] β
≡ λz σ∧τ .let hxσ , y τ i be z in M
≡ λz ∃X.τ .let hX, y τ i be z in M
RELATIONAL PARAMETRICITY AND CONTROL
7
3.2. Soundness. The type soundness follows from a straightforward induction.
Proposition 3.1 (type soundness).
Γ ⊢ M : σ | ∆ =⇒ ¬Γ◦ , ∆◦ ⊢ [[M ]] : ¬σ ◦
◦ when Γ is x : σ , . . . , x
◦
where ¬Γ◦ is x1 : ¬σ1◦ , . . . , xm : ¬σm
1
1
m : σm , and ∆ = α1 :
◦
◦
σ1 , . . . , αn : σn for ∆ = α1 : σ1 , . . . , αn : σn .
Note that (σ[τ /X])◦ ≡ σ ◦ [τ ◦ /X], [[M [N/x]]] ≡ [[M ]][[[N ]]/x], and also [[M [σ/X]]] ≡
[[M ]][σ ◦ /X] hold. Then we have the equational soundness:
Proposition 3.2 (equational soundness).
Γ ⊢ M = N : σ |∆ =⇒ ¬Γ◦ , ∆◦ ⊢ [[M ]] = [[N ]] : ¬σ ◦
In addition, we have the definability result:
Proposition 3.3 (fullness).
¬Γ◦ , ∆◦ ⊢ N : ¬σ ◦ =⇒ N = [[M ]] for some Γ ⊢ M : σ | ∆
This can be proved by providing an inverse translation of the CPS translation, so that
¬Γ◦ , ∆◦ ⊢ P : ¬σ ◦ =⇒ Γ ⊢ P −1 : σ | ∆
¬Γ◦ , ∆◦ ⊢ C : σ ◦ =⇒ Γ ⊢ C −1 [−σ ] : ⊥ | ∆
¬Γ◦ , ∆◦ ⊢ A : R =⇒ Γ ⊢ A−1 : ⊥ | ∆
hold, where
Program : ¬σ ◦
P ::= x | λk.A
◦
Continuation : σ C ::= k | hP, Ci | hσ ◦ , Ci | let hx, ki be C in C | let hX, ki be C in C
Answer : R
A ::= P C | let hx, ki be C in A | let hX, ki be C in A
as follows.
x−1 = x
= µkσ .A−1
◦
(λkσ .A)−1
k−1
hP, Ci−1
hσ ◦ , Ci−1
(let hx, ki be C1 in C2 )−1
(let hX, ki be C1 in C2 )−1
=
=
=
=
=
[k][−]
C −1 [− P −1 ]
C −1 [− σ]
C1−1 [λx.µk.C2−1 [−]]
C1−1 [ΛX.µk.C2−1 [−]]
(P C)−1 = C −1 [P −1 ]
(let hx, ki be C in A)−1 = C −1 [λx.µk.A−1 ]
(let hX, ki be C in A)−1 = C −1 [ΛX.µk.A−1 ]
This can be considered as a “continuation-grabbing style transformation” in the sense of
Sabry [35]. It follows that for any ¬Γ◦ , ∆◦ ⊢ M : ¬σ ◦ there exists ¬Γ◦ , ∆◦ ⊢ P : ¬σ ◦
generated by this grammar such that P = M — it suffices to take the β-normal form [7].
Moreover we can routinely show that [[P −1 ]] = P . Thus the CPS translation enjoys fullness:
all terms are definable modulo the provable equality. This definability is important for
relating the parametricity principles for the source and target calculi.
8
MASAHITO HASEGAWA
3.3. A semantic explanation. Here is a short explanation of why this CPS translation
works, intended for readers with category theoretic background — on the “categories of
continuations” construction [16, 39, 36], and on fibrations for polymorphic type theories
[17]. As a response category C with a response object R induces a control category RC
with RC (X, Y ) = C(RX , RY ), a fibred response category with finite products and simple
coproducts (for existential quantifiers) induces a fibred control category with finite products
and simple products (for universal quantifiers). Let us write CΓ for the response category
over the type-context Γ. We assume that the weakening functor π ∗ : CΓ → CΓ×A has a left
adjoint ∃A : CΓ×A → CΓ subject to the Beck-Chevalley condition. Thus
CΓ×A (X, π ∗ (Y )) ≃ CΓ (∃A (X), Y )
We then have
RCΓ×A (π ∗ (X), Y ) =
≃
≃
≃
≃
=
∗
CΓ×A (Rπ (X) , RY )
π ∗ (X)
)
CΓ×A (Y, RR
X
∗
R
CΓ×A (Y, π (R ))
X
CΓ (∃A (Y ), RR )
CΓ (RX , R∃A (Y ) )
RCΓ (X, ∃A (Y ))
Hence π ∗ , regarded as the weakening functor from RCΓ to RCΓ×A , has a right adjoint given
by ∃A , which can be used for interpreting the universal quantifier. Our CPS transformation
is essentially a syntactic interpretation of this semantic construction.
4. CPS semantics with the parametric target calculus
4.1. Parametricity for the target calculus. As the target calculus can be seen as a
subset of λ2 (via the standard encoding of the conjunctions and existential types), we can
define the relational parametricity for the target calculus in the same way as for λ2, e.g.
logic for parametricity [31, 40], system R [1], or system P [5]. One may directly define the
parametricity principle (often called the simulation principle) for the existential type, see
for example [31].
In this paper we only consider the relations constructed from the graphs of termsin-context, identity, and σ ∗ ’s obtained by the following construction, which we shall call
“admissible relations”.
Among admissible relations, the most fundamental are the graph relations. Given a
term f (x) : τ2 with a free variable x : τ1 we define its graph relation hx ⊢ f (x)i : τ1 ↔ τ2
(hf i for short) by u hf i v iff f (u) = v.
Given a type τ whose free type variables are included in X1 , . . . , Xn and admissible
relations s1 : τ1 ↔ τ1′ , . . . , sn : τn ↔ τn′ , we define an admissible relation τ ∗ as follows.
• Xi∗ = si : τi ↔ τi′
• R∗ is the identity relation on the terms of type R
• (¬τ )∗ : ¬τ [τ1 /X1 , . . . ] ↔ ¬τ [τ1′ /X1 , . . . ] is the relation so that f (¬τ )∗ g iff x τ ∗ y
implies f x R∗ g y (hence f x = g y)
• (τ ∧τ ′ )∗ : (τ ∧τ ′ )[τ1 /X1 , . . . ] ↔ (τ ∧τ ′ )[τ1′ /X1 , . . . ] is the relation so that u (τ ∧τ ′ )∗ v
iff u = hx, x′ i, v = hy, y ′ i and x τ ∗ y, x′ τ ′∗ y ′
• (∃X.τ )∗ : ∃X.τ [τ1 /X1 , . . . ] ↔ ∃X.τ [τ1′ /X1 , . . . ] is the relation so that u (∃X.τ )∗ v iff
u = hτ ′ , xi, v = hτ ′′ , yi and x τ [r/X]∗ y for some admissible r : τ ′ ↔ τ ′′
RELATIONAL PARAMETRICITY AND CONTROL
9
In the last case, the relation τ [r/X]∗ : τ [τ ′ /X] ↔ τ [τ ′′ /X] is defined as τ ∗ with X ∗ = r.
One may further define admissible relations ¬r, r ∧ s and ∃X.r for admissible r, s, so that
(¬τ )∗ = ¬τ ∗ , (τ ∧ τ ′ )∗ = τ ∗ ∧ τ ′∗ and (∃X.τ )∗ = ∃X.τ ∗ hold.
Let idτ : τ ↔ τ be the identity relation on the terms of type τ . The relational parametricity asserts that, for any τ whose free type variables are included in X1 , . . . , Xn and
τ1 , . . . , τn , M : τ [τ1 /X1 , . . . , τn /Xn ] implies M τ ∗ M with si = idτi .
Its consistency follows immediately from that of the parametricity for λ2.
Proposition 4.1. As consequences of the parametricity, we can derive:
(1) ∃X.X gives a terminal object ⊤ with a unique inhabitant ∗, so that for any M : ⊤
we have M = ∗.
(2) ∃X.¬(τ ∧ X) ∧ X (which could be rewritten as ∃X.(X → ¬τ ) ∧ X) gives a final
coalgebra νX.¬τ of ΛX.¬τ where X only occurs negatively in τ .
(3) (as an instance of the last case) the isomorphism ∃X.¬(τ ∧ X) ∧ X ≃ ¬τ holds if X
does not occur freely in τ .
Their proofs are standard, cf. papers cited above [31, 1, 40, 5].
Below we will see the implications of these parametricity results on the target calculus.
We refer to the λµ2-theory induced by the CPS translation into this parametric target
calculus as λµ2P .
4.2. The falsity type. As a first example, let us consider the falsity type ⊥ = ∀X.X in
λµ2. We have
⊥◦ = (∀X.X)◦ = ∃X.X ≃ ⊤
and
(σ → ⊥)◦ = ¬(σ ◦ ∧ ⊥◦ ) ≃ ¬(σ ◦ ∧ ⊤) ≃ ¬σ ◦
Since ∃X.X is terminal (with a unique inhabitant ∗) in the parametric target calculus,
◦
we obtain [[µασ .M ]] = λασ .[[M ]]∗ and [[[β]M ]] = λu∃X.X .[[M ]]β, which coincide with Streicher’s translation. As a consequence, the following equations on the named terms and
µ-abstractions are all validated in λµ2P .
(µασ1 →σ2 .M ) N
(µα∀X.σ1 .M ) σ2
[α′ ](µασ .M )
[α⊥ ]M
=
=
=
=
µβ σ2 .M [[β](− N )/[α](−)]
µβ σ1 [σ2 /X] .M [[β](− σ2 )/[α](−)]
M [α′ /α]
M
Thus the type ⊥ serves as the falsity type as found in some formulation of the λµ-calculus.
In addition, we can show that (σ, Cσ : ((σ → ⊥) → ⊥) → σ) is an algebra of the doublenegation monad ((−) → ⊥) → ⊥ on the term model.
4.3. Initial algebra? A more substantial example is the “initial algebra” µX.F [X] =
∀X.(F [X] → X) → X, with X positive in F [X] (here we see an unfortunate clash of µ’s for
the name-binding and for the fixed-point on types, but this should not cause any serious
10
MASAHITO HASEGAWA
problem). We calculate:
(µX.F [X])◦ =
=
≃
≃
≃
=
≃
(∀X.(F [X] → X) → X)◦
∃X.¬(¬F [X]◦ ∧ X) ∧ X
νX.¬¬F [X]◦
¬¬F [X]◦ [νX.¬¬F [X]◦ /X]
¬¬F [X]◦ [(µX.F [X])◦ /X]
¬¬(F [µX.F [X]])◦
((F [µX.F [X]] → ⊥) → ⊥)◦
This suggests that µX.F [X] is isomorphic not to F [µX.F [X]] but to its double negation
(F [µX.F [X]] → ⊥) → ⊥. One might think that this contradicts the standard experience on
λ2 with parametricity, where we have an isomorphism in : F [µX.F [X]] → µX.F [X]. Since
λµ2 subsumes λ2, we have this in in λµ2 too; however, it should not be an isomorphism,
regarding the CPS interpretation above (otherwise it causes a degeneracy). The truth is
that, in λµ2P , the term
in♯ = λm.µα.m (λx.[α](in x)) : ((F [µX.F [X]] → ⊥) → ⊥) → µX.F [X]
is an isomorphism. It still is not an initial algebra of (F [−] → ⊥) → ⊥; we shall further
consider this issue later. For now, we shall emphasize that the parametricity principle for
λ2 should not be used for λµ2, at least without certain constraint — otherwise in would be
an isomorphism, hence a degeneracy follows (because we have (σ → ⊥) → ⊥ ≃ σ for every
σ).
4.4. Other impredicative encodings. Recall other impredicative encodings of logical
connectives:
⊤ = ∀X.X → X
σ1 ∧ σ2 = ∀X.(σ1 → σ2 → X) → X
σ1 ∨ σ2 = ∀X.(σ1 → X) → (σ2 → X) → X
∃X.σ = ∀Y.(∀X.(σ → Y )) → Y
Their CPS translations into the parametric target calculus satisfy:
⊤◦
(σ1 ∧ σ2 )◦
(σ1 ∨ σ2 )◦
(∃X.σ)◦
≃
≃
≃
≃
R
¬(¬σ1 ◦ ∧ ¬σ2◦ )
¬¬σ1◦ ∧ ¬¬σ2◦
¬∃X.¬σ ◦
As easily seen, these defined logical connectives in the source calculus do not obey the
standard universal properties as in the parametric models of λ2. In short, they are all
“double-negated”, hence amount to some classical encodings:
• σ1 ∧ σ2 is not a cartesian product of σ1 and σ2 , but isomorphic to (σ1 → σ2 → ⊥) →
⊥. It is possible to add cartesian product types σ1 ×σ2 to λµ2, but then we also need
to add coproduct types τ1 + τ2 to the target calculus, so that (σ1 × σ2 )◦ = σ1◦ + σ2◦
and σ1 ∧ σ2 ≃ ¬¬(σ1 × σ2 ).
• ⊤ is not a terminal object, but isomorphic to ⊥ → ⊥. We can add a terminal object
1 to λµ2 and an initial object 0 to the target, so that 1◦ = 0 and ⊤ ≃ ¬¬1.
• σ1 ∨ σ2 is not a coproduct of σ1 and σ2 , but isomorphic to (σ1 → ⊥) → (σ2 → ⊥) →
⊥. If there is a coproduct σ1 + σ2 , then it should follow that σ ∨ τ ≃ ¬¬(σ + τ ).
On the other hand, it is not possible to enrich λµ2 with an initial object without
a degeneracy, cf. Selinger’s note on control categories [37]. Alternatively we might
RELATIONAL PARAMETRICITY AND CONTROL
11
add the “classical disjunction types” σ1 ℘σ2 [36] with (σ1 ℘σ2 )◦ = σ1◦ ∧ σ2◦ — hence
σ1 → σ2 ≃ ¬σ1 ℘σ2 and σ1 ∨ σ2 ≃ ¬¬σ1 ℘¬¬σ2 . We note that ⊥ = ∀X.X serves as
the unit of this classical disjunction.
• ∃X.σ does not work as the existential type; it is isomorphic to ¬∀X.¬σ.
4.5. Answer-type polymorphism. Note that the answer type R has been considered
just as a constant with no specific property. In fact we could have used any type for R —
Everything is defined polymorphically regarding R. Thus we can apply the “answer-type
polymorphism” principle (cf. [43]): in particular, a closed term of type σ in λµ2 can be
considered to be sent to a λ2-term of type ∀R.¬σ ◦ . This way of reasoning goes behind the
parametricity principle for our target calculus, but it is justified by the parametricity of λ2.
For instance, consider the type ⊤ = ∀X.X → X of λµ2. We have
∀R.¬⊤◦ ≃ ∀R.¬R
≃ ∀R.R → R
≃ 1
in λ2 with parametricity. This means that, although ⊤ is not a terminal object in λµ2, it
has a unique closed inhabitant. Similarly, we have ∀R.¬⊥◦ ≃ ∀R.R ≃ 0, thus we see that
there is no closed inhabitant of ⊥ in λµ2.
However, such reasonings based on the answer-type polymorphism become much harder
for more complicated types. The force of answer-type polymorphism in this setting seems
still not very obvious.
5. Focal parametricity
We have seen that the CPS semantics with respect to the target calculus with relational parametricity induces a reasonable equational theory λµ2P . However, here the
parametricity is used rather indirectly, via the CPS translation. We now consider a notion
of parametricity which is directly available within λµ2.
5.1. CPS translating relations. The key of formulating the relational parametricity is
the use of graph relations of terms (considered as representing a functional relation): without
graph relations, relational parametricity reduces to just the basic lemma of the (secondorder) logical relations. On the other hand, it does not have to allow all terms to be used
for constructing relations. In fact, in linear parametricity [30] only linear (or strict) maps
are allowed to be used for constructing graph relations, and this choice allows a weaker
notion of parametricity which can accommodate recursion. Naturally, we are led to look
for a characterisation of λµ-terms which can be used for graph relations without breaking
the soundness with respect to the CPS semantics into the parametric target calculus.
Now suppose that we are allowed to use the graph relation hf i : σ1 ↔ σ2 of a term
x : σ1 ⊢ f (x) : σ2 . To ensure the soundness of the use of this graph relation, we shall consider
the CPS translation of such relations. For instance, we hope that hf i will be sent to a relation
between types σ2◦ and σ1◦ in the target calculus. However, since x : ¬σ1◦ ⊢ [[f (x)]] : ¬σ2◦ , we
have some relation σ2◦ ↔ σ1◦ only when [[f (x)]] = λk.x (g (k)) for some k : σ2◦ ⊢ g(k) : σ1◦ in
the target calculus. If there is such g, we can complete the translation of the relations and
reduce the parametricity principle on λµ2 to the parametricity on the target calculus.
12
MASAHITO HASEGAWA
Fortunately, there is a way to characterise such “translatable” f ’s in the λµ-calculus
without performing the CPS-translation (modulo a technical assumption on the CPS-target,
known as “equalising requirement” [22]). It is the notion of “focus”, which we now recall
below.
5.2. Focus.
Definition 5.1. A λµ2-term M : σ1 → σ2 is called focal if it is an algebra morphism from
(σ1 , Cσ1 ) to (σ2 , Cσ2 ), i.e. the following diagram commutes.
(σ1 → ⊥) → ⊥
Cσ1
(M →⊥)→⊥
-
Cσ2
?
σ1
(σ2 → ⊥) → ⊥
?
M
- σ2
That is:
M (µασ1 .k (λxσ1 .[α]x)) = µβ σ2 .k (λxσ1 .[β](M x)) : σ2
holds for any k : (σ1 → ⊥) → ⊥.2
In any λµ2-theory, focal terms compose, and the identity λxσ .x is obviously focal. So,
the (equivalence classes of) focal maps form a category. Hereafter we shall call it the focus
of the λµ2-theory.
While this characterisation of focal maps is concise and closely follows the semantic
considerations in [36, 19], there is a subtle problem; the βη-axioms of λµ2 are too weak to
establish the focality of some important terms. This is because we have used the polymorphic feature of λµ2 for expressing Cσ — it involves the falsity type ⊥ = ∀X.X, but the
axioms of λµ2 do not guarantee that ⊥ does work properly. If there were not sufficiently
many focal maps, the parametricity principle restricted on focal maps would be useless.
To see this issue more clearly, we shall look at another “classical” combinator (the
Peirce law)
Pσ1 ,σ2 = λm.µασ1 .[α](m (λxσ1 .µβ σ2 .[α]x)) : ((σ1 → σ2 ) → σ1 ) → σ1
which does not make use of polymorphism, and the “abort” map (Ex Falso Quodlibet)
Aσ = λx⊥ .x σ : ⊥ → σ
which is defined without the classical feature. It is well known that the double-negation
elimination is as expressible as the Peirce law together with Ex Falso Quodlibet, see e.g.
[2]. This is also the case at the level of (uniformity of) proofs. Let us say that M : σ1 → σ2
is repeatable if
(M →σ3 )→M
- (σ2 → σ3 ) → σ2
(σ1 → σ3 ) → σ1
Pσ1 ,σ3
Pσ2 ,σ3
?
σ1
?
M
- σ2
2In [16], a focal map from σ to ⊥ is called a “C-term of type σ ”. C-terms of type σ with a free name
1
1
1
of σ2 correspond to focal maps from σ1 to σ2 , thus these notions (and the associated constructions of the
CPS target categories via C-terms (C-maps) [16] and via focus [36]) are essentially the same.
RELATIONAL PARAMETRICITY AND CONTROL
13
commutes for each σ3 ; and discardable if
Aσ 1
σ1
⊥
M
Aσ
@ 2
R
@
- σ2
commutes.
Proposition 5.2. In a λµ2-theory, M : σ1 → σ2 is focal if and only if it is both repeatable
and discardable.
We note that the corresponding result in the call-by-value setting has been observed by
Führmann [8] as the characterisation of algebraic values as repeatable discardable expressions. Here we follow his terminology.
This reformulation allows us to see that only the second diagram of A’s involves the
polymorphically defined ⊥ and needs to be justified by additional conditions.3 On the other
hand, the first diagram of P ’s is not problematic, as it does not make use of polymorphism
at all.
5.3. Additional axioms. To this end, we add more axioms to λµ2 before thinking about
parametricity. They are
(1) λxσ1 →σ2 .x N : (σ1 → σ2 ) → σ2 is discardable for any N : σ1
(2) λx∀X.σ .x σ1 : ∀X.σ → σ[σ1 /X] is discardable for any σ and σ1
(3) λxσ .[α]x : σ → ⊥ is discardable for any α : σ
which are equivalent to asking
(1) M (σ1 → σ2 ) N = M σ2 for any M : ⊥ and N : σ1
(2) M (∀X.σ) σ1 = M (σ[σ1 /X]) for M : ⊥
(3) [α](M σ) = M for M : ⊥ and α : σ (α 6∈ F N (M ))
and also equivalent to
(1) (µασ1 →σ2 .M ) N = µβ σ2 .M [[β](− N )/[α](−)]
(2) (µα∀X.σ1 .M ) σ2 = µβ σ1 [σ2 /X] .M [[β](− σ2 )/[α](−)]
(3) [α′ ](µασ .M ) = M [α′ /α]
Note that λµ2P discussed in Section 4 satisfies these conditions. Also we shall note that
λxσ1 →σ2 .x N , λx∀X.σ .x σ1 , λx⊥ .[α](x σ) are all repeatable in λµ2. Together with these
additional axioms, they become focal. (Alternatively, we could have ⊥ as a type constant
and assume the standard axiomatization of λµ-calculus with the falsity type [16, 36] — in
that case C is defined without polymorphism, and this problem disappears.)
Below we develop the focal parametricity principle on top of λµ2 with these additional
axioms.
3This problem was overlooked in the preliminary version of this paper [13] where it was wrongly assumed
that repeatability alone would imply focality.
14
MASAHITO HASEGAWA
5.4. A parametricity principle for λµ2. Given a focal f : σ1 → σ2 we define its graph
relation hf i : σ1 ↔ σ2 by u hf i v iff f u = v. Also, let idσ : σ ↔ σ be the identity relation
on the terms of type σ. In this paper we only consider the relations given by the graphs
of focal maps, identity, and σ ∗ ’s obtained by the following construction, which we shall call
“focal relations”.
Given a type σ whose free type variables are included in X1 , . . . , Xn and focal relations
s1 : σ1 ↔ σ1′ , . . . , sn : σn ↔ σn′ , we define a focal relation σ ∗ as follows.
• Xi∗ = si : σi ↔ σi′
• (σ → σ ′ )∗ : (σ → σ ′ )[σ1 /X1 , . . . ] ↔ (σ → σ ′ )[σ1′ /X1 , . . . ] is the relation so that f (σ →
σ ′ )∗ g iff x σ ∗ y implies (f x) σ ′∗ (g y)
• (∀X.σ)∗ : ∀X.σ[σ1 /X1 , . . . ] ↔ ∀X.σ[σ1′ /X1 , . . . ] is the relation so that u (∀X.σ)∗ v
iff (u σ ′ ) σ[r/X]∗ (v σ ′′ ) holds for any focal relation r : σ ′ ↔ σ ′′
The focal relational parametricity asserts that, for any σ whose free type variables are
included in X1 , . . . , Xn , M : σ[σ1 /X1 , . . . , σn /Xn ] implies M σ ∗ M with si = idσi .
Thus the only departure from the standard parametricity principle is the condition that
the graph relation construction is allowed only on focal maps. Note that this restriction
is necessary; if we apply parametricity to polymorphic terms ΛX.CX or ΛX.PX,σ , we will
get the naturality diagrams above for any term which is allowed to be used for the graph
relation construction.
5.5. On consistency and soundness. The consistency of focal parametricity (in the sense
that the equational theory of λµ2 with focal parametricity is not trivial) follows from the
fact that there are non-trivial parametric models of λ2 in which there is an object R so
τ
that the continuation monad T τ = RR satisfies the “equalising requirement” [22], i.e. each
component ητ : τ → T τ of its unit is an equaliser of ηT τ and T ητ . (Here we employ the
syntax of the CPS target calculus as an internal language for such models, where the CPS
translation is considered to give a semantic interpretation.) In such models, for any focal
term f : σ1 → σ2 , there exists a unique y : σ2◦ ⊢ g(y) : σ1◦ such that [[f x]] = λy.x (g(y)) (cf.
[36]).
Using this fact, given a focal relation r : σ1 ↔ σ2 , we construct an admissible relation
r ◦ : σ2◦ ↔ σ1◦ as follows. For a graph relation hf i : σ1 ↔ σ2 , we let hf i◦ = hgi : σ2◦ ↔ σ1◦
where g is the unique map as given above. For σ ∗ , σ ∗◦ is defined by straightforward
induction: (σ → σ ′ )∗◦ = ¬σ ∗◦ ∧ σ ′∗◦ , (∀X.σ)∗◦ = ∃X.σ ∗◦ (where the parameter relations si
are replaced by s◦i ).
Theorem 5.3. In such a model, given a focal relation r : σ1 ↔ σ2 , M r N implies
[[N ]] ¬r ◦ [[M ]].
Theorem 5.4 (consistency). Focal parametricity is consistent.
We do not know if the term model of the parametric target calculus satisfies the equalising requirement — if so, by the definability result, the parametricity on the target and the
focal parametricity on λµ2 should agree. Alternatively we should consider a refined target
calculus with a construct ensuring the equalising requirement, as detailed in Taylor’s work
on sober space (“a lambda calculus for sobriety” [41]). For now, we only know that one
direction is true (thanks to the definability).
Theorem 5.5. An equality derivable in λµ2P is also derivable in λµ2 with focal parametricity.
RELATIONAL PARAMETRICITY AND CONTROL
15
6. Examples
We show that certain impredicative encodings in λµ2 satisfy universal properties with
respect to the focus using the focal parametricity principle.
6.1. Focal decomposition. We start with a remark on the following “focal decomposition”
[36] (analogous to the linear decomposition σ1 → σ2 =!σ1 ⊸ σ2 [10]): there is a bijective
correspondence between terms of σ1 → σ2 and focal terms of ¬¬σ1 → σ2 natural in σ1 and
focal σ2 .
f : ¬¬σ1 → σ2 focal
f ♭ = f ◦ ησ1 = λxσ1 .f (λk.k x) : σ1 → σ2
g♯
= Cσ2 ◦ ¬¬g =
g : σ1 → σ2
σ
λm.µβ 2 .m (λxσ1 .[β](g x))
: ¬¬σ1 → σ2 focal
♯
♭
Proposition 6.1. g♯ = g for any g : σ1 → σ2 , while f ♭ = f holds for f : ¬¬σ1 → σ2 if
and only if f is focal.
6.2. Falsity as a focally initial object. Now we shall proceed to reason about impredicative encodings in λµ2. The first example is the falsity ⊥ = ∀X.X.
First, we note that Aσ = λx⊥ .x σ : ⊥ → σ is focal. The parametricity on ⊥ says x ⊥∗ x
for any x : ⊥. Since hAσ i : ⊥ ↔ σ, we have x ⊥ hAσ i x σ, i.e. Aσ (x ⊥) = x ⊥ σ = x σ. By
extensionality we get x = x ⊥ for x : ⊥.
Now suppose that g : ⊥ → σ is focal. Again by the parametricity on ⊥ we know x ⊥∗ x
for any x : ⊥, hence x ⊥ hgi x σ. Thus g (x ⊥) = x σ; but x = x ⊥, so we have g x = x σ,
hence g = λx⊥ .x σ = Aσ .
So we conclude that Aσ is the unique focal map from ⊥ to σ. This means that ⊥ is
initial in the focus.
6.3. Focally initial algebra. As in λ2, there is a fairly standard encoding
µX.F [X] = ∀X.(F [X] → X) → X
foldσ = λaF [σ]→σ .λxµX.F [X] .x σ a : (F [σ] → σ) → µX.F [X] → σ
in = λy.ΛX.λkF [X]→X .k (F [foldX k] y) : F [µX.F [X]] → µX.F [X]
for which the following diagram commutes (just by β-axioms).
F [µX.F [X]]
in -
F [fold a]
µX.F [X]
fold a
?
F [σ]
a
?
- σ
Therefore in is a weak initial F -algebra. However, as we noted before, in is not an initial
F -algebra — in fact it is not even an isomorphism. By applying the focal decomposition
16
MASAHITO HASEGAWA
above, we obtain the commutative diagram
¬¬F [µX.F [X]]
¬¬F [fold a♭ ]
in♯-
µX.F [X]
fold a♭
?
¬¬F [σ]
?
- σ
a
for any focal a : ¬¬F [σ] → σ. We show that fold a♭ is the unique focal map making this
diagram commute, thus in♯ is an initial ¬¬F [−]-algebra in the focus.
We sketch a proof which is fairly analogous to that for the corresponding result in
parametric λ2 as given in [1]. First, from the parametricity on µX.F [X] we obtain that
F [σ1 ]
a-
F [h]
h
?
F [σ2 ]
b
µX.F [X]
σ1
?
- σ2
implies
folda
||
σ1
h
?
µX.F [X]
?
- σ2
fold b
whenever h is focal. We also have M (µX.F [X]) in = M for any M : µX.F [X] as a corollary
(thanks to extensionality). By combining these observations, now we have the desired result.
That is, if h : µX.F [X] → σ is focal and satisfies h ◦ in♯ = a ◦ ¬¬F [h], then
foldσ a♭ x = h (foldµX.F [X] in x)
= h (x (µX.F [X]) in)
= hx
so by extensionality we conclude foldσ a♭ = h. This also implies that in♯ is an isomorphism,
with the inverse given by fold¬¬µX.F [X] (¬¬F [in]).
As a special case, by letting F be a constant functor, we obtain isomorphisms between
(σ → ⊥) → ⊥ and ∀X.(σ → X) → X where X is not free in σ. With some further
calculation we see that in♯ = λm.ΛX.λkσ→X .µαX .m (λxσ .[α](k x)) is the inverse of λn.n ⊥ :
(∀X.(σ → X) → X) → (σ → ⊥) → ⊥. We will see more about this isomorphism in Section
7.
6.4. The type of Church numerals. We conclude this section by a remark on the type
of Church numerals N = ∀X.X → (X → X) → X. Recall that, in λ2 with parametricity,
N is an initial algebra of ∀X.X → (− → X) → X ≃ 1 + (−), i.e. a natural numbers object,
whose closed inhabitants are equal to the Church numerals Sn O which can be given by, as
usual,
O = ΛX.λxX f X→X .x
: N
S = λnN .ΛX.λxX f X→X .f (n X x f ) : N → N
It is no longer true in λµ2, as observed by Parigot, as there are closed inhabitants which
are not equal to Church numerals, e.g.
µαN .[α](S (µβ N .[α]O)) = ΛX.λxX f X→X .µαX .[α](f (µβ X .[α]x)) : N
In contrast, N in λµ2 with focal parametricity is a focally initial algebra of ∀X.X → (− →
X) → X ≃ ⊥ → (− → ⊥) → ⊥; this can be shown in the same way as the case of focally
RELATIONAL PARAMETRICITY AND CONTROL
17
initial algebras. Spelling this out, we have a focal map in : (⊥ → (N → ⊥) → ⊥) → N,
and for any focal g : (⊥ → (σ → ⊥) → ⊥) → σ there exists a unique focal foldσ g : N → σ
making the following diagram commute.
⊥ → (N → ⊥) → ⊥
in -
N
⊥→(foldσ g→⊥)→⊥
foldσ g
?
⊥ → (σ → ⊥) → ⊥
g
?
- σ
To see this, it is useful to observe the following bijective correspondence (a variant of the
focal decomposition): given focal g : (⊥ → (σ → ⊥) → ⊥) → σ we have
go = g (λx⊥ kσ→⊥ .x)
: σ
gs = λy σ .g (λx⊥ kσ→⊥ .k y) : σ → σ
and conversely, for a : σ and f : σ → σ we have a focal map
ϕa,f = λm⊥→(σ→⊥)→⊥ .µασ .m ([α]a) (λy σ .[α](f y)) : (⊥ → (σ → ⊥) → ⊥) → σ
It follows that (ϕa,f )o = a and (ϕa,f )s = f hold for any a and f , while ϕgo ,gs = g for any
focal g. Now we define
foldA g = λnN .n A go gs : N → A
in
= ϕO,S
: (⊥ → (N → ⊥) → ⊥) → N
It then follows that the diagram above commutes — and the focal parametricity implies
that foldA g is the unique such focal map.
7. A general characterisation
So far, we concentrated on the relational parametricity for λµ2. One may feel that
this story is very specific to the case of λµ2, or of the first-class continuations, and is not
immediately applicable to other computational effects.
In this section we describe an alternative characterisation of the focus, which makes
sense in any extension of λ2. Namely, we show that, any λ2-theory is equipped with a monad
L, such that each type is equipped with an algebra structure — and then see that, in the
case of λµ2 with focal parametricity, this monad L is isomorphic to the double-negation
(continuation) monad, and focal maps are precisely the algebra maps of the monad L.
This suggests a natural generalisation of this work to a theory of parametricity for general
computational effects.
7.1. A monad on λ2. Let Lσ = ∀X.(σ → X) → X (with no free X in σ), and define
ησ
= λxσ .ΛX.λkσ→X .k x
: σ → Lσ
2
µσ
= λz L σ .ΛX.λkσ→X .z X (λy Lσ .y X k) : L2 σ → Lσ
L(f ) = λy Lσ1 .ΛX.λhσ2 →X .y X (h ◦ f )
: Lσ1 → Lσ2 (f : σ1 → σ2 )
Proposition 7.1. On the term model of any λ2-theory, (L, η, µ) forms a monad.
18
MASAHITO HASEGAWA
One might think that this is trivial as Lσ is isomorphic to σ when we assume the
standard parametricity. This is not always the case however, as we have already seen,
Lσ ≃ (σ → ⊥) → ⊥ in the focally parametric λµ2.
Proposition 7.2. ασ = λy Lσ .y σ (λxσ .x) : Lσ → σ is an algebra of the monad (L, η, µ).
Thus each σ is canonically equipped with an algebra structure ασ . Again one may think
that this is trivial, as under the standard parametricity ασ is just an isomorphism with ησ
being an inverse. However, again it is not the case in a non-trivial λµ2-theory.
Now we define the notion of linear maps in terms of the monad L and the canonical
algebras ασ — this is close to what we do in (axiomatic) domain theory for characterising
the strict maps, and also in control categories for characterising the focal maps.
Definition 7.3. f : σ1 → σ2 is linear when it is an algebra morphism from ασ1 to ασ2 , i.e.
f ◦ ασ1 = ασ2 ◦ L(f ) holds.
Lσ1
L(f )
- Lσ2
ασ1
ασ2
?
σ1
That is, f is linear when
f
?
- σ2
f (M σ1 (λxσ1 .x)) = M σ2 f
holds for any M : Lσ1 . We may write f : σ1 ⊸ σ2 for a linear f : σ1 → σ2 . Under the
standard parametricity every f : σ1 → σ2 is linear, while for focal parametricity on λµ2
we have that linear maps are precisely the focal maps (see below). In passing, we note the
following interesting observation.
Proposition 7.4. In a λ2-theory, the following conditions are equivalent.
(1) algebras on σ1 → σ2 and ∀X.σ are determined in the pointwise manner, i.e.
ασ1 →σ2 = λf L(σ1 →σ2 ) .λxσ1 .ασ2 (L(λgσ1 →σ2 .g x) f )
α∀X.σ = λxL(∀X.σ) .ΛX.ασ (L(λy ∀X.σ .y X) x)
(2) λxσ1 →σ2 .x N is linear for any N : σ1 , and λx∀X.σ .x σ1 is linear for any σ and σ1 .
Note that they are very close to the “additional axioms” for λµ2 discussed in Section 5.
Also note that, if a λ2-theory satisfies one of these conditions, µσ and αLσ agree for every
σ. and we have a “linear decomposition” correspondence between the maps of σ1 → σ2 and
the linear maps of Lσ1 → σ2 .
7.2. Focal maps as algebra maps. Now we shall consider the double-negation monad
¬¬σ = (σ → ⊥) → ⊥ on λµ2 with focal parametricity.
Proposition 7.5. In a focally parametric λµ2-theory, Cσ : ¬¬σ → σ is an algebra of the
double-negation monad.
Corollary 7.6. f : σ1 → σ2 is focal if and only if it is an algebra map from Cσ1 to Cσ2 .
Proposition 7.7. The monad (L, η, µ) is isomorphic to the double negation monad in the
≃
focally parametric λµ2, with λxLσ .x ⊥ : Lσ → ¬¬σ.
RELATIONAL PARAMETRICITY AND CONTROL
19
Proposition 7.8. The following diagram commutes in a focally parametric λµ2-theory:
λxLσ .x ⊥-
Lσ
@
@
R
@
ασ
¬¬σ
Cσ
σ
Corollary 7.9. f : σ1 → σ2 is linear if and only if it is focal.
Thus a focal map in a focally parametric λµ2-theory can be characterised just in terms
of the monad L which is defined for arbitrary λ2-theory.
We believe that the monad L deserves much attention. It has been considered trivial,
but now we know that it does characterise an essential notion (focus) in the case of relational
parametricity under the presence of control feature. In fact, the story does not end here;
under the presence of non-termination or recursion, L behaves like a lifting monad — indeed
it is a lifting in the theory of linear parametricity, because
Lσ = ∀X.(σ → X) → X = ∀X.!(!σ ⊸ X) ⊸ X ≃ !σ
where the last isomorphism follows from the fact that ∀X.!(F [X] ⊸ X) ⊸ X gives an
initial algebra of F , cf. [4].
These observations suggest that there exists a general framework similar to (axiomatic
or synthetic) domain theory where the lifting monad can be replaced by any strong monad
— a continuation monad for example — on which a theory of parametricity for general
computational effects can be built. Recently, Alex Simpson has made a progress in this
direction, by developing a two-level polymorphic type theory (for interpreting types and
algebras of a monad) in a constructive universe [38]. His work fits very well with the case
of linear parametricity for recursion; it is plausible that it also explains the case of focal
parametricity for first-class control.
8. Conclusion and future work
We have studied the relational parametricity for λµ2, first by considering the CPS
translation into a parametric fragment of λ2, and then by directly giving a constrained
parametricity for λµ2. The later, which we call “focal parametricity”, seems to be a natural
parametricity principle under the presence of first-class controls — in the same sense that
linear parametricity works under the presence of recursion and non-termination.
There remain many things to be addressed in future. In the previous section, we already
discussed a research direction towards a relational parametricity for general effects. Below
we shall briefly mention some future work more closely related to the main development of
this paper.
Firstly, we are yet to complete the precise comparison between focal parametricity on
λµ2 and the parametricity on the CPS target calculus. This involves some subtle interaction
between parametricity and a technical condition (equalising requirement).
Secondly, we should study focal parametricity for extensions of λµ2. As we observed,
λµ2 with focal parametricity does not have many popular datatypes, e.g. cartesian products,
and classical disjunction types which however can be added with no problem. Adding
general initial algebras is problematic (having an initial object already means inconsistency),
but it might be safe to add certain carefully chosen instances. On the other hand, final
coalgebras seem less problematic, though a generic account for them in λµ2 is still missing.
20
MASAHITO HASEGAWA
Perhaps we also need to consider the CPS translation of such datatypes (cf. [3]) in a
systematic way.
An interesting topic we have not discussed in this paper is the Filinski-Selinger duality
[6, 36] between call-by-name and call-by-value calculi with control primitives. In fact it is
straightforward to consider its second-order extension: in short, universal quantifiers in callby-name (as studied in this paper) amount to existential quantifiers in call-by-value. We
are not sure if the call-by-value calculus with existential quantifiers itself is of some interest.
However, it can be a good starting point to understand the call-by-value parametric polymorphism (possibly with computational effects), from both syntactic and semantic aspects.
In particular, it should provide new insights on the famous difficulty of accommodating
first-class continuations in ML type system [11].
Finally, we also should consider if there is a better (ideally semantic) formulation of
focal relations. In this paper we only consider those coming from focal maps, but it seems
natural to regard a subalgebra (of the double-negation monad) of Cσ1 ×σ2 as a focal relation
between σ1 and σ2 , where we assume the presence of cartesian product σ1 × σ2 . This looks
very closely related to Pitts’ ⊤⊤-closed relations for λ2 with recursion [28].
Acknowledgement
I thank Ken-etsu Fujita for discussions and cooperations related to this work. I am also
grateful to Ryu Hasegawa, Paul-André Melliès and Alex Simpson for comments, discussions
and encouragements.
References
[1] M. Abadi, L. Cardelli, and P.-L. Curien. Formal parametric polymorphism. Theoret. Comput. Sci.
121:9–58, 1993.
[2] Z.M. Ariola and H. Herbelin. Minimal classical logic and control operators. In Proc. Automata, Languages and Programming, Springer Lecture Notes in Comput. Sci. 2719, pp.871–885, 2003.
[3] G. Barthe and T. Uustalu. CPS translating inductive and coinductive types. In Proc. Partial Evaluation
and Semantics-Based Program Manipulation, pp.131–142, 2002.
[4] G.M. Bierman, A.M. Pitts, and C.V. Russo. Operational properties of Lily, a polymorphic linear lambda
calculus with recursion. In Proc. Higher Order Operational Techniques in Semantics, Electronic Notes
in Theoretical Computer Science 41, 2000.
[5] B.P. Dunphy. Parametricity as a Notion of Uniformity in Reflexive Graphs. PhD thesis, University of
Illinois, 2002.
[6] A. Filinski. Declarative continuations: an investigation of duality in programming language semantics.
In Proc. Category Theory and Computer Science, Springer Lecture Notes in Comput. Sci. 389, pp.224–
249, 1989.
[7] K. Fujita. Galois embedding from polymorphic types into existential types. In Proc. Typed Lambda
Calculi and Applications, Springer Lecture Notes in Comput. Sci. 3461, pp.194–208, 2005.
[8] C. Führmann. Varieties of effects. In Proc. Foundations of Software Science and Computation Structures,
Springer Lecture Notes in Comput. Sci. 2303, pp.144–159, 2002.
[9] J.-Y. Girard. Interprétation fonctionnelle et élimination des coupures de l’arithmétique d’ordre
supérieur. Thèse d’Etat, Université Paris VII, 1972.
[10] J.-Y. Girard. Linear logic. Theoret. Comp. Sci. 50:1–102, 1987.
[11] B. Harper, B.F. Duba and D. MacQueen. Typing first-class continuations in ML. J. Funct. Programming
3(4):465–484.
[12] M. Hasegawa. Semantics of linear continuation-passing in call-by-name. In Proc. Functional and Logic
Programming, Springer Lecture Notes in Comput. Sci. 2998, pp.229–243, 2004.
RELATIONAL PARAMETRICITY AND CONTROL
21
[13] M. Hasegawa, Relational parametricity and control (extended abstract). In Proc. Logic in Computer
Science, pp.72–81, 2005.
[14] M. Hasegawa and Y. Kakutani. Axioms for recursion in call-by-value. Higher-Order and Symbolic Comput. 15(2/3):235–264, 2002.
[15] R. Hasegawa. Categorical data types in parametric polymorphism. Math. Struct. Comp. Sci. 4(1):71–
109, 1994.
[16] M. Hofmann and T. Streicher. Completeness of continuation models for λµ-calculus. Inf. Comput.
179(2):332–355, 2002.
[17] B. Jacobs. Categorical Logic and Type Theory. Elsevier, 1999.
[18] Y. Kakutani. Duality between call-by-name recursion and call-by-value iteration. In Proc. Computer
Science Logic, Springer Lecture Notes in Comput. Sci. 2471, pp.506–521, 2002.
[19] Y. Kakutani and M. Hasegawa. Parameterizations and fixed-point operators on control categories.
Fundam. Inform. 65(1/2):153–172, 2005.
[20] J. Lambek and P. Scott (1986) Introduction to Higher-order Categorical Logic. Cambridge University
Press.
[21] R. Matthes. Parigot’s second order λµ-calculus and inductive types. In Proc. Typed Lambda Calculi and
Applications, Springer Lecture Notes in Comput. Sci. 2044, pp.329–343, 2001.
[22] E. Moggi. Notions of computation and monads. Inf. Comput. 93(1):55–92, 1991.
[23] K. Nakazawa and M. Tatsuta. Strong normalization proof with CPS-translation for second order classical
natural deduction. J. Symb. Log. 68(3):851–859, 2003. Corrigendum: J. Symb. Log. 68(4):1415–1416,
2003.
[24] C.-H.L. Ong. A semantic view of classical proofs: type-theoretic, categorical, and denotational characterizations (preliminary extended abstract). In Proc. Logic in Computer Science, pp.230–241, 1996.
[25] C.-H.L. Ong and C.A. Stewart. A Curry-Howard foundation for functional computation with control.
In Proc. Principles of Programming Languages, pp.215–227, 1997.
[26] M. Parigot. λµ-calculus: an algorithmic interpretation of classical natural deduction. In Proc. Logic
Programming and Automated Reasoning, Springer Lecture Notes in Comput. Sci. 624, pp.190–201,
1992.
[27] M. Parigot. Proofs of strong normalisation for second order classical natural deduction. J. Symb. Log.
62(4):1461–1479, 1997.
[28] A. Pitts, Parametric polymorphism and operational equivalence. Math. Struct. Comp. Sci. 10:321–359,
2000.
[29] G.D. Plotkin. Call-by-name, call-by-value, and the λ-calculus. Theoret. Comput. Sci. 1(1):125–159,
1975.
[30] G.D. Plotkin. Type theory and recursion (extended abstract). In Proc. Logic in Computer Science,
pp.374, 1993.
[31] G.D. Plotkin and M. Abadi. A logic for parametric polymorphism. In Proc. Typed Lambda Calculi and
Applications, Springer Lecture Notes in Comput. Sci. 664, pp.361–375, 1993.
[32] A.J. Power and E. Robinson. Premonoidal categories and notions of computation. Math. Struct. Comp.
Sci. 7(5):453–468, 1997.
[33] J.C. Reynolds. Towards a theory of type structure. In Proc. Colloque sur la Programmation, Springer
Lecture Notes in Comput. Sci. 19, pp.408–425, 1974.
[34] J.C. Reynolds. Types, abstraction and parametric polymorphism. In Proc. IFIP 9th World Computer
Congress, Information Processing 83, pp.513–523, 1983.
[35] A. Sabry. Note on axiomatizing the semantics of control operators. Tech. Rep. CIS-TR-96-03, University
of Oregon, 1996.
[36] P. Selinger. Control categories and duality: on the categorical semantics of the lambda-mu calculus.
Math. Struct. Comp. Sci. 11(2):207–260, 2001.
[37] P. Selinger. Some remarks on control categories. Manuscript, 2003.
[38] A. Simpson. Relational parametricity for computational effects. Manuscript, 2006.
[39] T. Streicher and B. Reus. Classical logic, continuation semantics and abstract machines. J. Funct.
Program. 8(6):543–572, 1998.
[40] I. Takeuti. An axiomatic system of parametricity. Fundam. Inform. 33(4):397–432, 1998.
[41] P. Taylor. Sober spaces and continuations. Theory and Applications of Categories 10(12):248–300, 2002.
22
MASAHITO HASEGAWA
[42] H. Thielecke. Categorical Structure of Continuation Passing Style. PhD thesis, University of Edinburgh,
1997.
[43] H. Thielecke. Answer type polymorphism in call-by-name continuation passing. In Proc. European Symposium on Programming, Springer Lecture Notes in Comput. Sci. 2986, pp.279–293, 2004.
[44] P. Wadler. Theorems for free! In Proc. Functional Programming Languages and Computer Architecture,
pp.347–359, 1989.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
| 6cs.PL
|
An Analysis on the Influence of Network Topologies on
Local and Global Dynamics of Metapopulation Systems
Daniela Besozzia Paolo Cazzanigab
Dario Pescinib Giancarlo Maurib
a Università degli Studi di Milano
Dipartimento di Informatica e Comunicazione
Via Comelico 39, 20135 Milano, Italy
[email protected]
b Università degli Studi di Milano-Bicocca
Dipartimento di Informatica, Sistemistica e Comunicazione
Viale Sarca 336, 20126 Milano, Italy
cazzaniga/pescini/[email protected]
Metapopulations are models of ecological systems, describing the interactions and the behavior of
populations that live in fragmented habitats. In this paper, we present a model of metapopulations
based on the multivolume simulation algorithm tau-DPP, a stochastic class of membrane systems,
that we utilize to investigate the influence that different habitat topologies can have on the local and
global dynamics of metapopulations. In particular, we focus our analysis on the migration rate of
individuals among adjacent patches, and on their capability of colonizing the empty patches in the
habitat. We compare the simulation results obtained for each habitat topology, and conclude the
paper with some proposals for other research issues concerning metapopulations.
1
Introduction
The field of metapopulations ecology deals with the study of spatial systems describing the behavior of
interacting populations that live in fragmented habitats [17]. The purpose of these models is to understand
how the local and global dynamics of metapopulation systems, usually balanced between local extinctions and new colonizations of unoccupied patches, depend on the spatial arrangement of the habitat.
Consequently, relevant insights into related fields of ecological research, such as evolutionary ecology or
conservation and landscape management, can be achieved. Indeed, the topology of fragmented habitats
potentially holds relevant implications for the persistence of populations, and their robustness against
natural or anthropogenic disturbance [36].
Recently, in addition to ever increasing applications of graph-based methods for the analysis of complex networks in cell biology [1, 2], graph theory has also been applied to the study of metapopulations
systems. In graph models of metapopulations, nodes are used to represent habitat patches, and graph
edges are used to denote some functional connections between patches (typically related to the dispersal
of individuals). Attributes can be associated to nodes, describing the quality or dimension of patches,
while different types of edges can be exploited to represent the distance between connected patches, the
rate of dispersal between a couple of patches, or simply whether two patches are connected or not.
Metapopulation models using graph-based methods [36, 15] are simple to implement and require
relatively few data for their definition, while individual-based models implement more detailed aspects,
P. Milazzo and M.J. Pérez Jiménez (Eds.): Applications of Membrane Computing,
Concurrency and Agent-based Modelling in Population Biology (AMCA-POP 2010)
EPTCS 33, 2010, pp. 1–17, doi:10.4204/EPTCS.33.1
c D. Besozzi et al.
The influence of network topologies on metapopulations dynamics
2
concerning the nature and the interaction of populations [34, 4]. Both types of modeling approaches are
useful for the analysis of specific features of metapopulations but, while the first focuses on the properties
of the habitat topology, the second is more concerned with the emergent dynamics. In this paper, we
present a stochastic multivolume model of metapopulations, which integrates the explicit representation
of interactions between the individuals of the populations – and therefore allows to simulate the emergent
local and global dynamics – with a graph description of the habitat topology – which allows to investigate
the influence of distinct spatial structures on the dynamics.
This model, which represents a simplified extension of a previous metapopulation model that we
introduced in [7, 6], is based on the multivolume stochastic simulation algorithm tau-DPP [11, 8], a
stochastic class of membrane systems. Membrane systems, or P systems, were introduced in [27] as a
class of unconventional computing devices of distributed, parallel and nondeterministic type, inspired
by the compartmental structure and the functioning of living cells. The basic model consists of a membrane structure where multisets of objects evolve according to given evolution rules. A comprehensive
overview of P systems and of its many applications in various research areas, ranging from Biology to
Linguistics to Computer Science, can be found in [28, 12, 29].
In tau-DPP, the distinct compartments of any multivolume model can be arranged according to a
specified hierarchy (e.g., a membrane structure), under the additional assumption that the topological
structure and the volume dimensions do not change during the system evolution (each volume is assumed
to satisfy the standard requirements of the classical stochastic simulation algorithm, see [16] and [5] for
more details). Inside each volume, two different types of rules can be defined: the internal rules, which
modify the objects contained inside the volume where they take place (in the case of metapopulation,
they describe the growth and death of population individuals according to the Lotka-Volterra model of
preys and predators), and the communication rules, which are used to move the objects between adjacent
volumes (in the case of metapopulation, they describe the migration of population individuals).
In this paper, tau-DPP is exploited to analyze the emergent dynamics of metapopulation systems,
where the focus is on the influence that the topology of patches has on the migration of individuals, and
their capability to colonize other patches in the habitat. To this purpose, we consider six different habitat
topologies, formally described by graph structures, and analyze how the topological structure of patchto-patch connections, and the rate of individual dispersal between connected patches, influence the local
and global dynamics of a metapopulation. In particular, we will first consider how a given topology and
a fixed dispersal rate between patches can influence the prey-predators dynamics, and then we will focus
on the colonization of empty patches, starting from the dispersal of predators that live in a few patches
which occupy peculiar positions in the given network topology.
The paper is structured as follows: in Section 2 we present the concept of metapopulations in Ecology, and then describe the multivolume model of metapopulations by focusing, in particular, to the
different habitat topologies. In Section 3 we will show the simulation results concerning the influence of
these habitat topologies on the emergent dynamics of metapopulations, considering the effects of predators dispersal and colonization. Finally, in Section 4 we conclude the paper with some final remarks and
several proposals for further research issues concerning metapopulations.
2
Metapopulations
In this section, we first provide a brief introduction to the most relevant features of metapopulations,
concerning both the topology of the habitats and the emergent dynamics. Then, we describe the modeling
approach used in this paper, that is based on a stochastic class of membrane systems, which will be used
D. Besozzi et al.
3
in Section 3 to analyze the influence of different network topologies on the dynamics of metapopulations.
2.1 Dynamical models of interacting populations in Ecology
Since its introduction in [22], the concept of metapopulations (also called multi-patch systems) has been
extensively applied in Ecology to analyze the behavior of interacting populations, to the purpose of determining how fragmented habitats can influence various aspects of these systems, such as local and global
population persistence, or the evolution of species [18]. Lately, this topic has been largely employed for
other populations species, living in both natural and artificial/theoretical fragmented landscapes [17].
A metapopulation consists of local populations, living in spatially separated habitats called patches
– which can be characterized by different areas, quality or isolation – connected each other through a
dispersal pool, which is the spatial place where individuals from a population spend some lifetime during
the migration among patches. In multi-patch systems, two principal types of dynamics exist: on the one
hand, the individuals of the different populations can have local interactions inside each patch (according
to a given dynamical model, e.g., the Lotka-Volterra system of interaction between preys and predators
[25]); on the other hand, the dispersal of individuals among mutually connected patches can influence
the global behavior of the whole system [20, 21, 33, 37]. The dispersal of individuals, which is usually
dependent on the distance between patches, may reduce the local population growth, and thus increase
the extinction risk, which can be due also to environmental and demographical stochasticity. Hence,
the persistence of populations is assumed to be balanced between local extinctions and the process of
colonization, that is, the establishment of new populations in empty patches [17].
Several theoretical frameworks for metapopulation analysis have been defined up to now, remarking
specific properties of multi-patch systems which have been either explicitly or implicitly considered in
these modeling methods (see, e.g., [14, 17, 24, 19] for further details). For instance, referring to the
landscape, most theoretical models take care of the spatial structure of the habitat, the local quality of
the environment, the patch areas and their mutual connectivity (or isolation), in order to capture the
effect of habitat fragmentation on species persistence. In fact, good local conditions can determine the
growth and the survival of populations inside the patches, and high patch connectivity can decrease local
extinction risk. Moreover, as dispersal and colonization are distance-dependent elements, they can be
used to account for the importance of real landscape structures. Referring to population interactions and
dynamics, colonization can depend or not on the cooperation of migrating individuals (in the first case, it
is called “Allee effect”). Models not accounting for within-patch dynamics – but only assuming whether
a patch is occupied or not – usually consider local dynamics on a faster time scale with respect to the
global dynamics, and also neglect the dependence of colonization and extinction rates on population
sizes. Finally, regional stochasticity can account for “bad” or “good” years over the local environmental
quality, which depends on, e.g., the weather conditions which affect sustenance resource availability and,
once more, they can influence the growth and survival of populations.
Recently, graph-based models for metapopulations have started to be more and more defined because of the intuitive and visual way they hold for the representation of these ecological systems (see
[36, 23, 35] and references therein). In these models, nodes represent habitat patches and graph edges
denote functional connections between patches (typically related to the dispersal of individuals). In addition, attributes can be associated to nodes, describing the quality or dimension of patches, and different
types of edges can be adopted to represent the distance between connected patches, the rate of dispersal
between a couple of patches, or simply whether two patches are connected or not. These models allow
to make insights into the features of habitat distribution, such as the predominant importance of some
nodes or clusters of nodes with respect to other characteristics of metapopulation, like their dynamics, the
4
The influence of network topologies on metapopulations dynamics
vulnerability to disturbance, the persistence of populations according to dispersal, and so on. These results open promising perspective in related research fields as evolutionary ecology, conservation biology,
epidemiology, management and design of natural reserves.
2.2 A P system–based model of metapopulations: focusing on network topologies
Most of the issues discussed in Section 2.1 were explicitly considered in our previous model for metapopulations [6, 7]. In those works, metapopulation models were based on a class of membrane systems
called DPP [31, 30], which were used to execute qualitative stochastic simulations of the local and
global dynamics of metapopulations. In particular, in [7] we introduced a model of metapopulations
with predator-prey dynamics, where additional features were used in order to catch and better describe
relevant properties of the modeled system. For instance, the regions of the membrane structure were
represented as nodes of a weighted graph with attributes, where the weight associated to edges corresponds to the “distance” among connected regions, while attributes specify their surface dimension.
These new features are necessary in order to outline the spatial distribution of patches and the relevant
additional features associated to them: the dimension of a patch is needed to define the density of the
populations living inside that patch, while the distance is needed to identify isolated patches, as well as to
define the dispersal rates of migrating individuals. Moreover, by using some rules which do not modify
the objects on which they act (the so-called “mute rules”), we modified the classical view of maximal
parallelism, by allowing the maximal application of rules but, at the same time, reducing the maximal
consumption of objects. The model was applied to investigate some emergent metapopulation behaviors,
such as the influence of patch dimension, patch-to-patch distance, stochastic breeding, the dynamics underlying migration and colonization, the effects due to isolated patches, etc. Then, in [6] we extended
the analysis of that model by focusing on periodic resource feeding strategies, and compared different
systems where either increasing, decreasing, stationary or purely feeding stochastic phases were defined
inside each patch. We have shown there, for instance, how the seasonal variance can transform the basic
Lotka-Volterra dynamics inside each patch into a more complex dynamics, where the different phases of
a feeding cycle can be identified through the effect that they have on the standard oscillations of preys
and predators.
In this section, we present a simplified model of metapopulations, which exploits the multivolume
stochastic simulation algorithm tau-DPP [11, 5]. With respect to the previous model, here we will not
need to use the concept of mute rules, as the probabilistic choice and applications of rules is already
embedded in the tau leaping algorithm [10], on which tau-DPP is based. Moreover, we will not consider
the presence of the dispersal pool, but we will instead focus our analysis on the direct communication
of individuals among interconnected patches, according to some fixed network topologies. In order to
compare the influence of each network, we have decided to perform our analysis on a total of 6 patches,
spatially arranged in different ways. Namely, we assume that these network topologies can be described
by graphs having the same number of nodes, but distinct connections, such as the chain, grid, star, ring,
complete or random structure (see graphs a, b, c, d, e, f , respectively, in Fig. 1). From now on, we will
refer to the formal data structure by using the term ‘graph’, and use the term ‘network’ to denote the
topological relationship on each graph.
Formally, each network topology ν ∈ {a, b, c, d, e, f }, can be generally described by a weighted
undirected graph Gν = (N∆ν , E ν , wν ) where:
• N∆ν is the set of nodes, such that each node pi ∈ N∆ν , i=1, . . ., 6, is characterized by a value δ (pi ) ∈ ∆
(with ∆ being a set of attributes of some kind);
• E ν ⊆ {(pi , p j ) | pi , p j ∈ N∆ν } is the set of (undirected) edges between nodes;
D. Besozzi et al.
5
Figure 1: Network topologies.
• wν : E ν → R+ is the weight function associating a cost to each edge.
In the case of metapopulations, the set of nodes N∆ν coincides with the set of patches, the attribute
of a node represents the area of the patch, the edges characterize which patches are directly reachable
from any patch (self-edges might exist as well but will not be considered in this work), and the weight
wνi, j of an edge (pi , p j ) represents a cost to measure the effort that individuals have to face when moving
from patch pi to p j . Given a network topology ν , we denote by Ad j(pi )ν the set of nodes that are
directly connected to any node pi , that is, Ad j(pi )ν = {p j ∈ N∆ν | ∃ (pi , p j ) ∈ E ν }. We also denote
by deg(pi )ν the degree of patch pi , that is, the number of patches directly connected to pi (formally,
deg(pi )ν = card(Ad j(pi )ν )). We outline that, in what follows, we will assume that: (1) wνi, j = 1 for
each (pi , p j ) ∈ E ν and each ν ∈ {a, b, c, d, e, f }, that is, all edges have the same cost; (2) δ (pi ) = 1 for
each pi ∈ N∆ν and each ν ∈ {a, b, c, d, e, f }, that is, all patches have the same dimension. The rational
behind this is that, in this paper, we focus our attention on the influence that different topologies of the
habitat network can have on the local and global dynamics of metapopulations, regardless of the local
features of each patch, or of the distances between patches. These features might be naturally added
in further works related to this model, where real data can be used to define a specific model of some
metapopulation systems.
In addition to the chosen network topology, this model of metapopulations also considers the presence of species individuals, which locally interact according to a chosen dynamics, and give rise to global
dynamics thanks to the dispersal processes. To this purpose, in this paper we assume that each patch is
characterized by the Lotka-Volterra (LV) model describing the interaction between the individuals of two
populations, namely preys and predators. Inside each patch, the LV model is described by the following
set of internal rules:
r1 : AX → X X
r2 : XY → YY
r3 :
Y →λ
where X denotes the preys, Y denotes the predators, A denotes the sustenance resources and λ is the
The influence of network topologies on metapopulations dynamics
6
empty symbol. Rules r1 and r2 model the growth of preys and predators, respectively, while rule r3
models the death of predators. Each rule is also characterized by a stochastic constants (expressed in
time−1 ), that is used – together with the current amounts of individuals occurring in the patch – to evaluate
its application probability step by step, according to the tau leaping algorithm (see [10, 11, 8] for more
details). All the simulations shown hereafter have been executed using the following values of stochastic
constants and of initial amount of preys, predators, and sustenance resources: c1 =0.1, c2 =0.01, c3 =10,
X0 =Y0 =1000, A0 =200 (the value of A is fixed for the entire duration of each simulation). The simulations
have been performed with the software BioSimWare [5], that implements different stochastic simulation
algorithms for both single and multivolume systems. The software is available for free download at
http://bimib.disco.unimib.it/index.php/Software.
In Fig. 2 we show the oscillating dynamics (left side) of preys and predators in the single patch,
obtained with this choice of parameters, and the corresponding phase space (right side). These figures
can be considered as reference to compare and discuss the dynamics obtained in the multi-patch model,
as described in Section 3.
5000
5000
X
Y
4500
4500
4000
4000
3500
3500
3000
2500
Y
Individuals
3000
2500
2000
2000
1500
1500
1000
1000
500
0
500
0
2
4
6
8
10
0
500
Time [a.u.]
1000
1500
2000
2500
3000
3500
X
Figure 2: The Lotka-Volterra dynamics in the single patch: oscillations in preys, X , and predators, Y (left
side), and corresponding phase space (right side).
The single patch model is then extended to a multi-patch model where, inside each patch pi of each
network topology ν , we add as many communication rules as the number of patches connected to pi
(that is, a total of deg(pi )ν rules inside each patch). These rules are needed to move population individuals among the various patches of the network, thus allowing to analyze the effects of migration and
colonization in the metapopulation. This is done by attaching a destination target to each communication
rule, specifying the destination patch, as it is usually done in P systems. Formally, in each patch pi of
network ν , we add the so-called dispersal rules
rd p j : Y → (Y,target(p j )),
for each p j ∈ Ad j(pi )ν . Similarly to the local rules r1 , r2 , r3 , the probability of applying each dispersal
rule is determined by using its stochastic constant cd p j , whose values will be given in the next section to
consider different migration rates.
D. Besozzi et al.
3
7
The influence of network topologies on metapopulation dynamics
In this section we analyze how the topological structure of patch-to-patch connections, and the rate of
individual dispersal between connected patches, influence the local and global dynamics of a metapopulation. In particular, in Section 3.1 we consider how a given topology and a fixed dispersal rate can
influence the prey-predators dynamics, while in Section 3.2 we focus on the capability of colonization
of empty patches, starting from the dispersal of predators living in a few patches which occupy peculiar
positions in the given network topology.
3.1 Network topologies and migration
In this section, we analyze the role of migration and compare the six network topologies with respect to
four different conditions for the dispersal rules. Namely, we assume that each patch of each topology is
initialized with a complete LV model as given in Section 2.2, where the value of the stochastic constant
cd p j for the dispersal of predators, in each patch pi ∈ N∆ν , can assume one of the following values:
1. cd p j =1, for each p j ∈ Ad j(pi )ν ;
2. cd p j =10, for each p j ∈ Ad j(pi )ν ;
3. cd p j =20, for each p j ∈ Ad j(pi )ν ;
10
4. cd p j = deg(p
, for each p j ∈ Ad j(pi )ν .
i)
By considering the first condition as reference, the power of dispersal in the second (third) condition is
ten-fold (twenty-fold) the first one, irrespective of the position that patch pi occupies in the considered
network. In other terms, the flux of dispersal from each patch, in the first three conditions, results
amplified by the number of connections that each patch has with respect to the other patches in the
network. On the contrary, the fourth condition corresponds to the situation when, for each patch p j ∈
Ad j(pi )ν , the sum of the values of constants of dispersal rules in pi is always equal to 10, but the rate
of dispersal along each edge from pi to p j depends on the degree of pi . For instance, in the network
topology a (Fig. 1), the value of cd p j in patches p0 and p5 is equal to 10, while the value of cd p j in
patches p1 , . . ., p4 is equal to 5; in the network topology c (Fig. 1), the value of cd p j in patch p0 is equal
to 2, while the value of cd p j in all other patches is equal to 10, and so on. So doing, we can weigh the
dispersal of predators according to the position of each patch in the network, and simulate a situation
where the flux of dispersal from each patch towards its adjacent patches is uniform throughout the whole
network.
For space limits, in Fig. 3 we present the phase spaces of all network topologies, obtained from
simulations of the fourth condition only. For each network, in particular, we show the phase space of the
local dynamics of each patch. The graphics show that, in the case of the chain graph (phase space (a)), the
patches having different degrees are characterized by different dynamics: in fact, patches p0 and p5 show
a different behavior with respect to the other patches. In addition to the role of patch degree, we can see
that also the position of patches in the graph plays a central role: despite the fact that patches p1 , p2 , p3
and p4 have all the same degree, the dynamics inside p1 and p4 differs from that of patches p2 and p3 .
This is due to the different power of dispersal rules of their two neighbors, namely cd p j = 10 in patches
p0 , p5 , while cd p j = 5 in patches p2 , p3 , which cause a larger flux of predators dispersal towards patches
p1 and p4 . The global effect is the presence of three different dynamics (one in p0 , p5 , another one in p1 ,
p4 , and a third one in p2 , p3 ), all of which are characterized by oscillations in X and Y with no regular
amplitudes (compare these phase spaces with the standard LV phase space in the single patch model
The influence of network topologies on metapopulations dynamics
8
(a) chain
Y
(b) grid
4000
3500
3000
2500
2000
1500
1000
500
Y
4000
3500
3000
2500
2000
1500
1000
500
3000
3000
2500
2500
0
0
2000
1
2
Patch
1500
2
Patch
1000
3
2000
1
X
1500
500
4
5
X
1000
3
500
4
0
5
0
(d) ring
(c) star
Y
8000
7000
6000
5000
4000
3000
2000
1000
0
Y
5000
4500
4000
3500
3000
2500
2000
1500
1000
500
3500
3500
3000
0
3000
0
2500
1
2
Patch
1500
2500
1
2000
X
2
Patch
1000
3
2000
500
4
5
500
4
0
5
0
(e) complete
Y
4000
3500
3000
2500
2000
1500
1000
500
X
1500
1000
3
(f) random
Y
5000
4500
4000
3500
3000
2500
2000
1500
1000
500
2500
2000
0
1500
X
1
2
Patch
1000
3
500
4
5
0
3500
3000
0
2500
1
2000
2
Patch
1500
X
1000
3
500
4
5
0
Figure 3: The power of migration: LV dynamics in the phase space of each network topology.
D. Besozzi et al.
9
given in Fig. 2, right side, and also with the phase spaces in Fig. 3, graphics (d) and (e)). Furthermore,
we can evidence that these oscillations are characterized by an initial wider amplitude, which is reduced
during time.
Similarly, the dynamics of the patches in the grid graph (phase space (b)) is influenced only by the
number of edges; in this phase space, we can identify two different types of dynamics: one for the patches
with three edges (p1 , p4 ) and another one for those with two connections.
In the star graph (phase space (c)), the LV dynamics endures in all patches apart from p0 , where the
number of preys X collapses to an attractor in zero, and no oscillations according to the LV dynamics
in both X and Y can be established. In this patch, the number of predators fluctuates in a certain range,
because of their dispersal from/to the other patches. Basically, in this condition patch p0 , that represents
the center of the star, becomes a local area of the habitat where only dispersal occurs.
The simulations for the ring and complete graphs (phase spaces (d), (e)) show very similar results:
in both cases, all patches in each graph have the same degree (two in the first configuration and five in
the second one), leading to regular oscillations in X and Y with almost constant amplitude.
The results concerning the last configuration, the random graph (phase space (f)), show a combination
of the effects described above. In particular, the dynamics of the patches differ each other depending on
the degree of the patches themselves; moreover, in p4 , which is characterized by the highest degree, the
high number of incoming predators (migrating from the four adjacent patches) leads to the extinction of
preys (similarly to what happens in patch p0 of the star graph).
We also tested, for each network topology, the other three conditions listed above. In these cases,
the results have shown that the amplification of the power of dispersal with respect to the patch degree
gives rise to a balance between the incoming and migrating individuals, which leads to comparable LV
dynamics for all networks, with regular oscillations inside each patch (data not shown).
3.2 Network topologies and colonization
In this section, we compare the six network topologies with respect to the capability of colonizing the
empty patches that each network contains, starting from the patches that contain a complete LV model
and that occupy a peculiar position in that network . We recall that in this work we are considering only
the migration of predators, hence the empty patches are hereby assumed to contain no predators but only
an initial amount of preys. In each network ν , the set of patches initialized with the complete LV model
will be denoted as pνLV . To test the feature of colonization, we consider four different initial conditions,
hereby denoted as ICk, k=1, . . . , 4, where Y0 =0 and:
1. IC1 is characterized by cd p j =1 and X0 =10;
2. IC2 is characterized by cd p j =1 and X0 =100;
3. IC3 is characterized by cd p j =10 and X0 =10;
4. IC4 is characterized by cd p j =10 and X0 =100.
In each given network, all empty patches are initialized with the same chosen condition ICk, besides the
patches in the set pνLV that are initialized with a standard LV model, having the communication constant
cd p j equal to the one given in the chosen ICk, and all other parameters as given in Section 2.2.
With this type of analysis, we expect to determine which features of the network topologies are more
relevant with respect to the colonization of empty patches, under a given initial condition. All conditions
have been tested for each network and, for each fixed initial condition, different sets of pνLV have been
considered. In the following, for space limits, we present only some results of these simulations, and
The influence of network topologies on metapopulations dynamics
10
briefly discuss the results obtained in the other analyzed conditions. In each of the following graph,
preys (X ) are represented with solid lines, while predators (Y ) are represented with dashed lines.
We start by considering the network ν = a, that is, the chain graph. In this case, we present the
results obtained in all the initial conditions IC1, IC2, IC3, IC4, considering three sets of LV patches,
namely paLV ={p0 , p5 }, paLV ={p2 } and paLV ={p0 }. In the first case (paLV ={p0 , p5 }, shown in Fig. 4) we
can see that, when the power of dispersal is low (IC1, IC2), the time required by the predators to reach
the patches p2 and p3 , which are at the highest distance from p0 and p5 , allows an initial uncontrolled
growth of the preys in p2 and p3 , which subsequently undergo extinction as soon as the predators enter
the patch. Such “delay” in the local establishment of a population of predators is the effect that prevent
the formation of the LV dynamics; this effect, as shown hereafter, is a common aspect of all network
topologies. Concerning the chain network, this is more evident in condition IC2, where the initial amount
of preys inside the empty patches is higher than IC1: in this case, the LV dynamics can be established
only in four of the six patches. On the other hand, with the initial conditions IC3 and IC4, the power
of dispersal is sufficient to colonize all of the patches, irrespectively of the numbers of preys that are
initially present in the empty patches and of the position of the LV complete patch. Similar results for
the chain network have been obtained in the second analyzed case (paLV ={p2 }, shown in Fig. 5) and in
the third case (paLV ={p0 }, data not shown).
12000
14000
12000
10000
Individuals
Individuals
10000
8000
6000
8000
6000
4000
4000
2000
00
1
2
3
Patch
4
5
0
2
4
6
8
10
12
14
16
18
2000
20
00
Time [a.u.]
1
2
3
Patch
4500
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
6000
4000
5000
3000
Individuals
Individuals
3500
2500
2000
1500
4000
3000
2000
1000
500
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
Time [a.u.]
20
1000
25
20
00
15
1
2
Patch
10
3
4
Time [a.u.]
5
5
0
Figure 4: Colonization in the chain topology, with paLV ={p0 , p5 } and initial conditions IC1 (top left), IC2
(top right), IC3 (bottom left), IC4 (bottom right).
D. Besozzi et al.
11
16000
25000
14000
20000
Individuals
Individuals
12000
10000
8000
15000
10000
6000
4000
5000
2000
00
1
2
3
Patch
4
5
0
2
4
6
8
10
12
14
16
18
20
00
Time [a.u.]
1
2
3
Patch
6000
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
9000
8000
5000
Individuals
Individuals
7000
4000
3000
2000
6000
5000
4000
3000
2000
1000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
Time [a.u.]
20
1000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
Figure 5: Colonization in the chain topology, with paLV ={p2 } and initial conditions IC1 (top left), IC2
(top right), IC3 (bottom left), IC4 (bottom right).
For the network topology ν = b, that is, the grid graph, we show the results obtained in the cases IC1,
when pbLV ={p0 } (Fig. 6, left side) and pbLV ={p1 } (Fig. 6, right side). According to the position of the LV
complete patches in this network topology, we can see that, in the first case, the predators are capable to
colonize patches p1 and p3 , that are directly connected to p0 , and patch p4 , that is directly connected to
both p1 and p3 . However, patches p2 and p5 cannot be colonized. In the second case, the higher degree
of the LV complete patch p1 , allows the colonization of all patches. With the initial condition IC2 (data
not shown), in the other tested cases pbLV ={p0 } and pbLV ={p1 }, only the patches directly connected to p0
and p1 , respectively, are colonized by the predators.
For the network topology ν = c, that is, the star graph, we show the results obtained in the cases
IC1, when pcLV ={p1 } (Fig. 7, left side) and pcLV ={p1 , p3 } (Fig. 7, right side). According to the position
of the LV complete patches in this network topology, we can see that, in the first case, no patches are
colonized because of the high degree of p0 (which is the only patch connected to p1 ) that spreads the
predators over the other patches, thus preventing the formation of the LV dynamics. In the second case,
the combined effect of migration from p1 and p3 allows the colonization of patch p0 , which is directly
connected with both of them. We then performed other simulations starting with conditions IC3 and
IC4: in these cases, the higher value of cd p j allows the colonization of every patch (except from patch p0 )
independently from the initial position of the LV complete patch (data not shown). On the contrary, when
The influence of network topologies on metapopulations dynamics
12
16000
12000
14000
10000
Individuals
Individuals
12000
10000
8000
8000
6000
6000
4000
4000
2000
00
1
2
3
Patch
4
5
0
2
4
6
8
10
12
14
16
18
2000
20
00
Time [a.u.]
1
2
3
Patch
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
Figure 6: Colonization in the grid topology, with initial condition IC1 and pbLV ={p0 } (left), pbLV ={p1 }
(right).
25000
14000
12000
20000
Individuals
Individuals
10000
15000
8000
6000
10000
4000
5000
00
1
2
3
Patch
4
5
0
2
4
6
8
10
12
14
16
18
2000
20
00
Time [a.u.]
1
2
3
Patch
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
Figure 7: Colonization in the star topology, with initial condition IC1 and pcLV ={p1 } (left), pcLV ={p1 , p3 }
(right).
18000
20000
16000
18000
16000
14000
12000
Individuals
Individuals
14000
10000
8000
12000
10000
8000
6000
6000
4000
4000
2000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
Time [a.u.]
20
2000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
Figure 8: Colonization in the ring topology, with pdLV ={p0 } and initial condition IC1 (left) and IC2
(right).
D. Besozzi et al.
13
we assume pcLV ={p0 }, that is, the center of the star, then all patches are fully colonized, independently
from the considered initial condition.
For the network topology ν = d, that is, the circular graph, we show the results obtained in the cases
IC1 and IC2, when pdLV ={p0 } (Fig. 8, left and right sides, respectively). Starting with the initial condition
IC2, the predators are capable of colonizing only the patches directly connected to the LV complete patch
p0 , while in the case IC1, also patch p4 (being at distance 2 from the LV complete patch) is colonized.
These results highlight, in particular, another aspect that was more marginal in the other simulations: the
stochastic nature of the communication process and of the growth of preys, which leads to the extinction
of preys in patch p2 , while in patch p4 it drives the local behavior to an oscillatory dynamics.
For the network topology ν = e, that is, the complete graph, we show the results obtained in the cases
IC1, when peLV ={p0 } (Fig. 9, left side) and peLV ={p0 , p3 } (Fig. 9, right side). While in the second case
– where the LV dynamics is initially placed in two patches – the predators can colonize all patches, in
the first case the colonization of all empty patches fails. Once more, this is an effect of the stochastic
noise combined with the low amounts of predators, which is in turn caused by the fact that the higher the
number of adjacent patches, the lower the number of predators that persist inside each patch. In all other
simulations performed with initial conditions IC3 and IC4, all patches have always been colonized, as
the higher values of dispersal rules assure a more uniform spread of predators throughout the network,
and thus flattens the influence of migration delay (data not shown).
For the network topology ν = f , that is, the random graph, we show the results obtained in the cases
f
f
IC1, when pLV ={p0 } (Fig. 10, left side) and pLV ={p2 } (Fig. 10, right side). According to the position
of the LV complete patches in this network topology, we can see that, in the first case, all patches are
colonized by predators (similar results are obtained by placing the LV complete model in patch p4 – data
not shown). In the second case, patch p5 is not colonized because there is only one path of length 2 which
connects it to the initial complete LV patch p2 ; the same holds for patch p3 , which has distance from p2
equal to 3. For similar reasons, considering the case of initial condition IC1, with the LV complete model
in patch p3 , the only patch that is not colonized by predators is p2 (data not shown). In all the simulations
performed with the initial condition IC2, some of the patches have not been colonized because of the high
amount of preys initially occurring in the patches. On the other hand, with the initial conditions IC3, IC4,
the power of dispersal allows the colonization of all patches (data not shown).
12000
8000
7000
10000
Individuals
Individuals
6000
8000
6000
5000
4000
3000
4000
2000
2000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
Time [a.u.]
20
25
1000
20
00
15
1
2
Patch
10
3
4
Time [a.u.]
5
5
0
Figure 9: Colonization in the complete topology, with initial condition IC1 and peLV ={p0 } (left),
peLV ={p0 , p3 } (right).
The influence of network topologies on metapopulations dynamics
14
12000
16000
14000
10000
Individuals
Individuals
12000
8000
6000
10000
8000
6000
4000
4000
2000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
Time [a.u.]
20
2000
00
1
2
Patch
3
4
5
0
2
4
6
8
10
12
14
16
18
20
Time [a.u.]
f
Figure 10: Colonization in the random topology, with initial condition IC1 and pLV
={p0 } (left),
f
pLV ={p2 } (right).
4
Discussion
The fragmented habitats of real metapopulations are usually characterized by complex network topologies. In this paper, we have analyzed six small topologies that can be considered representative of local
areas in a structured habitat, and we have investigated the influence that the degree and the position of
each patch in the topology can have on the migration of individuals, as well as on the capability of colonizing empty patches. Our analysis suggests that, with respect to the power of migration (Section 3.1),
we can identify different behaviours that depend on two characteristics of the topology: on a first level,
the local behaviour inside each patch is influenced by its degree. This is especially evident if we compare
the network topology described by the circular or complete graphs, with the topology described by the
star graph: while in the first case (where all nodes have the same degree) all patches are characterized by
a similar (regular) oscillating dynamics, in the second case the most critical node is the center of the star
(which has a much higher degree than all other nodes in the same graph). In the latter case, this patch is
likely to undergo a local modification of its initial dynamics, due to a more higher incoming migration
of individuals from all other adjacent patches. On a second level, assuming in this case that the degree
of nodes is equal, then also the position of each patch in the topology matters: for instance, we have
seen that in the network topology described by the chain graph – where all nodes, besides the ones at
the extremes of the chain, have the same degree – the local dynamics is also influenced by the dynamics
of the adjacent patches in the graph. Therefore, in hypothetical habitats where there exist many patches
connected in a linear way, our results suggest that the length of the chain might have a negative role in
the establishment and in the maintenance of local dynamics.
Considering the feature of colonization (Section 3.2), we have evidenced that, in most network
topologies, the lack of colonization can be due to the delay of migrating predators with respect to the
(uncontrolled) local growth of prey, which then leads to the extinction of preys and the prevention of
the LV dynamics. To effectively measure how strong is the power of the delay, it would be interesting
to understand whether the local growth of preys can be controlled by inducing their death and thus potentially allowing the establishment of oscillations. Besides this aspect deserving further investigations,
our analysis have evidenced that the colonization of empty patches occurs more easily in those patches
that are adjacent to the patch(es) initialized with the LV complete model. Once more, this highlights the
relevance of the position of the patch(es) where standard oscillations in preys and predators are already
D. Besozzi et al.
15
settled at the beginning of the simulation. Indeed, the power of colonization is stronger in the circular
and complete networks – where the position of the LV complete patch is irrelevant (as the spread of migrating individuals throughout the network results uniform), and it is weaker in the star network – where
the position of the LV complete patch is of primary importance (as the spread of migrating individuals
throughout the network strongly depends on whether the patch is placed at the center or at the tips of the
star).
In addition to the investigations that we have presented in this work, further types of analysis that
we plan to perform on metapopulation systems concern, for instance, the study of the aspects considered
in this paper (migration, colonization, network topologies, etc.) by assuming other local and global
dynamics, e.g., the population growth according to the logistic function. Moreover, an interesting issue
that might be investigated is the synchronization of local population dynamics (e.g. by considering the
establishment and decay of oscillations in preys and predators) during migration through a given network
topology, or in the process of colonization.
Concerning the use of graphs, other relevant questions regard the analysis of the dynamics with
respect to graph properties, such as different measures of habitat connectivity (centrality indexes) [13,
26]. In this context, for example, the star graph can resemble the notion of hub (a node with high degree)
in a typical scale-free network, a structure that is known to be robust to random disturbances but highly
vulnerable to deliberate attacks on the hubs [32, 3].
Another topic of interest concerns the fact that various populations can coexist in a common habitat,
but have distinct (inter)species dynamics or different dispersal capabilities in that habitat [9]. In cases
like this, it would be interesting to construct and analyze different metapopulation models, one for each
target species, according to both the patch-to-patch connections and to the specific population dynamics.
By comparing and intersecting the results obtained on the distinct network topologies of the common
habitat derived in this way, it would be possible to determine the locations of the habitat that are most
important for each species, and thus aid the design of natural reserve systems where we can have the
most appropriate solution for all species in terms of the maximal improvement of dispersal (reduction
of species isolation) and the minimal spread of disturbances (diseases, pathogens, invasive species, etc.)
[36].
We believe that our modeling approach opens interesting perspectives and can represent an useful tool
for the investigation of a wide range of properties in metapopulation systems. We expect that applications
of this model to real cases – characterized by complex habitat networks (where each patch possesses its
own features of quality, occupancy, connectivity) and different population dynamics – will aid in the
achievement of important results and new perspective in Ecology.
References
[1] T. Aittokallio & B. Schwikowski (2006): Graph-based methods for analysing networks in cell biology. Briefings in bioinformatics 7(3), pp. 243–255. Available at http://dx.doi.org/10.1093/bib/bbl022.
[2] R. Albert (2005): Scale-free networks in cell biology. Journal of Cell Science 118(21), pp. 4947–4957.
Available at http://dx.doi.org/10.1242/jcs.02714.
[3] R. Albert & A. L. Barabási (2002): Statistical mechanics of complex networks. Reviews of Modern Physics
74(1), pp. 47–97. Available at http://dx.doi.org/10.1103/RevModPhys.74.47.
[4] L. Berec (2002): Techniques of spatially explicit individual-based models: construction, simulation, and
mean-field analysis. Ecological Modelling 150(1-2), pp. 55–81. Available at http://dx.doi.org/10.
1016/S0304-3800(01)00463-X.
16
The influence of network topologies on metapopulations dynamics
[5] D. Besozzi, P. Cazzaniga, G. Mauri & D. Pescini (2010): BioSimWare: A P Systems–based Simulation
Environment for Biological Systems. Accepted for presentation at CMC11, Jena, Germany, 2010.
[6] D. Besozzi, P. Cazzaniga, D. Pescini & G. Mauri (2007): Seasonal variance in P system models for metapopulations. Progress in Natural Science 17(4), pp. 392–400. Available at http://www.informaworld.com/
smpp/content~db=all~content=a790271810~tab=linking.
[7] D. Besozzi, P. Cazzaniga, D. Pescini & G. Mauri (2008): Modelling metapopulations with stochastic
membrane systems. BioSystems 91(3), pp. 499 – 514. Available at http://www.sciencedirect.com/
science/article/B6T2K-4PD4XHR-1/2/0903081b39759345708d5fc97aeee6bc.
[8] D. Besozzi, P. Cazzaniga, D. Pescini & G. Mauri (2009): Algorithmic Bioprocesses, chapter A Multivolume Approach to Stochastic Modelling with Membrane Systems, pp. 519–542. Springer Verlag. Available at http://www.springer.com/computer/theoretical+computer+science/book/
978-3-540-88868-0.
[9] A. Bunn (2000): Landscape connectivity: A conservation application of graph theory. Journal of Environmental Management 59(4), pp. 265–278. Available at http://dx.doi.org/10.1006/jema.2000.0373.
[10] Y. Cao, D. T. Gillespie & L. R. Petzold (2006): Efficient step size selection for the tau-leaping simulation
method. Journal of Chemical Physics 124, p. 044109. Available at http://www.ncbi.nlm.nih.gov/
pubmed/16460151.
[11] P. Cazzaniga, D. Pescini, D. Besozzi & G. Mauri (2006): Tau leaping stochastic simulation method in P
systems. In: H. J. Hoogeboom, G. Păun, G. Rozenberg & A. Salomaa, editors: Proc. of the 7th International
Workshop on Membrane Computing, 4361. LNCS, pp. 298–313. Available at http://www.springerlink.
com/content/y055172665v12t2k/?p=0a193a302bda4824b742ec37b8cda9d9&pi=18.
[12] G. Ciobanu, G. Păun & M. J. Pérez-Jiménez, editors (2005): Applications of Membrane Computing. SpringerVerlag, Berlin.
[13] S. N. Dorogovtsev & J. F. F. Mendes (2002):
Evolution of networks.
Advances in
Available at http://www.informaworld.com/smpp/
Physics 51(4), pp. 1079–1187.
content~db=all~content=a713801291~tab=linking.
[14] J. B. Dunning Jr., D. J. Stewart, B. J. Danielson, B. R. Noon, T. L. Root, R. H. Lamberson & E.E. Stevens
(1995): Spatially Explicit Population Models: Current Forms and Future Uses. Ecological Applications 5(1),
pp. 3–11. Available at http://www.esajournals.org/doi/abs/10.2307/1942045.
[15] A. Fall, M. Fortin, M. Manseau & D. O’brien (2007): Spatial Graphs: Principles and Applications
for Habitat Connectivity. Ecosystems 10, pp. 448–461. Available at http://dx.doi.org/10.1007/
s10021-007-9038-7.
[16] D. T. Gillespie (1977): Exact stochastic simulation of coupled chemical reactions. Journal of Physical Chemistry 81(25), pp. 2340–2361. Available at http://dx.doi.org/10.1021/j100540a008.
[17] I. Hanski (1998): Metapopulation dynamics. Nature 396, pp. 41–49. Available at http://www.nature.
com/nature/journal/v396/n6706/abs/396041a0.html.
[18] A. Hastings & S. Harrison (1994): Metapopulation dynamics and genetics. Annual Review of Ecology
and Systematics 25, pp. 167–188. Available at http://arjournals.annualreviews.org/doi/abs/10.
1146/annurev.es.25.110194.001123.
[19] A. Hastings & C. L. Wolin (1989): Within-patch dynamics in a metapopulation. Ecology 70(5), pp. 1261–
1266. Available at http://www.esajournals.org/doi/abs/10.2307/1938184.
[20] V. A. A. Jansen (2001): The dynamics of two diffusively coupled predator-prey populations. Theoretical
Population Biology 59, pp. 119–131. Available at http://dx.doi.org/10.1006/tpbi.2000.1506.
[21] V. A. A. Jansen & A. L. Lloyd (2000): Local stability analysis of spatially homogeneous solutions of
multi-patch systems. Journal of Mathematical Biology 41, pp. 232–252. Available at http://www.
springerlink.com/content/px0an76xh7551jew/.
D. Besozzi et al.
17
[22] R. Levins (1969): Some demographic and genetic consequences of environmental heterogeneity for biological
control. Bulletin of the Entomological Society of America 71, pp. 237–240.
[23] E. S. Minor & D. L. Urban (2008): A Graph-Theory Framework for Evaluating Landscape Connectivity and
Conservation Planning. Conservation Biology 22(2), pp. 297–307. Available at http://dx.doi.org/10.
1111/j.1523-1739.2007.00871.x.
[24] A. Moilanen (2004): SPOMSIM: software for stochastic patch occupancy models of metapopulation dynamics. Ecological Modelling 179, pp. 533–550. Available at http://dx.doi.org/10.1016/j.ecolmodel.
2004.04.019.
[25] J. D. Murray (2002): Mathematical Biology. I: An introduction. Springer-Verlag, New York.
[26] M. E. J. Newman (2003): The Structure and Function of Complex Networks. SIAM Review 45(2), pp. 167–
256. Available at http://scitation.aip.org/getabs/servlet/GetabsServlet?prog=normal&
id=SIREAD000045000002000167000001&idtype=cvips&gifs=yes.
[27] G. Păun (2000): Computing with membranes. Journal of Computer and System Sciences 61(1), pp. 108–143.
Available at http://dx.doi.org/10.1006/jcss.1999.1693.
[28] G. Păun (2002): Membrane Computing. An Introduction. Springer-Verlag, Berlin.
[29] G. Păun, G. Rozenberg & A. Salomaa, editors (2010): The Oxford Handbook of Membrane Computing.
Oxford University Press.
[30] D. Pescini, D. Besozzi, G. Mauri & C. Zandron (2006): Dynamical probabilistic P systems. International
Journal of Foundations of Computer Science 17(1), pp. 183–204. Available at http://dx.doi.org/10.
1142/S0129054106003760.
[31] D. Pescini, D. Besozzi, C. Zandron & G. Mauri (2006): Analysis and simulation of dynamics in probabilistic
P systems. In: N. Pierce A. Carbone, editor: Proc. of 11th International Workshop on DNA Computing,
DNA11, 3892. LNCS, London, ON, Canada, pp. 236–247. Available at http://www.springerlink.com/
content/7p7653442111r273/?p=832da88ca13a4d75be122f548c3b0df6&pi=18.
[32] S. H. Strogatz (2001): Exploring complex networks. Nature 410, pp. 268–276. Available at http://tam.
cornell.edu/SS_exploring_complex_networks.pdf.
[33] A. D. Taylor (1990): Metapopulations, dispersal, and predator-prey dynamics: an overview. Ecology 71(2),
pp. 429–433. Available at http://www.esajournals.org/doi/abs/10.2307/1940297.
[34] J. M. J. Travis & C. Dytham (1998): The evolution of dispersal in a metapopulation: a spatially explicit, individual-based model. Proceedings of the Royal Society of London. Series B: Biological Sciences
265(1390), pp. 17–23. Available at http://dx.doi.org/10.1098/rspb.1998.0258.
[35] D. Urban & T. Keitt (2001): Landscape Connectivity: A Graph-Theoretic Perspective. Ecology 82(5), pp.
1205–1218. Available at http://dx.doi.org/10.2307/2679983.
[36] D. L. Urban, E. S. Minor, E. A. Treml & R. S. Schick (2009): Graph models of habitat mosaics. Ecology
Letters 12, pp. 260–273. Available at http://dx.doi.org/10.1111/j.1461-0248.2008.01271.x.
[37] W. W. Weisser, V. A. A. Jansen & M. P. Hassell (1997): The effects of a pool of dispersers on host-parasitoid
systems. Journal of Theoretical Biology 189, pp. 413–425. Available at http://dx.doi.org/10.1006/
jtbi.1997.0529.
| 5cs.CE
|
Аналитический подход к синтезу многополосных
фильтров и его сравнение с другими подходами
arXiv:1612.01753v1 [cs.SY] 6 Dec 2016
c 2016
Богатырев А.Б., Горейнов С.А., Лямаев С.Ю.
∗
В настоящее время многополосные фильтры широко применяются в аналоговой и цифровой технике. Например, в современной аппаратуре связи, работающей в СВЧ и радиодиапазонах, часто требуются фильтры с несколькими полосами пропускания, соотвествующими различным стандартам беспроводных коммуникаций (таким, как IEEE 806.11, IEEE
806.16, GSM, LTE, GPS, CDMA, DVBT/T2 и др.). Разработка фильтров с многополосными
спецификациями представляет собой сложную инженерную проблему, особенно если требуется одновременно обеспечить и высокую производительность компонента, и его компактную
реализацию, что и представляет интерес на практике. К текущему моменту в литературе не
появилось полностью удовлетворительных и универсальных решений этой проблемы.
Очевидным подходом является модульный: многополосный фильтр получается в результате соединения нескольких однополосных. Недостаток его – существенная неоптимальность
порядков синтезируемых фильтров, а значит и их массогабаритных характеристик. Этим
объясняется интерес к разработке других подходов, идейно более сложных, но позволяющих
получать фильтры, имеющие меньшие порядки при той же спецификации. Большую часть
таких подходов можно условно разделить на следующие категории: 1) методы синтеза СВЧ
фильтров, основанные на использовании многочастотных резонаторов; 2) методы, основанные на использовании частотных преобразований (например, [2], [3]); 3) оптимизационные
методы, основанные на алгоритме Ремеза ([4], [5], [6] и др.).
Подходы первого типа предполагают синтез фильтра на основе мультимодальных резонаторов и позволяют достигать высокой степени миниатюрности синтезируемых устройств. При
таком решении в формировании полос пропускания и задержки фильтра одновременно участвуют резонансы нескольких мод колебаний от каждого резонатора, за счет чего уменьшается
число резонаторов, а потому и массо-габаритные характеристики фильтра. Однако подходы
этого типа специфичны только для СВЧ устройств, причем дизайн и реализация фильтров
высокого порядка с их помощью представляет существенную трудность. Методы второго
типа – синтез фильтра на основе многополосного частотного преобразования – являются
аналитическими либо полуаналитическими, и осуществляют приведение однополосного (широкополосного) фильтра-прототипа к многополосной конфигурации за счет использования
замен частотной переменной. Сложность таких преобразований быстро растет с увеличением числа полос пропускания, а обслуживаемые спецификации, как правило, ограничиваются
симметричными.
Большое внимание в технической литературе уделяется развитию методов третьего типа,
связанных с синтезом оптимальных фильтров и близких к ним. Под оптимальным для заданной многополосной спецификации в данном случае понимается фильтр, который обладает наименьшим порядком среди всех физически осуществимых фильтров, удовлетворяющих
требованиям этой спецификации. Скажем также, что под спецификацией (маской) фильтра
здесь имеется в виду совокупность следующих параметров: 1) значения желаемых граничных
частот всех полос пропускания и задержки; 2) максимальная допустимая величина неравномерности АЧХ на полосах пропускания и минимальная допустимая величина подавления в
полосах задержки. Эти данные определяют желаемый вид амплитудно-частотной характеристики фильтра, не затрагивая других его характеристик (фазо-частотной, импульсной и
т.д.).
∗ Поддержано
грантом РНФ 16-11-10349
1
Все существующие методы синтеза оптимальных многополосных фильтров основываются на прямой численной оптимизации с помощью алгоритмов типа Ремеза, и, ввиду присущей для такой оптимизации принципиальной неустойчивости, имеют ограниченную область
применимости: с их помощью возможен синтез фильтров с числом полос пропускания, как
правило, не превышающим трех, и порядками, как правило, не выше 20.
Авторами разработан аналитический подход к синтезу оптимальных многополосных фильтров, который может быть применен для спецификаций с общим числом полос (на текущий
момент) вплоть до 23 и порядками вплоть до 1000. Целью работы является сравнение нового подхода к синтезу оптимальных многополосных фильтров с методом прямой оптимизации, основанном на алгоритме типа Ремеза, а также сравнение оптимальных фильтров с
неоптимальными составными фильтрам, получаемыми модульным подходом. В первом разделе ставится оптимизационная задача, лежащая в основании проблемы синтеза оптимального
многополосного фильтра. Второй раздел содержит краткое описание нового аналитического подхода к решению этой задачи. В третьем разделе приводятся рецепты, использованные нами при решении той же задачи методом прямой оптимизации. В четвертом разделе
описывается модульный подход к синтезу неоптимальных многополосных фильтров. Пятый
раздел содержит результаты численных экспериментов, посвященных сравнению этих трех
подходов. В приложении приводятся графики фазо-частотной и импульсной характеристик
оптимального четырехполосного фильтра.
Авторы благодарны проф. Л. Баратшару за обсуждение задач равномерного рационального приближения и д-рам Ф. Зайферту и В. Люно (INRIA, Sophia-Antipolis) за предоставленный код, позволяющий рассчитывать оптимальные фильтры алгоритмом типа Ремеза.
1
Постановка задачи
Поиск оптимального фильтра по заданной спецификации может быть сведен к решению серии
задач о наименьшем уклонении типа сформулированных ниже. Для этого на каждом шаге
фиксируется порядок фильтра и максимизируется величина подавления в полосах задержки
при фиксированных остальных параметрах спецификации. Приведем две встречающиеся в
литературе формулировки.
Пусть E — множество, состоящее из m непересекающихся отрезков действительной оси
(частотных диапазонов), разделенное на два подмножества: полосы пропускания E+ и полосы
задержки E− . Скажем, что идеальная функция пропускания F равна 1 на отрезках из E+ и
−1 на отрезках из E− .
Задача 1.Найти вещественную рациональную функцию Rn степени не выше n, для
которой величина отклонения от функции пропускания F минимальна в равномерной норме
на E:
||Rn − F ||C(E) := max |Rn (x) − F (x)| → min =: µ.
(1)
x∈E
Задача 2. Найти вещественную рациональную функцию Qn (x) степени не выше n, минимизирующую величину θ при условиях
min |Qn (w)| ≥ θ−1 ,
w∈E−
max |Qn (w)| ≤ θ.
w∈E+
(2)
Нетрудно показать, что эти задачи эквивалентны, и их решения отличаются дробнолинейной подстановкой, то есть Qn = f ◦Rn , где f – дробно-линейное преобразование, а обратные к минимальным отклонениям связаны преобразованием Жуковского µ−1 = (θ + θ−1 )/2.
Отметим, что сформулированная задача о рациональной аппроксимации есть третья задача Золотарева для конденсатора (E+ , E− ) [12]. Она является многоэкстремальной: все множество рациональных функций разбивается на 2m−2 непересекающихся класса [7], [19], в
каждом из которых решение существует, единственно и допускает альтернансную характеризацию: на E имеется 2n + 2 точек альтернанса, в которых для решения степени n достигается
величина отклонения с последовательной переменой знака [20].
2
До недавнего времени аналитическое решение этой задачи было известно только для случая m = 2 (для одной полосы пропускания и одной полосы задержки), найденное в 1870-х
годах учеником П.Л. Чебышёва Е.И. Золотарёвым [17]. Именно оно около полувека спустя
было использовано немецким учёным-электротехником В. Кауэром [18] для расчёта передаточных функций оптимальных фильтров низких частот, получивших название фильтров
Кауэра-Золотарёва (эллиптических). К настоящему времени эти фильтры нашли широкое
применение в аналоговой и цифровой технике, а метод аппроксимации АЧХ по ЗолотарёвуКауэру стал классическим.
2
Новый аналитический подход
Подход к решению задач (1), (2), предлагаемый в данной статье, опирается на явные аналитические формулы, полученные в [1] и обобщающие решение Золотарева на случай m > 2.
Идея этого подхода следующая: как было отмечено выше, все решения степени n задачи о
наименьшем уклонении имеют 2n+2 точки альтернанса на E, причем каждая из них, находящаяся во внутренности множества E, с необходимостью будет критической точкой функциирешения R(x) со значением во множестве из четырех элементов ±1 ± µ для формулировки
(1) и ±θ, ±θ−1 для формулировки (2). Всего же рациональная функция степени n имеет
с учетом кратности 2n − 2 критических точки, поэтому решения рассматриваемых задач о
наименьшем уклонении удовлетворяют следующему определению с малым параметром g:
Определение 1 Рациональная функция R(x) называется g−экстремальной относительно
4-элементного множества значений Q, если все ее критические точки – за исключением g
из них – простые со значениями в Q. Число исключительных критических точек подсчитывается по формуле
X
X 1
ord dR(x) ,
(3)
g =1+
ord dR(x) +
2
x: R(x)6∈Q
x: R(x)∈Q
где суммирование производится по всем точкам сферы Римана; ord dR(x) – порядок нуля
дифференциала голоморфного отображения R : CP 1 → CP 1 в точке x (например, в простых
полюсах функции R(x) это число равно нулю), и [·] – целая часть числа.
Функции с малым числом экстремальности весьма специфичны и заполняют многообразия малой размерности в пространстве всех рациональных функций. При решении задач о
наименьшем уклонении возможно перейти от поиска по всему пространству рациональных
функций к поиску по этим многообразиям. Экстремальные рациональные функции обладают
следующим эффективным малопараметрическим представлением [1]:
!
Z
(x,w)
R(x) = sn
dζ + A(e) τ
,
Q = {±1, ±1/k(τ )},
(4)
(e,0)
где dζ – голоморфный дифференциал на римановой поверхности рода g, периоды которого
лежат в решетке периодов эллиптического синуса, и фазовый сдвиг A(e) также соизмерим с
этой решеткой. Эта поверхность определяется по рациональной функуции R(x) как двулистная накрывающая сферы Римана с ветвлением в точках, где R(x) принимает значения из Q с
нечётной кратностью. Возникаюшая риманова поверхнось не произвольна, это так называема кривая Калоджеро-Мозера: она разветвленно накрывает тор, определяемый по множеству
выделенных значений Q.
Используемая параметризация экстремальных функций позволяет контролировать поведение решения в переходных полосах фильтра (выбор класса) и непосредственно решать задачу о фильтре наименьшей степени для заданной спецификации без рассмотрения цепочки
задач о наименьшем уклонении. Вычислительные средства, используемые для нахождения
экстремальных рациональных функций по явной аналитической формуле, включают разработанный ранее [1, 9, 10, 11] аппарат эффективных вычислений на римановых поверхностях,
3
и позволяют устойчиво вычислять решения степени n равной по меньшей мере нескольким
сотням.
С использованием нового аналитического подхода были рассчитаны примеры решений
оптимизационной задачи 1. На Рис. 1 представлен график функции-решения степени 654
для спецификации с числом отрезков m = 23. На Рис. 2 изображен график решения (для
спецификации с m = 7) из класса, допускающего наличие полюсов в первой, второй, третьей
и шестой переходных полосах.
1.0
0.5
0.0
-0.5
-1.0
0.00
0.05
0.10
0.15
0.20
0.25
0.30
Рис. 1: График решения оптимизационной задачи (для спецификации с m = 23)
3
2
1
0
-1
-2
-3
0.00
0.05
0.10
0.15
Рис. 2: График решения оптимизационной задачи (для спецификации с m = 7)
Детальное изложение алгоритма, основанного на использовании явной аналитической формулы (4) будет приведено в отдельной работе, поскольку требует сложного математического
аппарата.
3
Прямая оптимизация
Решения оптимизационной задачи могут быть вычислены приближенно, например, алгоритмами типа Ремеза [13, 14, 7]. Отправной точкой снова служит теорема Ахиезера: оптимальная
4
рациональная функция φ(w)/ψ(w) степени n должна иметь на E по крайней мере 2n + 2 точки альтернанса. Предположим, что мы имеем приближенную дробь, заданную, например,
положением нулей и полюсов или коэффициентами числителя и знаменателя по некоторому
полиномиальному базису; и, кроме того, приближенные точки альтернанса, множество которых обозначим за A. Рассматриваемые алгоритмы состоят в поочередном уточнении каждого
из этих двух объектов:
1. имея множество точек A ⊂ E, можно изменить дробь так, чтобы на дискретном множестве A соблюдался альтернанс;
2. имея дробь, можно изменить положение точек с целью увеличения на A равномерной
нормы ошибки приближения – или же убедиться, что это невозможно, то есть равномерная норма ошибки не уменьшается при ограничении последней с E на A.
Кроме того, требуется начальное приближение, выбор которого из-за локальной сходимости
алгоритмов типа Ремеза весьма важен. Изложим кратко использованные нами рецепты для
всех указанных шагов.
Выбор начального приближения. Мы начинаем с выбора множества A, именно, используются 1/(2n + 2)-квантили равновесной меры для множества E, полученной при помощи
численного решения интегрального уравнения с логарифмическим ядром [16].
Шаг 2 (уточнение A по заданной дроби φ/ψ). Рассматривается сначала более широкое
множество, составленное из концов наших m интервалов и критических точек функции ошибки φ(w)/ψ(w)−F (w), лежащих на E. Точки, в которых отклонение меньше полученного ранее
на шаге 1, отбрасываются. Новое множество A должно подчиняться правилу перемены знаков и содержать 2n + 2 точки; наличие меньшего или большего их количества исправляется с
помощью введенной ранее равновесной меры; например, выбрасываются точки, соответствующие меньшим значениям плотности этой меры. Поиск критических точек осуществляется
методом Брента или, если приближение достаточно хорошее, проще — методом Ньютона.
Шаг 1 (уточнение дроби φ/ψ по заданному A). В отличие от полиномиальных приближений, условие на альтернанс на множестве A, по всей видимости, невозможно представить
в виде линейной системы; в лучшем случае возникают обобщенные задачи на собственные
значения, вида
V+ −V+
φ
0 Σ+ V+
φ
=µ
(5)
V− V−
ψ
0 Σ− V−
ψ
для первой формулировки экстремальной задачи, или вида
V−
0
φ
0 Σ− V−
φ
=θ
0 Σ + V+
ψ
V+
0
ψ
(6)
для второй формулировки (см. раздел 1). Последнюю задачу можно найти в [7]. В уравнениях (5–6) использованы обозначения: V± — матрица Вандермонда по узлам A, принадлежащим E± ; Σ± — диагональные матрицы с элементами ±1, фиксированными для каждого
из интервалов пропускания или задержки, по узлам A, принадлежащим E± ; φ, ψ — векторы коэффициентов числителя и знаменателя рациональной дроби; µ, θ — экстремальные
альтернирующие значения. Наличие матриц Σ± связано с возможным изменением знака альтернанса в соседних интервалах.
Как видно, матричные задачи (5–6) несимметричны, и соответствующие пучки могут оказаться сингулярными; даже в случае двух интервалов может оказаться, что стандартной мантиссы не хватит для вычисления хотя бы одного знака коэффициентов φ, ψ (не говоря уже
о корнях). Кроме того, перебор всех вариантов знаков альтернанса в соседних интервалах
означает экспоненциальную по m сложность.
Некоторым преимуществом, по сравнению с задачами на собственные значения, обладает редукция задачи (1), сформулированной относительно заданного множества A, к задаче
линейного программирования, причем и здесь возникает экспоненциальный по m перебор,
5
связанный с выбором знака знаменателя дроби на каждом интервале. Вот эта задача: потребуем минимизации величины t ∈ R при ограничениях
−tσ(w)ψ(w) ≤ σ(w)φ(w) − σ(w)ψ(w)F (w) ≤ tσ(w)ψ(w),
w ∈ A,
j = 1, . . . , m,
причем мы предполагаем signψ(w) =: σ(w) априори заданным на каждом интервале. Формально от последнего предположения можно освободиться некоторым увеличением числа
переменных. Хорошие результаты для такой задачи показывает метод внутренней точки для
объединенных прямой и двойственной формулировок [15].
Линейная параметризация числителей и знаменателей дроби существенно ограничивает
устойчивость алгоритмов типа Ремеза [13]. Максимальная степень решения R, полученная
таким методом при вычислениях с двойной точностью зависит от конфигурации множества
E и не превосходит n = 20.
4
Модульный подход
Модульный подход не связан с решением сформулированной оптимизационной задачи и предполагает получение (неоптимального) многополосного фильтра в результате соединения нескольких однополосных. Архитектурно это может осуществляться, например, путем параллельного
включения полосно-пропускающих фильтров, каждый из которых обслуживает одну рабочую полосу из заданной многополосной спецификации, либо путем каскадного соединения
полосно-пропускающих и полосно-заграждающих фильтров.
Неоптимальные фильтры, получаемые модульным подходом, ниже мы называем составными. Нами использовался следующий рецепт их построения: по каждой полосе пропускания
рассчитывалась передаточная функция соотвествующего полосового эллиптического фильтра, затем полученные функции складывались. Параметры эллиптических фильтров оптимизировались ручным перебором для достижения как можно меньшего порядка итогового
многополосного фильтра при условии, чтобы его АЧХ была вписана в заданный спецификацией коридор.
5
Примеры синтеза
В этом разделе приводятся примеры синтеза цифровых многополосных фильтров с помощью
трех подходов: нового аналитического подхода, метода прямой оптимизации и модульного
подхода. Первые два подхода, основываются на решении сформулированной выше оптимизационной задачи, и дают оптимальные фильтры (то есть порядка, минимального достижимого
для заданной спецификации), последний подход заключается в соединении однополосных эллиптических фильтров и дает неоптимальные многополосные фильтры.
Оптимальные цифровые фильтры получались из оптимальных аналоговых путем стандартного, так называемого билинейного, преобразования частоты.
Прямая оптимизация ни для одного из рассмотренных примеров не дала удовлетворительных результатов, ввиду сложности спецификаций. Её результаты приведены для первого и
второго примера. Модульный подход, как видно из примеров, дает фильтры с существенно
большими порядками, в сравнении с оптимальными.
5.1
Однополосный фильтр
С помощью нового аналитического подхода нами был рассчитан однополосный оптимальный
фильтр 18-ого порядка с сильно несимметричными ширинам переходных полос, равными
0.016 и 2 · 10−5 . График его АЧХ приводится на Рис. 3.
6
Magnitude response in dB
-2.0
-10
-20
-30
-40
-46.9
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 3: АЧХ однополосного оптимального фильтра
Стандартное частотное преобразование, используемое при синтезе полосового эллиптического фильтра из НЧ-прототипа, не может обеспечить такой разницы в ширинах переходных
полос, и эллиптический фильтр, рассчитанный по той же спецификации, имеет больший порядок – 28. График его АЧХ представлен на Рис. 4
Magnitude response in dB
-2.0
-10
-20
-30
-40
-46.9
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 4: АЧХ однополосного эллиптического фильтра
Результаты прямой численной оптимизации для такой спецификации приведены в Табл.
5.1. При расчетах с помощью алгоритма типа Ремеза границы полос фильтра задавались спецификацией, неравномерность в полосе пропускания фиксировалась равной 2 dB, а затухание
в полосах задержки (второй столбец в таблице) определялось, исходя из порядка получаемого оптимального фильтра. Затухание в полосах задержки в −46.9 dB для метода прямой
оптимизации оказалось недостижимым ввиду того, что уже для расчета фильтра 10 порядка
7
требовались большие временные затраты.
Порядок
3
4
5
6
8
9
10
Затухание в полосе задержки
-3.3812
-4.2272
-7.5894
-10.612
-19.355
-22.378
—
Число итераций
102
154
211
308
457
652
>1000
-46.9
?
18
Таблица 1: Результаты прямой численной оптимизации для однополосного фильтра.
5.2
Двухполосный фильтр
В качестве примера оптимального двухполосного фильтра с помощью аналитического подхода был рассчитан фильтр 16-ого порядка с минимальным затуханием в полосах задержки
в 40 dB, и величиной пульсаций АЧХ в полосах пропускания в 2.6 dB. Ширины переходных
полос: 0.012, 0.012, 0.003, 0.008. График АЧХ приведен на Рис. 5.
Magnitude response (dB)
-2.6
-10
-20
-30
-40.0
-50
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 5: АЧХ двухполосного оптимального фильтра
Построенный по той же спецификации составной фильтр имеет порядок 23. График его
АЧХ представлен на Рис. 6.
8
Magnitude response (dB)
-2.6
-10
-20
-30
-40.0
-50
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 6: АЧХ двухполосного составного фильтра
Результаты прямой оптимизации для такой спецификации представлены в Табл. 2. При
расчетах границы полос фильтра задавались спецификацией, неравномерность в полосе пропускания фиксировалась равной 2 dB, а затухание в полосах задержки (второй столбец в таблице) определялось, исходя из порядка получаемого оптимального фильтра. В данном случае
алгоритм типа Ремеза оказался неприменимым для поиска оптимальных фильтров начиная
уже с восьмого порядка, соответственно, оптимальный фильтр 16-ого порядка, найденный с
помощью аналитического подхода, метод прямой оптимизицации рассчитать не позволяет.
Порядок
3
4
5
6
7
8
16
Затухание в полосах задержки
-5.3421
-9.1785
-12.649
-16.013
-19.215
—
Число итераций
151
273
351
408
594
> 1000
-40
?
Таблица 2: Результаты прямой оптимизации для двухполосного фильтра.
5.3
Четырёхполосный фильтр
С помощью аналитического подхода нами был построен четырёхполосный оптимальный фильтр
36-ого порядка с минимальным затуханием в полосах задержки в 42.8 dB, неравномерностью
АЧХ в полосах пропускания в 2.0 dB, и ширинами переходных полос от 0.002 до 0.005. График
АЧХ полученного фильтра приведен на Рис. 7.
9
Magnitude response (dB)
-2.0
-10
-20
-30
-42.8
-50
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 7: АЧХ четырёхполосного оптимального фильтра
Рассчитанный по той же спецификации составной фильтр имеет порядок 55. Его АЧХ
приводится на Рис. 8.
Magnitude response (dB)
-2.0
-10
-20
-30
-42.8
-50
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 8: АЧХ четырёхполосного составного фильтра
5.4
Пятиполосный фильтр
Для пятиполосной спецификации был рассчитан оптимальный фильтр 76-ого порядка с минимальным затуханием АЧХ в полосах задержки в 50 dB, и неравномерностью в полосах
пропускания в 2.0 dB. Ширины переходных полос лежат в промежутке от 0.002 до 0.005.
График АЧХ представлен на Рис. 9.
10
Magnitude response (dB)
-2.0
-10
-20
-30
-40
-50.0
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 9: АЧХ четырёхполосного оптимального фильтра
Построенный по той же спецификации составной фильтр будет имеет 121. Его АЧХ приводится на Рис. 10
Magnitude response (dB)
-2.0
-10
-20
-30
-40
-50.0
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 10: АЧХ четырёхполосного составного фильтра
5.5
Фильтр режекторного типа с двумя полосами задержки
График АЧХ оптимального фильтра 16-ого порядка режекторного типа, осуществляющего
точное вырезание двух заданных частот, представлен на Рис. 11. Участок АЧХ, содержащий
вырезаемые частоты, в большем масшатабе представлен на Рис. 12.
11
0
Magnitude response (dB)
-10
-20
-30
-40
-51.9
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 11: АЧХ оптимального фильтра режекторного типа с двумя полосами задержки.
0
Magnitude response (dB)
-10
-20
-30
-40
-51.9
-60
-70
-80
-90
-100
0.255
0.256
0.257
0.258
Normalized frequency
0.259
0.260
Рис. 12: Участок АЧХ, содержащий вырезаемые частоты.
Составной фильтр, обеспечивающий такое же качество аппроксимации АЧХ, имеет 62-ой
порядок.
5.6
Фильтр с двумя полосами пропускания, расположенными критически близко друг к другу
С помощью аналитического подхода был рассчитан оптимальный фильтр 24-ого порядка с
двумя полосами пропускания, расположенными критически близко друг к другу. Соответствующий график АЧХ приводится на Рис. 13. Подобные фильтры могут быть использованы,
12
если требуется выделить из спектра некоторую частотную полосу и при этом задержать одну
или несколько частот, сорежащихся внутри этой полосы.
Magnitude response (dB)
-2.0
-10
-20
-30
-40
-50.0
-60
-70
-80
-90
-100
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Рис. 13: АЧХ оптимального фильтра с двумя полосами пропускания, расположенными критически близко друг к другу.
Magnitude response (dB)
-2.0
-10
-20
-30
-40
-50.0
-60
-70
-80
-90
-100
0.250
0.251
0.252
0.253
Normalized frequency
0.254
0.255
Рис. 14: Участок АЧХ, содержащий вырезаемую частоту.
Составной фильтр, обеспечивающий такое же качество аппроксимации АЧХ, имеет 59-ый
порядок.
13
6
Заключение
В статье приведено сравнение трех подходов к синтезу многополосных фильтров: нового аналитического подхода, прямой численной оптимизации на основе метода Ремеза и полуаналитический модульный подход. Прямая численная оптимизация сегодня наиболее алгоритмически проработана: существуют и совершенствуются готовые пакеты для инженерных расчетов. К сожалению, неустранимая неустойчивость алгоритмов ремезовского типа не позволяет решать слишком сложные задачи: при использовании двойной точности (15 десятичных
знаков) порядок фильтра не доходит до 20, и при этом невозможно достичь хороших аппроксимационных свойств в случае сложных спецификаций, например, при большом числе полос
пропускания и задержки, узких переходных полосах, полосах пропускания, расположенных
критически близко друг к другу. Модульный подход состоит в разбиении сложной задачи
на ряд простых и их последовательном решении с использованием дроби Золотарева для
построения АЧХ полосового фильтра. Его преимущество состоит в том, что таким образом
можно гарантированно получить (эрзац-) решение с заданной спецификацией. Как правило,
оно далеко от оптимального: порядок составного фильтра может в несколько раз превышать
порядок оптимального с той же спецификацией. Ситуация усугубляется с ростом сложности
спецификации фильтра. Наиболее перспективным, с нашей точки зрения, как и наименее изученным с алгоритмической точки зрения, является аналитический подход, опирающийся на
сложный математический аппарат. Авторы намереваются продолжать исследования в этом
направлении.
7
Приложение
Phase response (rad)
На Рис. 15, 16, 17 приводятся графики фазо-частотной и импульсной характеристики, а также
группового времени задержки оптимального четырёхполосного фильтра из примера 5.3. Его
амплитудно-частотная характеристика представлена на Рис. 7.
5
0
-5
-10
-15
0.0
0.1
0.2
0.3
Normalized frequency
0.4
Рис. 15: Фазо-частотная характеристика четырёхполосного оптимального фильтра.
14
0.5
Group delay (samples)
250
200
150
100
50
0
0.0
0.1
0.2
0.3
Normalized frequency
0.4
0.5
Amplitude
Рис. 16: Групповое время задержки четырёхполосного оптимального фильтра.
0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
0
50
100
150
Sample number
200
250
300
Рис. 17: Импульсная характеристика четырёхполосного оптимального фильтра (точки графика соединены с квадратичной интерполяцией).
Список литературы
[1] A.B. Bogatyrev, Chebyshev representation for rational functions, Sb. Math., 201:11 (2010),
pp. 1579–1598.
[2] A. Mohan, S. Singh, A. Biswas, Generalized synthesis and design of symmetrical multiple
passband filters, Progress In Electromagnetics Research B, Vol. 42, pp. 115-139, 2012.
[3] J. Lee, and K. Sarabandi, Design of triple-passband microwave filters using frequency
transformation, IEEE Trans. On Microwave Theory and Tech., vol. MTT-56, No. 1, pp
187-193, Jan. 2008.
[4] V. Lunot, S. Bila, and F. Seyfert, Optimal synthesis for multi-band microwave filters, IEEE
MTT-S Int. Microw. Symp. Dig., pp. 115–118, Jun. 2007.
[5] D. Deslandes and F. Boone, An iterative design procedure for the synthesis of generalized
dual-bandpass filters, Int.J. RF Microw. Computer-Aided Eng., vol. 19, no. 5, pp. 607-614,
2009.
[6] G. Macchiarella, "Equi-ripple" Synthesis of Multiband Prototype Filters Using a Remez-Like
Algorithm, IEEE Microwave And Wireless Components Letters, Vol. 23, No. 5, pp. 231-233,
May 2013.
15
[7] R.A.-R. Амer, The approximation problem of electrical filters, Basel-Stuttgart, Birkhauser,
1964.
[8] V. Crnojevic-Bengin, Advances in Multi-Band Microstrip Filters, Cambridge University
Press, 2015.
[9] A.B. Bogatyrev, Computations in moduli spaces, Computational Methods and Function
Theory, 7 (2007), No. 2, 309-324.
[10] А.Б. Богатырев Экстремальные многочлены и римановы поверхности, М., МЦНМО,
2005. (монография). A.B. Bogatyrev, Extremal Polynomials and Riemann Surfaces, Springer
Monographs in Mathematics, ISBN 978-3-642-25633-2 (print).
[11] А.Б. Богатырев, Конформное отображение прямоугольных семиугольников, Математический сборник 203:12 (2012), стр 35-56.
[12] А.А. Гончар, О задачах Е.И.Золотарева, связанных с рациональными функциями, Математический сборник 78(120):4 (1969), стр 640-654.
[13] Е.Я. Ремез Основы численных методов чебышевского приближения, Киев, Наукова
думка, 1969.
[14] L. Veidinger, On the numerical determination of the best approximations in the Chebyshev
sense, Numer. Math., Vol. 2, pp. 99-105, 1960.
[15] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[16] W. Fuchs, On Chebyshev approximation on sets with several components // D.A. Brannan
and J.G. Clunie, eds., Aspects of Contemporary Complex Analysis, pp. 399-408, Academic
Press, 1980.
[17] Е.И. Золотарев, Приложение эллиптических функций к вопросам о функциях, наименее и наиболее отклоняющихся от нуля, Зап. Росс. Акад Наук, ХХХ стр. 37-41, 54-71,
1877.
[18] W. Cauer, Theorie der linearen Wechselstromschaltungen, Bd. 1. Becker und Erler, Leipzig,
1941; Bd. 2. Akademie, Berlin, 1960.
[19] В.Н. Малоземов, Задача синтеза многополосного электрического фильтра, Ж. вычисл.
матем. и матем. физ., 19:3 (1979), 601–609.
[20] Н.И. Ахиезер, Об одной задаче Е.И. Золотарева, Изв. Акад. наук СССР, VII серия,
1929, №10, стр. 919–931.
Институт вычислительной математики РАН,
Москва, 119991, ул. Губкина, 8
[email protected],
[email protected],
[email protected]
16
| 3cs.SY
|
Bloofi: Multidimensional Bloom Filters
Adina Crainiceanua,∗, Daniel Lemireb
b LICEF
a US Naval Academy, USA
Research Center, TELUQ University of Quebec, Canada
arXiv:1501.01941v3 [cs.DB] 21 Sep 2016
Abstract
Bloom filters are probabilistic data structures commonly used for approximate membership problems in many areas of
Computer Science (networking, distributed systems, databases, etc.). With the increase in data size and distribution of
data, problems arise where a large number of Bloom filters are available, and all them need to be searched for potential
matches. As an example, in a federated cloud environment, each cloud provider could encode the information using
Bloom filters and share the Bloom filters with a central coordinator. The problem of interest is not only whether a
given element is in any of the sets represented by the Bloom filters, but which of the existing sets contain the given
element. This problem cannot be solved by just constructing a Bloom filter on the union of all the sets. Instead, we
effectively have a multidimensional Bloom filter problem: given an element, we wish to receive a list of candidate
sets where the element might be.
To solve this problem, we consider 3 alternatives. Firstly, we can naively check many Bloom filters. Secondly, we
propose to organize the Bloom filters in a hierarchical index structure akin to a B+ tree, that we call Bloofi. Finally,
we propose another data structure that packs the Bloom filters in such a way as to exploit bit-level parallelism, which
we call Flat-Bloofi.
Our theoretical and experimental results show that Bloofi and Flat-Bloofi provide scalable and efficient solutions
alternatives to search through a large number of Bloom filters.
Keywords: Bloom filter index, multidimensional Bloom filter, federated cloud, data provenance
1. Introduction
of Bloom filters, and these filters need to be searched to
find the sets containing particular objects.
Bloom filters [3] are used to efficiently check whether
an object is likely to be in the set (match) or whether
the object is definitely not in the set (no match).
False positives are possible, but false negatives are
not. Due to their efficiency, compact representation,
and flexibility in allowing a trade-off between space
and false positive probability, Bloom filters are popular in representing diverse sets of data. They are
used in databases [23], distributed systems [5], web
caching [14], and other network applications [4]. For
example, Google BigTable [6] and Apache Cassandra [30] use Bloom filters to reduce the disk lookups for
non-existent data. As digital data increases in both size
and distribution, applications generate a large number
Our work is motivated by highly distributed data
provenance applications, in which data is tracked as it is
created, modified, or sent/received between the multiple
sites participating in the application, each site maintaining the data in a cloud environment. Bloom filters can
be maintained by each individual site and shared with
a central location. For each piece of data, we need to
find the sites holding the data. Thus, we may need to
search through a large number of Bloom filters stored at
the central location.
∗ Corresponding author.
Tel.: 00+1+ 410 293-6822; fax:
00+1+410 293-2686.
Email addresses: [email protected] (Adina Crainiceanu),
[email protected] (Daniel Lemire)
Preprint submitted to Information Systems
Indexing Bloom filters is different than indexing
generic objects to improve search time. There is one
level of indirection between the elements searched for,
and the objects directly indexed by the index structure.
In particular, each Bloom filter is a compact representation of an underlying set of elements. The question of
interest is an all-membership query: given a particular
element (not a Bloom filter), which underlying sets contain that element? The query subject is an element, but
September 22, 2016
the objects we are indexing and searching through are
Bloom filters, so what we are creating is a meta-index.
The traditional index structures, such as hash indexes,
B+trees, R trees etc. and their distributed versions [1]
do not directly apply in this case as we are indexing
Bloom filters and not the keys themselves. All we are
given from each site is a Bloom filter.
There has been significant work in using Bloom filters in various applications, and developing variations
of Bloom filters. Counting filters [14, 15] support deletions from the Bloom filter; compressed Bloom filters [21] are used with web caching; stable Bloom filters [11] eliminate duplicates in streams, spectral Bloom
filters [7] extend the applicability of Bloom filters to
multi-sets, multi-class Bloom Filter (MBF) [20] use perelement probabilities. Yet there has been few attempts
to accelerate queries over many Bloom filters, what we
call the multidimensional Bloom filter problem, even
though our problem is closely related to signature file
methods (see Section 8) where one seeks to index setvalue attributes.
To solve this problem, we propose Bloofi (Bloom
Filter Index), a hierarchical index structure for Bloom
filters. Bloofi provides probabilistic answers to allmembership queries and scales to tens of thousands of
Bloom filters. When the probability of false positives
is low, Bloofi of order d (a tunable parameter) can provide O(d logd N) search cost, where N is the number
of Bloom filters indexed. Bloofi also provides support
for inserts, deletes, and updates with O(d logd N) cost
and requires O(N) storage cost. In designing Bloofi,
we take advantage of the fact that the bitwise OR between Bloom filters of same length, constructed using
the same hash functions, is also a Bloom filter. The
resulting Bloom filter represents the union of the sets
represented by the individual Bloom filters. This property allows us to construct a tree where the leaf levels
are the indexed Bloom filters, and the root level is a
Bloom filter that represents all the elements in the system. This tree is used to prune the search space (eliminate Bloom filters as candidates for matches) while processing all-membership queries. Our performance evaluation shows that Bloofi performs best when the false
positive probability of the union Bloom filter (a Bloom
filter that is the union of all the indexed Bloom filters) is
low and provides O(d × logd N) search performance in
most cases, with O(N) storage cost and O(d × logd N)
maintenance cost. Bloofi could be used whenever a
large number of Bloom filters that use the same hash
functions need to be checked for matches.
Bloom filters are constructed over bitmaps, i.e., vector of Booleans. With bitmaps, we can exploit bit-level
parallelism: on a 64-bit processor, we can compute the
bitwise or between 64 bits using a single instruction. We
use bit-level parallelism with Bloofi to optimize the construction of the data structure. However, we have also
designed an alternative data structure that is designed
specifically to exploit bit-level parallelism (henceforth
Flat-Bloofi). Though not as scalable as Bloofi, it can be
fast when the number of Bloom filters is moderate.
This article is an extended version of “Bloofi: A
Hierarchical Bloom Filter Index with Applications to
Distributed Data Provenance” [9] published in the 2nd
International Workshop on Cloud Intelligence Cloud-I
2013. The paper was completely revised, and the new
version introduces an additional data structure, FlatBloofi, a new implementation for Bloofi that improves
the performance by an order of magnitude, and a new
performance evaluation.
The rest of this paper is structured as follows: Section 2 describes a distributed data provenance application for Bloofi. Section 3 briefly reviews the concept
of Bloom filter. Section 4 introduces Bloofi, a hierarchical index structure for Bloom filters. Section 5 introduces the maintenance algorithms and a theoretical
performance analysis. Section 6 introduces Flat-Bloofi,
a data structure for the multidimensional Bloom filter
problem, designed to exploit bit-level parallelism. Section 7 shows the experimental results. We discuss related work in Section 8 and conclude in Section 9.
2. Motivation:
provenance
application to distributed data
In this section we describe the distributed data provenance application that motivated our work on Bloofi.
Let us assume that a multinational corporation with hundreds of offices in geographically distributed locations
(sites) around the world is interested in tracking the documents produced and used within the corporation. Each
document is given a universally unique identifier (uuid)
and is stored in the local repository, in a cloud environment. Documents can be sent to another location
(site) or received from other locations, multiple documents can be bundled together to create new documents,
which therefore are identified by new uuids, documents
can be decomposed in smaller parts that become documents themselves, and so on. All these “events” that are
important to the provenance of a document are recorded
in the repository at the site generating the event. The
events can be stored as RDF triples in a scalable cloud
triple store such as Rya [25]. The data can be modeled as a Directed Acyclic Graph (DAG), with labeled
edges (event names) and nodes (document uuids). As
2
documents travel between sites, the DAG is in fact distributed not only over the machines in the cloud environment at each site, but also over hundreds of geographically distributed locations. The data provenance problem we are interested in solving is finding all the events
and document uuids that form the “provenance” path of
a given uuid (all “ancestors” of a given node in the distributed graph).
Storing all the data, or even all the uuids and their location, in a centralized place is not feasible, due to the
volume and speed of the documents generated globally.
Fully distributed data structures, such as Chord [29] or
P-Ring [10], require even more communication (messages) than the centralized solution, increasing the latency and bandwidth consumption, so they are also not
feasible due to the volume and speed of the documents
generated globally. Moreover, local regulations might
impose restrictions on where the data can be stored.
However, since all the global locations belong to the
same corporation, data exchange and data tracking must
be made possible.
Without any information on the location of a uuid,
each provenance query for a uuid must be sent to all
sites. Each site can then determine the local part of
the provenance path, and return it. However, the provenance might contain new uuids, so a new query needs to
be sent to each site for each new uuid connected to the
original uuid, until no new uuids are found. This recursive process could consume significant bandwidth and
latency in a geographically distributed system.
To minimize the number of unnecessary messages
sent to determine the full provenance of an object, each
local site maintains a Bloom filter of all the uuids in the
local system. Updates to the Bloom filter are periodically propagated to a centralized location (the headquarters for example). Since the Bloom filters are compact
representations of the underlying data, less bandwidth
is consumed when Bloom filters are sent versus sending the actual data. At the central location, a Bloofi
index is constructed from the Bloom filters, and every
time a provenance query for a uuid is made, the Bloofi
index is used to quickly determine the sites that might
store provenance information for the given uuid. If the
query load is too high and having the Bloofi index in a
single location affects performance and/or availability,
multiple Bloofi indexes could be constructed in several
locations.
structure is stored on disk. When we expect many requests for values that are not present, it is helpful to
have an auxiliary data structure that will quickly dismiss
these requests. Bloom filters can serve this purpose.
Each Bloom filter [3] is a bit array (or bitmap) of
length m constructed by using a set of k hash functions.
The empty Bloom filter has all bits 0. To add an element
to the filter, each of the k hash functions maps the new
element to a position in the bit array. The bit in that position is turned to 1. To check whether an element is a
member of the set represented by the Bloom filter, the
k hash functions are applied to the test element. If any
of the resulting k positions is 0, the test element is not
in the set, with probability 1. If all k positions are 1,
the Bloom filter matches the test element, and the test
element might be in the set (it might be a true positive
or a false positive). There is a trade-off between the size
of the Bloom filter and the probability of false positives,
pfalse , returned by it. We have that pfalse ≈ (1 − e−kn/m )k
assuming that there are n elements in the filter [22]; the
probability is minimal when k = m/n ln 2. The probability pfalse can be lowered by increasing the size of the
Bloom filter (m). For a fixed number of elements (n), the
probability goes to zero exponentially. In the rest of the
paper we assume that all the Bloom filters indexed have
the same length and use the same set of hash functions.
4. Indexing Bloom filters
We are given a collection of N Bloom filters B. Given
a query for a value, we want to find all Bloom filters that
are a match. We are most interested in the case where
there are relatively few such Bloom filters and when a
sequential search is potentially inefficient.
4.1. Bloofi: a hierarchical Bloom filter index
Bloofi, the Bloom filter index, is based on the following idea. We construct a tree: the leaves of the tree are
the Bloom filters to be indexed, and the parent nodes
are Bloom filters obtained by applying a bitwise OR on
the child nodes. This process continues until the root is
reached. The index has the property that each non-leaf
Bloom filter in the tree represents the union of the sets
represented by the Bloom filters in the sub-tree rooted
at that node. As a consequence, if an object matches a
leaf-level Bloom filter, it matches all the Bloom filters
in the path from that leaf to the root. Conversely, if a
particular Bloom filter in Bloofi does not match an object, there is no match in the entire sub-tree rooted at
that node.
Using Bloofi, a membership query starts by first
querying the root Bloom filter: if it does not match the
3. Bloom filters
Checking for the presence of a value in a set or similar data structure can be expensive, especially if the data
3
queried object, then none of the indexed sets contain
the object and a negative answer is returned. If the root
does match the object, the query proceeds by checking
which of the child Bloom filters matches the object. The
query continues down the path of Bloom filters matching the object, until the leaf level is reached. In a balanced tree, the height of Bloofi is logarithmic in the
number of Bloom filters indexed, and each step in the
query process goes down one level in the tree. In the
best case, a query with a negative answer is answered in
constant time (check the root only), and a query with a
positive answer is answered in logarithmic time. However, if multiple paths in the index are followed during
the query process, the query time increases. Section 5.1
introduces our heuristics for Bloofi construction such
that the number of “misleading” paths in the Bloofi is
reduced (similar Bloom filters are in the same sub-tree).
There are many possible implementations for Bloofi:
as a tree similar with binary search trees, AVL trees,
B+ trees, etc. Due to the flexibility allowed by having a
balanced tree with the number of child pointers higher
than two, we implement Bloofi like a B+ trees. We
start with an order parameter d, and each non-leaf node
maintains l child pointers: d ≤ l ≤ 2d for all non-root
nodes, and 2 ≤ l ≤ 2d for the root. Each node in Bloofi
stores only one value, which is different than general
search trees. For the leaves, the value is the Bloom filter
to be indexed. For all non-leaf nodes, the value is obtained by applying bitwise OR on the values of its child
nodes. Throughout the paper we use the usual definitions for tree, node in a tree, root, leaf, depth of a node
(the number of edges from root to the node), height of a
tree (maximum depth of a node in the tree), sibling, and
parent.
Fig. 1 shows an example of a Bloofi index of order 2.
Each internal node has between 2 and 4 child pointers.
The leaf level of the tree contains the original Bloom filters indexed by the Bloofi index, with identifiers 1, 2, 3,
4, 5, and 6. The node identifiers are shown here for ease
of presentation, but they are also used in practice during
updates and deletes to identify the node that needs to be
updated or deleted (see Section 5.2 and Section 5.3 for
details). In the rest of the paper, we often use “node X”
to refer to the node with identifier X. At the next higher
level, the values are obtained by applying bitwise OR
on the values of the children nodes, so the value of the
node 7 is the bitwise OR between values of nodes 1, 2,
3, and 4. The process continues until the root is reached.
We use the following notation. Given a node, its
parent is written node.parent. The ordered set of its
children is written by node.children. The number of
children is written node.nbDesc with the convention
Id:9
1 1 1 1 1 0 1 0
Id:7 1 1 1 1 0 0 0 0
1
0
0
0
0
0
0
0
Id:1
0
1
0
0
0
0
0
0
Id:2
0
0
1
0
0
0
0
0
Id:3
0
0
0
1
0
0
0
0
Id:4
0 0 0 0 1 0 1 0 Id:8
0
0
0
0
1
0
0
0
Id:5
0
0
0
0
0
0
1
0
Id:6
Figure 1: Bloofi Tree of Order 2
that node.nbDesc is zero for a leaf node. For i =
0, . . . node.nbDesc − 1, node.children[i] is the (i + 1)th
children of the node.
Each node has a corresponding bit array indicated by
node.val. We let | be the bitwise OR operation so that
when node is not a leaf, the bit array is just the aggregation of the bit array of the children:
node.val =node.children[0].val node.children[1].val · · ·
node.children[node.nbDesc − 1].val.
Individual bit values in the bit array are accessed as
node.val[i] for i = 0, . . . , m.
Since Bloofi is a balanced tree, with each internal
node having at least d child nodes, where d is the order of the tree, the height of a Bloofi index of order d is
at most blogd Nc, where N is the number of Bloom filters
to index.
4.2. Search
The search algorithm (Algorithm 1) returns the identifiers of all the leaf-level Bloom filters that match a
given object in the subtree rooted at the given node. It
first checks whether the current node value matches the
object (line 3). If not, then none of the Bloom filters
in that sub-tree match the object, so the empty set is
returned (line 4). If the current node does match the object, then either it is a leaf, in which case it returns the
identifier (line 8), or it is an inner node, in which case
the findMatches function is called recursively for all
of its child nodes (lines 12–13).
Example. Consider a query for object value 4 in
the Bloofi tree in Figure 1. The findMatches function in Algorithm 1 is invoked with arguments root
4
Algorithm 1 : findMatches(node,o)
5. Bloofi maintenance
1: //RETURN VALUE: the identifiers of leaves in the sub-
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
We introduce the algorithms for inserting, deleting,
and updating Bloom filters.
tree rooted at node with Bloom filters matching the object o
//if node does not matches the object, return empty set,
else check the descendants
if not match(node.val,o) then
return ∅;
else
//if this node is a leaf, just return the identifier
if node.nbDesc = 0 then
return node.id;
else
//if not leaf, check the descendants
returnList = ∅;
for i = 0; i < node.nbDesc; i++ do
returnList.add(findMatches(node.children[i],o));
return returnList;
5.1. Insert
Ideally, when inserting Bloom filters, we would like
to keep them in partitions so that the overlap between
different partitions is small. That is, we would like similar Bloom filters to be grouped together as much as
possible. And, conversely, we would like Bloom filters from different partitions to be as different as possible. Such problems are commonly NP-hard (e.g., the
Minimum Graph Bisection Problem) though they can be
sometimes approximated efficiently. We leave a more
formal investigation of this problem to future work and
use a heuristic.
The algorithm for inserting (Algorithm 2) finds a leaf
which is “close” to the input Bloom filter in a given metric space, and inserts the new Bloom filter next to that
leaf. The intuition is that similar Bloom filters should
be in the same sub-tree to improve search performance.
As distance metric, we use the Hamming distance. That
is, we count the number of bits that differ. This can
be computed quickly by computing the cardinality of
the bitwise exclusive OR of two bit arrays along with
fast functions to count the number of 1s in the resulting words (e.g., Long.bitCount in Java). We could
consider other distance metric, and we experiment with
Cosine and Jaccard metrics in Section 7.2.6.
The new Bloom filter is inserted by first updating the
value of the current node by computing the bitwise or
with the value of the filter to be inserted (since that node
will be in the sub-tree), and then recursively calling the
insert function on the child node most similar with the
new value (line 10). Once the most similar leaf node is
located, a new leaf is created for the new Bloom filter
(line 33) and is inserted as a sibling of the node by calling the insertIntoParent function (Algorithm 3).
This function takes as parameters the new node newEntry, and the most similar node to it, node. We insert
newEntry as a sibling of node. If the number of children
in the parent is still at most 2d, the insert is complete.
If an overflow occurs, the node splits (lines 9-16) and
the newly created node is returned by the function. The
splits could occasionally propagate up to the root level.
In that case, a new root is created, and the height of the
Bloofi tree increases (line 18).
Example. Consider the Bloofi tree in Fig. 1 and
assume that we insert the Bloom filter with value
“00100100”. The new node is inserted as a child of node
7 (Hamming distance between the new node and nodes
and 4. In line 3 of Algorithm 1, the algorithm checks
whether the value of the root matches 4. For simplicity of presentation, assume that there is only one
hash function used by the Bloom filters, the function
is h(x) = x mod 8, and the elements in the underlying set are integers. Since root.val[4] is 1, the root
matches the queried object 4 and the search proceeds
by invoking the findMatches function for each of
its child nodes. The first child node, node 7, does not
match the queried object, so findMatches for that
sub-tree returns ∅. The second child node of the root,
node 8, matches 4, so the search continues at the next
lower level. Node 5 matches the queried object, and is a
leaf, so the findMatches function returns the identifier 5. Leaf 6 does not match the queried object, so that
findMatches call returns ∅. Now, the recursive call
on node 8 returns with the value {5}, and finally the call
on the root returns {5}, which is the result of the query.
Search cost. The complexity of the search process is
given by the number of findMatches invocations.
In the best case, if there are no leaf-level Bloom filters
matching a given object, the number of Bloom filters to
be checked for matches is 1 (the root). To find a leaflevel Bloom filter that matches a query, the number of
findMatches invocations is O(d logd N) in the best
case (one path is followed, and at each node, all children are checked to find the one that matches) and O(N)
in the worst case (since the maximum number of nodes
in a Bloofi tree is dN + (N − 1)/(d − 1)e, the search cost
is O(N) if all nodes need to be checked for matches).
5
Algorithm 3 :
node)
Algorithm 2 : insert(newBloomFilter, node)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
// D is a distance function between bit arrays
//insert into the sub-tree rooted at the given node
//RETURN: null or pointer to new child if split occurred
//if node is not leaf, direct the search for the new filter
place
if node.nbDesc > 0 then
//update the value of the node to contain the new filter
node.val = node.val newBloomFilter;
//find the most similar child and insert there
find child C minimizing D(C.val, newBloomFilter)
newSibling = insert(newBloomFilter, C);
//if there was no split, just return null
if newSibling = null then
return null;
else
//there was a split; check whether a new root is
needed
if node.parent = null then
//root was split; create a new root
newRoot = new BFINode() ; // create new node
newRoot.val = node.val newSibling.val;
newRoot.parent = null;
newRoot.children.add(node);
newRoot.children.add(newSibling);
root = newRoot;
node.parent = newRoot;
newSibling.parent = newRoot;
return null;
else
newSibling
=
insertIntoParent(newSibling, node)
insertIntoParent(newEntry,
//insert into the node’s parent, after the node pointer
//RETURN: null or pointer to new child if split occurred
node.parent.children.addAfter(newEntry, node);
newEntry.parent = node.parent;
//check for overflow
if node.nbDesc > 2d then
return null;
else
//overflow, so split
P = node.parent;
P0 = new BFINode();
move last d children from P to P0 ;
update parent information for all children of P0 ;
re-compute P.val as the OR between its children values;
15:
compute P0 .val as the OR between its children values;
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
16:
return P0 ;
Id:9
1 1 1 1 1 1 1 0
Id:7
1 1 1 0 0 0 0 0
return newSibling;
//current node is root or not
//there was a split or not
else
//if node is leaf, need to insert into the parent
//create a node for newBloomFilter
newLeaf = new BFINode();
newLeaf.val = newBloomFilter;
//insert the new leaf into the parent node
newSibling = insertIntoParent(newLeaf, node);
1
0
0
0
0
0
0
0
Id:1
return newSibling;
//current node is leaf or not
0
1
0
0
0
0
0
0
Id:2
0
0
1
0
0
0
0
0
Id:3
Id:11
0 0 1 1 0 1 0 0
0
0
1
0
0
1
0
0
Id:10
0
0
0
1
0
0
0
0
Id:4
Id:8
0 0 0 0 1 0 1 0
0
0
0
0
1
0
0
0
Id:5
0
0
0
0
0
0
1
0
Id:6
Figure 2: Bloofi Tree After Insert and Split
6
7 and 8 is 4, so let’s assume that node 7 is chosen as the
closest node), which needs to split. The resulting Bloofi
tree is shown in Fig. 2.
Theorem 1 gives the cost of the insert algorithm for
Bloofi. The cost metric we use to measure the performance of an operation in Bloofi is the number of Bloofi
nodes accessed by that operation: either the Bloom filter
value is read/modified by that operation, or the parent or
children pointers in the node are read/modified.
a sibling. If redistribution is possible, the entries are redistributed, parent information in the moving nodes is
updated, and the Bloom filter values in the parent node,
sibling node, and all the way up to the root are updated
(lines 15–21). If redistribution is not possible, the parent node merges with a sibling, by giving all its entries
to the sibling (lines 23–29). The Bloom filter value in
the sibling node is updated, and the delete procedure is
called recursively for the parent node. Occasionally, the
delete propagates up to the root. If only one child remains in the root, the root is deleted and the height of
the tree decreases (lines 7–9).
Example. Assume that the node with id 5 and value
“00001000” is deleted from Fig. 1. The resulting tree,
after deletion of node 5 and redistribution between 8 and
7 is shown in Fig. 3.
Theorem 1 (Insert Cost). The number of Bloofi nodes
accessed during the insert operation in Bloofi is
O(d logd N), where d is the order of the Bloofi index and
N is the number of Bloom filters that are indexed.
P ROOF. The following components are part of an insert
operation:
Theorem 2 (Delete Cost). The number of Bloofi nodes
accessed during the delete operation in Bloofi is
O(d logd N), where d is the order of the Bloofi index and
N is the number of Bloom filters that are indexed.
1. The values of all nodes in the path from the root
to the new leaf are updated to reflect the newly inserted Bloom filter. Updating the value of a node
means computing the OR with the newly inserted
Bloom filter, so only the old and the new Bloom filter values need to be accessed. The cost of the update is therefore constant (2). The height of Bloofi
tree is at most blogd Nc and at each level we perform a constant amount of work, so the total cost
for that update is O(logd N).
P ROOF. (1) Once the reference to a node is deleted
from its parent in the Bloofi tree (constant cost operation), the values of the nodes from the deleted node
to the root need to be recomputed, so the total cost is
O(d logd N). (2) Occasionally, a redistribute or merge is
needed, with a cost in O(d). In the worst case, the merge
and delete propagates up to the root, so the worst case
cost for merge is O(d logd N). From (1) and (2) it follows that the cost of the delete operation is O(d logd N).
2. At each level in the tree, a search for the most similar child node is performed (line 9), and the cost of
that search is O(d), so total cost due to the search
for the placement of the new leaf is O(d logd N).
When subjected to many insertions and deletions,
though the Bloofi tree remains balanced, it could be that
the partition of the Bloom filters as per their Hamming
distance could degrade in quality. The performance of
Bloofi could diminish in such instances and it could become necessary to reconstruct the Bloofi data structure.
3. The cost of a split is O(d) since there are at most
2d + 1 children for a node. In the worst case, the
split propagates up to the root, and the height of
the tree is O(logd N), so the worst case cost for the
split operations is O(d logd N).
5.3. Update
Object insertions in the underlying set lead to updates
of the Bloom filters, so we expect the update operation
for Bloofi to be quite frequent. Instead of treating a
Bloom filter update as a delete followed by insert, we
use an “in-place” update. If the Bloofi tree becomes inefficient in routing due to updates (too many false positives during search) the Bloofi tree can be reconstructed
from scratch in batch mode. Algorithm 5 shows the
pseudo-code for the update algorithm. The algorithm
takes as parameters the leaf node corresponding to the
updated value and the new Bloom filter value for that
node. All the Bloom filters in the path from the leaf to
the root are updated by OR-ing with the new value.
From the three points, we see that the cost of the insert
operation is O(d logd N).
5.2. Delete
The delete algorithm (Algorithm 4) deletes the given
node from the Bloofi index. When the procedure is first
invoked with the leaf to be deleted as argument, the
pointer to that leaf in the parent is deleted. If the parent
node is not underflowing (at least d children are left),
the Bloom filter values of the nodes in the path from the
parent node to the root are re-computed to be the bitwise
OR of their remaining children (line 32) and the delete
procedure terminates. If there is an underflow (lines 11–
29), the parent node tries to redistribute its entries with
7
Algorithm 4 : delete(childNode)
//find the parent node
parentNode = childNode.parent;
//remove the reference to the node from its parent
parentNode.children.remove(childNode);
5: //check whether the tree height needs to be reduced
1:
2:
3:
4:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
Id:9
1 1 1 1 0 0 1 0
if parentNode = root AND parentNode.nbDesc = 1
then
root = parentNode.children.get(0);
root.parent = null;
return null
//if not, check if underflow at the parent
if parentNode.underflow then
//underflow, so try to redistribute first
sibling = sibling of parentNode;
if sibling.canRedistribute then
//redistribute with sibling
remove some children from sibling to even out
the number of children
insert new children into parentNode
update .parent information for all nodes moved
Id:7 1 1 1 0 0 0 0 0
1
0
0
0
0
0
0
0
Id:1
0
1
0
0
0
0
0
0
Id:2
0
0
1
0
0
0
0
0
Id:3
0 0 0 1 0 0 1 0 Id:8
0
0
0
1
0
0
0
0
Id:4
0
0
0
0
0
0
1
0
Id:6
Figure 3: Bloofi Tree After Delete and Redistribute
//update value of all nodes involved, up to the
root
sibling.val = OR of all children value;
recomputeValueToTheRoot(parentNode);
else
//merge with sibling
move all children from parentNode to sibling;
update .parent information for all nodes moved
Id:9
1 1 1 1 1 0 1 1
26:
27:
28:
29:
30:
31:
32:
33:
//recompute sibling value
sibling.val =OR of all childrenValue
//delete the parentNode
delete(parentNode);
//merge or redistribute
else
//no underflow
//re-compute the value of all Bloom filters up to
the root
recomputeValueToTheRoot(parentNode);
Id:7 1 1 1 1 1 0 0 0
1
0
0
0
0
0
0
0
Id:1
Algorithm 5 : update(leaf,newBloomFilter)
//update all values on the path from leaf to the root
node = leaf;
repeat
4:
node.val = node.val newBloomFilter;
5:
node = node.parent;
6: until node = null
1:
2:
3:
0
1
0
0
0
0
0
0
Id:2
0
0
1
0
0
0
0
0
Id:3
0
0
0
1
0
0
0
0
Id:4
0 0 0 0 1 0 1 1 Id:8
0
0
0
0
1
0
0
0
Id:5
0
0
0
0
0
0
1
1
Id:6
Figure 4: Bloofi Tree After Update
8
Example. In the Bloofi tree in Fig. 1, assume that
we update the value of node 6 to be “00000011”. The
values of all the nodes in the path from node 6 to the
root are OR-ed with “00000011” and the resulting tree
is shown in Fig. 4.
Checking for membership in a Bloom filter is equivalent to checking the value of a few bits at random locations in a bitmap. Though fast, this operation does not
exploit bit-level parallelism: the processor’s ability to
do several bitwise operations in one instruction.
Let us assume that we have a 64-bit processor. We
propose a new approach (Flat-Bloofi) which stores the
data corresponding to 64 Bloom filters in a packed
data structure. Each Bloom filter is backed by a m-bit
bitmap. In their place, we construct a single array of
64-bit integers of length m (henceforth a Flat-Bloofi array). The first 64-bit integer corresponds to the first bit
of each of the 64 bitmaps. And so on. Thus, the value
of the ith bit of the jth bitmap is the value of the jth bit
of the ith integer in Flat-Bloofi array. Given N Bloom
filters, we create dN/64e Flat-Bloofi arrays. When N is
not a multiple of 64, some bits are unused in one of the
Flat-Bloofi arrays.
We organize Flat-Bloofi using the following data
structures:
Theorem 3 (Update Cost). The number of Bloofi
nodes accessed during the update operation in Bloofi is
O(logd N), where d is the order of the Bloofi index and
N is the number of Bloom filters that are indexed.
5.4. Improving pruning efficiency
In Bloofi, each non-leaf node value is the bitwise
OR of its children values. As the total number of objects in the underlying sets indexed by Bloofi increases,
the probability of false positive results returned by the
Bloom filters at the higher levels in the tree increases.
In the worst case, all bits in the Bloofi nodes at higher
levels in the tree could be one. This leads to decreased
pruning efficiency of the higher levels in the tree, as
more false positive paths are followed during search. To
improve upon the number of Bloom filters that need to
be checked for matches during a query, we propose the
following heuristic.
During the insert procedure, we do not split a node
that has all the bits set to one, even if the node is full.
This could stop the splitting a little too early, but it
avoids creating multiple levels in the tree with all the
bits set to one. Our experimental results in Section 7.2
show that the search cost is indeed improved by using
this heuristic when the root level value has all bits set to
one. Effectively, Bloofi can be viewed as a forest instead
of a tree.
Alternatively, we could dynamically change the size
of the Bloom filters when the false positive probability
at the root reaches 1. In such a case, if the application
allows, we could reconstruct the base Bloom filters to
have a lower false positive probability, and reconstruct
the Bloofi tree from bottom-up. The index construction
time for 100 000 Bloom filters was only about 15 seconds in our experiments, so periodic reconstruction of
the index is a viable solution.
• We maintain ζ Flat-Bloofi arrays. With this data,
we can index L = ζ × 64 Bloom filters.
• We use an array β of L bits, of which exactly N are
set to true. This indicates which index locations are
in use. Thus if a Bloom filter is deleted from the
index and then a new one inserted, we can reuse
the space.
• We use a hash table which maps from Bloom filter identifiers to internal index values (in the range
[0, L)). We also maintain an array of identifiers of
length up to L which gives us the identifier of the
Bloom filter stored at a given index. Combined together, the hash table and the array of identifiers
provide a two-way index from Bloom filter identifiers to index locations.
Queries. Given k hash functions, we map a given value
to k index locations in the range [0, m). For each FlatBloofi array, we retrieve the corresponding k 64-bit integers and compute their bit-wise AND aggregate. We
then iterate over the bits having a value of true: each
one corresponds to a matching Bloom filter. We use our
array of identifiers to recover the corresponding Bloom
filter identifiers. Thus if we have N Bloom filters, we
will access no more than k × ζ 64-bit integers from
the Flat-Bloofi arrays. Iterating through the set bits
can be done quickly using fast functions such as Java’s
Long.bitCount.
6. Bit-level parallelism
Bloofi keeps the Bloom filters as they are, and only
adds new (aggregated) Bloom filters to accelerate the
queries. However, in a worst case scenario, Bloofi may
need to check many Bloom filters. In such a case, Bloofi
may not be faster, and could even be slower, than a naive
approach which merely checks every Bloom filter.
9
Insertion. When inserting a new Bloom filter, using the
bit array β, we first seek an available index. If none
is found, then we create a new array of 64-bit integers,
thus, effectively, making available 64 new index positions. The bit array β, the array of identifiers, and the
hash table mapping to index values are updated. Finally,
we iterate through all of the set bits in the bitmap of the
new Bloom filter and set the corresponding bits in the
Flat-Bloofi array using a bit-wise OR operation.
size of Bloom filters, different underlying data distributions, and, for Bloofi, different similarity metrics used
for Bloofi construction and different order values. We
also compare Bloofi and Flat-Bloofi’s performance with
the “naive” case, where the Bloom filters are searched
linearly, without any index. We show that in most cases,
Bloofi achieves logarithmic search performance, with
low maintenance cost, regardless of the underlying data
distribution.
Deletion. When deleting a Bloom filter, we use the
hash table to recover its index using its identifier. The
key is then removed from the hash table. The corresponding bit in β is set to false. There are then two
possibilities.
7.1. Experiments setup
We implemented Bloofi and Flat-Bloofi in Java
and ran our experiments using Oracle’s JDK version
1.7.0 45. For Bloom filters, we use the implementation
provided by Skjegstad [28], modified to use faster hashing and a more efficient BitSet implementation [19].
The experiments were run on a HP Z820 Workstation
with an Intel Xeon E5-2640 processor (2.50 GHz) having 6 cores. Our test machine had 32 GB of RAM
(DDR3-1600, 4 × 8 GB). We did not use parallelism and
all our data structures are in RAM.
• If this Bloom filter was stored alone in a Flat-Bloofi
array, then the Flat-Bloofi array is removed. We
also remove the corresponding 64 bits in β as well
as the corresponding 64 entries in the array of identifiers. We scan the values of the hash table, and
deduct 64 from all index entries exceeding index
of the deleted Bloom filter.
• Otherwise, we go through the Flat-Bloofi array and
unset the bits corresponding to the Bloom filter using a bit-wise AND operation. Because we do not
keep a copy of the original Bloom filter, we need to
update every single component of the Flat-Bloofi
array.
7.1.1. Performance metrics
As performance metrics we use:
• search bf-cost: the number of Bloom filters
checked to find the one(s) matching a queried object, averaged over 50 000 searches
We believe that this compaction approach should provide reasonable performance and memory usage in a
context where deletions and insertions are frequent.
In the worst case scenario, however, and after many
deletions, we could have N Flat-Bloofi arrays indexing
N Bloom filters. To guard against such inefficiencies,
we would need more aggressive compaction strategies,
but we leave them to future work.
Note We could further accelerate the queries by replacing many (ζ) Flat-Bloofi arrays, with a single array
containing words of ζ × 64 bits. This would improve
memory locality. However, it would make compaction
more expensive.
• search time: the average time, in milliseconds, to
find all the Bloom filters matching a queried object
• storage cost: space required to store the Bloom
filters and/or the associated index structure, if applicable. For Bloofi, we estimate this cost as the
number of bytes for a Bloom filter, multiplied by
the number of nodes in the Bloofi tree, including
the leaves; for Flat-Bloofi, the storage cost is estimated as the number of bytes for a Bloom filters,
multiplied by number of Bloom filters rounded up
to a multiple of 64 (Flat-Bloofi uses longs for storage); for the “naive” case, the storage cost is estimated as the number of bytes for a Bloom filters,
multiplied by number of Bloom filters
Update. Updating a Bloom filter is easiest: we go
through the corresponding Flat-Bloofi array and set the
corresponding bits as we did with an insertion.
7. Experimental evaluation
• maintenance bf-cost: the average number of Bloofi
nodes accessed during an insert, delete, or update
operation
We evaluate Bloofi’s and Flat-Bloofi’s search performance and maintenance cost for different number and
• maintenance time: the average time, in milliseconds, for an insert, delete, or update operation
10
7.1.2. Parameters varied
As described in Section 3, a Bloom filter is a bit array of length m constructed using a set of k hash functions. In practice, when constructing a Bloom filter,
one does not specify the length and the number of hash
functions. It is unlikely that the average engineer would
know which values to pick. Rather, one specifies the expected maximal number of elements to be stored, nexp ,
(a possibly large number) as well as the desired probability of having a false positive, ρfalse , (e.g., 1%). One
can reasonably expect an engineer to be able to set these
values from domain knowledge. Then the number of
hash functions and the size of the bitmaps can be computed using
l the following
m formulas k = d− ln ρfalse / ln 2e
and m = k/ ln 2 ∗ nexp .
In the experiments we vary the following parameters:
parameter
value
N — Number of Bloom filters indexed
d — Bloofi order
m — Bloom filter size (bits)
n — Nb of elements in each Bloom filter
ρfalse — Desired probability of false positives
Construction method
Similarity measure
Data distribution
1000
2
100 992
100
0.01
iterative
Hamming
nonrandom
Table 1: Default Values for Parameters
run each experiment 10 times, and we report the averages over the last 5 runs. Since our values x are integers, we picked hash functions at random of the form
h(x) = ax mod m where a is an odd integer defining
the hash function. We found that this choice gave good
performance in our case.
• N: the number of Bloom filters indexed;
• d: Bloofi order;
7.2. Performance results
7.2.1. Varying number of Bloom filters indexed N
Fig. 5a shows the increase in search time as N increases from 100 to 100 000 for Bloofi, Flat-Bloofi, and
Naive case. Note the logarithmic scale on both X and
Y axes. For Bloofi, the increase in search time is logarithmic as long as the false positive probability (pfalse )
at the root is less than one (N ≤ 1000), but the increase
is higher than logarithmic after that, due to high value
for pfalse at the high levels of the tree. However, even
for 100 000 filters, Bloofi still performs orders of magnitude better than the “naive” case. The difference between “naive” case and Bloofi is not so big when only a
few Bloom filters are checked, likely because the search
time depends not only on the number of Bloom filters
checked, but also on the locality of nodes in memory.
We see that the curves for Bloofi and Flat-Bloofi intersect each other. For small numbers of Bloom filters,
Flat-Bloofi performs better, due to its superior memory
locality and exploitation of bit-level parallelism. However as the number of Bloom filters increases, Bloofi
performs better, due to its superior pruning abilities.
For Bloofi, using bulk construction leads to improved
search performance, since a global sort of all the Bloom
filters is performed before the bulk insert, while the
incremental construction is greedy and might not lead
to optimal placement. However, the cost of sorting is
O(N 2 ) in our implementation, which leads to high index construction time for the bulk construction.
To evaluate the effect of the in-place update for Bloofi
introduced in Section 5.3, we performed experiments
where the Bloofi tree is built incrementally using only
• m: the size, in bits, of the Bloom filters indexed: m
is varied indirectly, by specifying nexp ;
• n: the number of elements in each Bloom filter indexed;
• ρfalse : the desired probability of false positives in
the Bloom filters indexed;
• index construction method: iterative, where we insert Bloom filters one by one using the algorithm in
Section 5.1, or bulk. For the bulk construction, we
first sort all Bloom filters such that the first Bloom
filter is the one closest to the empty Bloom filter,
the second is the filter closest to the first Bloom
filter, etc., and then construct the Bloofi tree by always inserting next to the right-most leaf;
• similarity measure: the measure used to define
“closeness” during insert. We consider Hamming,
Cosine and Jaccard distances;
• data distribution:
nonrandom, with nonoverlapping ranges for each Bloom filter: each
Bloom filter i contains n integers in the range
[i × n, (i + 1) × n) and random, with overlapping
ranges for data in the Bloom filters: each Bloom
filter i contains n random integers in a randomly
assigned range.
For each experiment we vary one parameter and use
the default values shown in Table 1 for the rest. We
11
100
10-1
105
Flat
Bloofi
Bloofi-Bulk
Naive
Flat-AU
Bloofi-AU
Naive-AU
10-2
10-3
10-4
102
103
104
Number of Bloom Filters Indexed
(a) Search Time vs. N
105
104
107
Ideal
Bloofi
Bloofi-Bulk
Bloofi-AU
Naive
Storage Cost (kB)
101
Search BF-Cost
Search Time (ms)
102
103
102
101
102
103
104
105
106
Flat
Bloofi
Bloofi-Bulk
Naive
105
104
103
102
Number of Bloom Filters Indexed
(b) Search BF-Cost vs. N
103
104
105
Number of Bloom Filters Indexed
(c) Storage Cost vs. N
Figure 5: Varying Number of Bloom Filters N
half of the elements in each Bloom filter. The rest of
the elements are inserted then in each Bloom filter and
Bloofi is updated in-place. We perform a similar experiment for Flat-Bloofi and the Naive case. The AU (After Updates) curves in Fig. 5a show the search time in
the final data structure. As we expected, for Flat-Bloofi
and Naive, the After Updates and normal curves are almost identical, since the properties of these data structures are not affected by updates. However, the BloofiAU and Bloofi curves are also almost identical, which
shows that the in-place update maintains the search performance of the Bloofi tree.
This shows that using the heuristic increases the search
performance of Bloofi index when the false positive
probability at the high levels in the tree is high.
Fig. 5c shows the increase in the storage cost for
Bloofi, Flat-Bloofi, and Naive, as the number of Bloom
filters indexed increases. In all cases, the storage cost
increases linearly with N. The storage cost is lowest
for Naive case, as no extra information besides the actual Bloom filters is maintained. Flat-Bloofi has a small
overhead, only because sometimes rounds up the space
to multiples of 64, so it can take advantage of the bitlevel parallelism. The storage cost for Bloofi is also
quite low, as the number of non-leaf nodes in the Bloofi
tree is less than N. The storage cost for bulk construction (Bloofi-bulk) is slightly higher than for incremental
construction (Bloofi), because constructing a tree by always inserting in the right-most leaf leads in general to
skinnier trees, with more levels and more nodes.
Fig. 6a shows the cost, in terms of time, of maintaining Bloofi, Flat-Bloofi, and the Naive data structure. The maintenance cost for Naive is negligible,
as the Bloom filters are just maintained in a list. For
Flat-Bloofi, the maintenance cost increases only slightly
with the number of Bloom filters, as the cost of the different operations in Flat-Bloofi depends mainly on the
size of the Bloom filters and the number of bits turned
on, and does not depend much on the number of Bloom
filters. For Bloofi, the trend is increasing for the insert
and delete, as there are more nodes in the Bloofi tree
that get impacted by insert or delete. The cost of updates does not increase after the root becomes all one
and does not split, because we use in-place updates and
the height of the tree does not increase after N > 10000.
We find the relative cost of insert, delete, and update
for Bloofi and Flat-Bloofi quite interesting. For Bloofi,
update is the cheapest operation, both as time and number of Bloom filters accessed (see Fig. 6b), since only
the values of the nodes from the leaf to the root need
to be updated (OR-ed with the new value). However,
Fig. 5b shows similar trends for the search bf-cost
(average number of Bloom Filter nodes accessed for
a search). The search bf-cost performance metric is
not used for Flat-Bloofi and Naive, as for these data
strictures, all the Bloom filters are checked during a
search. In the “ideal” case for Bloofi, when exactly
one path from root to the leaf is followed during search,
the search bf-cost is approximately l logl N + 1 if each
non-leaf node has l children, and the search bf-cost increases logarithmically with N. In our experiments, the
increase in search bf-cost is logarithmic as long as the
false positive probability (pfalse ) at the root is less than
one (N ≤ 1000), but the increase is higher than logarithmic after that, due to high value for pfalse at the high levels of the tree. However, even for 100 000 filters, Bloofi
still performs two orders of magnitude better than the
“naive” case. If the size of the Bloom filters increases,
the search cost decreases to the “ideal” cost, as shown
in Fig. 8a.
To evaluate the effects of the heuristic introduces in
Section 5.4, we run the same experiment without using
the heuristic, so always splitting the root even if all the
bits were set to 1. When there are Bloofi nodes with all
bits 1, both the search bf-cost and the search time when
the heuristic is used are lower than when the heuristic is not used (search bf-cost is 104.29 vs. 110.17 for
N = 10 000 and 876.33 vs. 974.92 for N = 100 000).
12
10-1
Flat-Insert
Bloofi-Insert
Naive-Insert
Flat-Delete
Bloofi-Delete
Naive-Delete
Flat-Update
Bloofi-Update
Naive-Update
10-2
10-3
10-4
10-5
10-6
102
103
104
Number of Bloom Filters Indexed
(a) Maintenance Time vs. N
105
103
Average Nb Nodes Accessed
100
Average Nb Nodes Accessed
Operation Time (ms)
101
Bloofi-Insert
Bloofi-Delete
Bloofi-Update
102
101
102
103
104
Number of Bloom Filters Indexed
(b) Maintenance BF-Cost vs. N
105
140
Bloofi-Insert
Bloofi-Delete
Bloofi-Update
120
100
80
60
40
20
0
0
5
10
15
20
25
Bloofi Order
(c) Maintenance BF-Cost vs. d
Figure 6: Maintenance Cost
for Bloofi, the insert operation is the most expensive
as bf-cost, but delete becomes the most expensive as
time, with the increased number of indexed Bloom filters. The reason for this difference is that while during
a delete operation fewer nodes are accessed (no need
to search for the place in the tree), we do more work
at each node — we need to re-compute the values of
each node in the path from the leaf to root, by OR-ing
all the children value. During insert, the values of the
nodes get updated by OR-ing with the newly inserted
Bloom filter. For Flat-Bloofi, when inserting a sparse
Bloom filter, only a few words need to change, so it is
fast. When deleting, we do not know which bit changed,
so we have to set them all to zero, which takes longer.
If the Bloom filters would be more dense, it is possible
that deletions could be faster than insertions. During updates, we do not know how the new Bloom filter differs
from the old one, and we cannot compute the difference
(with XOR) since we no longer have the old one, so we
have to go through the set bits in the Bloom filter, and
set them all to 1. The running time is close to an insert,
which is what we see in Fig. 6a. Comparing Bloofi and
Flat-Bloofi, Bloofi’s updates are cheaper, as Bloofi only
needs to OR the new value with a few Bloom filters. For
inserts, Flat-Bloofi is faster, as it does not need to search
(and compute distances) the tree for the best place to insert the new node. For deletes, Bloofi is faster when
the number of Bloom filters is below 10 000 (and the
pfalse at the root is not 1), but Flat-Bloofi’s cost does not
depend much on the number of Bloom filters, so it becomes faster for large number of Bloom filters indexed.
tributions are needed. The update bf-cost does not increase after the root of the Bloofi tree becomes one, as
the root does not split in that case, so the height of the
Bloofi tree does not increase with the increased number
of indexed Bloom filters.
7.2.2. Varying Bloofi order d
Fig. 7a and Fig. 7b show the search time and search
bf-cost as the order of Bloofi increases. The search bfcost and search time increase as d increases, since the
search cost is proportional to d logd N. For a constant
N, the function f (x) = x log x N is convex, with minimum value when x = 2.718(e). The best search performance is achieved for Bloofi tree of low order. The
search bf-cost obtained from experiments is close to the
ideal search cost for full trees (2d children per non-leaf
node), which shows that our algorithms for tree construction perform well, and not many nodes outside of
the path from the root to the answer leaf are checked
during search. Bulk construction performs better than
incremental construction, due to the global sort. The
Bloofi and Bloofi-AU lines are almost identical, which
shows that our in-place update maintains the performance of the Bloofi tree.
The storage cost decreases with order (Fig. 7c), as the
number of non-leaf nodes in a Bloofi tree of order d is
N−1
and N−1
between 2d−1
d−1 , so higher the order, the lower the
overhead of storing the constructed Bloofi tree. While
the storage cost is lower for higher orders, the search
cost is higher (Fig. 7b), so there is a trade-off between
search and storage cost. We expect search performance
to be more important than storage in practice, and the
storage cost of Bloofi is similar with the Naive case
even in the worst case (at most twice as much space
is needed for Bloofi as it is needed for the Naive case),
so we believe that Bloofi trees of low order will be used
more in practice. The storage cost for bulk construction
is slightly higher than for incremental construction, because always inserting into the right-most leaf leads to
The maintenance bf-cost for Bloofi, as the average
number of Bloom filters accessed during a maintenance
operation, is shown in Fig. 6b. The maintenance bf-cost
increases logarithmically with N (Fig. 6b), as expected
from Theorems 1–3. The bf-cost of insert is higher than
the delete, as the place for the new node needs to be
found during insert. The update cost is the lowest, as
we use in-place updates, and no splits, merges, or redis13
Search Time (ms)
0.007
Search BF-Cost
Bloofi
Bloofi-Bulk
Bloofi-AU
0.008
0.006
0.005
0.004
0.003
0.002
90
20
80
18
70
16
60
50
40
20
10
0
0
5
10
15
20
25
Ideal-Skinny
Ideal-Full
Bloofi
Bloofi-Bulk
Bloofi-AU
30
0.001
0
0
5
Bloofi Order
(a) Search Time vs. d
Storage Cost (MB)
0.009
10
15
20
Bloofi
Bloofi-Bulk
Bloofi-AU
Naive
14
12
10
8
6
4
2
25
Bloofi Order
(b) Search BF-Cost vs. d
0
0
5
10
15
20
25
Bloofi Order
(c) Storage Cost vs. d
Figure 7: Varying Bloofi Order d
when pfalse increases. pfalse < 1 at the root for all the
ρfalse values tested in the experiment, so Bloofi’s performance is close to the ideal case.
skinnier, taller trees, with more nodes.
Fig. 6c shows the variation of maintenance bf-cost
with the order of Bloofi tree. The insert and delete
costs (O(d logd N)) increase with d, while the update
cost (O(logd N)) decreases with d.
A more surprising result is shown in Fig. 9b. The
search time for the Naive case, Bloofi, and Flat-Bloofi
is decreasing in the experiment, even if the search bfcost is increasing. The main reason is the memory locality. As the ρfalse increases, the size of the Bloom filters decreases, leading to better memory locality properties. Flat-Bloofi, with its efficient exploitation of the bitlevel parallelism benefits the most from the low memory
footprint. This is consistent with the results shown in
Fig. 8b. A second reason for the decrease in the search
time for all cases is that the number k of hash functions
used for the Bloom filters decreases as ρfalse increases.
k = d− ln ρfalse / ln 2e and decreases from 24 to 4 in the
experiment. As fewer hash functions are used, fewer
bits in the Bloom filters need to be checked for matches.
This contributes to the faster search time.
7.2.3. Varying Bloom filter size
We vary the Bloom filter size m by varying the number of expected elements
in the Bloom
filter nexp from
l
m
100 to 100 000. m = k/ ln 2 ∗ nexp .
Fig. 8a shows the search bf-cost variation with the
Bloom filter size in a system with 100 000 total elements (1000 filters with 100 elements each). The search
bf-cost for Bloofi is always below Naive cost, and it decreases to the ideal cost as the size of the Bloom filters
increases and the false positive probability at the high
level nodes in Bloofi decreases. In fact, O(d logd N)
search cost is achieved as soon as pfalse < 1 for the root,
even if pfalse is close to 1 (0.993 in our experiments for
size 100 992).
The search time (Fig. 8b) also decreases for Bloofi
as pfalse at the higher levels is reduced, while the search
time for Naive case is almost constant. For small Bloom
filter sizes, the search time for the Naive case is slightly
lower than using Bloofi, likely due to memory locality.
The search time for Flat-Bloofi is lower than for Bloofi
when the pruning capabilities of Bloofi are reduced, due
to the efficient bit level parallelism exploited by FlatBloofi. As the size of the Bloom filters increases, the
memory locality of Flat-Bloofi is reduced, and the trend
for the search time is upward, while the trend is downward for Bloofi.
7.2.5. Varying the number of elements n
Fig. 9c shows how the search time varies with the
increase in n, the number of elements in the indexed
Bloom filters, as the size of the Bloom filters remains
constant. Flat-Bloofi provides the best performance in
this case, as it benefits from the memory locality and
bit-level parallelism, and the performance is almost constant regardless of the actual number of elements in
the Bloom filter. Search time for Bloofi increases only
slightly as long as pfalse < 1 at the root, even when the
number of elements in the system is ten times larger
than the expected number of elements in a Bloom filter
(for n = 100, there are 100 000 elements in the system,
and nbElExp = 10 000). Bloofi’s performance degrades
as its pruning capabilities are reduced when pfalse = 1 at
the root. The search bf-cost for Bloofi varies in a similar
way (not shown).
7.2.4. Varying the probability of false positives
Fig. 9a shows how the search bf-cost for Bloofi increases as we increase the desired probability of false
positives in the Bloom filters indexed. As expected, the
bf-cost slowly increases as ρfalse and implicitly pfalse increases, as the pruning capabilities of Bloofi are reduced
14
900
10-1
Ideal
Bloofi
Naive
102
Search BF-Cost
800
Search Time (ms)
Search BF-Cost
103
Flat
Bloofi
Naive
10-2
700
Hamming
Cosine
Jaccard
600
500
400
300
200
100
101
103
104
105
106
10-3
107
103
104
Bloom Filter Size (bits)
105
106
0
102
107
Bloom Filter Size (bits)
(a) Search BF-Cost vs. Filter Size
103
104
105
Number of Bloom Filters Indexed
(b) Search Time vs. Filter Size
(c) Search BF-Cost vs. Metric vs. N
Figure 8
102
101
10-7
10-6
10-5
10-4
10-3
10-2
10-1
100
100
Flat
Bloofi
Naive
10-2
10-3
10-7
10-6
Desired False Positive Probability
(a) Search BF-Cost vs. ρfalse
Search Time (ms)
10-1
Ideal
Bloofi
Naive
Search Time (ms)
Search BF-Cost
103
10-5
10-4
10-3
10-2
Desired False Positive Probability
(b) Search Time vs. ρfalse
10-1
100
10-1
Flat
Bloofi
Naive
10-2
10-3
100
101
102
103
104
Number of Elements in Bloom Filters
(c) Search Time vs. n
Figure 9
filters, which is set to 100 for these experiments. In
the nonrandom case, there is no overlap between the
sets represented by different Bloom filters, while in the
random case, we could have overlap between the sets.
We expect that in a real scenario, the random case is
more common. The search cost is shown in Fig. 10b
and the search time is shown in Fig. 10c. We expected
the performance in the random case to be a little worse
than in the nonrandom case, since multiple Bloom filters might match a queried object. However, the overlap
of ranges in the nonrandom case was not enough to lead
to a substantial increase in the search cost or search time
for the nonrandom case, even if the number of results
was indeed larger for nonrandom case. The search performance is similar for both distributions used, which
shows that Bloofi is robust to the underlying data distribution.
7.2.6. Varying similarity metric
During Bloofi construction, we use a similarity metric to measure the similarity between two Bloom filters,
and this metric determines the location of the Bloom filters in Bloofi. In this section we experiment with several
metrics:
• Hamming(A, B) = |A xor B|,
• Jaccard(A, B) = 1 − |A and B|/|A or B|,
• and Cosine(A, B) = 1 − |A and B|/kAk2 × kBk2 ,
where |A| is the number of 1s in the bit array A. The
search bf-cost is similar for all the metrics (Fig. 8c).
When Jaccard is used, the search cost is a little lower,
but the differences are small, and they might be just due
to chance. A similar trend was obtained for search time,
as shown in Fig. 10a. The search time is the lowest for
Jaccard, and highest for Hamming.
7.3. Experimental results conclusion
7.2.7. Varying underlying data distribution
We use the following distributions for the underlying data in the Bloom filters: nonrandom, in which we
insert in each Bloom filter 0 ≤ i ≤ N the n integers
in the range [i × n, i × n + n), and random, in which
we insert in each Bloom filter n integers randomly chosen from a random range assigned to that Bloom filter. n is the actual number of elements in the Bloom
Our experimental evaluation shows that:
• For Bloofi, the search cost increases logarithmically with N, but gets worse if the false positive
probability at the root is 1. For optimal performance, the size of the Bloom filters should be
based on the estimate for the total number of elements in the entire system. If the false positive
15
10-1
10-2
10-3
102
103
104
Number of Bloom Filters Indexed
(a) Search Time vs. Metric vs. N
105
104
102
Bloofi-Random
Bloofi-Nonrandom
Naive
Search Time (ms)
105
Hamming
Cosine
Jaccard
Search BF-Cost
Search Time (ms)
100
103
102
101
102
103
104
Number of Bloom Filters Indexed
105
101
100
10-1
Flat-Nonrandom
Bloofi-Nonrandom
Naive-Nonrandom
Flat-Random
Bloofi-Random
Naive-Random
10-2
10-3
10-4
102
103
104
105
Number of Bloom Filters Indexed
(b) Search BF-Cost vs. Data Distribution (c) Search Time vs. Data Distribution vs. N
vs. N
Figure 10
•
•
•
•
•
•
probability at the root is below 1, even if close to
1, the search cost is O(d logd N).
For Flat-Bloofi, the search time increases with N.
For low number of Bloom filters, search time for
Flat-Bloofi is lower than for Bloofi.
Search costs for Bloofi increase with order, since
search cost is in O(d logd N), so low order is preferred for Bloofi.
Storage cost for Bloofi is O(N + N/d)
For Bloofi, insert cost and delete cost are
O(d logd N), and update cost is O(logd N)
Bulk construction gives slightly better search results, since the trees constructed with bulk construction are skinnier. However, differences in performance between bulk and incremental construction are small. The cost of sorting, which is the
first step in bulk construction, is O(N 2 ) in our implementation, so the operation is expensive. Our
experimental results show that the Bloofi algorithm
for insertion produces a tree close to the tree obtained by using a global ordering of Bloom filters.
The distance metric used to compare the “closeness” between Bloom filters does not have a big
effect on the search performance of the resulting
Bloofi tree, but the Jaccard distance seems to lead
to best search performance.
based on Bloom filters, and uses a data structure similar
to Flat-Bloofi. Mullin [23] uses Bloom filters to reduce
the cost of semijoins in distributed databases. Most applications of Bloom filter [4, 6], use the Bloom filters
directly and do not search through a large number of
Bloom filters.
B+ trees [8] inspired the implementation of the Bloofi
tree. However, each node in Bloofi has only one value,
and the children of a node do not represent completely
disjoined sets, so multiple paths might be followed during search.
A closely related idea to that of the multidimensional Bloom filter problem dates back to superimposed
codes [26] and descriptor files [24]: descriptors (effectively bit arrays) are searched through large files with
a hierarchical data structure constructed by OR-ing the
descriptions. When applied to set-valued attributes,
these descriptors are called “signatures”: each possible value is mapped to a fixed number of bit positions
that are set to true if the value is present, sets are constructed by setting all the bit positions corresponding
to all values. The S-tree [12, 17] is an implementation
of this idea which resembles Bloofi: signatures are organized in a similar tree structure using the Hamming
distance to aggregate similar signatures. S-trees are primarily used to index set-valued attributes. Our problem
differs in that we are provided Bloom filters in a distributed setting, and must quickly locate which Bloom
filter matches a given query.
8. Related work
Part of the existing work related to Bloom filters [7,
11, 13, 14, 21] is concerned with extending or improving the Bloom filters themselves and does not deal with
the problem of searching through a large set of Bloom
filters.
Applications of the Bloom filters to web caching [14]
use multiple Bloom filters, but their number is in general small and a linear search through the Bloom filters
is performed. [16] introduces an XML filtering system
Our Flat-Bloofi approach is similar to the wordparallel, bit-serial (WPBS) approach used for scanning
signatures [2, 18], except that our data structure is inmemory, supports moderately fast deletion and does not
require dedicated hardware. It is also similar to bitsliced
signature files [27, 31] as they apply to text indexing.
16
9. Conclusions and future work
[7] S. Cohen and Y. Matias. Spectral Bloom filters. In Proceedings
of the 2003 ACM SIGMOD International Conference on Management of Data, SIGMOD ’03, pages 241–252, New York, NY,
USA, 2003. ACM.
[8] D. Comer. Ubiquitous b-tree. ACM Comput. Surv., 11(2):121–
137, June 1979.
[9] A. Crainiceanu. Bloofi: A hierarchical bloom filter index with
applications to distributed data provenance. In Proceedings of
the 2Nd International Workshop on Cloud Intelligence, Cloud-I
’13, pages 4:1–4:8, New York, NY, USA, 2013. ACM.
[10] A. Crainiceanu, P. Linga, A. Machanavajjhala, J. Gehrke, and
J. Shanmugasundaram. P-ring: An efficient and robust p2p
range index structure. In Proceedings of the 2007 ACM SIGMOD International Conference on Management of Data, SIGMOD ’07, pages 223–234, New York, NY, USA, 2007. ACM.
[11] F. Deng and D. Rafiei. Approximately detecting duplicates for
streaming data using stable Bloom filters. In Proceedings of
the 2006 ACM SIGMOD International Conference on Management of Data, SIGMOD ’06, pages 25–36, New York, NY, USA,
2006. ACM.
[12] U. Deppisch. S-tree: A dynamic balanced signature index for
office retrieval. In Proceedings of the 9th Annual International
ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’86, pages 77–87, New York, NY,
USA, 1986. ACM.
[13] S. Dutta, S. Bhattacherjee, and A. Narang. Towards “intelligent compression” in streams: A biased reservoir sampling
based Bloom filter approach. In Proceedings of the 15th International Conference on Extending Database Technology, EDBT
’12, pages 228–238, New York, NY, USA, 2012. ACM.
[14] L. Fan, P. Cao, J. M. Almeida, and A. Z. Broder. Summary cache: a scalable wide-area web cache sharing protocol.
IEEE/ACM Trans. Netw., 8(3):281–293, 2000.
[15] D. Ficara, S. Giordano, G. Procissi, and F. Vitucci. Blooming
trees: Space-efficient structures for data representation. In Proceedings of the 2008 IEEE International Conference on Communications, pages 5828–5832, May 2008.
[16] X. Gong, W. Qian, Y. Yan, and A. Zhou. Bloom filter-based
XML packets filtering for millions of path queries. In In Proc.
ICDE, pages 890–901, 2005.
[17] D. L. Lee, Y. M. Kim, and G. Patel. Efficient signature file methods for text retrieval. IEEE Trans. Knowl. Data Eng., 7(3):423–
435, Jun 1995.
[18] D. L. Lee and F. Lochovsky. HYTREM—a hybrid text-retrieval
machine for large databases. IEEE Trans. Comput., 39(1):111–
123, Jan 1990.
[19] D. Lemire. JavaEWAH — a compressed alternative to the
Java BitSet class (version 0.8.3). https://github.com/
lemire/javaewah [last checked 10-03-2014], 2014.
[20] D. Li, H. Cui, Y. Hu, Y. Xia, and X. Wang. Scalable data center multicast using multi-class Bloom filter. In Proceedings of
the 2011 19th IEEE International Conference on Network Protocols, ICNP ’11, pages 266–275, Washington, DC, USA, 2011.
IEEE Computer Society.
[21] M. Mitzenmacher. Compressed Bloom filters. In Proceedings
of the Twentieth Annual ACM Symposium on Principles of Distributed Computing, PODC ’01, pages 144–150, New York, NY,
USA, 2001. ACM.
[22] M. Mitzenmacher and E. Upfal. Probability and Computing:
Randomized Algorithms and Probabilistic Analysis. Cambridge
University Press, New York, NY, USA, 2005.
[23] J. K. Mullin. Optimal semijoins for distributed database systems. IEEE Trans. Softw. Eng., 16(5):558–560, May 1990.
[24] J. L. Pfaltz, W. J. Berman, and E. M. Cagley. Partialmatch retrieval using indexed descriptor files. Commun. ACM,
23(9):522–528, Sept. 1980.
We introduced Bloofi, a hierarchical index structure
for Bloom filters. By taking advantage of intrinsic properties of Bloom filters, Bloofi reduces the search cost
of membership queries over thousands of Bloom filters
and efficiently supports updates of the existing Bloom
filters as well as insertion and deletion of filters. Our
experimental results show that Bloofi scales to tens of
thousands of Bloom filters, with low storage and maintenance cost. In the extreme worst case, Bloofi’s performance is similar with not using any index (O(N) search
cost), and in the vast majority of scenarios, Bloofi delivers close to logarithmic performance even if the false
positive probability at the root is close to (but less than)
one. When there are fewer Bloom filters, we found that
an alternative designed to exploit bit-level parallelism
(Flat-Bloofi) fared better.
We could pursue more advanced applications. For
example, Bloofi and Flat-Bloofi could be applied when
only current data from a moving window needs to be
maintained. If the window of interest is multiple days,
separate Bloom filters can be constructed for data collected each day. Old Bloom filters can be deleted from
Bloofi, and new filters inserted, so only the objects in
the current window are represented in the Bloofi index.
10. Acknowledgements
Special thanks to C. Crainiceanu for insightful
discussions while developing Bloofi, and D. Rapp,
A. Skene, and B. Cooper for providing motivation and
applications for this work.
11. References
[1] M. K. Aguilera, W. Golab, and M. A. Shah. A practical scalable distributed b-tree. Proc. VLDB Endow., 1(1):598–609, Aug.
2008.
[2] S. R. Ahuja and C. S. Roberts. An associative/parallel processor
for partial match retrieval using superimposed codes. In Proceedings of the 7th Annual Symposium on Computer Architecture, ISCA ’80, pages 218–227, New York, NY, USA, 1980.
ACM.
[3] B. H. Bloom. Space/time trade-offs in hash coding with allowable errors. Commun. ACM, 13(7):422–426, July 1970.
[4] A. Broder and M. Mitzenmacher. Network applications of
Bloom filters: A survey. In Internet Math., pages 636–646,
2002.
[5] F. Chang, J. Dean, S. Ghemawat, W. C. Hsieh, and D. A. Wallach. BigTable: A distributed storage system for structured data.
In OSDI, pages 205–218, 2006.
[6] F. Chang, J. Dean, S. Ghemawat, W. C. Hsieh, D. A. Wallach,
M. Burrows, T. Chandra, A. Fikes, and R. E. Gruber. BigTable:
A distributed storage system for structured data. ACM Trans.
Comput. Syst., 26(2):1–26, June 2008.
17
[25] R. Punnoose, A. Crainiceanu, and D. Rapp. Rya: A scalable
RDF triple store for the clouds. In Proceedings of the 1st International Workshop on Cloud Intelligence, Cloud-I ’12, pages
4:1–4:8, New York, NY, USA, 2012. ACM.
[26] C. Roberts. Partial-match retrieval via the method of superimposed codes. Proceedings of the IEEE, 67(12):1624–1642, Dec
1979.
[27] R. Sacks-Davis, A. J. Kent, and K. Ramamohanarao. Multikey access methods based on superimposed coding techniques.
ACM Trans. Database Syst., 12(4):655–696, Nov. 1987.
[28] M. Skjegstad. Java-BloomFilter — a stand-alone Bloom
filter implementation written in java (version 1.0). https:
//github.com/magnuss/java-bloomfilter [last
checked 10-03-2014], 2011.
[29] I. Stoica, R. Morris, D. Karger, M. F. Kaashoek, and H. Balakrishnan. Chord: A scalable peer-to-peer lookup service for
internet applications. In Proceedings of the 2001 Conference
on Applications, Technologies, Architectures, and Protocols for
Computer Communications, SIGCOMM ’01, pages 149–160,
New York, NY, USA, 2001. ACM.
[30] H. T. Vo, C. Chen, and B. C. Ooi. Towards elastic transactional
cloud storage with range query support. Proc. VLDB Endow.,
3(1-2):506–514, Sept. 2010.
[31] J. Zobel, A. Moffat, and K. Ramamohanarao. Inverted files versus signature files for text indexing. ACM Trans. Database Syst.,
23(4):453–490, Dec. 1998.
18
| 8cs.DS
|
A neuromorphic hardware architecture using the
Neural Engineering Framework for pattern
recognition
Runchun Wang, Chetan Singh Thakur, Tara Julia Hamilton, Jonathan Tapson, André van Schaik
The MARCS Institute, University of Western Sydney, Sydney, NSW, Australia
[email protected]
Abstract—We present a hardware architecture that uses the
Neural Engineering Framework (NEF) to implement largescale neural networks on Field Programmable Gate Arrays
(FPGAs) for performing pattern recognition in real time.
NEF is a framework that is capable of synthesising large-scale
cognitive systems from subnetworks. We will first present the
architecture of the proposed neural network implemented
using fixed-point numbers and demonstrate a routine that
computes the decoding weights by using the online
pseudoinverse update method (OPIUM) in a parallel and
distributed manner. The proposed system is efficiently
implemented on a compact digital neural core. This neural
core consists of 64 neurons that are instantiated by a single
physical neuron using a time-multiplexing approach. As a
proof of concept, we combined 128 identical neural cores
together to build a handwritten digit recognition system using
the MNIST database and achieved a recognition rate of
96.55%. The system is implemented on a state-of-the-art
FPGA and can process 5.12 million digits per second. The
architecture is not limited to handwriting recognition, but is
generally applicable as an extremely fast pattern recognition
processor for various kinds of patterns such as speech and
images.
millions of pattern recognitions in real time. The NEF,
which was first introduced in 2003, is a framework that is
capable of building large systems from subnetworks with a
standard three-layer neural structure (the first layer contains
the input neurons; the second layer is a hidden layer, which
consists of a large number of non-linear neurons; and the
third layer is the output layer, which consists of linear
neurons). The NEF has been used to construct SPAUN,
which is the first brain model, implemented in software and
is capable of performing cognitive tasks (Eliasmith et al.,
2012). This demonstrates that the NEF is a powerful tool for
synthesising large-scale cognitive systems.
We have previously presented a compact neural core
architecture specifically for FPGA implementation of large
NEF networks (Wang et al., 2014a). In this paper, we
present an application that uses this neural core to build
pattern recognition systems. The outline for this paper is as
follows: Section 2.1 introduces the basic concepts of the
NEF; the algorithm and theory is presented in Section 2.2;
the hardware implementation is presented in Section 2.3;
the performance for different design choices will be
thoroughly compared in Section 3; in section 4 we compare
our work with other solutions and discuss future works.
Keywords: neural engineering framework; time-multiplexing;
pattern recognition; pseudo inverse; MNIST; neuromorphic
engineering
2. Materials and methods
1. Introduction
2.1 Background
Neural networks have been proved to be powerful tools
for real world tasks, such as pattern recognition,
classification, regression, and prediction. However, their
high computational demands are not ideally suited to
modern computer architectures. This constraint has so far
often prohibited their use in applications that need real-time
control, such as interactive robotic systems. On the other
hand, scientists have been developing hardware platforms
that are optimised for neural networks over the past two
decades (Vogelstein et al., 2007; Boahen, 2006; Pfeil et al.,
2013; Wang et al., 2014d). However, these systems are not
capable of synthesising large-scale neural networks for
these real world tasks from subnetworks and therefore are
not very suitable, as pointed out by Tapson et al. (Tapson et
al., 2013).
In this section, we review the theoretical framework of a
Here, we present a generic hardware architecture that
uses the Neural Engineering Framework (NEF) (Eliasmith
and Anderson, 2003) to implement large-scale neural
networks on FPGAs, which are capable of processing up to
Figure 1 | A typical NEF network. The stimulus X(t) is
encoded into a large number of nonlinear hidden layer
neurons N using randomly initialised connection weights.
The output of the system, Y(t), is the linear sum of the
weighted spike trains from the hidden neurons.
1
trained using the training dataset only and is subsequently
validated using the test dataset.
The proposed digit recognition system is a three-layer
feed forward neural network, consisting of 784 input layer
neurons (pixels), 8192 (8k) hidden layer neurons and ten
output layer neurons. The input layer neurons are connected
to the hidden layer neurons using randomly weighted all-toall connections. The hidden layer neurons are also
connected to the output-layer neurons using all-to-all
connections but with weights calculated using a
pseudoinverse operation.
In the digit recognition system, a single input digit
(28x28=784 pixels) is mapped onto a layer of input
neurons, which we refer to as a vector Img with a dimension
of 784× 1. The Img matrix is multiplied by a matrix,
Random_weights, with a dimension of 8192 × 784. The
resultant vector, referred to as Vin with a dimension of
8192×1, is thus given by:
Figure 2 | Tuning curves maps input stimuli to spike
rates. For clarity, this figure only shows the tuning curve of
16 neurons. Each neuron in the neural layer has a distinct
tuning curve.
𝑉𝑉𝑉𝑉𝑉𝑉 = 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅_𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤ℎ𝑡𝑡𝑡𝑡×𝐼𝐼𝐼𝐼𝐼𝐼
(1)
𝑌𝑌 = 𝑊𝑊×𝐻𝐻
(2)
Each value in Vin is the sum of the randomly weighted
pixels, and is the stimulus for the corresponding neuron in
the hidden layer. Each neuron of the hidden layer responds
to its Vin value according to a distinct tuning curve (Figure
2). The output of the hidden layer neurons for each input
digit is collected in a matrix referred to as H with a
dimension of 8192×1. Finally, the response of the output
layer neuron is given by:
typical NEF system, which encodes an input stimulus into a
spiking rate of neurons of a heterogeneous population and
decodes the desired function by linearly combining the
responses of these neurons. The topology of the NEF
network is illustrated in Figure 1. A NEF network performs
three tasks to calculate a desired function f(X):
1. Encoding: An encoder will have a fixed random
weight (RW) for each hidden layer neuron, and multiplies
the input stimulus by this weight. The firing rate of
individual neurons is a nonlinear function of the input
stimulus weighted by the random weights. The parameters
of the neurons are also randomised, so that each neuron in
the hidden layer exhibits a distinct tuning curve. An
example of such tuning curves is shown in Figure 2.
where, W is the decoding weight (a matrix with a dimension
of 10×8192, ten columns for ten digits: 0-9) and Y (a
Boolean matrix with a dimension of 10×1) represents the
corresponding value of the input digit. For example, if the
input digit represents 2, then, during training, Y[2] will be
set to 1 and the other values in Y will be set to 0. Since this
is a linear system, the weights can be found by
calculating W = H+Y, where H+ is the pseudo-inverse of H.
2. Decoding: The activity, H, of the hidden neurons (i.e.
the spike rate of each neuron) can be measured over the
desired range of input values X. The output of each neuron
will be multiplied by their decoding weights such that WH =
f(X) = Y. Since this is a linear system, these weights can be
found by calculating W = YH+, where H+ is the MoorePenrose pseudo-inverse (Penrose and Todd, 1955) of H.
The above description is for one single digit. For
training purposes, we used 60000 sample digits and hence
the dimensions of Img, Vin, H and Y will change to
784×60,000, 8192×60,000, 8192×60,000 and 10×60,000,
respectively. When we use the digits from the test dataset
with 10,000 digits, the dimensions of Img, Vin, H and Y will
change to 784×10,000, 8192×10,000, 8192×10,000 and
10×10,000, respectively. In the testing phase, the predicted
output Y will be the product of W*H and will be compared
with the expected output to obtain the error rate (the number
of unrecognised digits among 10000 test digits). We will
address the details of testing in Section 3.
3. Averaging: The output of the system, Y(t), is the
linear sum of the weighted spike trains from the neurons.
2.2 Algorithm and Theory
2.2.1 Methodology
Recognition or classification of handwritten digits is a
standard machine learning problem, and in the form of the
MNIST database (Lecun et al., 1998) it has become a
benchmark problem. Hence, as a proof of concept, we have
used the proposed design framework to implement a digit
recognition system (Figure 3). Importantly, the same system
could be used for other pattern recognition applications. In
the MNIST database, the digits are represented as 28 × 28 =
784 pixels, and the training and testing dataset contain
60,000 and 10,000 digits, respectively. The system is
2.2.2 Modelling
Our aim is to develop a fast hardware pattern
recognition system running in real time, rather than aiming
for the lowest test error. Thus, we have adopted a hardwaredriven method to implement our system, which will achieve
the best trade-off between performance and hardware
resources. This method will first consider the hardware
2
Figure 3 | System Topology. The inputs are the pixels; they are connected to a higher-dimensional hidden layer with 8k
neurons, using randomly weighted connections. The output layer consists of linear neurons and the output layer weights are
solved analytically using the pseudoinverse operation.
constraints, and then all the building blocks will be
optimised.
conversion will reduce the hardware cost significantly with
a negligible performance loss, and will be presented in
detail in Section 2.3.2. We will compare the performance
differences in section 3.1. This conversion is carried out by
comparing the grey scale value with 0 - if it is larger than 0,
that pixel will be set 1; else it will be set to 0.
For FPGA implementations, there will be a significant
difference in the hardware cost between fixed-point and
floating-point implementations, as the latter requires many
more digital signal processors (DSPs). More importantly,
the floating-point number is represented by 64-bits, which
would lead to a huge data storage requirement, which would
be a bottleneck for the system. Thus, we have implemented
our system using fixed-point numbers.
To guarantee that the pixels of each digit from the input
layer will be nonlinearly projected to the high dimensional
hidden layer, for each neuron in the hidden layer, the
encoder will first generate a uniformly distributed random
weight for each pixel of one input digit and then sum these
weighted pixels up for generating the stimulus. For
verification of our hardware system, the random weights
used in the software and in the hardware models should be
the same and produce identical results. In a software model,
random weights are generated using special routines, which
is difficult to implement on hardware.
Before implementing the design in hardware, we have
modelled our system in Python, which is a popular software
programming
language,
using
the
fixed-point
representation. This will ensure that the software and the
hardware results are the same, and avoid any performance
drop or malfunctioning of the system in hardware due to
conversion from floating to fixed point numbers. The
models presented in the remaining part of this section were
all software models unless otherwise specified.
One option is to use a look up table (LUT) in the FPGA
to store the random weights generated by the software
model. The major drawback of this solution is that it
requires a significant amount of memory, which scales
linearly with number of input neurons and hidden layer
neurons. For FPGA implementations, the most efficient way
to generate random numbers is to use linear feedback shift
registers (LFSRs), as we have previously used to implement
a randomly weighed all-to-all connectivity in a spiking
neural network (Wang et al., 2014c). Based on that work,
we have developed an encoder, which uses LFSRs to
perform the nonlinear projection. We have implemented the
2.2.3 Input layer
The input layer will read digits from the MNIST
database and map them into the input layer pixels (one by
one). This task consists of not only converting the
dimension from 28×28 to 784×1 but also converting the
grey scale value (an 8-bit number that ranges from 0 to 255)
of the pixels to a binary value. The latter is a major
difference between our system and existing algorithms
(Tapson and van Schaik, 2013) (Lecun et al., 1998). This
3
ELSE:
T = Stim + 4×N_index
F_rate = max(2 × N_index × T / N_A , 0)
END
Here F_rate represents the firing rate of the neuron as a
result of the input stimulus, N_index represents the index of
the neuron in the neural core, and T is calculated as shown
for the different neurons. N_A represents the size of the
hidden layer, Max_Stim represents the maximum value of
the stimulus and Stim represents the current value of the
input stimulus using an integer in the range of [0,
Max_Stim) to code for an input range of [-1, 1). Figure 4
shows the tuning curves of a set of N_A = 64 of the
proposed fixed-point neurons, using Max_Stim = 255. The
transfer function is thus a nonlinear function of the stimulus
since the value of F_rate cannot go negative. Our system
requires the stimulus to be nonlinearly encoded into the
firing rate of the neuron and it is hardware intensive to use
digital circuits to implement conventional nonlinear
functions such as tanh. Instead, this piecewise linear
function can be easily implemented using a single 9-bit
fixed-point multiplier. We will present its implementation in
detail in section 2.3.3.
Figure 4 | The tuning curves of the proposed fixed-point
non-spiking neuron. This figure shows the tuning curve of
64 neurons.
same LFSR encoder in software to ensure that the random
weights are identical in both implementations. We have
highly optimised the encoder for hardware implementation,
and details of this will be presented in Section 2.3.
2.2.4 Rate neuron
The NEF intrinsically uses spike rates to calculate the
weights, and low-pass filters to sum the weighted output
spikes to implement the desired function. In contrast, we
have implemented our neurons as non-spiking neurons that
compute their firing rate directly. If these neurons were to
be implemented as leaky-integrate-and-fire neurons on
FPGA, as we have done previously (Wang et al., 2014c),
their average firing rates would have to be measured for
each value of the input stimulus to compute the decoding
weights. This method is quite inefficient and inflexible, as
we would have to repeat the measurements each time the
parameters of the neurons change. Another drawback is that
spiking neurons running in real time would not be able to
accurately communicate their firing rate in a short time
period, e.g., 1ms. This would significantly limit their usage
in real time applications. Using non-spiking neurons, their
actual firing rate can be communicated immediately after
presenting the stimulus to the neurons. This feature is quite
important for applications that need real-time control, such
as interactive robotic systems.
2.2.5 Hidden layer
We refer to the set of 64 neurons as a neural core, which
will be used as the standard building block for our digit
recognition system. Multiple neural cores can easily be
combined to build real-time large-scale neural networks
using our design framework. Furthermore, the development
cycle of large-scale neural networks will be significantly
shortened as there is no requirement for measurement of the
firing rate anymore, since each neural core has the same set
of known tuning curves.
The hidden layer was implemented with 128 identical
neural cores, for a total of 8192 (8k) neurons and
8192×(784+10) ≈ 6.5M synaptic connections. This hidden
layer size has achieved the best trade-off between
performance and memory usage and we will compare the
performance differences in Section 3.2. Given an input
image, the encoder will generate, via the random weight
projection, a different Vin for each neuron in each core,
even if each core contains identical neurons. In other words,
even though neuron[0] in neural core[0] and neuron[0] in
neural core[1] have the same tuning curve as a function of
Vin, the are highly likely to get different Vin so that their
firing rates will be different too.
In a system with non-spiking neurons, the system will
not compute correctly if these neurons cannot reproduce the
same firing rate as the one used to calculate the decoding
weights. In other words, the computed firing rate must be
repeatable for a given input value. Based on these
requirements, we proposed to compute the firing rate of
each neuron using its index in the array together with the
stimulus value to produce a ‘broken-stick’ nonlinearity
using the following algorithm:
2.2.6 Regression
The
decoding
weights
are
obtained
by
calculating W = H+Y, where H+ is the pseudoinverse of H.
However, the pseudo-inverse of the matrix H of size 60000
× 8192 requires a huge amount of memory and
computational time. We have previously developed an
online pseudoinverse update method (OPIUM) (Tapson and
van Schaik, 2013), which is an incremental method to
compute the pseudoinverse solution to the regression
FOR N_index in (0, N_A-1):
IF N_index < N_A/2:
T = Max_Stim - (Stim + 4×N_index)
4
problem, which requires significantly less memory. Hence,
we use this method here to compute the decoding weights.
We chose to use a 6-bit resolution for the decoding weights,
to obtain the best trade-off between performance and
memory usage. We will address this in details in section
3.1.
The pseudoinverse method only gives the best solution
with the lowest square root error for any given H matrix,
i.e., any given set of random weights; it does not necessarily
achieve the lowest test error for the MNIST data set. So we
adopted a regression method to find the best seed, which
will be used by the encoder to generate random weights,
and will in turn change the H matrix. In this way, we can
obtain the lowest possible test error in our system. Figure 5
shows the flow of this regression method. It uses a
simplified version of OPIUM, called OPIUM lite (Tapson
and van Schaik, 2013), which is a fast online method for
calculating an approximation to the pseudoinverse. It is
significantly quicker than the full-scale OPIUM, but will
find output weights resulting in a slightly worse test error.
OPIUM lite is used with different random seeds, i.e., for
different random weight vectors, until a seed is found with a
target error below a desired threshold. After that, the full
scale OPIUM is used to compute the decoding weights with
that seed. As there is no guarantee that OPIUM lite will be
able to achieve a target error below the desired threshold, a
time-out mechanism is introduced. In our system, this timeout will be activated when the regression has run for 1000
seeds. If a time-out happens we simply use the seed that has
so far resulted in the lowest error and then use the full scale
OPIUM to compute the decoding weights.
2.3 Hardware implementation
2.3.1 Topology
To efficiently implement the system on an FPGA, we
use a time-multiplexing approach (Cassidy et al., 2011;
Wang et al., 2013, 2014d, 2014c, 2014b, 2015; Thakur et
al., 2014), which leverages the high-speed digital circuit.
State-of-the-art FPGAs can easily run at a clock speed of
266MHz (clock period 3.75ns). Thus, we can exploit timemultiplexing approach to simulate 218 neurons (256k,
powers of two are preferable as they optimise memory use
for storage) in ~1 millisecond by only implementing one
physical neuron on an FPGA. We refer to these neurons as
time-multiplexed (TM) neurons. This means that on every
clock cycle, a TM neuron will be processed. Each TM
neuron is updated every 256k/266MHz ≈ 943 µs while a
sub-millisecond resolution is generally acceptable for neural
simulations.
The time-multiplexing approach is however constrained
by its data storage requirement. The on-chip SRAM is
limited in size (usually only tens of MBs). Due to
bandwidth constraints it is difficult to use off-chip memory
with the time-multiplexing approach, as new values need to
be available from memory every clock cycle to provide
real-time simulation. Furthermore, the architecture of the
system will be more complex when using off-chip memory
because it needs a dedicated memory controller.
Nevertheless, using off-chip memory promises the ability to
implement much larger networks and we will investigate
this option for future designs. However, we chose to use onchip memory for the current work to keep the architecture
simple.
Figure 5 | The flow of the proposed regression method.
5
Figure 6 | FPGA implementation of the proposed system. (a) The system topology;(b) The internal structure of the timemultiplexed system.
Figure 6 shows the topology of the FPGA
implementation of the system, which consists of an input
layer (the encoder), a hidden layer with 128 neural cores
and an output layer with 10 neurons. The encoder and the
hidden layer are both implemented with the timemultiplexing approach and Figure 6b shows their internal
structure. It consists of a physical encoder, a physical
neuron, a global counter and a weight buffer. The global
counter processes the time-multiplexed (TM) encoders and
neurons sequentially. The decoding weights of the physical
neuron are stored in the weight buffer. For simplicity, let us
assume that each TM encoder and TM neuron are processed
in only one clock cycle. This means that in every clock
cycle, a TM encoder will generate the stimulus for an input
digit, and the corresponding TM neuron will generate a
firing rate with that stimulus and then multiply it with the
decoding weights (ten numbers for ten digits obtained by
using the OPIUM). The input digit will not change and will
remain static until all the TM neurons finish their
processing. The output of every TM neuron will be ten
weighted firing rates, each of which will be accumulated by
its corresponding output neuron. Using a pipelined
architecture, the result from calculating one time step for a
TM encoder and neuron only has to be available just before
the turn of that TM encoder and TM neuron comes around
again. The above description assumes that it only takes one
clock cycle to process one TM encoder and TM neuron,
while this timing requirement is quite difficult to meet in a
practical design. We will address this issue in detail in next
section.
binary value. This saves significant hardware resources in
the FPGA, since otherwise we would need 784 multipliers
to compute the multiplication between all pixels and their
corresponding random weights. Each binary pixel is used to
control a 2-input multiplexer, one is connected to its
corresponding random weight and the other is tied down to
zero. If the value of a pixel is high, that corresponding
random weight will be accumulated for the generation of
stimulus for a hidden layer neuron.
The major challenge in implementing the encoder in
hardware using the time-multiplexing approach is to meet
the timing requirement. We need to sum all the 784
weighted pixels in 3.75 ns, since each TM neuron needs to
be processed in one clock cycle. Moreover, this operation
will require 784 adders, which will cost a significant
amount of hardware resources. The introduction of
pipelines will mitigate the critical timing requirement, but
will need even more adders. As a compromise we chose to
process each TM encoder and TM neuron in a time slot of
four clock cycles. So the encoder will perform this sum
operation in four cycles, each of which will sum 784/4=196
weighted pixels. This modification not only mitigates the
critical timing requirement, but also reduces the number of
adders that are needed. The price paid is that the timemultiplexing rate has to be divided by four. Hence, we can
only time-multiplex 64k neurons rather than 256k neurons.
Figure 7 shows the structure of the physical encoder,
which consists of an input buffer, a global counter, 49
random weight (RW) generators (each implemented with an
20-bit LFSR), 196 2-input multiplexers and a sum up
module. When an input digit arrives, it is stored in the input
buffer. In each time slot, the global counter sends that stored
digits to multiplexers for generating the weighted pixels.
The lowest 196 bits are sent in the first clock cycle (of that
time slot) and then the higher 196 bits in the next clock
2.3.2 Physical encoder
The encoder will generate a uniformly distributed
random weight for each pixel of the input digit, and then
sum these weighted pixels to generate the stimulus for each
neuron in the hidden layer. We have pre-processed the input
digit by converting grey-scale value of each pixel to a
6
Figure 7. The structure of the physical encoder
cycle, one by one, and highest 196 bits in the fourth clock
cycle.
implemented with a single 9-bit multiplier, which computes
the F_rate and multiplies it with one and only one decoding
weight. In the digit recognition system implemented here,
the neuron needs to multiply F_rate with ten decoding
weights (for ten digits: 0-9). A naïve implementation would
instantiate ten identical neurons, each with one decoding
weight (for each output neuron), and would cost 10
multipliers. The whole operation would require 11
multiplications. Since the time slot consists of four clock
cycles, we can distribute these 11 multiplications to these
four clock cycles so that only 11/4=3 multipliers will be
needed. Based on this strategy, the neuron has been
efficiently implemented with three identical 9-bit
multipliers as shown in Figure 8. The number of the
implementable multipliers is usually one of the bottlenecks
Each RW generator generates a 20-bit random number,
which is divided into four 5-bit random signed numbers.
Hence, 49 RW generators will provide totally 49x4 = 196 5bit random weights, each is sent to its corresponding
multiplexer. All these LFSRs will reload their own initial
seed (obtained using the pseudoinverse method) on the
arrival of an input digit. After that, it keeps generating
random numbers until a new input digit arrives. In this way,
we can guarantee that the encoder will generate the exact
same set of random weights (for each incoming digit) with
any given seed. This “on the fly” generation scheme
reduces the usage of the memory significantly, as there is no
requirement for storing the random weights anymore – only
the seeds need to be stored.
The accumulator module sums the 784 weighted pixels
(in four clock cycles) for generating the stimulus for that
TM neuron. A naive implementation would need a 196input 5-bit parallel adder and create a large delay (~20 ns).
To mitigate this critical timing requirement, we use a 2stage pipeline, which consists of fourteen 14-input 5-bit
parallel adders and one 14-input 9-bit parallel adder. Since
it is a pipelined design, the stimulus (for each TM neuron) is
still being generated every time slot (with a latency of two
clock cycles).
2.3.3 Physical neuron
The rate neuron achieves a significant reduction in
memory usage, since it computes its firing rate with its
index, the input stimulus and fixed parameters, none of
which need memory access. Memory access is only needed
to read the decoding weights. In our previous work (Wang
et al., 2014a), the physical neuron has already been
Figure 8. The structure of the physical neuron
7
TABLE I
Device utilisation Altera Cyclone 5CGXFC5C6F27C7
Adaptive Logic Modules
(ALMs)
RAMs
DSPs
2162/29080
480k/4.5M
3/450
of large-scale FPGA/ASIC design.
The multiplier’s inputs A and B are 9 bits wide and the
output result is 18 bits wide. All of the three multipliers will
need four clock cycles to process the algorithm. For
multiplier [0], the first cycle computes the F_rate, which is
represented by a 7-bit number, by multiplying N_index and
T; the second cycle latches F_rate at input A of the
multiplier; the third and fourth cycle multiplies F_rate with
the decoding weight [0] and [1], respectively. For multiplier
[1], the first, second, third and fourth cycle multiplies F_rate
with the decoding weight [2],[3],[4] and [5] respectively.
For multiplier [2], the first, second, third and fourth cycle
multiplies F_rate with the decoding weight [6],[7],[8] and
[9] respectively. Again, since it is a pipelined design, the
output of each TM neuron is updated only once in its time
slot (with a latency of four clock cycles).
2.3.4 Output layer
The output layer consists of ten neurons (Figure 6) that
will linearly sum the results of all the 8k TM neurons. Since
it is a time-multiplexed system, this sum is just an
accumulation of the outputs of the TM neurons of each time
slot and the computational cost can be reduced in
magnitudes. Hence, the implementation of each output
neuron will only need a register and an adder. When all the
8k neurons have all been processed, the index of the output
neuron with the maximum value will be sent out as the
result, which indicates the most likely input digit. After
that, the values of the ten output neurons are cleared.
2.3.5 Utilisation
The system was developed using the standard ASIC
design flow, and can thus be easily implemented with stateof-the-art manufacturing technologies, should an integrated
circuit implementation be desired. A bottom-up design flow
was adopted, in which we designed and verified each
module separately. Once the module level verification was
complete, all the modules were integrated together for toplevel verification. We have successfully implemented 128
proposed neural cores, yielding 8k neurons, on an Altera
Cyclone V FPGA (on a Terasic Cyclone GX starter kit).
The design uses less than 6% of the hardware resources
(with the exception of the RAMs, Table I). Note that this
utilisation table includes the circuits that carry out other
tasks such as the JTAG interface.
Figure 9. (a) and (b) The histogram of the error rate for
configuration 1 and configuration 2; (c) the normalised
histogram of the difference between the paired errors
(blue) and sample T distributions modelling the data
(red); (d) the distribution of the estimated mean of the
difference data.
system as our goal is to develop a hardware system running
in real time, rather than exploiting an algorithm that is as
accurate as possible. The performance results were obtained
using the full test set of 10,000 handwritten digits after
training on the full 60,000 digit training set, unless
otherwise specified. The results presented in Section 3.1-3.2
were all obtained using the software (Python) models. The
results presented in section 3.3 were obtained from the
hardware implementation.
3. Results
The results presented here will focus on how different
design choices will affect the performance of the proposed
3.1 Comparison across different configurations
8
Compared to our previous work (Tapson and van Schaik,
2013), we have made three major modifications: the greyscale pixel in the input images were replaced by black &
white (binary) pixels; tanh neurons in the hidden layer were
replaced by rate neurons; and 64-bit floating-point numbers
for the decoding weights were replaced by 6-bit fixed-point
numbers. We investigated the effects of these modifications
using four configurations: configuration 1 was the
configuration used in our previous work (Tapson and van
Schaik, 2013); configuration 2 used black and white images;
configuration 3 used black and white images and rate
neurons instead of tanh neurons; and configuration 4 had all
three modifications. The hidden layer consisted of 8k
neurons in all four configurations.
lite to calculate the decoding weights and the test error. This
significantly reduces the simulation time needed for these
tests while still providing a fair comparison between the
four configurations.
We first investigated the effect of using the binary
values in the input layer. We compared the performance
result between the one using the grey-scale values and
binary values (see Figure 9). The top two panels show a
histogram of the number of errors out of 10,000 test
patterns. Given the skewed nature of the two error
distributions, rather than simply reporting p-values to
indicate the statistical significance of this difference, we
have chosen to display the full distribution here. Because
the same set 100 random weight vectors was used for each
configuration, we can determine a paired difference
between the two configurations, shown as a histogram in
Figure 9c. We then modelled the distribution of the
difference of errors using a non-central T distribution,
which is optimal for modelling distributions that are
approximately Gaussian but contain outliers. We followed
the Bayesian estimation method according to Kruschke
(Kruschke, 2012) using Markov Chain Monte Carlo
For each configuration, 100 test runs were conducted,
each with a different random seed. The same set of 100
seeds was used for all four configurations, so that the
encoder will generate the same random weights. Since the
goal of this exercise was simply to investigate the impact of
the three modifications on performance, rather than to find
the best possible performance, we only used the first five
steps of the regression method, i.e., we only used OPIUM
Figure 11. (a) The histogram of the error rate for
configuration 4; (b) the normalised histogram of the
difference between the paired errors (blue) and sample
T distributions modelling the data (red); (c) the
distribution of the estimated mean of the difference
data.
Figure 10. (a) The histogram of the error rate for
configuration 3; (b) the normalised histogram of the
difference between the paired errors (blue) and sample
T distributions modelling the data (red); (c) the
distribution of the estimated mean of the difference
data.
9
simulation. We simulated the Markov Chain for 110,000
steps and discarded the first 10,000 steps as a burn in
period. Figure 9d shows the distribution of the 100,000
mean values for the T distribution modelling the data, and
the red curves in Figure 9c show 50 examples of the T
distribution with parameters (mean, standard deviation, and
a normality parameter – see (Kruschke, 2012)) taken at
random from the Markov Chain.
From the distribution of the mean value for the
difference data (Figure 9d), we can see that configuration 2
results in 59.5 more errors on average. If we define a
difference of 10 or fewer errors as a region of practical
equivalence (ROPE), or, in other words, we consider as
insignificant a change of 10 or fewer errors out of 10,000
tests, i.e., a change of less than 0.1%, we note that the 95%
highest density interval (HDI) of the distribution of the
mean of the difference of errors is outside the ROPE, and
therefore we conclude that changing the input images from
grey scale to binary values results in a small but significant
increase in error of around 0.6%.
Figure 12. Error rates as a function of the number of
neurons in the hidden layer.
respectively. It is clear that the error decreases with the
number of hidden layer neurons, although with a
diminishing return. Since the system used the timemultiplexing approach and rate neurons, the hardware cost
of a single TM neuron is almost negligible. The memory
required by the decoding weights is linearly proportional to
size of the hidden layer and is thus the bottleneck of the
system. To achieve a good balance between the desired
accuracy and memory, we chose to implement the hidden
layer with 8k rather than 16k neurons.
Next, we investigated the effect of using the rate
neurons in the hidden layer. The distribution of errors for
this configuration (configuration 3) is shown in Figure 10a.
This should be compared with configuration 2 (Figure 9b)
and their paired difference is shown in Figure 10b. Figure
10c shows the distribution of the mean of the difference in
errors between configuration 3 and configuration 2. It
shows that changing from tanh neurons to rate neurons
increases the number of errors by approximately 18.5.
However, this difference is not strongly significant, as the
95% HDI is not entirely outside the ROPE, indicating that a
difference within the region of practical equivalence is
amongst the possible mean values. Finally, we investigated
the effect of using limited-resolution decoding weights.
Figure 11a shows the distribution of errors for this
configuration and the difference between configuration 3
and configuration 4 is close to zero (Figure 11b). In fact the
distribution of the mean of the error difference is entirely
within the ROPE, indicating that somewhat surprisingly
there is no significant loss in performance when using 6-bit
fixed-point output weights instead of floating point weights.
3.2 System performance
To explore the best performance that the proposed
system can achieve, 1000 runs were carried out using the
full regression method (Figure 5) with different random
seeds. The lowest error achieved with lite and full version
of OPIUM is 4.52% and 3.45%, respectively. After that, the
decoding weights (obtained with full version of OPIUM)
were loaded into the FPGA board for real time digit
recognition. The pixels of input digits were converted to
binary values in software and a Python-based front-end
client software sent the selected test digit to the FPGA via
JTAG interface. Since the system runs at 266MHz and the
hidden layer contains 8k neurons, each of which has a time
slot of four clock cycles, the processing time for one input
digit will be 8k×4/266MHz ≈ 120 µs, yielding 1s/120µs ≈
8k digit recognitions per second. Due to the fact that our
system only used 8k out of 64k neurons in one single TM
neuron layer, the maximum number of the digit recognitions
that can be processed by one TM neuron layer is ~64k per
second. The system used less than 6% of the hardware
resources (with the exception of the RAMs), multiple TM
neuron layers can be instantiated to run in parallel. It is
practical to scale this system to process millions of digit
recognitions in one second. We will address this in details
in section 4.2.
The performance drop between configuration 1 and 4
was merely 0.8%. We can therefore conclude that, in this
digit recognition system, the modifications that we made
achieved significant reductions in terms of hardware cost
with a minimal drop in performance.
3.2 Size of the hidden layer
In this scenario, we used configuration 4 from the
previous section and changed the hidden layer size in the
range from 1k to 16k neurons. For each size, ten test runs
(each with a different random seed) were conducted. Again,
to reduce the testing time, we used OPIUM lite to calculate
the decoding weights and then calculate the test error.
4. Discussion
4.1 Comparison with other solutions
The work reported here constitutes the basis for building
real-time, large-scale, general purpose hardware pattern
recognition systems using the NEF, hence we are mainly
interested in the trade-off between the scale, the
The median error over 10 runs (Figure 12) for the
hidden layer with 1k, 2k, 4k, 8k, 12k and 16k neurons was
14.5%, 10.4%, 6.96%, 5.01%, 4.47% and 4.33%
10
performance and the hardware cost. We will concentrate on
comparing our work with the solutions that were developed
for similar goals, rather than the solutions that are extremely
optimised for achieving the lowest error rate of MNIST
although they cannot be efficiently implemented on
hardware.
TABLE II
Comparison with other solutions
The IBM TrueNorth system is a general-purpose system
for building large-scale neural networks running in real time
(Merolla et al., 2014). When it was programmed for digit
recognition, it achieved a result of 8.06% error rate in the
10000 test set of the MNIST with 13 cores, each of which
consisted of TM 256 spiking neurons and needs ~96k bits
memories (Esser et al., 2013). Hence, our system achieved a
much lower error rate while with significantly fewer
hardware resources, especially the memories (Table II).
Regarding the processing speed, their system needs 20 time
steps (each one is 1 ms) to process one digit, whereas our
system needs only 120 µs (approximately 167 times
speedup). Moreover, while their system consists of a feature
extractor that clusters and extracts features from data, our
system is feature-less, hence can be easily configured for
different input data without feature extractions. The
TrueNorth system however has much more applications
besides pattern recognition task, as compared to our system.
Error
Computation
time
Resources
Minitaur
8%
0.152 s
155k bits
TrueNorth
8.06%
20 ms
1.248M bits
This work
3.45%
120 µs
480k bits
is mainly used by the encoders. Hence, the maximum
number of the physical hidden neurons that can be
implemented is 230k/1600 ≈ 143. The memory requirement
of one single TM hidden neuron layer is 64k×60bits =
3840k bits. The on-chip SRAM, which is 52M bits, can be
used to implement up to 13 TM hidden neuron layers. To
further scale up the system, we need to use external
memories. The bandwidth requirement is indeed a
bottleneck for the time-multiplexing approach, as new
values need to be available from memory every four clock
cycles.
The maximum theoretical bandwidth of one DDR3
SDRAM memory and one QDRII+ SRAM memory on the
DE5 board is 512 bits and 72 bits @266MHz, respectively.
The DDR3 memory, in general, can only achieve an
efficiency of 70% (of the theoretical bandwidth) as it will
need flow control, which takes into consideration the bus
turn around time, refresh cycles, and so on. The maximum
number of neuron arrays is ((512bits × 2 × 70% +
72bits×4)×4)/60bits ≈ 67. Adding the ones using the onchip SRAM, the theoretical maximum number of neuron
layers is 80, yielding 64k×80 = 5.12M neurons. As the
maximum number of the digit recognitions that can be
processed by one TM neuron layer is ~64k per second, the
maximum number of the digit recognitions that can be
processed by the system with 80 parallel layers is therefore
5.12M per second.
The Minitaur, which is an event–based neural network
accelerator, achieved an error rate of 8% on a deep spiking
network with 1785 neurons (Neil and Liu, 2014). Since the
scheme it used is a variant of the time-multiplexing
approach, which only needs very few neurons to be
physically implemented, the cost of one single neuron is
also negligible and the bottleneck again is the memory.
Each of the neuron used by the Minitaur needs 73 bits
memories and the connection weight needs 16 bit
memories. Our neuron needs 60 bit memories for the
decoding weights. The processing time of the Minitaur for
one digit is 0.152s (table II), which is approximately 1300
times slower than our system.
4.2 Future work
Since the larger the scale is, the more pattern
recognitions can be carried out, our future work will focus
on scaling up the network that we have presented here. It is
a scalable design as it is a fully digital implementation. The
number of TM hidden neurons implemented by a single
physical neuron will increase linearly with the amount of
available memory, as long as the multiplexing scale keeps
the time resolution within the biological time scale. The
number of physical neurons will increase linearly with the
number of available ALMs.
The programmability of the FPGA, especially the
decoding weights, makes the integration of the system with
the desired pattern recognition applications seamless.
However, the advantages of running large-scale networks in
real-time are strongly reduced if such neural networks take
a long time to compute the decoding weights. Hence,
another major improvement is to speed up this
computationally extensive task. One promising solution is
to implement the OPIUM on FPGA, since this algorithm is
an adaption procedure without the requirement of hundreds
of Gigabyte RAMs and is quite friendly for hardware
implementation. Running OPIUM in real time makes it
possible to upgrade the system to be a true turnkey solution
for pattern recognition in real world. In addition, since the
proposed system does not need feature extraction, it could
be used for any other pattern recognition tasks such as
speaker recognition, natural language processing and so on.
In the following calculation, we will use the digits
recognition system as a metric and different applications
will require different amounts of hardware resources while
still using the same topology. We can calculate the
theoretical maximum network size on a state-of-the-art
FPGA board, such as the Terasic DE5 board containing an
Altera Stratix V (5SGXEA7N2F45C2) FPGA with ~230k
ALMs, two DDR3 SDRAMs and four QDRII+ SRAMs.
One single TM hidden layer requires ~1600 ALMs, which
11
5. Acknowledgment
Tapson, J., and van Schaik, A. (2013). Learning the pseudoinverse
solution to network weights. Neural Netw. 45, 94–100.
doi:10.1016/j.neunet.2013.02.008.
This work has been supported by the Australian
Research Council Grant DP140103001. The support by the
Altera university program is gratefully acknowledged. This
work was inspired by the Capo Caccia Cognitive
Neuromorphic Engineering Workshop 2013, 2014 and
Telluride Neuromorphic workshop 2013.
Thakur, C. S., Hamilton, T. J., Tapson, J., van Schaik, A., and Lyon, R. F.
(2014). FPGA Implementation of the CAR Model of the Cochlea.
in IEEE International Symposium on Circuits and Systems, 1853–
1856. doi:10.1109/ISCAS.2014.6865170.
Vogelstein, R. J., Mallik, U., Vogelstein, J. T., and Cauwenberghs, G.
(2007). Dynamically reconfigurable silicon array of spiking
neurons with conductance-based synapses. IEEE Trans. Neural
Netw. 18, 253–65. doi:10.1109/TNN.2006.883007.
6. References
Boahen, K. (2006). Neurogrid: emulating a million neurons in the cortex.
Conf. Proc. IEEE Eng. Med. Biol. Soc. Suppl, 6702.
doi:10.1109/IEMBS.2006.260925.
Wang, R., Cohen, G., Stiefel, K. M., Hamilton, T. J., Tapson, J., and van
Schaik, A. (2013). An FPGA Implementation of a Polychronous
Spiking Neural Network with Delay Adaptation. Front. Neurosci.
7, 14. doi:10.3389/fnins.2013.00014.
Cassidy, A., Andreou, A. G., and Georgiou, J. (2011). Design of a one
million neuron single FPGA neuromorphic system for real-time
multimodal scene analysis. 2011 45th Annu. Conf. Inf. Sci. Syst.,
1–6. doi:10.1109/CISS.2011.5766099.
Wang, R., Hamilton, T. J., Tapson, J., and van Schaik, A. (2014a). A
compact neural core for digital implementation of the Neural
Engineering Framework. in BIOCAS2014
doi:10.1109/BioCAS.2014.6981784.
Eliasmith, C., and Anderson, C. (2003). Neural Engineering:
Computation, Representation, and Dynamics in Neurobiological
Systems. Boston: MA: MIT Press.
Wang, R., Hamilton, T. J., Tapson, J., and van Schaik, A. (2014b). A
compact reconfigurable mixed-signal implementation of synaptic
plasticity in spiking neurons. in 2014 IEEE International
Symposium on Circuits and Systems (ISCAS) (IEEE), 862–865.
doi:10.1109/ISCAS.2014.6865272.
Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y.,
Tang, C., and Rasmussen, D. (2012). A large-scale model of the
functioning brain. Science 338, 1202–5.
doi:10.1126/science.1225266.
Wang, R., Hamilton, T. J., Tapson, J., and van Schaik, A. (2014c). An
FPGA design framework for large-scale spiking neural networks.
in 2014 IEEE International Symposium on Circuits and Systems
(ISCAS) (Melboune: IEEE), 457–460.
doi:10.1109/ISCAS.2014.6865169.
Esser, S. K., Andreopoulos, A., Appuswamy, R., Datta, P., Barch, D.,
Amir, A., Arthur, J., Cassidy, A., Flickner, M., Merolla, P., et al.
(2013). Cognitive computing systems: Algorithms and
applications for networks of neurosynaptic cores. in The 2013
International Joint Conference on Neural Networks (IJCNN)
(IEEE), 1–10. doi:10.1109/IJCNN.2013.6706746.
Wang, R. M., Hamilton, T. J., Tapson, J. C., and van Schaik, A. (2014d).
A mixed-signal implementation of a polychronous spiking neural
network with delay adaptation. Front. Neurosci. 8, 51.
doi:10.3389/fnins.2014.00051.
Kruschke, J. K. (2012). Bayesian Estimation Supersedes the t Test. J. Exp.
Psychol. Gen. 142, 573–603. doi:10.1037/a0029146.
Lecun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based
learning applied to document recognition. Proc. IEEE 86, 2278–
2324. doi:10.1109/5.726791.
Wang, R. M., Hamilton, T. J., Tapson, J. C., and van Schaik, A. (2015). A
neuromorphic implementation of multiple spike-timing synaptic
plasticity rules for large-scale neural networks. Front. Neurosci. 9,
1–17. doi:10.3389/fnins.2015.00180.
Merolla, P. A., Arthur, J. V., Alvarez-Icaza, R., Cassidy, A. S., Sawada, J.,
Akopyan, F., Jackson, B. L., Imam, N., Guo, C., Nakamura, Y., et
al. (2014). A million spiking-neuron integrated circuit with a
scalable communication network and interface. Science (80-. ).
345, 668–673. doi:10.1126/science.1254642.
Neil, D., and Liu, S. (2014). Minitaur, an Event-Driven FPGA-Based
Spiking Network Accelerator. IEEE Trans. Very Large Scale
Integr. Syst., 1–1. doi:10.1109/TVLSI.2013.2294916.
Penrose, R., and Todd, J. A. (1955). A generalized inverse for matrices.
Math. Proc. Cambridge Philos. Soc. 51, 406–413.
doi:10.1017/S0305004100030401.
Pfeil, T., Grübl, A., Jeltsch, S., Müller, E., Müller, P., Petrovici, M. A,
Schmuker, M., Brüderle, D., Schemmel, J., and Meier, K. (2013).
Six networks on a universal neuromorphic computing substrate.
Front. Neurosci. 7, 11. doi:10.3389/fnins.2013.00011.
Tapson, J. C., Cohen, G. K., Afshar, S., Stiefel, K. M., Buskila, Y., Wang,
R. M., Hamilton, T. J., and van Schaik, A. (2013). Synthesis of
neural networks for spatio-temporal spike pattern recognition and
processing. Front. Neurosci. 7, 153. doi:10.3389/fnins.2013.00153.
12
| 9cs.NE
|
ON THE BINOMIAL EDGE IDEALS OF BLOCK GRAPHS
arXiv:1402.2041v1 [math.AC] 10 Feb 2014
FARYAL CHAUDHRY, AHMET DOKUYUCU, RIDA IRFAN
A BSTRACT. We find a class of block graphs whose binomial edge ideals have minimal
regularity. As a consequence, we characterize the trees whose binomial edge ideals have
minimal regularity. Also, we show that the binomial edge ideal of a block graph has the
same depth as its initial ideal.
1. I NTRODUCTION
In this paper we study homological properties of some classes of binomial edge ideals.
Let G be a simple graph on the vertex set [n] and let S = K[x1 , . . ., xn , y1 , . . . , yn ] be the
polynomial ring in 2n variables over a field K. For 1 ≤ i < j ≤ n, we set fi j = xi y j − x j yi .
The binomial edge ideal of G is defined as JG = ( fi j : {i, j} ∈ E(G)). Binomial edge
ideals were introduced in [8] and [12]. Algebraic and homological properties of binomial
edge ideals have been studied in several papers. In [5], it was conjectured that JG and
in< (JG ) have the same extremal Betti numbers. Here < denotes the lexicographic order
in S induced by x1 > x2 > · · · > xn > y1 > y2 > · · · > yn . This conjecture was proved in [3]
for cycles and complete bipartite graphs. In [6], it was shown that, for a closed graph G,
JG and in< (JG ) have the same regularity which can be expressed in the combinatorial data
of the graph. We recall that a graph G is closed if and only if it has a quadratic Gröbner
basis with respect to the lexicographic order.
In support of the conjecture given in [5], we show, in Section 3, that if G is a block
graph, then depth(S/JG ) = depth(S/ in< (JG )); see Theorem 3.2. By a block graph we
mean a chordal graph G with the property that any two maximal cliques of G intersect in
at most one vertex.
Also, in the same section, we show a similar equality for regularity. More precisely, in
Theorem 3.3 we show that reg(S/JG ) = reg(S/ in< (JG )) = ℓ if G a Cℓ -graph. Cℓ -graphs
constitute a subclass of the block graphs; see Section 3 for definition and Figure 1 for an
example.
In [11] it was shown that, for any connected graph G on the vertex set [n], we have
ℓ ≤ reg(S/JG ) ≤ n − 1,
where ℓ is the length of the longest induced path of G.
2010 Mathematics Subject Classification. 13D02, 05E40.
Key words and phrases. Binomial edge ideals, regularity, depth.
The first and the third author were supported by the Higher Education Commission of Pakistan and the
Abdus Salam School of Mathematical Sciences, Lahore, Pakistan.
1
The main motivation of our work was to answer the following question. May we characterize the connected graphs G whose longest induced path has length ℓ and reg(S/JG ) =
ℓ? We succeeded to answer this question for trees. In Theorem 4.1, we show that if T is
a tree whose longest induced path has length ℓ, then reg(S/JT ) = ℓ if and only if T is
caterpillar. A caterpillar tree is a tree T with the property that it contains a path P such
that any vertex of T is either a vertex of P or it is adjacent to a vertex of P.
In [12], the so-called weakly closed graphs were introduced. This is a class of graphs
which includes closed graphs. In the same paper, it was shown that a tree is caterpillar if
and only if it is a weakly closed graph. Having in mind our Theorem 4.1 and Theorem 3.2
in [6] which states that reg(S/JG) = ℓ if G is a connected closed graph whose longest induced path has length ℓ, and by some computer experiments, we are tempted to formulate
the following.
Conjecture 1.1. If G is a connected weakly closed graph whose longest induced path has
length ℓ, then reg(S/JG) = ℓ.
2. P RELIMINARIES
In this section we introduce the notation used in this paper and summarize a few results
on binomial edge ideals.
Let G be a simple graph on the vertex set [n] = {1, . . ., n}, that is, G has no loops
and no multiple edges. Furthermore, let K be a field and S = K[x1 , . . . , xn , y1 , . . ., yn ] be
the polynomial ring in 2n variables. For 1 ≤ i < j ≤ n, we set fi j = xi y j − x j yi . The
binomial edge ideal JG ⊂ S associated with G is generated by all the quadratic binomials
fi j = xi y j − x j yi such that {i, j} ∈ E(G). Binomial edge ideals were introduced in the
papers [8] and [12].
We first recall some basic definitions from graph theory. A vertex i of G whose deletion
from the graph gives a graph with more connected components than G is called a cut
point of G. A chordal graph is a graph without cycles of length greater than or equal to
4. A clique of a graph G is a complete subgraph of G. The cliques of a graph G form
a simplicial complex, ∆(G), which is called the clique complex of G. Its facets are the
maximal cliques of G. A graph G is a block graph if and only if it is chordal and every
two maximal cliques have at most one vertex in common. This class was considered in
[5, Theorem 1.1].
The clique complex ∆(G) of a chordal graph G has the property that there exists a leaf
order on its facets. This means that the facets of ∆(G) may be ordered as F1 , . . ., Fr such
that, for every i > 1, Fi is a leaf of the simplicial complex generated by F1 , . . . , Fi . A leaf
F of a simplicial complex ∆ is a facet of ∆ with the property that there exists another facet
of ∆, say G, such that, for every facet H 6= F of ∆, H ∩ F ⊆ G ∩ F.
2.1. Gröbner bases of binomial edge ideals. Let < be the lexicographic order on S
induced by the natural order of the variables. As it was shown in [8], the Gröbner basis
of JG with respect to this order may be given in terms of the admissible paths of G.
2
Definition 2.1. [8] Let i < j be two vertices of G. A path i = i0 , i1 , . . . , ir−1, ir = j from i
to j is called admissible if the following conditions are fulfilled:
(1) ik 6= il for k 6= l;
(2) for each k = 1, . . . , r − 1 on has either ik < i or ik > j;
(3) for any proper subset { j1 , . . . , js} of {i1 , . . . , ir−1}, the sequence i, j1, . . . , js, j is
not a path in G.
Given an admissible path π in G from i to j, we set uπ = (∏ik > j xik )(∏il <i yil ).
Theorem 2.2. [8] The set of binomials
Γ=
[
{uπ fi j : π is an admissible path f rom i to j}
i< j
is the reduced Gröbner basis of JG with respect to the lexicographic order on S induced
by x1 > · · · > xn > y1 > · · · > yn .
2.2. Primary decomposition of binomial edge ideals. Theorem 2.2 shows, in particular, that in< (JG ) is a radical monomial ideal which implies that the binomial edge ideal JG
is radical as well. Hence JG is equal to the intersection of all its minimal primes. These
minimal prime ideals were determined in [8].
Let S ⊂ [n] be a (possible empty) subset of [n] and let G1 , . . . , Gc(S ) be the connected
components of G[n]\S , where G[n]\S denotes the restriction of G to the vertex set [n] \ S .
ei be the complete graph on the vertex set of Gi . Let
For 1 ≤ i ≤ c(S ), let G
PS (G) = (
[
{xi , yi }, JG̃1 , . . . , JG̃c(S ) ).
i∈S
Theorem 2.3. [8]
JG =
\
PS (G)
S ⊂[n]
In particular, the minimal primes of JG are among PS (G), where S ⊂ [n].
Theorem 2.4. [8] Let G be a connected graph on the vertex set [n], and S ⊂ [n]. Then
PS (G) is a minimal prime ideal of JG if and only if S = 0/ or S 6= 0/ and for each i ∈ S
one has c(S \ {i}) < c(S ).
A set S which satisfies the condition in the above theorem is called a cut-point set of
G.
3. I NITIAL
IDEALS OF BINOMIAL EDGE IDEALS OF BLOCK GRAPHS
In this section, we first show that, for a block graph G on [n] with c connected components, we have depth(S/JG ) = depth(S/ in< (JG )) = n + c, where < denotes the lexicographic order induced by x1 > · · · > xn > y1 > · · · > yn in the ring S = K[x1 , . . ., xn , y1 , . . . , yn ].
We begin with the following lemma.
3
Lemma 3.1. Let G be a graph on the vertex set [n] and let i ∈ [n]. Then
in< (JG , xi , yi ) = (in< (JG ), xi , yi ).
Proof. We have in< (JG , xi , yi ) = in< (JG\{i} , xi , yi ) = (in< (JG\{i} ), xi , yi ). Therefore, we
have to show that (in< (JG ), xi , yi ) = (in< (JG\{i} ), xi , yi ). The inclusion ⊇ is obvious since
JG\{i} ⊂ JG . For the other inclusion, let us take u to be a minimal generator of in< (JG ).
If xi | u or yi | u, obviously u ∈ (in(JG\{i} ), xi , yi ). Let now xi ∤ u and yi ∤ u. This means
that u = uπ xk yl for some admissible path π from k to l which does not contain the vertex
i. Then it follows that π is a path from k to l in G \ {i}, hence u ∈ in< (JG\{i} ).
Theorem 3.2. Let G be a block graph. Then depth(S/JG ) = depth(S/ in< (JG )) = n + c,
where c is the number of connected component of G.
Proof. Let G1 , . . . , Gc be the connected components of G and Si = K[{x j , y j } j∈Gi ]. Then
S/JG ∼
= S1 /JG1 ⊗ · · · ⊗ Sc /JGc , so that depth S/JG = depth S1 /JG1 + · · · + depth Sc /JGc .
Moreover, we have S/ in< (JG ) ∼
= S/ in< (JG1 ) ⊗ · · · ⊗ S/ in< (JGc ), thus depth S/ in< (JG ) ∼
=
depth S1 / in< (JG1 ) + · · · + depth Sc / in< (JGc ). Therefore, without loss of generality, we
may assume that G is connected. By [5, Theorem 1.1] we know that depth(S/JG ) = n + 1.
In order to show that depth(S/ in< (JG )) = n + 1, we proceed by induction on the number
of maximal cliques of G. Let ∆(G) be the clique complex of G and let F1 , . . . , Fr be a
leaf order on the facets of ∆(G). If r = 1, then G is a simplex and the statement is well
known. Let r > 1; since Fr is a leaf, there exists a unique vertex, say i ∈ Fr , such that
Fr ∩ Fj = {i} where Fj is a branch of Fr . Let Ft1 , . . ., Ftq be the facets of ∆(G) which
intersect the leaf Fr in the vertex {i}. Following the proof of [5, Theorem 1.1] we may
T
T
write JG = J1 ∩ J2 where J1 = i∈S
/ PS (G) and J2 = i∈S PS (G). Then, as it was shown
in the proof of [5, Theorem 1.1], it follows that J1 = JG′ where G′ is obtained from G
Sq
by replacing the cliques Ft1 , . . ., Ftq and Fr by the clique on the vertex set Fr ∪ ( j=1 Ft j ).
Also, J2 = (xi , yi ) + JG′′ where G′′ is the restriction of G to the vertex set [n] \ {i}.
We have in< (JG ) = in< (J1 ∩ J2 ). By [1, Lemma 1.3], we have in< (J1 ∩ J2 ) = in< (J1 ) ∩
in< (J2 ) if and only if in< (J1 + J2 ) = in< (J1 ) + in< (J2 ). But in< (J1 + J2 ) =in< (JG′ +
(xi , yi )+JG′′ ) =in< (JG′ +(xi , yi )). Hence, by Lemma 3.1, we get in< (J1 +J2 ) =in< (JG′ )+
(xi , yi ) = in< (J1 ) + in< (J2 ). Therefore, we get in< (JG ) = in< (J1 ) ∩ in< (J2 ) and, consequently, we have the following exact sequence of S-modules
S
S
S
S
−→
⊕
−→
−→ 0.
0 −→
in< (JG )
in< (J1 ) in< (J2 )
in< (J1 + J2 )
By using again Lemma 3.1, we have in< (J2 ) = in< ((xi , yi ), JG′′ ) = (xi , yi ) + in< (JG′′ ).
Thus, we have actually the following exact sequence
S
S
S
S
−→
⊕
−→
−→ 0.
(1) 0 −→
in< (JG )
in< (JG′ ) (xi , yi ) + in< (JG′′ )
(xi , yi ) + in< (JG′ )
Since G′ inherits the properties of G and has a smaller number of maximal cliques
than G, it follows, by the inductive hypothesis, that depth(S/JG′ ) = depth(S/ in< (JG′ )) =
4
n+1. Let Si be the polynomial ring S/(xi , yi ). Then S/((xi , yi )+in< (JG′′ )) ∼
= Si / in< (JG′′ ).
′′
Since G is a graph on n − 1 vertices with q + 1 connected components and satisfies our
conditions, the inductive hypothesis implies that depth S/((xi , yi ) + in< (JG′′ )) = n + q ≥
n + 1. Hence,
depth(S/ in< (JG′ ) ⊕ S/((xi , yi ) + in< (JG′′ ))) = n + 1.
Next, we observe that S/((xi , yi ) + in< (JG′ )) ∼
= Si / in< (JH ), where H is obtained from G′
Sq
by replacing the clique on the vertex set Fr ∪ ( j=1 Ft j ) by the clique on the vertex set
Sq
Fr ∪ ( j=1 Ft j ) \ {i}. Hence, by the inductive hypothesis, depth(S/((xi, yi ) + in< (JG′ ))) =
n since H is connected and its vertex set has cardinality n − 1. Hence, by applying the
Depth lemma to our exact sequence (1), we get depth S/JG = depth S/ in< (JG ) = n + 1.
Let G be a connected graph on the vertex set [n] which consists of
(i) a sequence of maximal cliques F1 , . . ., Fℓ with dim Fi ≥ 1 for all i such that |Fi ∩
Fi+1 | = 1 for 1 ≤ i ≤ ℓ −1 and Fi ∩Fj = 0/ for any i < j such that j 6= i +1, together
with
(ii) some additional edges of the form F = { j, k} where j is an intersection point of
two consecutive cliques Fi , Fi+1 for some 1 ≤ i ≤ ℓ − 1, and k is a vertex of degree
1.
In other words, G is obtained from a graph H with ∆(H) = hF1 , . . . , Fl i whose binomial
edge ideal is Cohen-Macaulay (see [5, Theorem 3.1]) by attaching edges in the intersection points of the facets of ∆(H). Therefore, G looks like the graph displayed in Figure 1.
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
F IGURE 1. Cℓ -graph
Such a graph has, obviously, the property that its longest induced path has length equal
to ℓ. If a connected graph G satisfies the above conditions (i) and (ii), we say that G is a
Cℓ -graph. In the case that dim Fi = 1 for 1 ≤ i ≤ ℓ, then G is called a caterpillar graph.
5
We should also note that any Cℓ –graph is chordal and has the property that any two distinct maximal cliques intersect in at most one vertex. So that any Cℓ -graph is a connected
block graph.
Theorem 3.3. Let G be a Cℓ -graph on the vertex set [n]. Then
reg(S/JG ) = reg(S/ in< (JG )) = ℓ.
Proof. Let G consists of the sequence of maximal cliques F1 , . . ., Fℓ as in condition (i) to
which we add some edges as in condition (ii). So the maximal cliques of G are F1 , . . . , Fℓ
and all the additional whiskers. We proceed by induction on the number r of maximal
cliques of G. If r = ℓ, then G is a closed graph whose binomial edge ideal is CohenMacaulay, hence the statement holds by [6, Theorem 3.2]. Let r > ℓ and let F1′ , . . . , Fr′ be
a leaf order on the facets of ∆(G). Obviously, we may choose a leaf order on ∆(G) such
that Fr′ = Fℓ . With the same arguments and notation as in the proof of Theorem 3.2, we
get the sequence (1).
We now observe that G′ is a Cℓ−1 -graph, hence, by the inductive hypothesis,
(2)
reg
S
S
= ℓ − 1.
= reg
JG′
in< (JG′ )
The graph G′′ has at most two non-trivial connected components. One of them, say
H1 , is a Cℓ′ -graph with ℓ′ ∈ {ℓ − 2, ℓ − 1}. The other possible non-trivial component, say
H2 , occurs if |Fℓ | ≥ 3 and, in this case, H2 is a clique of dimension |Fℓ | − 2 ≥ 1. By the
inductive hypothesis, we obtain
(3)
reg
S
S
S
S
= reg
= reg
+ reg
≤ ℓ − 1 + 1 = ℓ.
JG′′
in< (JG′′ )
JH1
JH2
Relations (2) and (3) yield reg(S/ in< (JG′ ) ⊕ S/((xi, yi ) + in< (JG′′ ))) ≤ ℓ. From the exact
sequence (1) we get
(4)
S
S
S
S
≤ max{reg
, reg
⊕
+ 1} ≤ ℓ.
reg
in< (JG )
in< (JG′ ) (xi , yi ) + in< (JG′′ )
in< (JG′ )
By [7, Theorem 3.3.4], we know that reg(S/JG ) ≤ reg(S/ in< (JG ), and by [10, Theorem 1.1], we have reg(S/JG ) ≥ ℓ. By using all these inequalities, we get the desired
conclusion.
4. B INOMIAL
EDGE IDEALS OF CATERPILLAR TREES
Matsuda and Murai showed in [10] that, for any connected graph G on the vertex set
[n], we have ℓ ≤ reg(S/JG ) ≤ n − 1, where ℓ denotes the length of the longest induced
path of G, and conjectured that reg(S/JG ) = n − 1 if and only if T is a line graph. Several
recent papers are concerned with this conjecture; see, for example, [6], [13], and [14].
One may ask as well to characterize connected graphs G whose longest induced path has
length ℓ and reg(S/JG ) = ℓ. In this section, we answer this question for trees.
6
A caterpillar tree is a tree T with the property that it contains a path P such that any
vertex of T is either a vertex of P or it is adjacent to a vertex of P. Clearly, any caterpillar
tree is a Cℓ -graph for some positive integer ℓ.
•
•
•
• • •
•
•
•
•
•
• •
•
•
•
•
•
•
• • •
F IGURE 2. Caterpillar
•
•
•
•
•
•
•
•
F IGURE 3. Induced graph H
Caterpillar trees were first studied by Harary and Schwenk [9]. These graphs have
applications in chemistry and physics [4]. In Figure 2, an example of caterpillar tree is
displayed. Note that any caterpillar tree is a narrow graph in the sense of Cox and Erskine
[2]. Conversely, one may easily see that any narrow tree is a caterpillar tree. Moreover,
as it was observed in [11], a tree is a caterpillar graph if and only if it is weakly closed in
the sense of definition given in [11].
In the next theorem we characterize the trees T with reg(S/JT ) = ℓ where ℓ is the length
of the longest induced path of T .
Theorem 4.1. Let T be a tree on the vertex set [n] whose longest induced path P has
length ℓ. Then reg(S/JT ) = ℓ if and only if T is caterpillar.
Proof. Let T be a caterpillar tree whose longest induced path has length ℓ. Then, by the
definition of a caterpillar tree, it follows that T is a Cℓ -graph. Hence, reg(S/JT ) = ℓ by
Theorem 3.3. Conversely, let reg(S/JT ) = ℓ and assume that T is not caterpillar. Then T
contains an induced subgraph H with ℓ + 3 vertices as in Figure 3.
7
Then, by [15, Theorem 27], it follows that reg(S/JH ) = ℓ + 1. Thus, since reg(S/JH ) ≤
reg(S/JG ) (see [10, Corollary 2.2]), it follows that reg(S/JG ) ≥ ℓ + 1, contradiction to our
hypothesis.
R EFERENCES
[1] A. Conca, Gorenstein ladder determinantal rings, J. London Math. Soc. 54(3) (1996), 453–474.
[2] D. A. Cox and A. Erskine, On closed graphs, arXiv:1306.5149.
[3] A. Dokuyucu, Extremal Betti numbers of some classes of binomial edge ideals, arXiv:1310.2903.
[4] S. El-Basil, Applications of caterpillar trees in chemistry and physics, J. Math. Chem. 1(2) (1987),
153–174.
[5] V. Ene, J. Herzog, T. Hibi, Cohen-Macaulay binomial edge ideals, Nagoya Math. J. 204 (2011), 57–68.
[6] V. Ene and A. Zarojanu, On the regularity of binomial edge ideals, to appear in Math. Nachr.
[7] J. Herzog and T. Hibi, Monomial Ideals, Graduate Texts in Mathematics 260, Springer, 2010.
[8] J. Herzog, T. Hibi, F. Hreinsdotir, T. Kahle, J. Rauh, Binomial edge ideals and conditional independence
statements, Adv. Appl. Math. 45 (2010), 317–333.
[9] F. Harary and A. J. Schwenk, The number of caterpillars, Discrete Math. 6(4), (1973), 359–365.
[10] K. Matsuda and S. Murai, Regularity bounds for binomial edge ideals, J. Commut. Algebra 5 (2013),
141–149.
[11] K. Matsuda, Weakly closed graphs and F-purity of binomial edge ideals, arXiv: 1209.4300.
[12] M. Ohtani, Graphs and Ideals generated by some 2-minors, Commun. Algebra 39 (2011), no. 3, 905–
917.
[13] S. Saeedi Madani and D. Kiani, Binomial edge ideals of graphs, Electron. J. Combin, 19 (2012), no.
2, # P44.
[14] S. Saeedi Madani and D. Kiani, The regularity of binomial edge ideals of graphs, arXiv: 1310.6126.
[15] S. Zafar and Z. Zahid, On the Betti numbers of some classes of binomial edge ideals, Electron. J.
Combin. 20 (2013) no. 4, # P37.
A BDUS S ALAM S CHOOL OF M ATHEMATICAL S CIENCES , GC U NIVERSITY, L AHORE . 68-B, N EW
M USLIM T OWN , L AHORE 54600, PAKISTAN
E-mail address: [email protected]
FACULTY OF M ATHEMATICS AND C OMPUTER S CIENCE , OVIDIUS U NIVERSITY B D . M AMAIA 124,
900527 C ONSTANTA , AND L UMINA -T HE U NIVERSITY OF S OUTH -E AST E UROPE S OS . C OLENTINA
NR . 64 B , B UCHAREST, ROMANIA
E-mail address: [email protected]
A BDUS S ALAM S CHOOL OF M ATHEMATICAL S CIENCES , GC U NIVERSITY, L AHORE . 68-B, N EW
M USLIM T OWN , L AHORE 54600, PAKISTAN
E-mail address: ridairfan [email protected]
8
| 0math.AC
|
S. Khor
12/5/2010
Static and dynamic characteristics of protein contact networks
Susan Khor1
Abstract The principles underlying protein folding remains one of Nature’s puzzles with important practical
consequences for Life. An approach that has gathered momentum since the late 1990’s, looks at protein heteropolymers and their folding process through the lens of complex network analysis. Consequently, there is now a
body of empirical studies describing topological characteristics of protein macro-molecules through their contact
networks and linking these topological characteristics to protein folding. The present paper is primarily a review of
this rich area. But it delves deeper into certain aspects by emphasizing short-range and long-range links, and
suggests unconventional places where “power-laws” may be lurking within protein contact networks. Further, it
considers the dynamical view of protein contact networks. This closer scrutiny of protein contact networks raises
new questions for further research, and identifies new regularities which may be useful to parameterize a network
approach to protein folding. Preliminary experiments with such a model confirm that the regularities we identified
cannot be easily reproduced through random effects. Indeed, the grand challenge of protein folding is to elucidate
the process(es) which not only generates the specific and diverse linkage patterns of protein contact networks, but
also reproduces the dynamic behavior of proteins as they fold.
Keywords: network analysis, protein contact networks, protein folding
1. Introduction
Breaking the code underlying protein folding has remained an intellectually tantalizing puzzle as
well as a problem of great practical significance. Everything a protein requires for correct folding under
normal circumstances appears to be embedded in its amino-acid sequence (Afinsen 1973), although a
minority rely on the aid of chaperone molecules to fulfill their destiny. Due to the large sizes that
amino-acid sequences can take, a random search approach to protein folding is deemed infeasible for
practical biological purposes (Levinthal 1969). However, an argument based on separability of the
protein folding problem, i.e. that the problem can be separated into parts which can be solved
independently and assembled into an optimal solution2, has been conceived as a way out of Levinthal’s
paradox (Zwanzig et al 1992; Karplus 1997). This argument is supported by the observation that some
sections of protein sequences have a propensity to fold to their secondary structures. In medicine,
protein misfolding has been identified as a causative factor in diseases such as cystic fibrosis, ALS and
Alzheimer’s (Chen et al 2008). Proteins play many roles in biological cells, e.g. as structural material,
1
Contact info.: [email protected]
For a more whimsical description, see Herbert A. Simon’s parable of the two watchmakers in The Architecture
of Complexity in The Sciences of the Artificial, 1969 MIT Press.
2
1
S. Khor
12/5/2010
catalysts, adaptors, hormones, transporters and regulators (Tramontano 2006). Proteins attain their
functionality through their unique (though not necessarily static) native three dimensional states and are
the ultimate expression of genes. Thus, the ability to predict the three dimensional structure of a protein
sequence is useful to comprehend genomic data. Current approaches to the protein structure prediction
problem such as comparative modeling, fold-recognition methods (profile-based and sequence
threading) and fragment-based methods rely heavily on existing knowledge of protein sequences and
their folds (Tramontano 2006). Notwithstanding the success of these methods, understanding the protein
folding process from first principles is a more complete and satisfying solution, and may prove
invaluable for protein design and therapies targeting protein misfolding.
In general, protein folding is a process that occurs in stages. What essentially begins as a linear
hetero-polymer (organized as a backbone with protruding side chain groups) obtains local structure in
the form of secondary alpha helices and beta sheets and finally global structure as the secondary
structures arrange themselves compactly in three dimensions. For a long time, this spontaneous
biological self-organization has been attributed to various inter-atomic physical forces and chemical
constraints impacting a protein molecule. However, in the last decade or so, another theory based on the
network topology of a protein’s native state has blossomed. Alm and Baker (1999) even suggest that
“protein folding mechanisms and landscapes are largely determined by the topology of the native state
and are relatively insensitive to details of the interatomic interactions.” In this other theory, a network
view of protein molecules (mostly in their native states) is adopted. The general recipe to transform a
protein molecule into a network is to represent amino acid residues (Cα or Cβ) as nodes, and contact
(spatial, non-covalent) distances between pairs of amino acid residues below a certain threshold as links.
Such protein contact networks or maps (Vendruscolo et al 2002) are constructed from the Cartesian
coordinates of amino acid residues of protein molecules stored in the Protein Data Bank (PDB) (Berman
et al 2000).
By examining protein contact networks (PCN), researchers have compiled a list of topological
characteristics shared by a diverse (in terms of structural class, homology and taxon) set of proteins and
speculated on the reasons for the observed topological characteristics in relation to the protein folding
mechanism. For example, a common feature of protein contact networks is their small-world nature, i.e.
they have lattice like clustering coefficients but random graph like diameters and characteristic path
lengths (Watts and Strogatz 1998). The need for rapid communication between amino acid residues to
facilitate interaction cooperativity crucial for protein folding is frequently cited as the reason for the
small-world feature of PCNs (Vendruscolo et al 2002; Dokholyan et al 2002; Atilgan et al 2004; Del
Sol et al 2006). Protein contact networks are also reported to exhibit high assortativity values which can
be related to protein folding speeds (Bagler and Sinha 2007).
2
S. Khor
12/5/2010
In this paper, we collected the PCN structural characteristics that have been reported in literature
on various sets of proteins and using different methods of PCN construction, and applied them on a
single set of proteins – the GH64 dataset – with a single PCN construction method. We find that in
general, the reported network characteristics hold for GH64. We also carry a line of inquiry found in the
literature (Green and Higman 2003) of probing the effects of short-range links and of long-range links
on network characteristics of PCNs throughout the paper. Our investigations led us to new research
questions, and to new quantitative and qualitative observations about PCNs which could be helpful to
finally realize a network approach to protein folding. These new observations include: about 30% of all
links are long-range, a power-law like distribution of link sequence distance, and for proteins which are
not too large, the maximum link sequence distance is about 70% of a protein’s size. However our model
of a network approach to protein folding is far from complete as the problem of choosing endpoints for
long-range links remains elusive (about 60% of a protein’s amino acids are involved in at least one
long-range interaction).
2. Protein Contact Map (PCN)
A protein contact network or PCN (Vendruscolo et al 2002) is an adjacency matrix representing a
network whose nodes are the Cα atom of amino acid residues of a protein, and where a link is placed
between a pair of nodes when the node pair is situated less than some distance threshold (TH) apart
from each other. Distance between node pairs is the Euclidean distance between their 3D Cartesian
coordinates obtained from the Protein Data Bank or PDB (Berman et al 2000). For this paper, TH is 7Å.
Structural characteristics of PCNs do not seem overly sensitive to the choice of the threshold value
(Bartoli et al 2007). We build a PCN for each protein in the GH64 dataset. Examples of four such PCNs
can be found in Figure 1. The proteins were randomly chosen for their differences in size: 1aep, 1agd,
1psd and 1CVJ have 153, 385, 813 and 1261 nodes in their PCNs respectively.
So far, we have referred to PCNs as uniform entities. In actuality, there are variations in the way
PCNs are constructed. For example, a single PCN may represent several non-homologous proteins
rather than a single protein. PCNs may also represent different aspects (e.g. surface or core), states (e.g.
native or transitional), structural classes (e.g. α, β, α+β or α/β), or types (e.g. globular or fibrous) of
proteins. Further, the nodes and links of PCNs may carry different meanings, e.g. atoms of the side
chain group of an amino acid may be included so that a node may represent more than one atom and
multiple links between nodes or weighted links are allowed. Examples of these alternative PCN
constructions can be found in the references for this paper. Incidentally, an earlier use of a matrix to
depict
protein
folding
can
be
found
3
in
(Tanaka
and
Scheraga
1977).
S. Khor
12/5/2010
Figure 1 Left: Protein Contact Networks (7Å) for four proteins in the GH64 dataset. A shaded cell
(x, y) denotes a link between nodes x and y. Right: Networks produced from two of the PCNs:
1aep (top) and 1psd (bottom). SE links are pink, LE links are turquoise, yellow highlights nodes
with more than 8 direct neighbours. Sections 4 and 5 of the text provide further explanation.
3. The GH64 Protein Dataset
The 64 proteins in the GH64 dataset (Table 1) were selected from literature surveyed, specifically
(Green and Higman 2003). The dataset encompass proteins from different protein classes, fold types and
branches of life. Their sizes range from 69 to 1734 residues (Figure 2). Proteins which did not form a
single component (1cuk and 1ho4), or had unusually high link density (1feo) in the PCNs we
constructed were excluded from the dataset. So too were proteins with more nodes in their PCNs than
their DSSP output (Kabsch and Sander 1983) (2hmz and 1epf). For the reverse situation, we truncated
the DSSP output. Another protein dataset, EVA132, is described in Appendix A and it is used as a
check to increase confidence in the main observations made in this paper.
Table 1 Proteins in the GH64 Dataset
1mjc 1gvp 1ten 1ris 2acy 1tlk 1ayc 1sha 1CD8 1d4t 1e86 2fgf 1eif 1pdo 1h7i 1amx 1bj7 1aep 1gm6
3rab 1wba 1rbp 1eyl 153L 1fap 1nsj 1hro 1jr8 256b 1ICE 1arb 1vlt 1urn 1amp 1j8m 1cjl 1beb 1OBP
1b7f 1hng 1agd 1aye 1g4t 1eov 1bmt 7tim 1ce7 1hwn 2AAI 1fbv 1bf5 1jly 1dar 1eun 1rpx 1bbp 1bih
1psd 1b8a 1ava 1CVJ 3eca 3kbp 1dio
4
12/5/2010
2000
1800
1600
1400
1200
1000
800
600
400
200
0
1mjc
1ten
2acy
1ayc
1CD8
1e86
1eif
1h7i
1bj7
1gm6
1wba
1eyl
1fap
1hro
256b
1arb
1urn
1j8m
1beb
1b7f
1agd
1g4t
1bmt
1ce7
2AAI
1bf5
1dar
1rpx
1bih
1b8a
1CVJ
3kbp
Number of residues (Nodes)
S. Khor
GH64 Proteins
Figure 2 Size of GH64 PCNs in terms of the number of Cα atoms documented in their PDB files.
4. Links: short-range (SE) and long-range (LE)
A link is placed between a pair of nodes representing the Cα atom of amino acids when the node
pair is situated less than 7Å apart from each other (section 2). We partitioned the set of links in a PCN
into two sets: long-range links (LE) and short-range links (SE). A link between nodes x and y is
classified as long-range if their absolute distance on the amino acid sequence chain is more than some
distance threshold (LE_TH). Long-range links connect amino acids which are far apart from each other
in the primary structure but are in close spatial proximity in the tertiary structure. In this study, LE_TH
is 9 (Green and Higman 2003), although we also consider LE_TH = 4. Unless otherwise stated, LE links
refer to long-range links with LE_TH = 9. Figures 3(a)-(c) examine various aspects of PCN links in the
GH64 dataset. Similar observations are made on the EVA132 dataset (Appendix A).
The number of links M in a PCN increases linearly with protein size (number of nodes in a PCN)
(Figure 3a). The number of long-range links |LE|, regardless of LE_TH, also increases linearly with
protein size, albeit at a slower rate (Figure 3a). For comparison, we cut in half the sequence distance
threshold for long-range links (LE_TH) from 9 to 4 (Figures 3a &b). On average, 30% of all links in a
PCN are LE (when LE_TH = 9, mean |LE| / M = 0.2929 with standard deviation of 0.0834; when
LE_TH = 4, mean |LE| / M = 0.3549 with std. dev. = 0.1005).
Instead of the entire PCN, Gaci and Balev (2009) considered the subgraph induced by the set of
amino acids participating in secondary structures (links belonging to the less evolutionary conserved
loop regions between SSEs are excluded). They call such a subgraph Secondary Structure Elements
Interaction Network or SSE-IN. Gaci and Balev (2009) observed that the ratio of the number of
interactions between secondary structures to the total number of interactions in SSE-INs does not
exceed 20%.
5
S. Khor
12/5/2010
8000
PCN Links (< 7A)
7000
Long-range links (> 9)
Count
6000
Long-range links (> 4)
5000
4000
3000
2000
1000
0
0
200
400
600
800
1000
1200
1400
1600
1800
Nodes
Figure 3a Link count M, by protein size.
Proportion of all links
0.6
0.5
0.4
0.3
0.2
Long-range links (> 9)
0.1
Long-range links (> 4)
0
0
200
400
600
800
1000
1200
1400
1600
1800
Nodes
Figure 3b Long-range links as a proportion of all links (|LE|/M), by protein size. When LE_TH is
9, the fraction of links in LE averages at 0.2929 with a standard deviation of 0.0834. When
LE_TH is 4, the average is 0.3549 with a standard deviation of 0.1005
0.12
PCN Links (< 7A)
Lattice 8
Lattice 4
Link Density
0.10
0.08
0.06
0.04
0.02
0.00
0
200
400
600
800
1000
1200
1400
Nodes
Figure 3c Link density by protein size.
6
1600
1800
S. Khor
12/5/2010
The GH64 PCNs have low link densities, and link density decreases rapidly with protein size
(Figure 3c). Link density is the fraction of actual links out of all possible links: 2M / [N (N-1)]. On a
log-log plot, link density for PCN in Figure 3c straightens out into a line (section 9). For reasons which
will become clearer in section 5, Figure 3c compares the link densities of GH64 PCNs with those of two
linear lattices: Lattice8 and Lattice4. LatticeV is a linear lattice with V/2 nearest neighbours to the left
and to the right where possible (V/2 nodes at each of the two ends of the lattice chain will have fewer
links than the rest of the nodes in the middle which will have V links each). For a given network size
(number of nodes), PCN link densities are indistinguishable from Lattice8 link densities (Figure 3c).
5. Node Degree and Degree Distribution
Typically, a node in a PCN represents the Cα atom of amino acids (section 2). The degree of a
node is the number of links attached to the node. In the present context, node degree measures the
number of contacts a node has in a PCN. Gaci and Balev (2009) remarked on the homogeneity of node
degree in their SSE-INs. The mean degree of SSE-INs increases very slightly with protein size and falls
within the range of 5 and 8. The absence of nodes with much higher degrees is attributed to the excluded
volume effect which imposes a physical limit on the number of residues that can reside within a given
radius around another amino acid.
Node degree distribution is the probability that a node has a particular degree. Node degree
distributions for a variety of PCNs (i.e. PCNs constructed in different ways and from different protein
sets) are bell-shaped, characteristic of a Gaussian distribution (Green and Higman 2003; Atilgan et al
2004; Bagler and Sinha 2005). Although, Green and Higman (2003) found that node degree
distributions of PCNs which consider only long-range interactions exhibit if not scale-free tendencies,
then exponential or highly right skewed distributions, irrespective of fold type. However, their PCN is
unlike our PCN and other PCNs referenced in this paper. Each amino acid residue is still represented by
a single node in the Green and Higman PCN. However, all atoms of an amino acid are considered when
deciding on contact or link establishment, and multiple links between nodes are allowed. Therefore the
Green and Higman PCN can register much larger node degrees, with reported modes of 100 (Green and
Higman 2003, p. 783). Gaci and Balev (2009) also observed scale-free tendencies in the node degree
distributions of their SSE-INs, with the caveat of a sharp cut-off above the mean node degree. Finding
scale-free connectivity patterns in PCNs is attractive since such connectivity pattern has direct
consequences on efficiency of information transfer and robustness to random attacks in proteins (Green
and Higman 2003; Atilgan et al 2004).
Node degree statistics for PCNs in GH64 is presented in Figure4a. The mean node degrees are
independent of protein size and hover around 8 (mean node degree averaged over all PCNs in GH64 is
7.9696 with a standard deviation of 0.3126). Little difference is observed between the corresponding
7
S. Khor
12/5/2010
mean node degrees and the median node degrees. The standard deviation values are about four times
smaller than the corresponding mean node degrees. All these statistics point towards bell shaped degree
distributions as reported in the literature, and as shown in Figure 4b for four proteins in GH64.
18
16
Node Degree
14
12
max
10
mid
8
mean
6
std.dev.
min
4
2
0
0
200
400
600
800
1000 1200 1400 1600 1800
Nodes
Figure 4a Node degree summary statistics for GH64 PCNs. The average mean node degree (K) is
7.9696 with a standard deviation of 0.3126.
0.35
0.30
Probability
0.25
1aep
0.20
1agd
0.15
1psd
1CVJ
0.10
0.05
0.00
0
2
4
6
8
10
Degree
12
14
16
Figure 4b Node degree distributions of four PCNs in the GH64 dataset.
Since we have partition the set of links into SE and LE (section 4), we can similarly classify the set
of nodes into two subsets (although these two subsets are not necessarily disjoint): SE_nodes for the set
of nodes which are endpoints of links in SE, and LE_nodes for the set of nodes which are endpoints of
links in LE. Figure 5a compares the sizes of these subsets and we find that LE_nodes is generally
slightly smaller than SE_nodes for proteins in GH64. This is not entirely surprising since we expect
almost all nodes to be in SE_nodes. For GH64, we found that the mean |SE_nodes| / N is 0.9970 with a
std. dev. of 0.0039, and the mean |LE_nodes| / N is 0.6659 with a std. dev. of 0.1093 where N = number
of nodes. Thus there is significant overlap between SE_nodes and LE_nodes, almost a subset
8
S. Khor
12/5/2010
relationship. For GH64, we found that mean |SE_nodes ∩ LE_nodes| / N is 0.6628 with a std. dev. of
0.1103. Green and Higman (2003) found that between 5% to 25% of nodes within each of their PCNs
do not have long-range contacts, i.e. are not in LE_nodes. For the GH64 PCNs, the mean (N |LE_Nodes|) / N is 0.3341 with a std. dev. of 0.1093 (Figure 5b). Similar observations are made on the
EVA132 dataset (Appendix A).
Node Count
1800
1600
SE
1400
LE
1200
1000
800
600
400
200
0
200
400
600
800
1000
1200
1400
1600
1800
Nodes
Figure 5a Number of nodes which are endpoints for links in SE (SE_nodes), and number of nodes
which are endpoint for links in LE (LE_nodes), by protein size for GH64. SE_nodes as a fraction
of all nodes averages at 0.9970 with a standard deviation of 0.0039. LE_nodes as a fraction of all
nodes averages at 0.6659 with a standard deviation of 0.1093.
SE_nodes
SE_ and LE_nodes
Proportion of all nodes
1.2
LE_nodes
Not in LE_nodes
1.0
0.8
0.6
0.4
0.2
0.0
0
200
400
600
800 1000 1200 1400 1600 1800
Nodes
Figure 5b Sizes of different subsets of nodes as a proportion of all nodes for GH64 PCNs. On
average, 0.6628 ± 0.1103 of all nodes are in both SE_nodes and LE_nodes, while 0.3341 ± 0.1093 of
all nodes are not in LE_nodes.
Being able to distinguish LE_nodes for a protein without prior knowledge of its 3D structure could
be an important step in predicting the 3D structure of proteins. Unfortunately, as we have observed
above, more than half of the amino acids in a protein sequence are expected to have at least one long-
9
S. Khor
12/5/2010
range interaction (in the protein’s native state). Perhaps a better strategy is to try to identify
distinguishing characteristics of the 30% or so nodes which are not involved in long-range interactions.
What about the node degree characteristics of SE_nodes and of LE_nodes? We use two ways to do
this comparison. First, we compute node degree statistics for nodes in SE_nodes (LE_nodes) using the
node degree values from the PCN constructed with SE (LE) only. The result of this first comparison is
displayed in Figure 6a. It yielded no significant difference at the 95% confidence level with the twosided paired t test (p-value = 0.871) or with the two-sided paired Wilcoxon test (p-value = 0.3305).
Nonetheless, we observed some differences between the degree distributions of PCN_SE (PCN
constructed with SE only) and PCN_LE (PCN constructed with LE only). The narrow degree ranges of
our PCNs limits our ability to observe long-tailed behavior over several scales (Figure 6b).
PCN SE avg
7
PCN LE avg
6
Degree
5
4
3
2
1
1mjc
1ten
2acy
1ayc
1CD8
1e86
1eif
1h7i
1bj7
1gm6
1wba
1eyl
1fap
1hro
256b
1arb
1urn
1j8m
1beb
1b7f
1agd
1g4t
1bmt
1ce7
2AAI
1bf5
1dar
1rpx
1bih
1b8a
1CVJ
3kbp
0
GH64 Proteins
Figure 6a Degree statistics for SE_nodes and for LE_nodes based respectively on PCNs
constructed from short-range links (SE) only and from long-range links (LE) only. Error bars
mark ± standard deviation.
0.6
1aep
0.7
1agd
0.6
1psd
0.5
1CVJ
1aep
0.5
Probability
Probability
0.9
0.8
0.4
0.3
0.2
1agd
1psd
0.4
1CVJ
0.3
0.2
0.1
0.1
0
0
0
2
4
6
0
8
2
4
6
8
10
PCN_LE Degree
PCN_SE Degree
Figure 6b Degree distributions of PCNs constructed with short-range (long-range) only links.
10
S. Khor
12/5/2010
Second, we compute node degree statistics for nodes in SE_nodes (LE_nodes) using the node
degree values from the entire PCN. Figure 6c displays the result of this second comparison. According
to the two-sided paired t test, and the two-sided paired Wilcoxon test, there is a significant difference at
the 95% confidence level between the two vectors. However, given the high overlap between SE_nodes
and LE_nodes, this difference though statistically significant, may be slight in magnitude.
14
12
Degree
10
8
6
4
SE avg
2
LE avg
1mjc
1ten
2acy
1ayc
1CD8
1e86
1eif
1h7i
1bj7
1gm6
1wba
1eyl
1fap
1hro
256b
1arb
1urn
1j8m
1beb
1b7f
1agd
1g4t
1bmt
1ce7
2AAI
1bf5
1dar
1rpx
1bih
1b8a
1CVJ
3kbp
0
GH64 Proteins
Figure 6c Degree statistics for SE_nodes and for LE_nodes based on node degree values for the
entire PCN. Error bars mark ± standard deviation.
6. Clustering
Clustering or transitivity reflects the cliquishness of nodes in a network: if node X connects to node
Y and to node Z, how likely is it that nodes Y and Z are connected to each other? There are different
ways of measuring such transitive relationships, strongly indicated in graphs by the presence of
triangles. One way is by taking the average clustering of all nodes in a network to yield the clustering
coefficient: C =
1
N
N
2ei
where ki is the degree of node i, and ei is the number of links that exist
i − 1)
∑ k (k
i =1
i
amongst the ki nodes (Watts and Strogatz 1998). Vendruscolo et al (2002) reports C values of around
0.58 ± 0.04 for their PCN (which has a different construction than ours), and Bagler and Sinha (2005)
found C values of around 0.553 ± 0.027, irrespective of protein structural class as defined in SCOP
(Murzin et al 1995).
For the GH64 PCNs, the C values averaged at 0.5484 with a std. dev. of 0.0233, and they are
independent of protein size. Figure 7a compares them with the C values for Lattice8, Lattice4 and the
theoretical C values for regular (CREGULAR) and random (CRANDOM) networks of the same size (same
number of nodes). CRANDOM ~ K / N, and CREGULAR = 3 (K-2) / [4 (K-1)], where K is average degree and
N is number of nodes (Watts 1999). The theoretical C values for 1D regular networks (CREGULAR) are
indistinguishable from the C values for Lattice8 (not surprising since mean degree of 8 is used). The C
11
S. Khor
12/5/2010
values for GH64 PCNs are slightly smaller than CREGULAR, and closer to the C values for Lattice4. This
last observation is telling as SE includes protein backbone links, and the protein backbone is the main
source of high clustering (Bartoli et al 2007). Further, Green and Higman (2003) observed that PCNs
built with only long-range links (LE) have relatively lower C values than those for PCNs with all links.
The same phenomenon is observed with GH64 proteins. Figure 7b shows the effect of randomizing
PCN links. Unlike randomizing only long-range links (randLE), randomizing only short-range links
(randSE) reduces clustering to levels similar to those of random graphs or randomizing all links
(randAll).
Clustering Coefficient
1.0
0.8
Regular
0.6
Lattice8
Lattice4
0.4
pcn
Random
0.2
0.0
0
200
400
600
800 1000 1200 1400 1600 1800
Nodes
Figure 7a Clustering coefficient values (± one standard deviation) for GH64 PCNs. Clustering
coefficients average at 0.5484 with standard deviation of 0.0233.
Clustering Coefficient
1.0
0.8
pcn
randLE
randSE
randAll
0.6
0.4
0.2
1mjc
1ten
2acy
1ayc
1CD8
1e86
1eif
1h7i
1bj7
1gm6
1wba
1eyl
1fap
1hro
256b
1arb
1urn
1j8m
1beb
1b7f
1agd
1g4t
1bmt
1ce7
2AAI
1bf5
1dar
1rpx
1bih
1b8a
1CVJ
3kbp
0.0
GH64 Proteins
Figure 7b Randomizing SE reduces clustering more than randomizing LE.
Links are randomized in the usual manner (Maslov and Sneppen 2002), by rewiring nodes while
preserving node degrees and without introducing multiple links between nodes. For example, to
randomize the set of long-range links (randLE), two LE links e1(a, b) and e2(c, d) are picked uniformly
12
S. Khor
12/5/2010
at random with replacement from LE. If both e3(a, c) and e4(b, d) do not already exist in the network,
then e3 and e4 replace e1 and e2.
High clustering in PCNs is almost always discussed in terms of the small-world property (Watts
and Strogatz 1998). Since the small-world property also involves the characteristic path length statistic
(section 8), the importance of short inter-node distances to protein folding inevitably dominates the
discussion. As a result, the role of high clustering (at the local interaction level) to protein folding is
somewhat under explained. A small exception to this is Bagler and Sinha (2005; 2007). They suggest
that high clustering in PCNs is indicative of a modular hierarchical organization as a result of the
protein folding process, but called for further investigation into this hypothesis (Bagler and Sinha 2005).
In (Bagler and Sinha 2007), they report a negative correlation between protein folding speed and high
clustering levels in PCN_LE (PCNs that only take into account long-range links). However, they did not
find any significant relationship between protein folding speed and clustering levels of entire PCNs.
Hence, the exact role clustering plays in protein folding is still an open question. Since most of the
clustering comes from short-range links (SE), which comprise backbone and secondary structure links,
it may be that high clustering of short-range links is necessary merely as a scaffolding device to create
the “all-important” long-range contacts. Alternatively, it may be that the high clustering levels found in
PCNs is a side effect of another network characteristic with a clearer relationship with some aspect of
protein folding, for example assortativity and protein folding speed (section 7). Or, it could be that the
local organization of short-range links really does play a role in some aspect of protein folding. In social
networks, transitive relationships are assumed to be part of the nature of human social behavior, and the
problem is to explain the establishment and subsequent efficient reuse of “random” short-cuts. An
interesting example of such a study is Kleinberg (2000), although how far the analogy between human
social networks and protein contact networks (or any other kind of residue “interaction” network)
applies to protein folding remains to be seen.
7. Assortativity
Network assortativity refers to the extent that nodes associate or connect with their own kind. A
common form of assortativity measured for PCNs is node degree. In this section (and throughout this
paper), positive assortativity refers to the proclivity of nodes with small (large) degree to link with other
nodes of small (large) degree. Using the method in (Newman 2002), Bagler and Sinha (2007) report
degree-degree correlation coefficients up to 0.58, which is considered unusual for networks with
biological origins (we discuss this point further ahead). Nonetheless, the positive assortativity values
could be correlated in a positive manner to protein folding speeds (Bagler and Sinha 2007).
Similarly, we find that the GH64 PCNs have positive degree-degree correlations (Figure 8a). Their
assortativity values average at 0.3387 with a standard deviation of 0.0536, and are independent of
13
S. Khor
12/5/2010
protein size. Figure 8b uses an alternative method (which is less sensitive to the effects of superconnected nodes) to assess degree assortativity (Pastor-Satorras et al 2001). In this other method, a
positive correlation between degree k and the average degree of nodes directly neighbouring nodes of
degree k is interpreted as evidence of positive degree-degree assortativity. In general, the plots in Figure
8b agree with the summary in Figure 8a. Although, the plots show some weakening (even reversal) in
the relationship as node degree increases, which could be due to the limits on node degree (Bagler and
Sinha 2007). This coupled with assortativity values close to those for Lattice4 (Figure 8a) is telling.
Like clustering (section 6), assortativity is more sensitive to random rewiring of short-range links
(randSE) than to random rewiring of long-range links (randLE) (Figure 8c). And when analyzed
separately, long-range links (LE) show less positive assortativity than short-range links (SE). This
means that like clustering, the high positive assortativity values of PCNs also mainly stem from SE.
0.8
Assortativity
0.6
Lattice8
0.4
Lattice4
pcn
0.2
randAll
0.0
-0.2
0
200
400
600
800 1000 1200 1400 1600 1800
Nodes
Figure 8a The GH64 PCNs have positive degree assortativity, with values close to Lattice4.
Mean Assortativity is 0.3387 with a standard deviation of 0.0536.
Average Neighbourhood Degree
12.0
10.0
8.0
1aep
1agd
6.0
1psd
1CVJ
4.0
2.0
0.0
0
2
4
6
8
10
Degree
12
14
16
Figure 8b Average neighbourhood degree for degree k is the average degree of all nodes adjacent
to all nodes of degree k. A positive correlation implies positive degree-degree assortativity.
14
S. Khor
12/5/2010
0.8
Assortativity
0.6
pcn
randLE
randSE
randAll
0.4
0.2
0.0
1mjc
1ten
2acy
1ayc
1CD8
1e86
1eif
1h7i
1bj7
1gm6
1wba
1eyl
1fap
1hro
256b
1arb
1urn
1j8m
1beb
1b7f
1agd
1g4t
1bmt
1ce7
2AAI
1bf5
1dar
1rpx
1bih
1b8a
1CVJ
3kbp
-0.2
GH64 PCMs
Figure 8c Effect of randomizing links on Assortativity.
randSE produces a larger effect than randLE.
0.6
Assortativity
0.4
0.2
0.0
SE
-0.2
LE
-0.4
0
200
400
600
800
1000
1200 1400
1600
1800
Nodes
Figure 8d Assortativity for long-range links (LE) and short-range links (SE).
Previously, a number of biological networks including gene regulatory networks were described as
disassortative, and it was suggested that this negative assortativity (specifically nodes of high degree or
hubs are not directly connected with each other in the networks studied) is advantageous in the sense
that the effects of harmful perturbations could be better localized (Maslov and Sneppen 2002). Brede
and Sinha (2005) conclude that the stability of networks with assortative mixing by node degree (i.e.
positive assortativity) declines more rapidly with increases in network size. A system with fragile
stability is easily dislodged from its equilibrium state by small disturbances. However, the positive
assortativity of PCNs may actually work to the benefit of the protein folding process by helping
information to flow quickly and thus facilitate coordinated action crucial for correct and rapid protein
folding. Further, native state proteins “adjust” themselves during ligand-binding, and thus some
vulnerability to perturbations or marginal stability may be advantageous (Taverna and Goldstein 2002).
15
S. Khor
12/5/2010
Indeed if percolation is important, Vazquez and Moreno (2003) conclude that node degree assortativity
can make networks robust to random node or edge removals. Additionally, there could be some
delegation of responsibilities between interacting biological networks in the sense that the amino acid
chain is the product of gene translation and concerns about containing unwanted perturbations in a
folding protein may be taken care of by the disassortative structure of genetic networks. Additionally,
since LE is less assortative than SE (Figure 8d), there may still be some degree of isolation between
parts (node subsets) of a PCN to buffer against unexpected perturbations. It appears that there are yet
interesting discussions to be had on this topic.
8. Average Path Length
Average path length of a network is the average shortest path between all node-pairs. It represents
the number of links that needs to be traversed on average when trying to move between nodes in a
network. PCNs are known to have average path lengths characteristic of random graphs of the same
size. For their respective protein sets and PCN constructions, Vendruscolo et al (2002) report an average
path length of 4.1 ± 0.9, while Bagler and Sinha (2005) report an average path length of 6.88 ± 2.61.
Path length is widely mentioned as a topological feature of PCNs that is highly relevant to protein
folding (Vendruscolo et al 2002; Dokholyan et al 2002; Atilgan 2004; Del Sol et al 2006). The
argument is: Short path lengths are conducive to rapid communication between amino acid molecules of
a protein, which facilitates interaction cooperativity crucial for protein folding. We revisit this point
elsewhere (manuscript submitted). Briefly, we find that shorter inter-node distances is no guarantee of
finding a global optimum more easily if the shorter average path lengths is the result of more links and
the fitness function is such that additional links increase frustration.
Figure 9a gives various path length statistics for GH64 PCNs. Both diameter (maximum shortest
path length found) and average path length increase logarithmically as protein size (nodes) increases.
When compared with average path lengths of other canonical networks, the average path lengths of
PCNs are much shorter than regular graphs and approximate the average path lengths expected of
random graphs of the same size (Figure 9b). Although the standard deviations indicate rather large
dispersions of shortest path lengths in the PCNs, the average path length and median path length values
are very close to each other (Figure 9a). We gather from this that there is little skew in the distribution
of shortest path lengths for PCNs, and it is indeed the case. Figure 9c shows the Gaussian like
distributions of shortest path lengths for four GH64 PCNs.
16
S. Khor
12/5/2010
32
28
Path Length
24
20
maximum
16
mean
12
median
8
4
0
0
200
400
600
800 1000 1200 1400 1600 1800
Nodes
Figure 9a Diameter, average (± one standard deviation) and median path length for GH64 PCNs.
300
Lattice4
Mean Path Length
250
Lattice8
Regular
200
pcn
150
Random
100
50
0
0
200
400
600
800
1000
1200
1400
1600
1800
Nodes
Figure 9b Average path lengths of PCNs are much closer to average path lengths of random
networks (LRANDOM) than to average path length of regular networks (LREGULAR). LRANDOM ~ ln N /
ln K, and LREGULAR = N (N + K-2) / [2K (N-1)], where K is average degree and N is number of
nodes (Watts 1999).
0.25
1aep
Probability
0.20
1agd
1psd
0.15
1CVJ
0.10
0.05
0.00
0
5
10
15
20
25
30
35
Shortest Path Length
Figure 9c Distribution of shortest path lengths for four GH64 PCNs.
17
S. Khor
12/5/2010
Figures 10a&b respectively show the effect of randomizing links on the average path lengths and
on the diameters of PCNs. Like Clustering (section 6) and Assortativity (section 7), randomizing only
long-range links (randLE) produces a smaller effect than randomizing only short-range links (randSE).
Both the two-sided paired t test and the two-sided paired Wilcoxon test report a significant difference at
the 95% confidence level between the randSE and the randLE vectors of mean path length values
(Figure 10a). Indirectly, this shows that a larger proportion of the magnitude of a PCN’s average path
length is due to SE.
28
Mean Path Length
24
pcn
randLE
randSE
randAll
20
16
12
8
4
0
0
200
400
600
800
1000
1200
1400
1600
1800
Nodes
44
40
36
32
28
24
20
16
12
8
4
0
pcn
randLE
randSE
randAll
1mjc
1ten
2acy
1ayc
1CD8
1e86
1eif
1h7i
1bj7
1gm6
1wba
1eyl
1fap
1hro
256b
1arb
1urn
1j8m
1beb
1b7f
1agd
1g4t
1bmt
1ce7
2AAI
1bf5
1dar
1rpx
1bih
1b8a
1CVJ
3kbp
Maximum Path Length
Figure 10a Effect of randomizing links on the average path lengths of GH64 PCNs. randSE has a
stronger effect than randLE.
GH64 Proteins
Figure 10b Effect of randomizing links on the diameters of GH64 PCNs. randSE has a stronger
effect than randLE.
Randomization of links also reveals that it is possible to rearrange the links of a PCN so that the
average path length is significantly reduced. Thus the question: if short inter-nodal distances are
important for protein folding, why aren’t the average path lengths of PCNs shorter than they are?
18
S. Khor
12/5/2010
Preserving high clustering and positive assortativity (if they are important for protein folding) is not an
adequate answer since it is still possible to maintain clustering and positive assortativity at levels higher
than would be in random graphs, while significantly reducing the average path lengths of PCNs by
randomizing the long-range links (randLE). We investigate this question elsewhere and find evidence
showing that randLE can be less conducive to search than PCN (manuscript submitted).
9. Power-laws in PCNs
In the more recent history of the science of complex networks, power-law distribution or the lack
of scale is often associated with the node degree variable. We have already touched upon this in section
5 in relation to PCNs. In this section we offer two other (non-node degree) places in PCNs where
power-law distribution can be found. The first resides across proteins (PCNs) in the relationship
between nodes and link density. This derives directly from the mean node degree constant, K (Figure
4a). The second resides within individual PCNs in terms of the distribution of link sequence distances.
9.1 Link Density Distribution across PCNs
Figure 11 (left) is the link density plot in Figure 3c on a log-log scale, and Figure 11 (right) is a
similar plot for Figure A3 in Appendix A. Both show clear power-law relationships, but with an atypical
exponent of -1. Number of links, M = K N / 2 where K is mean average node degree and N is number of
nodes. Link density = (2 M) / [N (N-1)] ≈ K / N. Empirical data tells us that K = 8 (section 5).
1.000
1.000
10
100
1000
0.100
0.010
1
10000
Link Density
Link Density
1
10
100
1000
10000
0.100
0.010
PCN Links (< 7A)
PCN Links (< 7A)
8/N
8/N
0.001
0.001
Nodes (EVA132)
Nodes (GH64)
Figure 11 Links density plots on logarithmic scales.
9.2 Link Sequence Distance Distribution for a PCN
Figure 12a summarizes the link sequence distance for PCNs in GH64. Sequence distance is the
absolute distance of a link on the protein sequence (section 4). Minimum sequence distance for all PCNs
is 1. We find for N < 4000, the mean link sequence distance generally increases with protein size as a
consequence of being pulled up by the maximum link sequence distance which generally increases with
protein size. However, the median link sequence distance remains fairly constant with protein size and
19
S. Khor
12/5/2010
well below the average link sequence distance per PCN. Similar observations are made for PCNs in the
EVA132 dataset (Figure A8). These summary statistics hint at a non-Gaussian distribution for link
sequence distances. The link sequence distance distributions for four proteins on a log-log scale can be
seen in Figures 12b and 13a. There is a straight line region with a negative slope and a sharp cut-off
point in each plot, revealing traces of a power-law. Taking the maximum link sequence distance as a
fraction of the number of nodes per PCN reveals that on average, links have a maximum span of about
0.7 N (Figures 12c and 13b).
Link Sequence Distance
160
Avg.
140
Std. dev.
120
Median
100
80
60
40
20
0
0
200
400
600
800
1000
1200
1400
1600
1800
1200
1400
1600
1800
Nodes
Link Sequence Distance
1600
Max
1400
1200
1000
800
600
400
200
0
0
200
400
600
800
1000
Nodes
Figure 12a Link sequence distance summary statistics for GH64 PCNs indicate right skewed
distributions for link sequence distance.
20
12/5/2010
Reversed Cumulative Probability
S. Khor
1
0.1
0.01
1aep
1agd
0.001
1psd
1CVJ
0.0001
1
10
100
1000
Link Sequence Distance
Figure 12b Link sequence distance distributions for 4 GH64 PCNs on logarithmic scales.
1.2
MaxLinkSeqDist/N
1.0
0.8
0.6
0.4
0.2
0.0
0
200
400
600
800
1000
1200
1400
1600
1800
Nodes
Figure 12c Maximum link sequence distance as a fraction of the number of nodes in a PCN.
MaxLinkSeqDist/N averages at 0.7874 ± 0.1402 for GH64.
Reversed Cumulative Probability
1.0000
0.1000
0.0100
1fp3
1S5L
0.0010
1m32
1sw6
0.0001
1
10
100
1000
Link Sequence Distance
10000
Figure 13a Link sequence distance distributions for 4 EVA132 PCNs on logarithmic scales.
21
S. Khor
12/5/2010
1.2
MaxLinkSeqDist/N
1.0
0.8
0.6
0.4
0.2
0.0
0
800
1600
2400
3200
4000
4800
5600
6400
Nodes
Figure 13b Maximum link sequence distance as a fraction of the number of nodes in a PCN.
MaxLinkSeqDist/N averages at 0.7068 ± 0.1860 for EVA132.
10. A Dynamic View
In this section, we study the effects of adding long-range links to a PCN already containing shortrange links. We assume links are independent of each other and that links of a shorter sequence distance
form before links of a longer sequence distance. Long-range links are added one at a time, and in order
of non-decreasing link sequence distance, and we do not pay attention to the order of links having the
same sequence distance but use random selection instead. This is admittedly a crude way of simulating
the folding process of a protein, but it captures its essence. We monitor the change over time of two
network characteristics: node betweeness and average path length. Based on the literature surveyed, we
expect a number of nodes to exhibit distinctively higher node betweeness, and we expect the average
path length to decrease over time. That average path length will decrease is a given (although occasional
small increases is possible), but what is more interesting and pertinent is the way the general decrease
happens.
Betweeness of a node is a measure of the node’s centrality, or importance in terms of path traversal
between nodes in a network. More specifically, the betweeness of a node is the fraction of shortest paths
found between all node pairs that pass through the node. In an effort to identify amino acids that play a
key role in a protein’s folding process, Vendruscolo et al (2002) observed that key residues often exhibit
significantly higher node betweeness in a protein’s transition state. However, this signal attenuates as a
protein reaches its native state (due to compactness). Hence, looking at node betweeness in a protein’s
native state PCN is more likely to produce false positives of a protein’s nucleation sites. Nonetheless, a
topological (network structure) understanding of the protein folding process may help us understand
how a protein’s unique native conformation is chosen (Vendruscolo et al 2002).
Dokholyan et al (2002) used topological characteristics of protein conformation graphs (PCN) to
distinguish conformations belonging to a protein’s transition state ensemble (TSE) that will cross the
22
S. Khor
12/5/2010
energy barrier and reach the protein’s native state versus those that will not. They found that posttransition conformations (TSE conformations which have their nucleus formed and are more likely to
reach the native state) are more small-world like than pre-transitions conformations (TSE conformations
which have not formed their nucleus and are less likely to reach the native state) in the sense that the
average path length of post-transition conformation graphs (PCN) is significantly shorter than the
average path length of pre-transition PCNs. In other words, there is a tightening of the small-world
structure of proteins as they approach their native conformations (become more compact).
Figure 14a shows that as the 1agd protein “folds” (using the scheme described above), average
node betweeness decreases slightly, but experiences a phase transition just after t=350, resulting in a
sudden and substantial increase in average node betweeness accompanied also by an increase in node
betweeness standard deviation, and a distinct shortening of average path length (Figure 14c). Both the
median and the maximum node betweeness also register sudden and substantial increases at the same
point in time (Figure 14b). These effects indicate a wider dispersion of node betweeness values amongst
the nodes with some nodes exhibiting much higher node betweeness as observed in (Vendruscolo et al
2002). In Figure 15a, we compare the distribution of node betweeness for 1agd at three points in time:
before the transition point (t=250), around the transition point (t=355) and at the end of the folding
process (t=482). All three plots are highly right-skewed (the use of a log-log plot helps with
visualization). Limiting the comparison to the three points in time, we find a definite widening of node
betweeness distribution from t=250 to t=355, followed by a narrowing of node betweeness distribution
from t=355 to t=482 (the tail of the plot swishes to the right then to the left). Betweeness of nodes
become more evenly spread out amongst nodes in 1agd’s native state.
However, the folding dynamics of 1aep shows a more subtle transition (Figure 14 right). At around
the transition point (t=29) where average path length shortens at a faster rate (Figure 14f), there is no
sudden or substantial increase in average node betweeness (Figure 14d) and maximum node betweeness
is actually dipping rapidly (Figure 14e). But there is a spike in median node betweeness (Figure 14e).
These effects are not indicative of a widening then narrowing range of node betweeness values (Figure
15b), but a convergence of some kind (average and median values get closer to each other in both 1aep
and 1agd). It appears that as with 1agd, betweeness of nodes become more evenly spread out amongst
nodes in 1aep’s native state (and therefore become less indicative of the key residues). Analogous to the
idea of crossing an energy barrier in protein folding kinetics, this change in node betweeness could be
seen as crossing some kind of communication barrier.
Within the nucleation model of protein folding, time prior to the transition point could be regarded
as a nucleation growth phase. After the nucleus is formed, the protein polymer becomes committed to its
native conformation and quickly reaches it as indicated by the rapid shortening of average path lengths.
It would be interesting to investigate this effect further: (i) to test the reliability of median node
23
S. Khor
12/5/2010
betweeness as an indicator of a nucleation event in protein folding, and (ii) to monitor the nucleation
rates for proteins of different fold types and sizes to see whether there are constants or specific
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.00
-0.02
-0.04
-0.06
(a)
0.25
avg
median
Node Betweeness
(d)
0.20
avg
0.15
median
0.10
0.05
0.00
-0.05
-0.10
0
50
0
100 150 200 250 300 350 400 450 500
10
20
30
0.006
max
0.005
0.004
0.003
0.002
0.001
0
0
70
0.60
(e)
0.025
0.50
0.020
0.40
0.015
0.30
0.010
0.20
median
0.005
0.10
max
0.000
50 100 150 200 250 300 350 400 450
0.00
0
10
20
Time
30
40
50
60
Time
6.5
(c)
20
Shortest Path Length
25
Shortest Path Length
60
0.030
Median
median
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
Maxim um
Median
0.007
(b)
50
Time
Time
0.008
40
Maxim um
Nodes Betweeness
relationships.
hpl
15
10
5
0
(f)
6.0
hpl
5.5
5.0
4.5
4.0
3.5
3.0
0
50 100 150 200 250 300 350 400 450 500
Time
0
10
20
30
40
50
60
70
Time
Figure 14 Left: plots for 1agd. Right: plots for 1aep. Each time step marks the addition of a longrange link to the PCN. At time=0, the PCN only has short-range links. (a) and (d): Average node
betweeness ± standard deviation. Median node betweeness lie below average values throughout.
(b) and (e): Median and maximum node betweeness values. (c) and (f): Average path length
measured using the harmonic mean method (Newman, 2003) to allow for possibility of
disconnected PCNs. hpl is more reliable than apl when dealing with disconnected networks, e.g.
1agd (Figure 16).
24
S. Khor
12/5/2010
Reversed Cumulative Probability
1.000
0.100
1agd
0.010
t=250
t=355
t=482
0.001
0.0001
0.001
0.01
0.1
1
Node Betweeness
Reversed Cumulative Probability
Figure 15a Distribution of node betweeness for 1agd at three points in time: 250, 355 and 482, on
logarithmic scales. The tail ends of the plots swishes to the right then to the left.
1.000
0.100
1aep
0.010
t=15
t=29
t=62
0.001
0.0001
0.001
0.01
0.1
1
Node Betweeness
Figure 15b Distribution of node betweeness for 1aep at three points in time: 15, 29 and 62, on
logarithmic scales. The tail ends of the plots swishes only in one direction, towards the left.
For comparison, Figures 16 and 17 show how average path length changes if long-range links are
added in random order, i.e. not added in order of non-decreasing sequence distance. For both proteins,
the drop in average path length occurs earlier when long-range links are added in random order, than
when long-range links are added in order of non-decreasing link sequence distance order. This kind of
decrease is similar to that observed in Figure 18 (section 11). Hence protein folding is a specific
rewiring process (Dokholyan et al 2002).
25
12/5/2010
350
300
Link Sequence Distance
Link Sequence Distance
S. Khor
1agd
250
200
150
100
50
0
0
350
300
250
200
150
100
50
0
50 100 150 200 250 300 350 400 450 500
0
Time
Time
25
hpl
20
Shortest Path Length
Shortest Path Length
25
15
10
5
0
hpl
20
15
10
5
0
0
50 100 150 200 250 300 350 400 450 500
0
Time
50
45
40
35
30
25
20
15
10
5
0
apl
0
50 100 150 200 250 300 350 400 450 500
Time
S hortest Path Length
Shortest Path Length
50 100 150 200 250 300 350 400 450 500
50 100 150 200 250 300 350 400 450 500
Time
50
45
40
35
30
25
20
15
10
5
0
apl
0
50 100 150 200 250 300 350 400 450 500
Time
Figure 16 Behavior of average path length as long-range links for 1agd are added in nondecreasing order (left) versus in random order (right).
26
12/5/2010
140
Link Sequence Distance
Link Sequence Distance
S. Khor
1aep
120
100
80
60
40
20
0
0
10
20
30
40
50
60
140
120
100
80
60
40
20
0
70
0
10
20
Time
50
60
70
6.5
Shortest Path Length
Shortest Path Length
40
Time
6.5
hpl
6.0
5.5
5.0
4.5
4.0
3.5
3.0
hpl
6.0
5.5
5.0
4.5
4.0
3.5
3.0
0
10
20
30
40
50
60
70
0
10
20
Time
30
40
50
60
70
Time
25
Shortest Path Length
25
Shortest Path Length
30
apl
20
15
10
5
0
apl
20
15
10
5
0
0
10
20
30
40
50
60
70
Time
0
10
20
30
40
50
60
70
Time
Figure 17 Behavior of average path length as long-range links for 1aep are added in nondecreasing order (left) versus in random order (right).
11. A Network Approach to Protein Folding
From a network perspective, the protein folding problem boils down to generating the right set of
links between amino acid residues of a protein. While many network characteristics of PCNs have been
discovered over the past decade, there is less (published) effort to integrate this knowledge into a single
model and use it for the protein folding problem. Perhaps this is due to the current difficulty of creating
networks to simultaneously fit a number of static and dynamic structural characteristics, still incomplete
description of PCNs, the very specific and diverse arrangement of links in PCNs, and/or even the
27
S. Khor
12/5/2010
relative success of knowledge-based protein structure prediction methods. In this section, we outline an
approach to protein folding that uses several of the network structural characteristics of PCNs
mentioned in this paper. This model is still in its infancy, but we hope that it is a good first step in
realizing a network approach to protein folding ab initio, or at least put to rest naïve conceptions of PCN
models.
The model begins with short-range links arranged on a linear lattice with N nodes. We use Lattice6
with N nodes, i.e. each node has 3 nearest neighbours to its left and to its right where possible. This
gives approximately
V
V
N − links, or ≈ 3N when V=6. Since we found in section 4 that |SE| ≈ 0.7 M
2
2
and M = 4 N, V=6 gives an empirically reasonable number of short-range links to start with. Long-range
links are then added to the lattice. The endpoints of a long-range link is selected uniformly at random
from the set of nodes such that the absolute distance between them are within [10, 0.7 N] (section 9.2).
We made two independent runs, and set N to 385. The statistics we monitored are link sequence
distance distribution, node degree distribution, average path length, clustering, and assortativity. Figure
18 (specifically plots lin_11 and lin_12) present the results of our experiment with the model. The
results are compared with statistics from the 1agd PCN, and runs made with the model on a ring
topology (plots ring_11 and ring_12). For the experiments on a ring topology, the initial lattice is still
linear, but long-range links are added as if the lattice is a ring. Results from both topologies showed no
significant differences in that both failed to approximate the statistics for 1agd. Sorting the links
generated by the model (plots slin_11 and slin_12) did not help to narrow this gap, although some
difference in average path length behavior is observed (the initial descend is slower).
11.1 The problem of identifying LE_nodes: some pitfalls
What our model lacks, amongst other elements, is a reliable way to identify the endpoints of longrange links, or the set of LE_nodes. The last observation in section 5 (Figure 6c) implies that for the
most part, long-range links (LE) more so than short-range links (SE) connect nodes with many contacts
in a PCN. At this point, given the assortative nature of PCNs (section 7) and the frequent description of
proteins with their domains or modules, and protein folding as a strict process from secondary to tertiary
structures and beyond (see Gō 1983 for a messier view), it may be tempting to picture long-range links
connecting hub nodes, i.e. nodes of high degree. However, this picture may not be entirely accurate for
two possibly related reasons.
First, for PCNs of proteins in their native state, hubs do not necessarily occupy a central position
(section 10). The correlation between node degree and node betweeness rapidly declines as protein size
increases (Figure 19, Appendix B). This implies that many shortest paths in a PCN do not go through
28
S. Khor
12/5/2010
hub nodes, and since long-range links provide important short-cuts links in PCNs, it is contradictory to
Reversed Cumulative Probability
think of long- range links connecting hub nodes.
1.000
0.100
0.010
ring_11
ring_12
lin_11
lin_12
0.001
1agd
0.000
1
10
100
1000
Link Sequence Distance
0.30
25
ring_11
ring_12
lin_11
lin_12
1agd
0.20
0.15
20
15
0.10
10
0.05
5
0.00
0
2
4
6
8
10
12
14
0
16
0
Degree
50
100
150
200
250
300
350
400
450
500
Time
0.7
0.5
0.6
ring_11
ring_12
lin_11
lin_12
slin_11
slin_12
0.4
0.5
Asso rtativity
C lu sterin g C o efficien t
ring_11
ring_12
lin_11
lin_12
1agd
slin_11
slin_12
HP L
P robability
0.25
0.4
ring_11
ring_12
lin_11
lin_12
slin_11
slin_12
0.3
0.2
0.1
0.3
0.2
0.1
0.0
0.0
0
50
100
150
200
250
300
350 400
450
500
Time
0
50
100
150
200
250
300
350
400
450
500
Time
Figure 18 Results from a network approach to protein folding compared with the network
statistics of the PCN for 1agd which is of the same size. The PCN for 1agd has a clustering
coefficient of 0.5397, an assortativity value of 0.3037, and about 72% of its nodes are endpoints of
long-range links. In the model, for both ring and linear topology, about 91% of nodes are in
LE_nodes.
29
12/5/2010
1.0
1.0
0.8
0.8
H ierarc h y
D eg ree-Cen trality Co rrelatio n
S. Khor
0.6
0.4
0.2
0.6
0.4
0.2
0.0
0.0
0
200
400
600
800
1000 1200 1400 1600 1800
0
200
400
600
800
1000 1200 1400 1600 1800
Nodes
Nodes
Figure 19 (Left) Pearson’s correlation between node degree and centrality for GH64 PCNs.
(Right) Gao’s hierarchy index for GH64 PCNs.
Further, and as a possible consequence also, PCNs actually show declining hierarchical structure
with increases in protein size (Figure 19). Following Trusina et al (2004), hierarchical organization is
gauged with the formulae in (Gao 2001) which defines a hierarchical path as one with that satisfies one
of the following conditions: monotonically increasing then monotonically decreasing node degree,
monotonically increasing node degree, or monotonically decreasing node degree. A network where a
high fraction of its shortest paths are also hierarchical paths will have a hierarchical index close to 1.0,
signaling pronounced hierarchical organization. However, declining average clustering levels with
increases in node degree has also been used to detect hierarchical organization (Ravasz et al 2002;
Ravasz and Barabasi 2003). Using this method, Aftabuddin and Kundu (2006) report evidence of
hierarchical structure for their weighted PCN, which is similar to the Greene and Higman (2003) PCN.
We also observe hierarchical tendencies in the four randomly chosen GH64 PCNs (Figure B3).
Evidently, there is some reconciliation to be performed between the two measures of hierarchy.
Second, there is evidence (depending on the scale used) of a positive correlation between node
degree and hydrophobicity (Alves and Martinez 2007; Aftabuddin and Kundu 2007) which implies that
hub nodes in PCNs tend to be hydrophobic. Hydrophobic residues shun water molecules and generally
lie in the core of a protein. As a result, these residues come into close proximity with many other
residues, and thus gain their hub status. The relationship(s) between hydrophobicity, high node degree,
assortativity and average path length needs further study, keeping in mind that the relationship(s) may
differ for transition versus native state proteins.
12. Summary and Outlook
Our main objective with this review paper is to integrate the many network characteristics of PCNs
which have been reported in the literature, and to understand their respective roles and also the interplay
between them. An attempt to apply the statistics to a network approach of protein folding was also
30
S. Khor
12/5/2010
made. In addition, we considered the dynamic aspects of PCNs, i.e. how their network statistics change
as a protein folds. Our model is woefully simple, but we hope it provides good hints for building better
models and integrating the dynamic and static aspects of PCNs. One can easily imagine several ways to
modify the model by adding more constraints, e.g. restrict the set of LE_nodes to 0.6 N (section 5), use
some kind of preferential selection mechanism to fit to the link sequence distance distribution (section
9) and node degree distribution (section 5). The model could also be extended to include link deletion
and/or rewiring (Dokholyan et al 2002 observed changes in node degrees between post- and pretransition states). It may also be interesting to extend the approach in section 11 to investigate how the
dynamics of PCNs couples with the conformational landscape (Scala et al 2001; Doye 2002; Kovacs et
al 2005; Auer et al 2007). Studies show that it is not only important for a protein to arrive at its final
destination, a.k.a. its native state, but how it gets there, its journey or folding pathway(s), is also
important (Chen et al 2008). In light of the exposition in this paper, a network approach may be the
vehicle to address key aspects of the protein folding problem in an integrated way: the existence of a
unique native structure, folding pathway(s) in conformation space to the native structure, protein folding
rates, and information in an amino acid sequence to propel the attainment of all this.
In Table 2, we summarize the network structural statistics mentioned in this paper for PCNs in the
GH64 and EVA132 data sets. We find a general agreement, both in qualitative and quantitative terms,
between the observed network structural characteristics of these two data sets, and with those reported
in literature. On this topic, there are of course other existing network characteristics and measurements,
and other ways of constructing PCNs, that can and need to be explored. However it could also be useful
to go further, and devise measures geared towards capturing the specific linkage patterns of PCNs.
Table 2 Summary of quantitative values: mean ± standard deviation reported.
Section Statistic
GH64
EVA132
|LE|/M
0.2929
0.2611
4
Fraction of links in LE (links with sequence distance > 9) ± 0.0834 ± 0.0986
K
7.9696
7.9147
5
Average node degree
± 0.3126 ± 0.4300
|SE_nodes|/N
0.9970
1.0000
5
Fraction of nodes which are endpoints to links in SE
± 0.0039 ± 0.0004
|LE_nodes|/N
0.6659
0.6348
5
Fraction of nodes which are endpoints to links in LE
± 0.1093 ± 0.1367
0.5484
0.5538
6
Clustering
± 0.0233 ± 0.0292
0.3387
0.3457
7
Assortativity
± 0.0536 ± 0.0743
0.7874
0.7068
9.2
(Maximum Link Sequence Distance) / N
± 0.1402 ± 0.1860
31
S. Khor
12/5/2010
Throughout this paper, we have maintained a separation of short-range (SE) and long-range (LE)
links. We did this to better understand their respective contributions to structural characteristics of
PCNs. But ultimately, the prize is to predict the long-range links of PCNs to within a level of accuracy.
Notwithstanding the physical, chemical and geometrical constraints, this is a challenging combinatorial
task. Even if we could identify LE_nodes accurately, the probability that a link chosen at random from
the set of all possible links with endpoints in LE_nodes is about 6/N (this includes links with sequence
distances less than 10 or some LE_TH). Hence a pure network model of protein folding, like the one we
attempted in section 11, is unlikely to be a winner. A more likely model could be one that incorporates
multiple approaches.
Given the importance of short inter-node distances between residues (section 8), the sparseness of
long-range links in PCNs could appear paradoxical. However, our model does show that in general, it
only takes a few long-range links to reduce the average path length significantly, and to levels that is not
improved upon by further addition of long-range links (Figure 18). According to Ngo and Marks (1992),
long-range links (non-local interactions) actually increases the computational complexity of the protein
folding problem (see also Ngo et al 1994). Thus, if a few well-placed long-range links is sufficient to
reduce the average path length to a good enough level, this “least effort” (albeit, it could have taken
evolution quite some time to figure the optimal wirings) strategy may well be adopted by proteins as a
way out of the paradox.
References
Aftabuddin, Md., and Kundu, S. (2006) Weighted and Unweighted Network of Amino Acids within Protein.
Physica A 396(2), pp. 895-904.
Aftabuddin, Md., and Kundu, S. (2007) Hydrophobic, Hydrophilic and Charged Amino Acids’ Network within
Protein. Biophysical Journal 93(1), pp. 225-231.
Alm, E., and Baker, D. (1999) Matching Theory and Experiment in Protein Folding. Curr. Opin. Struc. Biol. 9, pp.
189-196.
Alves, N.A., and Martinez, A. S. (2007) Inferring topological features of proteins from amino acid residue
networks. Physica A 375, pp. 336-344.
Anfinsen, C.B. (1973) Principles that Govern the Folding of Protein Chains. Science 181, pp. 223-230.
Atilgan, A.R., Akan, P., and Baysal, C. (2004) Small-world Communication of Residues and Significance for
Protein Dynamics. Biophysics Journal 86, pp. 85-91.
Auer, S., Miller, M.A., Krivov, S.V., Dobson, C.M., Karplus, M. and Vendruscolo, M. (2007) Importance of
Metastable States in the Free Energy Landscapes of Polypeptide Chains. Physical Review Letters 99, 178104.
Bagler, G., and Sinha, S. (2005) Network Properties of Protein Structures. Physica A: Statistics Mechanics and its
Apps. 346(1-2), pp. 27-33.
Bagler, G., and Sinha, S. (2007) Assortative Mixing in Protein Contact Networks and Protein Folding Kinetics.
Structural Bioinformatics 23(14), pp. 1760-1767.
Bartoli, L., Fariselli, P., and Casadio, R. (2007) The Effect of Backbone on the Small-world Properties of Protein
Contact Networks. Physical Biology 4, pp. L1-L5.
Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T.N., Weissig, H., Shindyalov, I.N., and Bourne, P.E.
(2000) The Protein Data Bank. Nucleic Acids Research 28, pp. 235-242. http://www.rcsb.org/pdb
Brede, M., and Sinha S. (2005) Assortative Mixing by Degree Makes a Network More Unstable. arXiv:condmat/0507710v1
32
S. Khor
12/5/2010
Chen, Y., Ding, F., Nie, H., Serohijos, A.W., Sharma, S., Wilcox, K.C., Yin, S., and Dokholyan, N.V. (2008)
Protein Folding: Then and Now. Archives of Biochemistry and Biophysics 469, pp. 4 -19.
Del Sol, A., Fujihashi, H., Amoros, D., and Nussinov, R. (2006) Residues Crucial for Maintaining Short Paths in
Network Communication Mediate Signaling in Proteins. Molecular Systems Biology 2.
Dokholyan, N.V., Li, L., Ding, F., and Shakhnovich, E.I. (2002) Topological Determinants of Protein Folding.
PNAS 99(13), pp. 8637-8641.
Doye, J.P.K. (2002) The Network Topology of a Potential Energy Landscape: A Static Scale-Free Network.
Physical Review Letters 88, 238701.
Gaci, O., and Balev, S. (2009) Node Degree Distribution in Amino Acid Interaction Networks. In: Proc.
Computational Structural Bioinformatics Workshop, Washington D.C. USA.
Gao, L (2001) On inferring autonomous system relationships in the Internet. IEEE/ACM Transactions on
Networking 9: 733-745.
Gō, N. (1983) Theoretical Studies of Protein Folding. Ann. Rev. Biophys. Bioeng. 12, pp. 183-210.
Greene, L.H., and Higman, V.A. (2003) Uncovering Network Systems within Protein Structures. Journal of
Molecular Biology 334, pp. 781-791.
Kabsch, W., and Sander, C. (1983) Dictionary of Protein Secondary Structure: Pattern Recognition of Hydrogenbonded and Geometrical Features. Biopolymers 22(12), pp. 2577-2637. http://swift/cmbi.ru.nl/gv/dssp
Karplus, M. (1997) The Levinthal Paradox: Yesterday and Today. Folding & Design 2, pp. S69-S75.
Kleinberg, J. (2000) Navigation in a Small World. Nature 406, 845
Kovacs, I.A., Szalay, M.S. and Csermely, P. (2005) Water and Molecular Chaperones Act as Weak Links of
Protein Folding Networks: Energy Landscape and Punctuated Equilibrium Changes Point towards a Game
Theory of Proteins. FEBS Letters 579(11), pp. 2254-2260.
Levinthal, C. (1969) How to Fold Graciously. In: Debrunner, J.T.P., and Munck, E. (eds.) Mossbauer
Spectroscopy in Biological Systems: Proceedings of a Meeting held at Allerton House. Monticello, Illinois.
University of Illinois Press, pp. 22-24.
Maslov, S., and Sneppen, K. (2002) Specificity and Stability in Topology of Protein Networks. Science 296 pp.
910-913.
Murzin, A.G., Brenner, S.E., Hubbard, T., and Chothia, C. (1995) SCOP: A Structural Classification of Proteins
Database for the Investigation of Sequences and Structures. J. Mol. Biol. 247, pp. 536-540. http://scop.mrclmb.cam.ac.uk/scop
Newman, M.E.J. (2002) Assortative Mixing in Networks. Physical Review Letters 89, 208701.
Newman, M.E.J. (2003) The Structure and Function of Complex Networks. SIAM Review 45, pp. 167-256.
Ngo, J.T. and Marks, J. (1992) Computational Complexity of a Problem in Molecular-Structure Prediction.
Protein Engineering 5(4), pp.313-321.
Ngo, J.T., Marks, J., and Karplus, M. (1994) Computational Complexity, Protein Structure Prediction and the
Levinthal Paradox. In: Merz K. Jr. and Le Grand, S. (eds.) The Protein Folding Problem and Tertiary
Structure Prediction, pp. 435 – 508, Birkhauser, Boston MA.
Pastor-Satorras, R., Vazquez, A., and Vespignani, A. (2001) Dynamical and Correlation Properties of the Internet.
Physical Review Letters 87, 258701-1-4.
Ravasz, E., Somera, A.L., Mongru, D.A., Oltvai, Z.N. and Barabasi, A.-L. (2002) Hierarchical Organization of
Modularity in Metabolic Networks. Science 297, pp. 1551-1555.
Ravasz, E., and Barabasi, A.-L. (2003) Hierarchical Organization In Complex Networks. Physical Review E 67,
026112.
Rost, B. (1999) Twilight Zone of Protein Sequence Alignments. Protein Engineering 12(2), pp. 85-94.
http://cubic.bioc.columbia.edu/ava/res/unique_list.html (Sep 23 2010).
Scala, A., Amaral, L.A.N., and Barthélémy, M. (2001) Small-world Networks and the Conformation Space of a
Lattice Polymer Chain. Europhysics Letters 55(4), 594
Tanaka, S. and Scheraga, H.A. (1977) Model of Protein Folding: Incorporation of a one-dimensional short-range
(Ising) model into a three-dimensional model. Proc. Natl. Acad. Sci. USA 74(4), pp. 1320-1323.
Taverna, D.M. and Goldstein, R.A. (2002) Why are Proteins Marginally Stable. Proteins 46, pp. 105-109.
Tramontano, A. (2006) Protein Structure Prediction:Concepts and Applications. Wiley-VCH Verlag GmbH &Co.
Trusina, A., Maslov, S., Minnhagen, P., and K. Sneppen. (2004) Hierarchy measures in complex networks.
Physical Review Letters 92, 178702.
Vazquez, A., and Moreno, Y. (2003) Resilience to damage of graphs with degree correlation. Physical Review E
33
S. Khor
12/5/2010
67, 015101-1-4(R).
Vendruscolo, M., Dokholyan, N.V., Paci, E., and Karplus, M. (2002) Small-world View of the Amino Acids that
Play a Key Role in Protein Folding. Physical Review E 65, 061910-1-4.
Watts, D.J., and Strogatz, S.H. (1998) Collective Dynamics of ‘small-world’ Networks. Nature 393, pp. 440-442.
Watts, D.J. (1999) Small Worlds: The Dynamics of Networks between Order and Randomness. Princeton
University Press, Princeton, NJ.
Zwanzig, R., Szabo A., and Bagchi, B. (1992) Levinthal’s Paradox. Proc. Natl. Acad. Sci. USA 89, pp. 20-22.
Appendix A. The EVA132 Protein Dataset
The EVA132 protein dataset was extracted from the list of 3477 unique chains archived by EVA
(Rost, 1999). 200 proteins were selected at random from this list, with no overlap with GH64. PCNs for
the proteins in EVA132 were constructed and selected in the manner as described in sections 2 and 3,
producing 132 valid PCNs (listed in Table A1) for us to work with.
Table A1. PID of proteins in the EVA132 dataset
7000
6000
5000
4000
3000
2000
1000
0
1dfe
1ctf
1apc
1AX8
1iko
1grj
1awp
1ylx
1mfq
1o9g
1sml
1e8o
2aql
1btk
1hh2
1a59
1h3e
1bkj
1fgg
1rhz
1adu
1e3h
2fmt
1a5k
1aoi
1upt
1ee4
2bpt
1eo8
1e6e
1knz
1m3
1h6d
Number of Residues (Nodes)
1dfe, 1bi6, 1a7i, 2ezh, 1ctf, 1cfw, 1tif, 1gh5, 1apc, 1cto, 1by2, 1d0d, 1AX8, 1gak, 1jgs, 1j74, 1iko, 2a1j,
1ECY, 1af3, 1grj, 1f8a, 1lf7, 1faj, 1awp, 1qhn, 1r5p, 1CZ4, 1ylx, 1j77, 1bwp, 1lkv, 1mfq, 1jfx, 1hek,
1gy5, 1o9g, 1e7k, 1F9Z, 1awc, 1sml, 1boo, 1tfr, 1ltq, 1e8o, 1pvl, 1bob, 1sbp, 2aql, 1xiz,
1pnf, 1fkm, 1btk, 1c28, 1nrj, 1e6c, 1hh2, 1dpt, 1g2q, 1a1r, 1a59, 1f45, 1i5e, 1esg, 1h3e, 1wdu, 1i85,
1gk4, 1bkj, 1dj8, 1oft, 1jzt, 1fgg, 1rh5, 1sw6, 1bfd, 1rhz, 1b44, 1g7r, 1aui, 1adu, 1h7z, 1zrl, 1dhp, 1e3h,
1ojh, 1bp7, 1kke, 2fmt, 1K3I, 1bwd, 1d8d, 1a5k, 1bih, 1bev, 1fp3, 1aoi, 1tmf, 1l1o, 1KQ4,
1upt, 1h4l, 1ngw, 1b4u, 1ee4, 1a2z, 1eqz, 1jct, 2bpt, 1e8t, 1zhq, 1ewt, 1eo8, 1eum, 1g2c, 1fp7, 1e6e,
1l8w, 1ndb, 1g8m, 1knz, 1gg1, 1b37, 1bcp, 1m32, 1fz6, 1jb0, 1hqm, 1h6d, 1S5L, 1mjg, 1M34
EVA132 Proteins
Figure A1 Size of EVA132 PCNs in terms of the number of Cα atoms documented their PDB files.
34
S. Khor
12/5/2010
30000
PCN Links (< 7A)
25000
Count
20000
15000
10000
5000
0
0
1000
2000
3000
4000
5000
6000
7000
Nodes
Link Density
Figure A2 Link count M, by protein size for EVA132 proteins.
0.20
0.18
0.16
0.14
0.12
0.10
0.08
0.06
0.04
0.02
0.00
PCN Links (< 7A)
0
1000
2000
3000
4000
5000
6000
Nodes
Figure A3 Link density by protein size for EVA132 proteins.
Link Count
20000
18000
SE
16000
LE
14000
12000
10000
8000
6000
4000
2000
0
0
800
1600
2400
3200
4000
4800
5600
6400
Nodes
Figure A4a Number of short-range (SE) and long-range (LE) links, by protein size for EVA132.
The fraction of links in LE is 0.2611 ± 0.0986. The fraction of links in SE is 0.7389 ± 0.0986.
35
S. Khor
12/5/2010
6400
SE
5600
LE
Node Count
4800
4000
3200
2400
1600
800
0
0
800
1600
2400
3200
4000
4800
5600
6400
Nodes
Figure A4b Number of nodes which are endpoints for links in SE (SE_nodes) and nodes which are
endpoint for links in LE (LE_nodes), by protein size for EVA132. The fraction of nodes in
LE_nodes is 0.6348 ± 0.1367.The fraction of nodes in SE_nodes is 1.0000 ± 0.0004.
SE_nodes
SE_ and LE_nodes
Proportion of all nodes
1.2
LE_nodes
Not in LE_nodes
1.0
0.8
0.6
0.4
0.2
0.0
0
800
1600
2400
3200
4000
4800
5600
6400
Nodes
Figure A4c Sizes of different subsets of nodes as a proportion of all nodes for EVA132 PCNs.
On average, 0.6347 ± 0.1367 of all nodes are in both SE_nodes and LE_nodes, while 0.3652 ±
0.1367 of all nodes are not in LE_nodes.
36
S. Khor
12/5/2010
18
16
Node Degree
14
12
max
10
mid
8
mean
6
std.dev.
4
min
2
0
0
800
1600
2400
3200
4000
4800
5600
6400
Nodes
Figure A5a Node degree summary statistics for EVA132 PCNs.
Mean node degree average at 7.9147 with a std. dev. of 0.4300.
0.30
Probability, P(k)
0.25
1sw6
0.20
1fp3
0.15
1m32
0.10
1S5L
0.05
0.00
0
2
4
6
8
10
12
14
16
Degree, k
Figure A5b Node degree distributions for four PCNs in the EVA132 dataset. The respective PCNs
of 1sw6, 1fp3, 1m32 and 1S5L has 508, 804, 2103 and 5130 nodes respectively.
Clustering Coefficient
1.0
0.8
0.6
0.4
0.2
1dfe
1ctf
1apc
1AX8
1iko
1grj
1awp
1ylx
1mfq
1o9g
1sml
1e8o
2aql
1btk
1hh2
1a59
1h3e
1bkj
1fgg
1rhz
1adu
1e3h
2fmt
1a5k
1aoi
1upt
1ee4
2bpt
1eo8
1e6e
1knz
1m3
1h6d
0.0
EVA132 Proteins
Figure A6 Clustering coefficients ± standard deviation. Mean is 0.5538 with a std. dev. of 0.0292.
37
S. Khor
12/5/2010
0.8
Assortativity
0.6
0.4
0.2
1dfe
1ctf
1apc
1AX8
1iko
1grj
1awp
1ylx
1mfq
1o9g
1sml
1e8o
2aql
1btk
1hh2
1a59
1h3e
1bkj
1fgg
1rhz
1adu
1e3h
2fmt
1a5k
1aoi
1upt
1ee4
2bpt
1eo8
1e6e
1knz
1m3
1h6d
0.0
EVA132 Proteins
Figure A7 Assortativity values. Mean is 0.3457 with a std. dev. of 0.0743.
Link Sequence Distance
300
Avg.
250
Std. dev.
Median
200
150
100
50
0
0
800
1600
2400
3200
4000
4800
5600
6400
4000
4800
5600
6400
Nodes
Link Sequence Distance
4800
Max
4000
3200
2400
1600
800
0
0
800
1600
2400
3200
Nodes
Figure A8 Link sequence distance summary statistics for EVA132 PCNs indicate right skewed
distributions for link sequence distance.
38
S. Khor
12/5/2010
Appendix B
Additional material for Figure 19 (section 11).
Figure B1 Degree-node
centrality scatter plots
for four GH64 PCNs.
Correlation
coefficients
are given in the table
below. Both methods
show weakening positive
correlation with increase
in N. We did not report
Spearman’s rho values as
their significance could
not be computed due to
ties.
Correlation method 1aep
1agd
1psd
1CVJ
Pearson
0.6691 0.3112 0.3010 0.1222
Kendall’s tau
0.5689 0.4661 0.4114 0.3616
All correlation coefficients are significant at the 95% confidence level.
0.10
Average Node Centrality
0.08
0.06
1aep
0.04
1agd
1psd
0.02
1CVJ
0.00
0
2
4
6
8
10
12
14
16
-0.02
-0.04
Degree
Figure B2 Node centrality for nodes with degree k are averaged and reported with their standard
deviations. The relationship is marginally positive and grows weaker with increase in protein size.
The large standard deviations point to wide dispersion in node centrality values, as seen in the
scatter plots in Figure B1.
39
S. Khor
12/5/2010
Average Clustering Coefficient
1.2
1.0
0.8
1aep
1agd
0.6
1psd
1CVJ
0.4
0.2
0.0
0
2
4
6
8
10
Degree
12
14
16
Figure B3 Some evidence of hierarchical organization in PCNs.
40
| 5cs.CE
|
THE EXISTENCE OF NON-INNER AUTOMORPHISMS OF
ORDER p IN FINITE THIN p-GROUPS
arXiv:1604.07267v1 [math.GR] 25 Apr 2016
MARCO RUSCITTI AND LEIRE LEGARRETA
Abstract. In this paper we study the existence of at least one noninner automorphism of order p of a finite thin p-group, whenever the
prime p is.
Marco Ruscitti DISIM
Università degli studi dell’Aquila
67100 L’Aquila, Italy
E-mail address: [email protected]
Leire Legarreta Matematika Saila
Euskal Herriko Unibertsitatea UPV/EHU
48080 Bilbao, Spain
E-mail address: [email protected]
1. Introduction
The main goal of this paper is to contribute to the longstanding conjecture
of Berkovich posed in 1973, that conjectures that every finite p-group admits
a non-inner automorphism of order p, where p denotes a prime number
[15, Problem 4.13]. The conjecture has attracted the attention of many
mathematicians during the last couple of decades, and has been confirmed
for many classes of finite p-groups. It is remarkable to put on record that,
in 1965, Liebeck [12] proved the existence of a non-inner automorphism of
order p in all finite p-groups of class 2, where p is an odd prime. However,
the fact that there always exists a non-inner automorphism of order 2 in all
finite 2-groups of class 2 was proved by Abdollahi [1] in 2007. The conjecture
was confirmed for finite regular p-groups by Schmid [14] in 1980. Indeed,
Deaconescu [8] proved it for all finite p-groups G which are not strongly
Frattinian. Moreover, Abdollahi [2] proved it for finite p-groups G such
that G/Z(G) is a powerful p-group, and Jamali and Visesh [11] did the
same for finite p-groups with cyclic commutator subgroup. In the realm of
finite groups, quite recently, the result has been confirmed for semi-abelian
p-groups by Benmoussa and Guerboussa [5], and for p-groups of nilpotency
2010 Mathematics Subject Classification. 20D15, 20D45.
Key words and phrases. Finite p-groups, non-inner automorphisms, derivation, thin
p-groups.
The first author would like to thank the Department of Mathematics at the University
of the Basque Country for its excellent hospitality while part of this paper was being
written. The second author is supported by the Spanish Government, grants MTM201128229-C02-02 and MTM2014-53810-C2-2-P, and by the Basque Government, grant IT75313 and IT974-16.
1
2
M.RUSCITTI AND L.LEGARRETA
class 3, by Abdollahi, Ghoraishi and Wilkens [3]. Finally, Abdollahi et al
[4] proved the conjecture for p-groups of coclass 2.
With the contribution of this paper we add another class of finite p-groups
to the above list, by proving that the above mentioned conjecture holds true
for all finite thin p-groups, whenever the prime p is.
The organization of the paper is as follows. In Section 2 we exhibit some
preliminary facts and tools that will be used in the proof of the main result of
the paper, and in Section 3 we recall elementary matters about thin p-groups
and we prove the main result as well.
Throughout the paper, most of the notation is standard and it can be found,
for instance, in [13].
2. Preliminaries
In this section, we recall some facts about derivations in the multiplicative
setting, and some related lemmas, which will be useful to prove the main
Theorem 3.5 of the paper. The reader could be referred to [10] for more
details and explicit proofs about derivations.
Definition 2.1. Let G be a group and let M be a right G-module. A derivation δ : G → M is a function such that
δ(gh) = δ(g)h δ(h), for all g, h ∈ G.
In terms of its properties, it is well-known that a derivation is uniquely
determined by its values over a set of generators of G. Let F be a free group
generated by a finite subset X and let G = hX : r1 , . . . , rn i be a group whose
free presentation is F/R, where R is the normal closure of the set of relations
{r1 , . . . , rn } of G. Then a standard argument shows that M is a G-module
if and only if M is an F -module on which R acts trivially. Indeed, if we
denote by π the canonical homomorphism π : F → G, then the action of F
on M is given by mf = mπ(f ), for all m ∈ M and all f ∈ F . Continuing
with the same notation, we have the following results.
Lemma 2.2. Let M be an F -module. Then every function f : X → M
extends in a unique way to a derivation δ : F → M .
Lemma 2.3. Let M be a G-module and let δ : G → M be a derivation.
Then δ̄ : F → M given by the composition δ(f ) = δ(π(f )) is a derivation
such that δ(ri ) = 0 for all i ∈ {1, . . . , n}. Conversely, if δ : F → M is
a derivation such that δ(ri ) = 0 for all i ∈ {1, . . . , n}, then δ(f R) = δ(f )
defines, uniquely, a derivation on G = F/R to M such that δ = δ ◦ π.
In the following lemma, we study the relationship between derivations
and automorphisms of a finite p-group.
Lemma 2.4. Let G be a finite p-group and let M be a normal abelian
subgroup of G viewed as a G-module. Then for any derivation δ : G → M ,
we can define uniquely an endomorphism φ of G such that φ(g) = gδ(g) for
all g ∈ G. Furthermore, if δ(M ) = 1, then φ is an automorphism of G.
In order to reduce some calculations in terms of commutators, we keep in
mind the following result.
NON-INNER AUTOMORPHISMS OF ORDER p IN FINITE THIN p-GROUPS
3
Lemma 2.5. Let F be a free group, p be a prime number and A be an
F -module. If δ : F → A is a derivation then,
(i) δ(F p ) = δ(F )p [δ(F ),p−1 F ],
(ii) if [A,i F ] = 1, we have δ(γi (F )) ≤ [δ(F ),i−1 F ] for all i ∈ N.
p−1
p−2
Proof. Let x ∈ F . We have δ(xp ) = δ(x)x +x +···+1 . Since (x − 1)p−1 ≡
xp−1 + xp−2 + · · · + 1 mod p, the first assertion follows. Now let us prove
the second assertion by induction on i. Clearly, the assertion holds when
i = 1. Let us suppose, by inductive hypothesis that if [A,k F ] = 1 then
δ(γk (F )) ≤ [δ(F ),k−1 F ] for some k ∈ N. Let us take any a ∈ F and any
b ∈ γk (F ), and let us suppose that [A,k+1 F ] = 1. Then
δ([a, b]) = [δ(a), b][a, δ(b)][a, b, δ(a)][a, b, δ(b)] ∈ [δ(F ),k F ].
3. Berkovich Conjecture for finite thin p-groups
To develop the section, we start enumerating some structural properties
concerned about thin p-groups. Firstly, let us recall that in a group G an
antichain is a set of mutually incomparable elements in the lattice of its
normal subgroups. It is well-known that, if G is a p-group of maximal class,
then the lattice of its normal subgroups consists of p + 1 maximal subgroups
and of the terms of the lower central series of G. Thus, a p-group of maximal
class has only one antichain, which consists of its maximal subgroups. The
necessity to extend the family of groups of maximal class to a bigger family
of p-groups with a bound on the antichains, leads us to introduce the formal
definition of thin p-group. Let us introduce the definition of thin p-groups
as in [7]. Forwards, we mention as well some already proved results about
the existence of non-inner automorphisms of order p in certain specific cases
to avoid repetitions.
Definition 3.1. Let G be a finite p-group. Then G is thin if every antichain
in G contains at most p + 1 elements.
The following results about finite thin p-groups are discussed in [6].
Lemma 3.2. Let G be a finite thin p-group, and let p be an odd prime.
(i) If N is a normal subgroup of G, then N is a term of the lower
central series of G if and only if N is the unique normal subgroup
of its order.
(ii) If G is not of maximal class, then |G| ≥ p5 and G/γ3 (G) is of
exponent p.
Lemma 3.3. Let G be a finite thin p-group and let us take any h ∈ γi (G) −
γi+1 (G). Then the following called coverty property holds,
[h, G]γi+2 (G) = γi+1 (G).
Remark 3.4. The properties shown in Lemma 3.2 give us a lot of information about thin p-groups. Since an elementary abelian p-group is thin if and
only if its order is p2 (see [7]), then every finite thin p-group is a two generator group. Secondly, by [6] we know that the lower and the upper central
series of this kind of groups coincide, and consequently that all quotients of
4
M.RUSCITTI AND L.LEGARRETA
these series are elementary abelian p-groups of order at most p2 . Consequently, Z(G) must be cyclic of order p. In particular, the quotients of the
lower and upper central of finite thin p-groups have exponent p. Moreover,
if G is a finite thin p-group, then Φ(G) = γ2 (G), Z2 (G) ≤ Z(Φ(G)) and the
property stated in the above Lemma 3.3 holds equivalently for terms of the
upper central series of G.
Since Liebeck in [12] proved the existence of at least a non-inner automorphism of order p in all finite p-groups of class 2, where p is an odd
prime, Abdollahi in [1] proved the existence of such an non-inner automorphism of order 2 in all finite 2-groups of class 2, and Abdollahi, Ghoraishi
and Wilkens in [3] did the same in the case of finite p-groups of nilpotency
class 3, in our study, we can deal with finite p-groups of nilpotency class
c ≥ 4. On the other hand, since Deaconescu in [8] proved the existence of
at least a non-inner automorphism of order p for all finite p-groups G which
are not strongly Frattinian, we may assume that the finite thin p-groups
G we are interested in, are strongly Frattinian, in other words, that the
groups of our interest satisfy CG (Φ(G)) = Z(Φ(G)). As a result due to
Abdollahi in [2], we have that if G is a fi nite p-group such that G has no
non-inner automorphisms of order p leaving Φ(G) elementwise fixed, then
d(Z2 (G)/Z(G)) = d(G)d(Z(G)). Thus, in view of this previous result, in order to prove the existence of a non-inner automorphism of order p in a finite
thin p-group G, we may assume without any loss of generality that the condition d(Z2 (G)/Z(G)) = d(G)d(Z(G)) holds. Thus, using the consequences
got in Remark 3.4 we have Z2 (G)/Z(G) is isomorphic to an elementary
abelian group of order p2 .
It is also showed in [9] (see Theorem B) that every finite thin 2-group is
a group of maximal class. Thus in the following we focus our attention on
finite thin p-groups, where p is an odd prime.
Now we are ready to prove the next theorem.
Theorem 3.5. Let G be a finite thin p-group, where p is an odd prime.
Then G has a non-inner automorphism of order p.
Proof. Let us denote c the nilpotency class of G. As it has been said before,
we can assume c ≥ 4. From Remark 3.4 and its consequences and assumptions we know that G is a two generator p-group, the lower and the upper
central series of G coincide, Φ(G) = γ2 (G), Z2 (G) ≤ Z(Φ(G)), Z(G) ∼
= Cp ,
d(Z2 (G)/Z(G))) = d(G)d(Z(G)) and Z2 (G)/Z(G) ∼
= Cp × Cp . In particular,
[Z2 (G), γ2 (G)] = 1 and Ω1 (Z2 (G)) is an elementary abelian subgroup of G.
Indeed, G/γ3 (G) has order p3 , class 2 and by Lemma 3.2 we may assume
that its exponent is p, i.e. G/γ3 (G) is an extraspecial p-group of exponent
p and order p3 .
Our goal is to obtain at least an automorphism of G of order p. Firstly,
we define an assignment on generators of the free group generated by two
elements sending them to Ω1 (Z2 (G)). By Lemma 2.2 it is possible to extend
these assignments to a derivation. Secondly, we show that this map preserves
the relations defining the quotient G/γ3 (G), and then we apply Lemma 2.3
to induce a derivation from G/γ3 (G) to Ω1 (Z2 (G)). Finally, we lift this
NON-INNER AUTOMORPHISMS OF ORDER p IN FINITE THIN p-GROUPS
5
found map to a derivation from G to Ω1 (Z2 (G)), applying Lemma 2.4. In
the following paragraphs we describe in detail each of the mentioned steps.
Let x → u, y → v be an assignment on generators x, y of the two generator free group F2 , with u, v ∈ Ω1 (Z2 (G)). By Lemma 2.3 this assignment
extends uniquely to a derivation δ : F2 → Ω1 (Z2 (G)) such that δ(x) = u
and δ(y) = v. Making perhaps some abuse of notation, we can assume
that G/γ3 (G) corresponds to the presentation hx, y | xp , y p , [y, x, x], [y, x, y]i.
Now let us see that the equalities δ(xp ) = 1, δ(y p ) = 1, δ([y, x, x]) = 1 and
δ([y, x, y]) = 1 hold. To do it, firstly, let us show that δ is trivial on the p-th
powers of elements of F2 . In fact, considering π ′ the canonical epimorphism
from F2 to G/γ3 (G) we have
δ(f p ) = δ(f )f
′
p−1
δ(f p−1 ) = · · · = δ(f )f
p−1 +···+1
= δ(f )π (f )
p−1 +···+1
′
= δ(f )π (f
p
= δ(f )p [δ(f ), π ′ (f )](2) = 1,
p−1 +···+1)
=
f or all f ∈ F2 .
Secondly, let us analyze the behaviour of δ on commutators. Since for any
g, h ∈ G it holds that gh = hg[g, h], then applying δ to this previous equality,
we get δ(gh) = δ(hg[g, h]) = δ(hg)[g,h] δ([g, h]), and consequently taking also
into account that [Z2 (G), γ2 (G)] = 1, we get δ([g, h]) = δ(gh)(δ(hg)[g,h] )−1 =
δ(gh)(δ(hg))−1 = [δ(g), h][g, δ(h)], which is an element of Z(G).
Moreover,
δ([y, x, y]) = [δ([y, x]), y][[y, x], δ(y)] = 1,
δ([y, x, x]) = [δ([y, x]), x][[y, x], δ(x)] = 1.
(Take into account that the previous two equalities can be obtained as well,
applying properly item (ii) of Lemma 2.5.)
By Lemma 2.3 we can induce a derivation δ from G/γ3 (G) to Ω1 (Z2 (G)).
The map δ′ : G → Ω1 (Z2 (G)) defined by the law δ′ (g) = δ(gγ3 (G)) for all
g ∈ G, is a derivation from G to Ω1 (Z2 (G)). By Lemma 2.4, δ′ induces
an automorphism φ of G by the law φ(g) = gδ′ (g) for all g ∈ G, leaving, in particular, Φ(G) elementwise fixed. Clearly, Z2 (G) ≤ Z(γ3 (G)). In
fact, the previous inclusion holds since Z2 (G) ≤ Z(Φ(G)), Φ(G) = γ2 (G)
and Z2 (G) = γc−1 (G) ≤ γ3 (G). This allows us to prove that the automorphism φ has order p. In fact, φp (g) = φp−1 (gδ′ (g)) = φp−1 (g)φp−1 (δ′ (g)) =
φp−1 (g)δ′ (g) = φp−2 (gδ′ (g))δ′ (g) = · · · = g(δ′ (g))p = g, for all g ∈ G.
Thus, doing this previous construction we produce a set of automorphisms
of G of order p, whose size is equal to |Ω1 (Z2 (G))|2 , in other words, whose
size is the number of possible choices for the images of the above generators.
Now, we distinguish the only two possible cases: Z2 (G) ∼
= Cp × Cp × Cp or
Z2 (G) ∼
= Cp2 × Cp . In the former case, when Z2 (G) ∼
= Cp × Cp × Cp , we
produce p6 automorphisms of G of order p. However, the number of inner
automorphisms of G induced by elements of Z3 (G) is at most p4 . Thus, a
counting argument is enough to say that G has a non-inner automorphism
of G of order p, and we get the statement of the Theorem in this case. Otherwise, in the latter case, when Z2 (G) ∼
= Cp2 × Cp , choose an assignment
x → u, y → v such that u, v ∈ Ω1 (Z2 (G)) − {1}, u is not central, and let φ
be the automorphism of G of order p obtained by this assignment. On the
other hand, we know that φ is inner if and only if there exists an element hφ
of Z3 (G) − Z2 (G) such that hpφ ∈ Z(G), φ(g) = ghφ and δ′ (g) = [g, hφ ], for
6
M.RUSCITTI AND L.LEGARRETA
all g ∈ G. Under these circumstances, since Z(G) ≤ Ω1 (Z2 (G)) we would
deduce that [hφ , G]Z(G) ≤ Ω1 (Z2 (G)) < Z2 (G), which is in contradiction
with the analogous coverty property of Lemma 3.3. Consequently, this second case does not happen and the statement of the Theorem is proved.
4. Acknowledgements
The first author would like to thank the Department of Mathematics at
the University of the Basque Country for its excellent hospitality while part
of this paper was being written; he also wish to thank Professors Gustavo
A. Fernández Alcober, Norberto Gavioli and Carlo Maria Scoppola for their
suggestions.
References
[1] A. Abdollahi, Finite p-groups of class 2 have non-inner automorphisms of order p, J.
Algebra 312 (2007), 876–879.
[2] A. Abdollahi, Powerful p-groups have non-inner automorphism of order p and some
cohomology, J. Algebra 323 (2010), 779–789.
[3] A. Abdollahi, S. M. Ghoraishi and B. Wilkens, Finite p-groups of class 3 have non-inner
automorphisms of order p, Beitr. Algebra Geom. 54 (2013), 363–381.
[4] A. Abdollahi, S. M. Ghoraishi, Y. Guerboussa, M. Reguiat and B. Wilkens, Non-inner
automorphism of order p for finite p-groups of coclass 2, J. Group Theory 17 (2014),
267–272.
[5] M. T. Benmoussa and Y. Guerboussa, Some properties of semi-abelian p-groups, Bull.
Austral. Math. Soc. 91 (2015), 86–91.
[6] R. Brandl, A. Caranti and C.M. Scoppola, Metabelian thin p-groups, Q. J. Math 43
(1992), 157–173.
[7] A. Caranti, S. Mattarei, F. Newman and C.M. Scoppola, Thin groups of prime powerorder and thin Lie algebras, Q. J. Math 47 (1996), 279–296.
[8] M. Deaconescu, Non-inner automorphisms of order p of finite p-groups, Journal of
Algebra 250 (2002), 283–287.
[9] R. brandl, The DIlworth number of subgroup lattices, Arch. Math. 50 (1988), 502–510.
[10] N. Gavioli, V. Pannone, Frobenius partitions in extraspecial p-groups Geometriae
Dedicata 74 (1999), 313–323.
[11] A. R. Jamali and R.Viseh, On the existence of non-inner automorphism of order two
in finite 2-groups, Bull. Aust. Math. Soc. 87 (2013), 278–287.
[12] H. Liebeck, Outer automorphisms in nilpotent p-groups of class 2, J London Math.
Soc. 40 (1965), 268–275.
[13] D. J. S. Robinson, A course in the theory of groups, second edition, Springer-Verlag,
New York, 1996.
[14] P. Schmid, A cohomological property for regular p-groups, Math. Z. 175 (1980),
1–3.
[15] Unsolved Problems in Group Theory, The Kourovka Notebook, No. 17, edited by V. D.
Mazurov and E. I. Khukhro, Russian Academy of Sciences, Siberian Division, Institute
of Mathematics, Novosibirisk, 2010.
DISIM, Università degli studi dell’Aquila, 67100 L’Aquila, Italy, E-mail address: [email protected]
Matematika Saila, Euskal Herriko Unibertsitatea UPV/EHU, 48080 Bilbao,
Spain, E-mail address: [email protected]
| 4math.GR
|
Reassembling Trees for the Traveling Salesman
Jens Vygen
arXiv:1502.03715v2 [cs.DM] 5 Jan 2016
University of Bonn
revised version – December 23, 2015
Abstract
Many recent approximation algorithms for different variants of the traveling salesman problem (asymmetric TSP, graph TSP, s-t-path TSP) exploit the well-known fact that a solution of
the natural linear programming relaxation can be written as convex combination of spanning
trees. The main argument then is that randomly sampling a tree from such a distribution and
then completing the tree to a tour at minimum cost yields a better approximation guarantee
than simply taking a minimum cost spanning tree (as in Christofides’ algorithm).
We argue that an additional step can help: reassembling the spanning trees before sampling. Exchanging two edges in a pair of spanning trees can improve their properties under
certain conditions.
We demonstrate the usefulness for the metric s-t-path TSP by devising a deterministic
polynomial-time algorithm that improves on Sebő’s previously best approximation ratio of 85 .
keywords: traveling salesman problem, s-t-path TSP, approximation algorithm, spanning
tree
1
Introduction
The traveling salesman problem (TSP) is probably the best-known NP-hard combinatorial optimization problem. Although for the general metric TSP Christofides’ [1976] algorithm with its
approximation ratio 23 is still unbeaten, we have seen progress for several variants and special
cases in particular since 2010. See Vygen [2012] for a detailed survey.
Many of the recent approximation algorithms begin by solving the natural linear programming
relaxation, which was first proposed by Dantzig, Fulkerson and Johnson [1954]. It was observed by
Held and Karp [1970] that a solution x∗ to this LP can (after scaling down by a factor n−1
n except
for the s-t-path case) be written as convex combination (or, equivalently, probability distribution)
of spanning trees. Of course, this distribution is far from unique.
Asadpour et al. [2010] and Oveis Gharan, Saberi and Singh [2011] improved the approximation ratio for the asymmetric TSP and graph TSP, respectively, by randomly sampling a spanning
∗
tree from a maximum entropy distribution describing n−1
n x and then completing it to a tour in an
optimal way. An, Kleinberg and Shmoys [2015] considered the metric s-t-path TSP and showed
that a spanning tree randomly chosen from any distribution describing x∗ is —in expectation—
good enough to improve on Christofides’ algorithm for this problem.
1
In this paper, we propose to modify the distribution before sampling. By exchanging two edges
in a pair of trees with certain properties we obtain two new trees and hence a new distribution.
We call this step reassembling trees. Under certain conditions the two new trees have better
properties than the old ones. For the s-t-path TSP we show that this step can indeed improve
the approximation ratio.
1.1
The s-t-path TSP
Let us define the (metric) s-t-path TSP formally. As in the classical version of the TSP, we want
to visit a set of cities at minimum total cost. However, rather than returning to the origin at the
end, we are given the origin and the destination as input. More precisely, we are given a finite
set V , two elements s and t of V , and a symmetric distance function c : V × V → R≥0 satisfying
the triangle inequality. Throughout this paper we will denote by n the numberPof elements of V .
n−1
We ask for a sequence V = {v1 , . . . , vn } with v1 = s and vn = t, such that i=1
c(vi , vi+1 ) is
minimized.
The classical metric TSP is the special case when s = t. We note that in both variants we can
either require that every city is visited exactly once, or, equivalently, we can drop the requirement
that c satisfies the triangle inequality but allow visiting cities more than once.
The s-t-path TSP (for s 6= t) is clearly no easier than the classical metric TSP (we can reduce
the latter to the former by guessing two cities that are adjacent in an optimum tour); in particular
123
unless P = NP
there is no polynomial-time approximation algorithm with better ratio than 122
(Karpinski, Lampis and Schmied [2013]).
We assume s 6= t henceforth. It is sufficient to compute a connected multi-graph with vertex
set V in which exacty s and t have odd degree; such a graph will be called an {s, t}-tour. Here
is why: given an {s, t}-tour, we can find an Eulerian walk from s to t (using every edge exactly
once) in linear time and shortcut whenever a vertex is visited not for the first time; this yields an
s-t-path with vertex set V , which —due to the triangle inequality— is no more expensive than
the {s, t}-tour.
1.2
Previous approximation algorithms
Christofides’ [1976] algorithm, originally designed for the classical metric TSP, works also for
the s-t-path TSP. It first computes a minimum cost spanning tree (V, S) in the complete graph
spanned by V , and then adds a minimum cost TS -join J, where TS is the set of vertices whose
degree has the wrong parity (even for s or t, odd for other vertices). The result is an {s, t}-tour.
Hoogeveen [1991] showed that Christofides’ algorithm has approximation ratio 35 for the s-tpath TSP, and that in fact this ratio is asymptotically attained by an infinite set of examples.
An, Kleinberg and Shmoys [2015] proposed the best-of-many
Christofides algorithm for the s√
1+ 5
t-path TSP and proved that it has approximation ratio 2 ≈ 1.619. The algorithm is quite
simple: it computes an optimum solution to the natural LP relaxation (see (1) below) and writes
2
it as convex combination of spanning trees. For each of these spanning trees, (V, S), it computes
a minimum weight TS -join. J, where TS is the set of vertices whose degree has the wrong parity,
obtains an {s, t}-tour S ∪ J, and finally outputs the best of these. We will go into details
in Subsection 1.4. This was the first improvement over Christofides’ algorithm that applied to
general metrics. Then Sebő [2013] improved the analysis, obtaining the approximation ratio 58 .
We will describe how in Subsection 1.8.
For the special case where c is the metric closure of an unweighted graph, better approximation
algorithms have been obtained by Mömke and Svensson [2011], Mucha [2014], and An, Kleinberg and Shmoys
[2015] (in this order). The best known approximation ratio 1.5, obtained first by Sebő and Vygen
[2014], matches the integrality ratio of the LP in this special case. Gao [2013] gave a simpler proof
of this result.
Some of the above-mentioned papers apply also to a generalization (to T -tours, for general
T ); see Section 6 for a brief discussion.
1.3
Notation
For a given instance (V, s, t, c) let n := |V |, and let E := V2 be the edge set of the complete
graph on V . For U ⊆ V , δ(U ) denotes the set of edges with exactly one endpoint in U , and E[U ]
the set of edges with both endpoints in U . We write δ(v) := δ({v}) for v ∈ V . For T ⊆ V with
|T | even, a T -join is a set J ⊆ E for which |δ(v) ∩ J| is odd for all v ∈ T and |δ(v) ∩ J| is even for
all v ∈ V \ T . A T -cut is a cut δ(U ) for which |U ∩ T | is odd. The intersection of a T -join and a
T -cut always contains an odd number of edges. Edmonds [1965] proved that a minimum weight
T -join can be computed in polynomial time.
P
P
For a vector x ∈ RE and F ⊆ E we write x(F ) := e∈F xe and c(x) := e={v,w}∈E c(v, w)xe ;
moreover, χF ∈ {0, 1}E denotes the incidence
vector of F (i.e., χFe = 1 for e ∈ F and χFe = 0
P
for e ∈ E \ F ), and c(F ) := c(χF ) = e={v,w}∈F c(v, w). By S we denote the set of edge sets of
spanning trees in (V, E). For S ∈ S, the set TS := {v ∈ V \ {s, t} : |δ(v) ∩ S| odd} ∪ {v ∈ {s, t} :
|δ(v) ∩ S| even} contains the vertices whose degree in S has the wrong parity.
1.4
Best-of-many Christofides
Our algorithm will be an extension of the best-of-many Christofides algorithm that was proposed
by An, Kleinberg and Shmoys [2015]. Therefore, we first describe this algorithm in more detail.
The algorithm begins by solving the LP relaxation
min c(x)
subject to
x(δ(U ))
x(δ(U ))
x(δ(v))
x(δ(v))
xe
≥
≥
=
=
≥
(∅ =
6 U ⊂ V, |U ∩ {s, t}| even)
(∅ =
6 U ⊂ V, |U ∩ {s, t}| odd)
(v ∈ V \ {s, t})
(v ∈ {s, t})
(e ∈ E)
2
1
2
1
0
3
(1)
Obviously, the integral solutions to (1) are precisely the incidence vectors of the edge sets of
the Hamiltonian s-t-paths in (V, E). So this LP is indeed a relaxation.
Following an idea of Heldand Karp [1970], An, Kleinberg and Shmoys [2015] observed that
the polytope defined by (1) is x ∈ RE
6 U ⊂ V, x(E[U ]) ≤
≥0 : x(E) = n−1, x(E[U ]) ≤ |U |−1 ∀ ∅ =
|U |−2 ∀ {s, t} ⊆ U ⊂ V and hence is contained in the spanning tree polytope of (V, E) (Edmonds
[1970]).
an optimum solution x∗ (in fact every feasible solution) can
P Therefore,
P be written as
∗
S
x = S∈S pS χ , where p is a distribution on S, i.e., pS ≥ 0 for all S ∈ S and S∈S pS = 1.
By Carathéodory’s theorem we can assume that pS > 0 for less than n2 spanning trees (V, S).
An optimum LP solution x∗ , such spanning trees, and such numbers pS can be computed in
polynomial time, as can be shown with the ellipsoid method (Grötschel, Lovász and Schrijver
[1981]) or using the splitting-off technique (cf. Genova and Williamson [2015]).
We will fix x∗ henceforth. We will also fix the distribution p for the rest of the introduction
but will modify it later.
For each S ∈ S with pS > 0, the best-of-many Christofides
algorithm then computes a
.
minimum weight TS -join J, and considers the {s, t}-tour S ∪ J; the output is the best of these.
Note that trying all S ∈ S with pS > 0 leads to a deterministic polynomial-time algorithm which
is at least as good as randomly picking S ∈ S with probability pS .
1.5
Basic Analysis
We follow the basic analysis of An, Kleinberg and Shmoys [2015]. The cost of the {s, t}-tour
that the best-of-many Christofides algorithm computes for a given instance depends on its choice
of
This is at most
P : J is a TS -join}).
P p only: it is BOMC(p) := minS∈S: pS >0 (c(S) + min{c(J)
∗
S∈S pS c(S) = c(x ).
S∈S pS (c(S) + min{c(J) : J is a TS -join}). Note that
A well-known result of Edmonds and Johnson [1973] says that the minimum weight of a TS -join
is the minimum c(y) over all y in the TS -join polyhedron
y ∈ RE
(2)
≥0 : y(C) ≥ 1 ∀ TS -cuts C .
Therefore
BOMC(p) ≤ c(x∗ ) +
X
pS c(y S )
(3)
S∈S
for any set of vectors (y S )S∈S such that y S is in the TS -join polyhedron (2). The difficulty in the
analysis lies in finding an appropriate set of vectors (y S )S∈S . Let us call them correction vectors
(although An, Kleinberg and Shmoys [2015] used this term with a different meaning), because
they bound the cost of parity correction.
1.6
Narrow cuts
Let C := {δ(U ) : ∅ =
6 U ⊂ V, x∗ (δ(U )) < 2}. The elements of C are called narrow cuts. By
∗
(1) they are {s, t}-cuts. If there are no narrow cuts except δ(s) and δ(t), then x2 is a valid
4
correction vector for all S ∈ S, leading to BOMC(p) ≤ 32 c(x∗ ), which is Wolsey’s [1980] analysis
of Christofides’ algorithm for the classical metric TSP. But in the s-t-path TSP we will in general
have narrow cuts.
The following property is useful:
Lemma 1 (An, Kleinberg and Shmoys [2015]) The narrow cuts form a chain: there are sets
{s} = L0 ⊂ L1 ⊂ · · · ⊂ Lℓ−1 ⊂ Lℓ = V \ {t} so that C = {δ(Li ) : i = 0, . . . , ℓ}.
Proof: We have x∗ (δ(s)) = x∗ (δ(t)) = 1 and hence δ(s) ∈ C and δ(V \ {t}) = δ(t) ∈ C. Suppose
we have U ′ ⊂ V and U ′′ ⊂ V with δ(U ′ ), δ(U ′′ ) ∈ C and s ∈ U ′ ∩ U ′′ ∈
/ {U ′ , U ′′ }. Then t ∈
/ U ′ ∪ U ′′
∗
′
∗
′′
∗
′
′′
∗
′′
′
and 2 + 2 > x (δ(U )) + x (δ(U )) ≥ x (δ(U \ U )) + x (δ(U \ U )) ≥ 2 + 2, a contradiction. ✷
We remark:
Proposition 2 Given x∗ , the set C of narrow cuts can be computed in polynomial time.
Proof: Lemma 1 implies that there is an order V = {v1 , v2 , . . . , vn } such that each narrow cut
has the form δ({v1 , . . . , vi }) and is the only {vi , vi+1 }-cut C with x∗ (C) < 2. Thus it suffices to
compute
a {v, w}-cut C with minimum x∗ (C) for all v, w ∈ V with v 6= w. This can be done by
n
2 applications of any polynomial-time max-flow algorithm (or more efficiently by computing a
Gomory-Hu tree).
✷
Similarly to Lemma 1 we have:
Lemma 3 For all C, C ′ ∈ C with C 6= C ′ we have
x∗ (C ∩ C ′ ) ≤
1
1 ∗
x (C) + x∗ (C ′ ) − 1.
2
2
Proof: Let C = δ(Li ) and C ′ = δ(Lj ) with 0 ≤ i < j ≤ ℓ. Then x∗ (C) + x∗ (C ′ ) − 2x∗ (C ∩ C ′ ) =
x∗ (δ(Lj \ Li )) ≥ 2.
✷
Moreover, An, Kleinberg and Shmoys [2015] and Sebő [2013] observed that
X
pC
pS ≤ x∗ (C) − 1
even :=
(4)
S∈S : |S∩C| even
and
pC
one :=
X
S∈S : |S∩C|=1
pS ≥ 2 − x∗ (C)
for every narrow cut C ∈ C. Note that (4) and (5) follow directly from x∗ (C) =
for all C ⊆ E and |S ∩ C| ≥ 1 for all C ∈ C and S ∈ S. From this we also get
X |S ∩ C| − 1
1 ∗
x (C) − 1 − pC
:=
pS
pC
=
even
many
2
2
S∈S
5
(5)
P
S∈S
pS |S ∩ C|
(6)
s
t
Figure 1: A spanning tree (V, S). Filled circles show vertices in TS . The edge set S is partitioned
into its s-t-path IS (turquoise) and its TS -join JS (red). The narrow cuts are those shown in grey
(dotted and solid). Those that have an even number of edges of S are shown in solid grey; each
of them contains (at least) one red and (at least) one blue edge.
for all C ∈ C.
For S ∈ S let IS denote the edge set of the s-t-path in (V, S). Let JS := S \ IS and note that
JS is a TS -join. See Figure 1 for an example. Sebő [2013] observed that
X
X
X
(7)
pS χS∩C ≤
p S χI S
C∈C S∈S : |S∩C|=1
S∈S
because IS (in fact every s-t-path) intersects every narrow cut (in fact every {s, t}-cut), and if
S ∩ δ(Li ) = S ∩ δ(Lj ) = {e} for some 0 ≤ i < j ≤ ℓ and e ∈ E, then S ∩ δ(Lj \ Li ) = ∅,
contradicting that (V, S) is connected.
1.7
Correction vectors
Consider
y S := βx∗ + (1 − 2β)χJS + z S
(8)
z S (C) ≥ β(2 − x∗ (C))
(9)
for S ∈ S, where 0 ≤ β ≤ 12 , and z S ∈ RE
≥0 is a nonnegative vector satisfying
for all C ∈ C with |S ∩ C| even.
Lemma 4 (An, Kleinberg and Shmoys [2015]) For every S ∈ S and every TS -cut C we have
y S (C) ≥ 1.
Proof: Let S ∈ S and C be a TS -cut. JS is a TS -join, so |JS ∩ C| ≥ 1.
If C ∈
/ C, then x∗ (C) ≥ 2 and hence y S (C) ≥ βx∗ (C) + (1 − 2β)|JS ∩ C| ≥ 2β + 1 − 2β = 1.
If C ∈ C, then |IS ∩ C| (the intersection of an s-t-path and an {s, t}-cut) is odd and |JS ∩ C|
(the intersection of a TS -join and a TS -cut) is odd, so |S ∩ C| is even. Hence y S (C) = βx∗ (C) +
(1 − 2β)|JS ∩ C| + z S (C) ≥ βx∗ (C) + 1 − 2β + z S (C) ≥ 1 due to (9).
✷
6
So y S is in the TS -join polyhedron (2) for all S ∈ S. For any distribution p and nonnegative
vectors (z S )S∈S with (9), we get with (3) and (8) the bound
X
X
BOMC(p) ≤ (1 + β)c(x∗ ) + (1 − 2β)
pS c(JS ) +
pS c(z S ).
(10)
S∈S
S∈S
Now the question is how to choose the vectors z S .
1.8
The analyses of An, Kleinberg, Shmoys, and Sebő
An, Kleinberg and Shmoys [2015], Sebő [2013], and then also Gao [2015], chose
X
z S = αχIS +
max 0, 2β − α − βx∗ (C) v C ,
(11)
C∈C : |S∩C| even
C
where v C ∈ RE
≥0 are vectors with v (C) ≥ 1 for all C ∈ C.
As IS ∩ C 6= ∅ (recall that IS is the edge set of an s-t-path), this choice implies z S (C) ≥
α + (2β − α − βx∗P
(C)) = 2β − βx∗ (C) for P
all C ∈ C with |S ∩ C| even, as required in (9).
I
S
Writing Ip := S∈S pS χ and Jp := S∈S pS χJS we get, with (10) and (11) and c(Ip ) +
c(Jp ) = c(x∗ ):
X
BOMC(p) ≤ (1 + β)c(x∗ ) + (1 − 2β)c(Jp ) +
pS c(z S )
S∈S
= (1 + β)c(x∗ ) + (1 − 2β)c(Jp ) + αc(Ip )
X
X
max 0, 2β − α − βx∗ (C) c(v C )
pS
+
S∈S
∗
C∈C : |S∩C| even
= (1 + α + β)c(x ) + (1 − α − 2β)c(Jp )
X
∗
C
+
pC
even max {0, 2β − α − βx (C)} c(v )
C∈C
∗
≤ (1 + α + β)c(x ) + (1 − α − 2β)c(Jp )
X
+
(x∗ (C) − 1) max {0, 2β − α − βx∗ (C)} c(v C ),
C∈C
where we used (4) in the last inequality. The three papers choose the vectors v C (C ∈ C) differently.
P
An, Kleinberg and Shmoys [2015] showed that v C (C ∈ C) can be chosen so that C∈C v C ≤
x∗ . Observing (x − 1) (2β − α − βx) ≤
√
(β−α)2
4β
for all x ∈ R and setting β =
√1
5
and α = 1 −
√2 ,
5
they obtained BOMC(p) ≤ 1+2 5 c(x∗ ).
P
Sebő [2013] chose v C := 2−x1∗ (C) S∈S : |S∩C|=1 pS χS∩C for C ∈ C. Then indeed v C (C) ≥ 1 for
√
all C ∈ C due to (5). Using (7) and observing (x−1)(2β−α−βx)
≤ α + β − 2 αβ for all x ≤ 2 one
2−x
7
√
gets BOMC(p) ≤ (1 + α + β)c(x∗ ) + (1 − α − 2β)c(Jp ) + (α + β − 2 αβ)c(Ip ). Setting β = 25 and
1
α = 10
, this yields BOMC(p) ≤ 58 c(x∗ ). (Sebő [2013] set β = 49 and α = 19 , which yields the same
bound if c(Jp ) ≥ 53 c(x∗ ); otherwise he used y S = χJS (S ∈ S) as correction vectors).
Gao [2015] simply chose v C to be the incidence vector of a cheapest edge in C, which is clearly
best possible in this framework, but he could not obtain a better approximation ratio.
1.9
New approach
We use the ideas of An, Kleinberg and Shmoys [2015] and Sebő [2013] but define z S differently;
see Section 2. Like Sebő [2013], we will boundP
c(z S ) by a constant fraction of c(Ip ). More precisely,
S
E
we will find vectors z ∈ R≥0 with (9) and S∈S pS c(z S ) ≤ (1 − 2β)c(Ip ). With (10) this will
immediately yield BOMC(p) ≤ (2 − β)c(x∗ ). The question is how large we can choose β.
In the Appendix we give an example that shows that it will in general not be possible to
choose β > 52 and hence directly improve on Sebő’s [2013] approximation ratio of 58 . Therefore
we will
P modify p first. By reassembling the trees that contribute to the convex combination
x∗ = S∈S pS χS we eliminate the most critical configurations. We will show how in Section 3.
Then we can complete the calculation in Section 4 and obtain an improved approximation ratio.
2
New correction vectors
Since we want to bound the weighted sum of the costs of the vectors z S by a multiple of c(Ip )
(the weighted sum of the costs of the s-t-paths IS , S ∈ S), each pair (S, e) with S ∈ S and e ∈ IS
′
will make a contribution to some of the vectors z S (where S ′ ∈ S can be S or a different tree).
As in Sebő’s [2013] analysis, there are two types of contributions. A pair (S, e) will contribute
′
γS,e (1 − 2β) to z S and a total of (1 − γS,e )(1 − 2β) to vectors z S for other trees S ′ . The latter
contribution is distributed as follows: if C ∈ C is the narrow cut with S ∩ C = {e} (there can be
′
only one such cut), then e will contribute to z S for all S ′ ∈ S with |S ′ ∩ C| even. Sebő’s [2013]
analysis is essentially equivalent to choosing γS,e = 21 for all S ∈ S and e ∈ IS . We will obtain an
improvement by choosing individual values.
2.1
The new vectors
For any S ∈ S and C ∈ C choose an edge eSC ∈ IS ∩ C. Moreover, for any C ∈ C let eC ∈ C be a
minimum cost edge in C. For any pair S ∈ S and e ∈ IS we will choose a number 0 ≤ γS,e ≤ 1
later. Then, for S ∈ S, we set
o
n
X
X
χ{eC } . (12)
z S :=
(1 − 2β)γS,e χ{e} +
max 0, β(2 − x∗ (C)) − (1 − 2β)γS,eS
C
e∈IS
C∈C : |S∩C| even
The first term in (12) is the direct contribution of the edges of IS to z S . The second term is
exactly what is still needed to obtain (9) for all C ∈ C with |S ∩ C| even; see the proof of Lemma
8
5 below. We will have
P to show
P that the total cost of the weighted sum of the second terms in (12)
is not more than S∈S pS e∈IS (1 − 2β)(1 − γS,e )c(e); see Section 2.2.
P
P
Lemma 5 If 0 ≤ β ≤ 12 and S∈S pS c(z S ) ≤ (1 − 2β) S∈S pS c(IS ) for the vectors z S (S ∈ S)
defined in (12), then BOMC(p) ≤ (2 − β)c(x∗ ).
Proof: Note that the vectors z S are nonnegative for all S ∈ S (as β ≤ 21 and γS,e ≥ 0 for all
e ∈ IS ). We have for every S ∈ S and C ∈ C with |S ∩ C| even:
z S (C) ≥ (1 − 2β)γS,eS + β(2 − x∗ (C)) − (1 − 2β)γS,eS = β(2 − x∗ (C)),
C
C
as required
by (9). So P
the bound immmediately follows from (10), using
P
2β) S∈S pS c(IS ) and S∈S pS (c(JS ) + c(IS )) = c(x∗ ).
P
S∈S
pS c(z S ) ≤ (1 −
✷
We will try to maximize β.
2.2
Bounding the cost
The cost of the vectors P
z S (S ∈ S) will of course depend
on the choice of the γS,e (S ∈ S, e ∈ IS ).
P
The desired inequality S∈S pS c(z S ) ≤ (1 − 2β) S∈S pS c(IS ) is implied by
o
n
X
X
∗
pS (1 − 2β)(1 − γS,eS ),
pS max 0, β(2 − x (C)) − (1 − 2β)γS,eS ≤
C
C
S∈S : |S∩C|=1
S∈S : |S∩C| even
as may be seen from (12) and will be formally shown in the proof of Lemma 7 below. To write
this inequality in a more compact form (see (13) below), we divide by 1 − 2β and use the following
notation:
Definition 6 Given
γS,e ≥o0 for S ∈ S and e ∈ IS , we define the benefit of (S, C) ∈ S×C
n numbers
β(2−x∗ (C))
, γS,eS if |S ∩ C| is even, bS,C := 1 − γS,eS if |S ∩ C| = 1, and
to be bS,C := min
1−2β
C
C
bS,C = 0 otherwise.
Now all we need to show is that we have enough benefit at every narrow cut:
Lemma 7 Let 0 ≤ β <
1
2
and 0 ≤ γS,e ≤ 1 for S ∈ S and e ∈ IS . If
X
S∈S
pS bS,C ≥
β(2 − x∗ (C))pC
even
1 − 2β
for all C ∈ C, then BOMC(p) ≤ (2 − β)c(x∗ ).
9
(13)
Proof: By (13) and Definition 6 we have for every C ∈ C:
X
β(2 − x∗ (C))pC
β(2 − x∗ (C))
even
≤
pS min
, γS,eS +
C
1 − 2β
1 − 2β
S∈S : |S∩C| even
X
pS (1 − γS,eS ).
C
(14)
S∈S : |S∩C|=1
We compute:
P
c(z S )
pS
1 − 2β
S∈S
=
X
S∈S
=
X
pS
X
γS,e c(e) +
e∈IS
pS
X
X
∗
β(2 − x (C))
− γS,eS
max 0,
c(eC )
C
1 − 2β
C∈C : |S∩C| even
γS,e c(e)
e∈IS
S∈S
∗ (C))
∗ (C))pC
X
β(2
−
x
β(2
−
x
even
−
pS min
, γS,eS c(eC )
+
C
1 − 2β
1 − 2β
C∈C
S∈S : |S∩C| even
X
X
X
X
≤
pS
γS,e c(e) +
pS (1 − γS,eS )c(eC )
X
C
e∈IS
S∈S
≤
≤
=
C∈C S∈S : |S∩C|=1
X
γS,e c(e) +
X
γS,e c(e) +
X
pS
X
pS
X
pS c(IS )
e∈IS
S∈S
S∈S
X
X
C∈C S∈S : |S∩C|=1
e∈IS
X
S∈S
pS
X
e∈IS
pS (1 − γS,eS )c(eSC )
C
(1 − γS,e )c(e)
S∈S
(we used (14) in the first, γS,eS ≤ 1 and c(eC ) ≤ c(eSC ) in the second, and (7) in the third
C
inequality). Now the assertion follows from Lemma 5.
✷
Let us quickly check the obvious (although we will not need it): if we simply choose γS,e = 21
P
β
for all S ∈ S and e ∈ IS , then we have for all C ∈ C: if 1−2β
(2 − x∗ (C)) ≤ 21 , then S∈S pS bS,C ≥
β
1−2β (2
− x∗ (C))pC
even , and otherwise
X
S∈S
1
1
1
C
∗
C
≥ ((2 − x∗ (C)) + pC
pS bS,C = pC
even ) ≥ 2(2 − x (C))peven
one + peven
2
2
2
(the first inequality follows from (5), and the second one follows from 0 ≤ (2 − x∗ (C)) + pC
even ≤ 1
(cf. (4)) and the fact that a + b ≥ 4ab for all a, b ≥ 0 with a + b ≤ 1).
So we have (13) for β = 52 and obtain again Sebő’s [2013]Papproximation ratio 58 . This is tight
1
1
C
for x∗ (C) = 23 and pC
one = peven = 2 , when the total benefit
S∈S pS bS,C is 2 .
10
2.3
More benefit
Let us first give an informal description of our idea. The approximation guarantee can be readily
1
1
C
improved if there are no cuts C with x∗ (C) ≈ 23 (and pC
one ≈ 2 and peven ≈ 2 ; see the previous
paragraph). We would like to obtain more benefit (i.e., more than 12 ) for “critical” cuts C (i.e.,
those with x∗ (C) ≈ 32 ), and to achieve this we will reduce the benefit for less critical cuts (but
still have enough). If C is a critical cut, S ∈ S, and eSC does not belong to any other critical cut,
we can choose γS,e > 12 if |S ∩ C| is even and γS,e < 21 if |S ∩ C| = 1. We will describe how exactly
at the beginning of Section 4. In this case we get bS,C > 21 .
However, this does not work if, for a critical cut C and every S ∈ S, the edge eSC also belongs
to another critical cut C ′ (and |S ∩ C| + |S ∩ C ′ | = 3). In this case we can only choose γS,e = 21
and get bS,C = 12 . But note that then S has at least one edge in C ∩ C ′ . However, due to Lemma
3, there cannot be too many edges in the intersection of two critical cuts: certainly less than one
per tree on average.
We will fix a constant ξ between 32 and 2, and consider C ξ := {C ∈ C : x∗ (C) < ξ}. These
cuts were called (ξ − 1)-narrow by An, Kleinberg and Shmoys [2015], but we prefer to call them
ξ-narrow. Cuts that are not ξ-narrow will not be critical. For any C ∈ C ξ \ {δ(s), δ(t)}, let C←
and C→ be the adjacent ξ-narrow cuts in both directions. More precisely, if C = δ(Lj ), then
C← := δ(Li ) and C→ := δ(Lk ), where i = max{ι : ι < j, x∗ (δ(Lι )) < ξ} and k = min{κ : κ >
j, x∗ (δ(Lκ )) < ξ} (cf. Lemma 1).
Let C be a critical cut. By Lemma 3, x∗ (C ∩ C← ) < 1 and x∗ (C ∩ C→ ) < 1, so. there must
be trees S ∈ S with pS > 0 with less than two edges in the disjoint union (C ∩ C← ) ∪ (C ∩ C→ ).
Therefore, if every tree S with pS > 0 has either larger benefit than 21 or (benefit exactly 12
.
P
and at least two edges in (C ∩ C← ) ∪ (C ∩ C→ )), we get S∈S pS bS,C > 21 , which leads to an
improvement. This will be essentially our argument.
As can be seen from Figure 2, there are however —in addition to trees with more than two edges
in C— four configurations (orange and purple) that do not have this property. Our reassembling
step, to be described next, aims at avoiding two out of these four configurations (one of the two
orange ones and one of the two purple ones). It will turn out that this is enough.
3
Reassembling Trees
In this
Psection we show how to reassemble the trees that contribute to the convex combination
x∗ = S∈S pS χS in order to remove certain bad configurations.
Given a constant 23 < ξ < 2 (the exact value will be chosen later), we number the ξ-narrow
cuts C ξ = {C0 , . . . , Cℓ′ } from left to right; i.e., for 0 ≤ i < k ≤ ℓ′ and Ci = δ(Lι ) and Ck = δ(Lκ )
we have ι < κ. Note that C0 = δ(s) and Cℓ′ = δ(t). Moreover, for C = Cj (i ∈ {1, . . . , ℓ′ − 1}) we
have C← = Cj−1 and C→ = Cj+1 .
11
C←
C→
C
type
010
011
110
111
020
021
120
022
220
121
121
Figure 2: Configurations at a ξ-narrow cut C; here C← and C→ are the adjacent ξ-narrow cuts.
The types, according to Definition 8, are shown on the right. In each configuration we see the (one
or two) edges of S ∩ C for some S ∈ S; if there are two, the edge eSC that belongs to the s-t-path is
shown on top. If C is critical (i.e., x∗ (C) ≈ 32 ), each of these configurations will get benefit at least
1
1
because eSC belongs
2 . Green configurations will get benefit strictly more than 2 , which is possible
.
to no other critical cut. Blue configurations have two edges in (C ∩ C← ) ∪ (C ∩ C→ ), which is
more than possible on average. Orange
and purple configurations are problematic. Configurations
.
with even more edges in (C ∩ C← ) ∪ (C ∩ C→ ) or with three or more edges in C (these will be
called good) are not shown.
3.1
Types of trees at ξ-narrow cuts
We now define the type of a spanning tree at a ξ-narrow cut. We remark that the type can depend
on the value of ξ. This is no problem since ξ is a fixed constant (chosen later).
Definition 8 Fix a constant 32 < ξ < 2. Then, for any tree S ∈ S and any cut C ∈ C ξ \{δ(s), δ(t)}
we define the type of S at C as follows. Let l = |S ∩C ∩C←| and m = |S ∩C| and r = |S ∩C ∩C→|.
If m ≥ 3 or l + r ≥ 3 or (l + r ≥ 1 and S ∩ C ′ 6= {e} for all e ∈ S ∩ C and all C ′ ∈ C ξ ), then the
type is “ good”, otherwise the type is “ lmr”.
Note that if the type of S at C is good, then |S ∩ C| ≥ 2. See Figure 2 for a list of all types
that are not good, and Figure 3 for the types of the tree in Figure 1 (assuming that all narrow
cuts are in C ξ ). All types are also listed in Table 1 on page 19.
12
s
t
021
111
120
010
good
121
Figure 3: The spanning tree (V, S) from Figure 1. Assuming that all narrow cuts (grey vertical
lines) are ξ-narrow, we list the type of S at each C ∈ C ξ \ {δ(s), δ(t)} according to Definition 8.
For the cut C marked good we have l = 0, m = 2, r = 1, and S ∩ C ′ 6= {e} for all e ∈ S ∩ C and
all C ′ ∈ C ξ .
3.2
Reassembling lemma
Here is our key lemma for reassembling trees.
Lemma 9 Let i ∈ {1, . . . , ℓ′ − 1}, so Ci ∈ C ξ \ {δ(s), δ(t)}. Let S1 , S2 ∈ S such that S1 has type
120 at Ci and S2 has type 011 at Ci .
Then there are two edges e1 ∈ S1 and e2 ∈ S2 such that S1′ := (S1 \ {e1 }) ∪ {e2 } ∈ S and
′
S2 := (S2 \ {e2 }) ∪ {e1 } ∈ S. Moreover we have:
(a) At each of C1 , . . . , Ci−1 , the new trees have the same type as the old trees.
(b) At Ci , S1′ has type 121, and S2′ has type 010.
(c) If S1′ has type 110 or 021 at Cj for some j > i, then S1 has the same type at Cj .
(d) If S2′ has type 110 or 021 at Cj for some j > i, then S2 has the same type at Cj or S1′ has
type good at each of Ci+1 , . . . , Cj .
See Figure 4 for an example.
Proof: Let e2 be the only edge in S2 ∩ Ci . As S2 has type 011 at Ci , there is an index k ∈
{i + 1, . . . , ℓ − 1} such that e2 belongs to the cuts Ci , . . . , Ck but neither Ci−1 nor Ck+1 .
Let S1 ∩ Ci = {e0 , e1 }, where e0 belongs to the s-t-path in S1 (so IS1 ∩ Ci = {e0 }). As S1 has
type 120 (and not good) at Ci , there must be an edge in S1 ∩ Ci that is the only edge of S1 in
some other ξ-narrow cut, and such an edge must belong to the s-t-path in S1 . So there exists an
index h ∈ {0, . . . , i − 1} such that Ch ∩ S1 = {e0 }; moreover e1 ∈
/ Ci−1 ∪ Ci+1 .
The graph (V, S1 ∪ {e2 }) contains a circuit A. Any circuit has even intersection with any cut.
As e2 ∈
/ Ch , the circuit A does not contain e0 . However, A must contain (at least) a second edge
in Ci besides e2 ; so A contains e1 .
13
v2
e2
e0
e1
S1
S1′
S2
S2′
v0
e0
e2
Ch
Ci
Ck
e1
Ch
Ci
Ck
Figure 4: Reassembling trees: exchanging two edges in trees of types 120 and 011 (Lemma 9).
We have S1′ ∈ S because e1 belongs to the circuit A in (V, S1 ∪ {e2 }). We have S2′ ∈ S because
S2 ∩ Ci = {e2 } and e1 ∈ Ci .
Property (a) is obvious as nothing changes for C0 , . . . , Ci−1 : note that e1 , e2 ∈ Ci \ Ci−1 .
Property (b) follows from S1′ ∩ Ci = {e0 , e2 } and S2′ ∩ Ci = {e1 }.
To show (c) and (d), we first prove:
Claim: S1′ has type good at Ci+1 , . . . , Ck .
To this end, observe that |S1′ ∩ Cj | ≥ |(S1 \ {e1 }) ∩ Cj | + 1 ≥ 2 for j = i, . . . , k.
Let v2 be the “right” endpoint of e2 , and v0 the “right” endpoint of e0 (i.e., if Ci = δ(Lι ),
then v0 , v2 ∈
/ Lι ; cf. Figure 4.) Let P be the v0 -v2 -path in S1′ . We observe that (∗) P crosses
every cut among Ci+1 , . . . , Ck an odd number of times and every cut Cj with j ≤ i or j > k an
even number of times. Therefore P contains neither e0 (it contains an even number of edges from
S1′ ∩ Ch = {e0 }) nor e2 (it contains an even number of edges from S1′ ∩ Ci = {e0 , e2 }).
Now let j ∈ {i + 1, . . . , k} and e ∈ IS1′ ∩ Cj . We have e 6= e0 because e0 ∈
/ Cj . Moreover,
e 6= e2 because e2 ∈
/ IS1′ (as this s-t-path contains an odd number of edges from S1′ ∩ Ch = {e0 }
/ Ci .
and from S1′ ∩ Ci = {e0 , e2 }). So e ∈
Suppose S1′ ∩ Cq = {e} for some q (otherwise S1′ has type good at Cj because e2 guarantees
l + r ≥ 1); then q > k (as e2 ∈ Ci ∪ · · · ∪ Ck ), and from (∗) we get e ∈
/ E(P ). Then we have
′
e, e2 ∈ S1 ∩ Cj and |E(P ) ∩ Cj | is odd and e, e2 ∈
/ E(P ). This implies |S1′ ∩ Cj | ≥ 3, so again S1′
has type good at Cj . The claim is proved.
The claim directly implies (c) and (d) for j ∈ {i + 1, . . . , k}.
Now let j ∈ {k + 1, . . . , ℓ′ − 1}. Suppose S1 and S1′ have different types at Cj . Since l, m, r (cf.
Definition 8) are identical at Cj , the only remaining reason for different types is that S1 ∩Cj ′ = {e}
(and thus S1′ ∩ Cj ′ = {e, e2 }) for some i < j ′ ≤ k and e ∈ S1 ∩ Cj = S1′ ∩ Cj . But then the new
type is good.
Now suppose S2 and S2′ have different types at Cj for some j ∈ {k + 1, . . . , ℓ′ − 1}. Again
l, m, r are identical at Cj , and the only remaining reason for different types is that S2′ ∩ Cj ′ = {e}
/ IS2 ,
(and thus S2 ∩ Cj ′ = {e, e2 }) for some i < j ′ ≤ k and e ∈ S2 ∩ Cj = S2′ ∩ Cj . But then e ∈
because IS2 contains an odd number of edges in Cj ′ and must contain e2 . This implies that S2 ∩Cj
contains another edge in addition to e, and hence |S2′ ∩ Cj | = |S2 ∩ Cj | ≥ 2. Therefore the new
14
type can only be 120 or 121 or 220.
3.3
✷
Resulting types
P
ξ
Let pC
τ :=
S∈S:S has type τ at C pS for every type τ and C ∈ C \ {δ(s), δ(t)}. We apply the
previous lemma from left to right, and at each cut as long as possible. In order to obtain a
polynomial-time algorithm, we will round all pS down to integer multiples of nǫ2 , for some small
positive constant ǫ.
Corollary 10 For any constants 23 < ξ < 2 and ǫ > 0 there is a polynomial-time algorithm which,
given an instance (V, s, t, c) and an optimum solution x∗ of (1), a set S+ ⊆ S and numbers pS
P
2
with nǫ pS ∈ N for S ∈ S+ and S∈S+ pS ≤ 1, computes another set S̄+ ⊆ S and numbers p̄S
P
P
2
with nǫ p̄S ∈ N for S ∈ S̄+ , such that S∈S+ pS χS = S∈S̄+ p̄S χS and
C
C
min{p̄011
, p̄120
}=0
and
C
C
p̄110
≤ pC
110 + p̄good
and
C
C
p̄021
≤ pC
021 + p̄good
(15)
for all C ∈ C ξ \ {δ(s), δ(t)}.
Proof: We first compute the ξ-narrow cuts (cf. Proposition 2) and process them from left to
right, ignoring δ(s) and δ(t). Initially, S̄+ := S+ and p̄ := p. At each Ci (1 ≤ i ≤ ℓ′ − 1) we
check the types of all trees S ∈ S̄+ . Whenever we have two such trees S1 and S2 of type 120 and
011, respectively, we set δ := min{p̄S1 , p̄S2 }, decrease p̄S1 and p̄S2 by δ, and increase p̄S1′ and p̄S2′
by δ, where S1′ and S2′ are chosen as in Lemma 9. If S2′ but not S2 has type 110 or 021 at Cj
for some j > i, then S1′ has type good at each of Ci+1 , . . . , Cj , and we maintain the properties
P
P
C
C
C
C
Ch
Ch
} = 0 for h < i and p̄110j ≤ p110j + S∈S ′ p̄S and p̄021j ≤ p021j + S∈S ′ p̄S for all j,
, p̄120
min{p̄011
where S ′ := {S ∈ S, S has type good at Cj and type 121 or good at Ci , . . . , Cj−1 }.
We remove a tree S from S̄+ if p̄S drops to zero. Note that at any stage, all p̄S (and δ) are
2
✷
integer multiples of nǫ2 , so there are never more than nǫ trees in S̄+ .
By symmetry, we also have:
Corollary 11 For any constants 32 < ξ < 2 and ǫ > 0 there is a polynomial-time algorithm which,
given an instance (V, s, t, c) and an optimum solution x∗ of (1), a set S+ ⊆ S and numbers pS
P
2
with nǫ pS ∈ N for S ∈ S+ and S∈S+ pS ≤ 1, computes another set S̄+ ⊆ S and numbers p̄S
P
P
2
with nǫ p̄S ∈ N for S ∈ S̄+ , such that S∈S+ pS χS = S∈S̄+ p̄S χS and
C
C
min{p̄110
, p̄021
}=0
and
C
C
p̄011
≤ pC
011 + p̄good
for all C ∈ C ξ \ {δ(s), δ(t)}.
and
C
C
p̄120
≤ pC
120 + p̄good
(16)
✷
We conclude:
15
Theorem 12 For any constants 23 < ξ < 2 and ǫ > 0 there is a polynomial-time algorithm which,
given an instance (V, s, t, c) and an optimum solution P
x∗ of (1), computes a set S+ ⊆ S of trees
∗
such that there exists a distribution p on S with x = S∈S pS χS and pS = 0 for all S ∈ S \ S+
and
C
C
C
C
C
C
C
min pC
≤ pC
(17)
120 + p021 , p011 + p110 , p011 + p021 , p120 + p110
good + ǫ
for all C ∈ C ξ \ {δ(s), δ(t)}.
P
Proof: First compute any distribution p′ with x∗ = S∈S p′S χS and with less than n2 trees S
′ of these trees, using the ellipsoid method (Grötschel, Lovász and Schrijver
with p′S > 0, and the set S+
2
[1981]) or the splitting-off technique (cf. Genova and Williamson [2015]).
Then set p′′S := nǫ2 ⌊ nǫ p′S ⌋
P
′′ be the set of trees with p′′ > 0. Note that
′
′′
for all S ∈ S; and let S+
S
S∈S (pS − pS ) < ǫ.
′′ and p′′ . We get S̄ and p̄ with min{p̄ C , p̄ C } = 0 for all
Then apply Corollary 11 to S+
+
110 021
ξ
C , p̄
C }=0
¯120
C ∈ C \ {δ(s), δ(t)}. Applying Corollary 10 to S̄+ and p̄ yields S̄¯+ and p̄¯ with min{p̄¯011
C
C } ≤ p̄
C , p̄
¯good
¯021
for all C ∈ C ξ \ {δ(s), δ(t)}. Then p := p̄¯ + p′ − p′′ is a distribution
and min{p̄¯110
′ .
✷
as required. We output S+ := S̄¯+ ∪ S+
2
Note that our algorithm only needs the set S+ , which contains less than nǫ + n2 trees. The
above proof shows that we can also compute the distribution p in polynomial time, but it is needed
only for the analysis.
4
Improved approximation ratio
We now show how to set the numbers γS,e for S ∈ S and e ∈ IS so that the total benefit (according
to Definition 6) is large.
.
For some constant β slightly larger than 52 (to be determined later), let f (x) := β(2−x)(x−1)
1−2β
By Lemma 7 and (4), an average benefit of f (x∗ (C)) from the trees will be sufficient for a narrow
cut C.
Ideally, we would like to have γS,e ≥ f (x∗ (C)) if e ∈ S ∩ C and |S ∩ C| even, and 1 − γS,e ≥
f (x∗ (C)) if e ∈ S ∩ C and |S ∩ C| = 1. But this may be impossible because f (x∗ (C)) can be
greater than 21 for β > 25 . Therefore we cut off at 12 . More precisely:
For S ∈ S and e ∈ IS we define two numbers 0 ≤ f1 , f2 ≤ 12 . If |S ∩ C| > 1 for all C ∈ C with
e ∈ C, then f1 := 0, otherwise f1 := min{ 12 , max{f (x∗ (C)) : e ∈ C ∈ C, |S ∩ C| = 1}}. If |S ∩ C|
is odd for all C ∈ C with e ∈ C, then f2 := 0, otherwise f2 := min{ 12 , max{f (x∗ (C)) : e ∈ C ∈
C, |S ∩ C| even}}. If f2 < f1 , then we set γS,e := f2 , otherwise we set γS,e := 1 − f1 .
4.1
Less critical cuts
Lemma 13 For all C ∈ C with f (x∗ (C)) ≤
1
2
we have (13).
16
Proof: By the above choice of γS,eS (S ∈ S), we have γS,eS ≥ f (x∗ (C)) if |S ∩ C| is even
C
C
P
′
S ∈ S : |S ∩
and 1 − γS,eS ≥ f (x∗ (C)) if |S ∩ C| = 1. Let q :=
′
S∈S pS where S :=
C
C| even,
β(2−x∗ (C))
1−2β
X
S∈S
< γS,eS . We have (by Definition 6)
pS bS,C ≥ q ·
∗
As pC
one ≥ 2 − x (C) ≥
X
pS bS,C
S∈S
C
β(2 − x∗ (C))
∗
C
∗
+ pC
even − q · f (x (C)) + pone · f (x (C)).
1 − 2β
2−x∗ (C)
x∗ (C)−1
pC
even ≥
2−x∗ (C)
x∗ (C)−1
(pC
even − q) by (5) and (4), we get
2 − x∗ (C)
β(2 − x∗ (C))
β(2 − x∗ (C))pC
even
∗
+ f (x (C)) 1 + ∗
.
≥ q·
pC
=
even − q
1 − 2β
x (C) − 1
1 − 2β
✷
So we need to analyze only the remaining cuts (with f (x∗ (C)) >
q
q
1
1
≤ x∗ (C) ≤ 32 + 54 − 2β
).
more precisely with 23 − 54 − 2β
4.2
1
2
and hence x∗ (C) ≈ 1.5,
Most critical cuts
Now we consider the cuts C ∈ C with f (x∗ (C)) >
Subsection 2.3:
1
2.
We first establish what we outlined in
Lemma 14 Let 0.4 ≤ β < 0.5 and 1.7 ≤ ξ ≤ 1.8. Let C ∈ C ξ \ {δ(s), δ(t)} with f (x∗ (C)) ≥ 21 ,
and S ∈ S with |S ∩ C| = 1 or |S ∩ C| even. Then:
• bS,C ≥ 12 .
• Moreover, if eSC ∈
/ C← ∪ C→ or S ∩ C ′ 6= {eSC } for all C ′ ∈ C ξ , then bS,C ≥ 1 − f (ξ).
Proof: First let |S ∩ C| = 1. Then bS,C = 1 − γS,eS . For S and e = eSC we have f1 =
C
γS,eS = f2 ≤ 12 . Moreover, if eSC ∈
/ C← ∪ C→ , then f2 ≤ f (ξ).
C
n
o
∗ (C))
Now let |S ∩ C| be even. Then bS,C := min β(2−x
.
,
γ
S
S,e
1−2β
1
2
≥ f2 , so
C
We first observe that x∗ (C) < 1.7 and f (ξ) ≥ f (1.8) ≥ 0.32 and hence
β(2−x∗ (C))
1−2β
=
f (x∗ (C))
x∗ (C)−1
>
f (x∗ (C))
0.7
≥ 75 > 1 − f (ξ).
Now consider the second term. We have f2 = 12 ≥ f1 , so γS,eS = 1 − f1 ≥ 12 . Moreover, if
C
S ∩ C ′ 6= {eSC } for all C ′ ∈ C ξ (this holds in particular if |S ∩ C| =
6 1 and eSC ∈
/ C← ∪ C→ ), then
f1 ≤ f (ξ).
✷
Now we can bound the benefit for most critical cuts, of course using a distribution according
to Theorem 12.
17
Lemma 15 Let 0.4 ≤ β < 0.5P
and 1.7 ≤ ξ ≤ 1.8 such that ν := 1 − f (ξ) > 21 . Let ǫ > 0. Let p be
∗
a distribution on S with x = S∈S pS χS and (17). Let C ∈ C ξ \ {δ(s), δ(t)} with f (x∗ (C)) > 12
and x∗ (C) ≥ 2 − 3ξ . Then
X
3 ∗
1
2
pS bS,C ≥ 1 + 5 − (x (C) + ξ) − ǫ
ν−
− (4ν − 1) pC
(18)
many .
2
2
S∈S
Proof: As before, let C← and C→ be the adjacent ξ-narrow cuts left and P
right.
∗
Let lS = |S ∩ C ∩ CP
|,
m
=
|S
∩
C|,
and
r
=
|S
∩
C
∩
C
|.
Note
that
←
S
S
→
S∈S pS mS = x (C)
and, using Lemma
3, S∈S pS lS = x∗ (C← ∩ C) ≤ 12 (x∗ (C← ) + x∗ (C)) − 1 ≤ 12 (ξ + x∗ (C)) − 1.
P
Analogously, S∈S pS rS ≤ 21 (ξ + x∗ (C)) − 1.
For all S ∈ S with mS ≥ 3, we have
2bS,C − (mS + 1)(ν − 12 ) + (4ν − 1)⌊ mS2−1 ⌋ ≥ 1.
(19)
(If mS is odd, then bS,C = 0 and thus 2bS,C −(mS +1)(ν − 21 )+(4ν −1)⌊ mS2−1 ⌋ = mS ν −3ν +1 ≥ 1.
If mS is even, then bS,C ≥ 12 and thus 2bS,C −(mS +1)(ν − 12 )+(4ν −1)⌊ mS2−1 ⌋ ≥ mS ν −5ν + 52 ≥ 23 .)
Now we distinguish four cases (cf. (17)).
C
C
Case 1: pC
120 + p021 ≤ pgood + ǫ.
For S ∈ S, let aS = 1 if SPhas type 120 or 021 at C, aS = −1 if S has type good at C, and
aS = 0 otherwise. Note that S∈S pS aS ≤ ǫ.
Then for all trees S ∈ S we have (cf. Lemma 14 and Table 1):
• mS ≥ 3, or
• bS,C ≥
1
2
and lS + rS − mS + aS ≥ 0, or
• bS,C ≥ ν and lS + rS − mS + aS ≥ −2,
and hence (using (19) for the case mS ≥ 3)
2bS,C + (lS + rS − mS + aS )(ν − 12 ) + (4ν − 1)⌊ mS2−1 ⌋ ≥ 1.
Taking the weighted sum, this implies (using (6))
X
2
pS bS,C ≥ 1 + (2 − ξ − ǫ) (ν − 12 ) − (4ν − 1)pC
many .
S∈S
As x∗ (C) ≥ 2 −
ξ
3
implies 2 − ξ ≥ 5 − 23 (x∗ (C) + ξ), we obtain (18).
C
C
Case 2: pC
011 + p110 ≤ pgood + ǫ.
For S ∈ S, let aS = 1 if SPhas type 011 or 110 at C, aS = −1 if S has type good at C, and
aS = 0 otherwise. Note that S∈S pS aS ≤ ǫ.
Then for all trees S ∈ S we have (cf. Lemma 14 and Table 1):
18
bS,C
benefit
≥
ν
type
010
011
110
111
020
021
120
022
220
121
good
good
good
1
2
1
2
1
2
ν
1
2
1
2
1
2
1
2
1
2
ν
1
2
0
l
m
r
number of edges in
C← ∩ C
0
0
1
1
0
0
1
0
2
1
l
l
≥0
C
1
1
1
1
2
2
2
2
2
2
2
2
m≥3
C ∩ C→
0
1
0
1
0
1
0
2
0
1
≥1−l
≥3−l
≥0
Case 1
Case 2
Case 3
Case 4
l+r−m
l+r+m
2l + r
l + 2r
−1
0
0
1
−2
−1
−1
0
0
0
−1
1
−m
≥
1
2
2
3
2
3
3
4
4
4
2
4
m
≥
0
1
2
3
0
1
2
2
4
3
1
3
0
≥
0
2
1
3
0
2
1
4
2
3
1
3
0
≥
Table 1: The different types and their contributions. The bounds on the benefit (second column)
follow from Lemma 14; here we use f (x∗ (C)) ≥ 21 . After reassembling, except for an ǫ fraction of
the trees, the two orange types (011 and 120) cannot occur simultaneously, and if the two purple
types (110 and 021) occur simultaneously, this is compensated by good types.
• mS ≥ 3, or
• bS,C ≥
1
2
and lS + rS + mS + aS ≥ 3, or
• bS,C ≥ ν and lS + rS + mS + aS ≥ 1,
and hence (using again (19) for the case mS ≥ 3)
2bS,C + (lS + rS + mS + aS − 3)(ν − 12 ) + (4ν − 1)⌊ mS2−1 ⌋ ≥ 1.
Taking the weighted sum, this implies
X
2
pS bS,C ≥ 1 + (5 − 2x∗ (C) − ξ − ǫ) (ν − 21 ) − (4ν − 1)pC
many ,
S∈S
implying (18) because ξ ≥ x∗ (C).
19
C
C
Case 3: pC
011 + p021 ≤ pgood + ǫ.
For S ∈ S, let aS = 1 if SPhas type 011 or 021 at C, aS = −1 if S has type good at C, and
aS = 0 otherwise. Note that S∈S pS aS ≤ ǫ.
Then for all trees S ∈ S we have (cf. Lemma 14 and Table 1):
• mS ≥ 3, or
• bS,C ≥
1
2
and 2lS + rS + aS ≥ 2, or
• bS,C ≥ ν and 2lS + rS + aS ≥ 0,
and hence (using once more (19) for the case mS ≥ 3)
2bS,C + (2lS + rS + aS − 2)(ν − 12 ) + (4ν − 1)⌊ mS2−1 ⌋ ≥ 1.
Taking the weighted sum, this implies
X
2
pS bS,C ≥ 1 + 5 − 32 (x∗ (C) + ξ) − ǫ ν − 12 − (4ν − 1)pC
many ,
S∈S
i.e., (18).
C
C
Case 4: pC
110 + p120 ≤ pgood + ǫ. This is symmetric to Case 3.
So (18) is proved in all cases.
4.3
✷
Setting the constants
We now obtain our main result easily:
Theorem 16 If ξ = 1.73 and ǫ = 0.01 and p is a distribution as obtained in Theorem 12, then
BOMC(p) ≤ (2 − β)c(x∗ ) for β = 0.401. In particular, we have an 1.599-approximation algorithm
for the s-t-path TSP, and the integrality ratio of (1) is at most 1.599.
Proof: We have ν = 1 − f (ξ) > 0.6. We use Lemma 7 and need to show (13).
Let C ∈ C be a narrow cut. If f (x∗ (C)) ≤ 12 , we have shown (13) in Lemma 13. So let now
C ∈ C be a narrow cut with f (x∗ (C)) > 12 . Note that 1.44 < x∗ (C) < 1.56 and thus in particular
C ∈ C ξ \ {δ(s), δ(t)} and x∗ (C) ≥ 2 − 3ξ . We apply Lemma 15 to C.
The constants ξ, ǫ, and β are chosen so that
3
1
2β
1 + 5 − (x∗ (C) + ξ) − ǫ
ν−
(x∗ (C) − 1)(2 − x∗ (C))
(20)
≥
2
2
1 − 2β
holds for all values of x∗ (C).
20
2β
3
∗
6+4ξ(2−ξ) , so 4β(2 − ξ) ≥ 3 − 6β − 4β(2 − ξ)(ξ − 1) and hence 1−2β (2 − x (C))
3−6β−4β(2−ξ)(ξ−1)
(ξ)
= 4ν−1
= 3−4f
2
2 . Therefore, using (6) and (4),
2(1−2β)
Moreover, β ≥
2β
1−2β (2
− ξ) ≥
≥
2β(2 − x∗ (C))pC
even
+ (4ν − 1)pC
many
1 − 2β
4ν − 1 ∗
2β(2 − x∗ (C)) 4ν − 1
C
−
(x (C) − 1)
= peven
+
1 − 2β
2
2
2β(2 − x∗ (C)) 4ν − 1
4ν − 1 ∗
∗
≤ (x (C) − 1)
−
(x (C) − 1)
+
1 − 2β
2
2
2β
(x∗ (C) − 1)(2 − x∗ (C)).
=
1 − 2β
Together with (18) and (20), this directly implies (13).
5
✷
Enhancements
The constants ξ and β in the previous section are not optimal, but they are close. The bounds
are almost tight for Case 3 (and 4) and x∗ (C) ≈ 1.52 and pC
many = 0. However, we now suggest
two ideas for a refined analysis that leads to a further improvement.
Firstly, since (in contrast to Sebő’s [2013] analysis) the worst case does not occur in x∗ (C) =
1.5, but in a slightly larger value, one can increase β and hence improve the approximation ratio
by increasing the γ-values slightly.
Secondly, the analysis in Case 3 (and 4) of Lemma 15 can be refined (and the analysis was not
tight in Case 1 and 2 anyway), as we will indicate now. Consider a critical cut C in Case 3, and
1
3
′
′
ξ′
′
assume for simplicity pC
many = 0. Let 2 ≤ ξ ≤ 2 with f (ξ ) = 2 , and let C← be the next ξ -narrow
′
ξ )+x(C ∩C ξ ) ≤ 3 x(C)+ 1 ξ+ 1 ξ ′ − 3
cut to the left of C. We have benefit 21 for at most a 21 x(C ∩C→
←
4
4
2
2
fraction of the trees, and larger benefit for the others. More generally, we have benefit at most
ξ ) + x(C ∩ C y ) ≤ 3 x(C) + 1 ξ + 1 y − 3 fraction of the trees, for
1 − f (y) for at most a 21 x(C ∩ C→
←
4
4
2
2
′
all y ∈ [ξ , ξ]. See Figure 5.
We have not performed the necessary calculations to obtain the best possible approximation
ratio with these ideas. However, it seems that the resulting improvements are rather small. For a
much better bound, we would probably need stronger reassembling results.
6
Discussion
Theorem 16 readily leads also to an improved approximation ratio for the prize-collecting s-t-path
TSP, simply by applying Theorem 6 of An, Kleinberg and Shmoys [2015]. See Guttmann-Beck et al.
[2000] for further applications.
21
benefit ≥
′
ξ
C←
y
C←
ξ
C←
C
ξ
C→
type
1 − f (ξ)
010
1
2
120
1 − f (y)
120
1
2
110
1 − f (y)
110
1 − f (ξ)
020
1
2
022
Figure 5: Refined analysis of Case 3.
Theorem 16 improves the best known upper bound on the integrality ratio of the LP (1). The
best known lower bound is 32 , shown by the metric closure of an unweighted circuit. For metric closures of unweighted graphs the integrality ratio is indeed 32 (even for T -tours), as Sebő and Vygen
[2014] proved. For general metrics, this remains open.
The most natural open question is of course to improve the approximation ratio further. Our
improvement was only small, but the reassembling technique could be more powerful than we
were able to prove. It seems that a stronger version of Lemma 9 would be needed. It would also
be interesting to generalize our algorithm to the T -tour problem for general T (Sebő’s [2013] 58 approximation algorithm and the previous algorithm of Cheriyan, Friggstad and Gao [2015] work
also for this more general problem). Finally, applying our technique to other TSP variants would
be very interesting.
Acknowledgement
The author thanks the three referees for careful reading and excellent suggestions.
References
An, H.-C., Kleinberg, R., and Shmoys, D.B. [2015]: Improving Christofides’ algorithm for the s-t path
TSP. Journal of the ACM 62 (2015), Article 34
Asadpour, A., Goemans, M.X., Mądry, A., Oveis Gharan, S., and Saberi, A. [2010]: An O(log n/ log log n)approximation algorithm for the asymmetric traveling salesman problem. Proceedings of the 21st Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA 2010), 379–389
22
Cheriyan, J., Friggstad, Z., and Gao, Z. [2015]: Approximating minimum-cost connected T -joins. Algorithmica 72 (2015), 126–147
Christofides, N. [1976]: Worst-case analysis of a new heuristic for the traveling salesman problem. Technical
Report 388, Graduate School of Industrial Administration, Carnegie-Mellon University, Pittsburgh 1976
Dantzig, G.B., Fulkerson, D.R., and Johnson, S.M. [1954]: Solution of a large scale traveling salesman
problem. Operations Research 2 (1954), 393–410
Edmonds, J. [1965]: The Chinese postman’s problem. Bulletin of the Operations Research Society of
America 13 (1965), B-73
Edmonds, J. [1970]: Submodular functions, matroids and certain polyhedra. In: Combinatorial Structures and Their Applications; Proceedings of the Calgary International Conference on Combinatorial
Structures and Their Applications 1969 (R. Guy, H. Hanani, N. Sauer, J. Schönheim, eds.), Gordon and
Breach, New York 1970, pp. 69–87
Edmonds, J., and Johnson, E.L. [1973]: Matching, Euler tours and the Chinese postman. Mathematical
Programming 5 (1973), 88–124
Gao, Z. [2013]: An LP-based 23 -approximation algorithm for the s-t path graph traveling salesman problem.
Operations Research Letters 41 (2013), 615–617
Gao, Z. [2015]: On the metric s-t path traveling salesman problem. SIAM Journal on Discrete Mathematics
29 (2015), 1133–1149
Genova, K., and Williamson, D.P. [2015]: An experimental evaluation of the best-of-many Christofides’
algorithm for the traveling salesman problem. In: Algorithms – ESA 2015; LNCS 9294 (N. Bansal,
I. Finocchi, eds.), Springer, Berlin 2015, pp. 570–581
Grötschel, M., Lovász, L., and Schrijver, A. [1981]: The ellipsoid method and its consequences in combinatorial optimization. Combinatorica 1 (1981), 169–197
Guttmann-Beck, N., Hassin, R., Khuller, S., and Raghavachari, B. [2000]: Approximation algorithms with
bounded performance guarantees for the clustered traveling salesman problem. Algorithmica 28 (2000),
422–437
Held, M., and Karp, R.M. [1970]: The traveling-salesman problem and minimum spanning trees. Operations Research 18 (1970), 1138–1162
Hoogeveen, J.A. [1991]: Analysis of Christofides’ heuristic: some paths are more difficult than cycles.
Operations Research Letters 10 (1991), 291–295
Karpinski, M., Lampis, M., and Schmied, R. [2013]: New inapproximability bounds for TSP. In: Algorithms and Computation. Proceedings of the 24th ISAAC Symposium; LNCS 8283 (L. Cai, S.-W. Cheng,
T.-W. Lam, eds.), Springer 2013, pp. 568–578
Mömke, T., and Svensson, O. [2011]: Approximating graphic TSP by matchings. Proceedings of the 52nd
Annual Symposium on Foundations of Computer Science (FOCS 2011), 560–569
Mucha, M. [2014]:
13
9 -approximation
for graphic TSP. Theory of Computing Systems 55 (2014), 640–657
23
Oveis Gharan, S., Saberi, A., and Singh, M. [2011]: A randomized rounding approach to the traveling
salesman problem. Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer
Science (FOCS 2011), 550–559
Sebő, A. [2013]: Eight fifth approximation for TSP paths. In: Integer Programming and Combinatorial
Optimization; Proceedings of the 16th IPCO Conference; LNCS 7801 (J. Correa, M.X. Goemans, eds.),
Springer 2013, pp. 362–374
Sebő, A., and Vygen, J. [2014]: Shorter tours by nicer ears: 7/5-approximation for graph-TSP, 3/2 for the
path version, and 4/3 for two-edge-connected subgraphs. Combinatorica 34 (2014), 597–629
Vygen, J. [2012]: New approximation algorithms for the TSP. OPTIMA 90 (2012), 1–12
Wolsey, L.A. [1980]: Heuristic analysis, linear programming and branch and bound. Mathematical Programming Study 13 (1980), 121–134
24
Appendix: an Example
The following example shows that we cannot get more benefit than Sebő [2013], no matter how
we choose the numbers γS,e , if we represent the LP solution x∗ in an arbitrary way as convex
combination of spanning trees.
We show one instance with 20 vertices and 30 edges. From this we obtain an infinite sequence
of instances by extending the wall-like part in the middle, inserting copies of the blue part, adding
4 vertices and 6 edges in each step.
The number next to an edge e is x∗ (e); then x∗ is a feasible solution of (1). Grey vertical lines
show the narrow cuts. We have x∗ (C) = 32 for all C ∈ C \ {δ(s), δ(t)}.
1
2
1
2
1
3
4
1
2
s
1
2
1
4
1
2
1
1
1
2
1
1
2
1
2
1
1
2
1
4
1
2
1
1
2
1
1
4
t
1
2
1
3
4
1
4
1
2
1
2
1
1
2
Consider the red sets in the following figure and all singletons; call this set of sets U . Then
the 30 vectors δ(U ), U ∈ U , are easily shown to be linearly independent. Since x∗ satisfies all
constraints of (1) that correspond to these sets with equality, x∗ is indeed a vertex of the polytope
defined by (1). Hence it is an optimum solution for some objective function.
1
2
1
2
1
3
4
1
1
2
s
1
4
1
4
1
2
1
1
2
1
1
2
1
2
1
2
1
1
2
1
1
4
1
2
1
1
2
1
4
t
1
2
1
3
4
1
2
1
2
1
2
1
∗
Now
P we willS show a (bad) 1decomposition of x into incidence vectors of trees. We have
= S∈S pS χ , where pS = 4 for the following four trees and pS = 0 for all other spanning
trees. In the wall-like part (grey cuts 6–9 from left in the top figure), the four trees have types 011,
110, 021, and 120. For each of these four trees S, each edge e of its s-t-path IS within the wall-like
part belongs to one narrow cut C with |C ∩ S| = 1 and one narrow cut C ′ with |C ′ ∩ S| = 2. No
matter how we choose γS,e , the total benefit of this edge to both cuts is 1, so the average benefit
x∗
25
that a narrow cut in the wall-like part receives is at most 12 . This shows that reassembling trees
is necessary.
We were unable to prove that the best-of-many Christofides algorithm has no better approximation ratio than 1.6, but to obtain a better ratio a completely different analysis would be
necessary. In fact, no better lower bound than 1.5 is known.
s
t
s
t
s
t
s
t
26
| 8cs.DS
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.